{"episode_id":"F000_T2024-01","field_index":0,"field":"federated learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Privacy and secure computation","description":"Federated learning research whose primary objective is preventing, measuring, or formally controlling disclosure of client data, updates, identities, or intermediate representations through privacy attacks, differential privacy, or cryptographic secure computation.","inclusion":"Include work centered on gradient or membership leakage, data reconstruction, privacy defenses, DP mechanisms/accounting, secure aggregation, homomorphic encryption, MPC, signatures, zero-knowledge, or privacy-preserving aggregation protocols.","exclusion":"If the main architecture is vertical or split FL rather than a reusable privacy primitive, use Partitioned-data and split architectures; if the threat is model corruption rather than data disclosure, use Adversarial integrity and robustness.","nearest_boundary":"Partitioned-data and split architectures","representative_topics":["federated learning defense via synthetic data generation to obscure client data from gradient leakage attacks","subject-level membership inference attacks in cross-silo federated learning","subject-level differential privacy in federated learning via hierarchical gradient averaging and group privacy"]},{"direction_index":1,"label":"D1","name":"Adversarial integrity and robustness","description":"Federated learning research whose primary objective is preserving model correctness, convergence, and integrity under malicious, poisoned, backdoored, free-riding, adversarial, or Byzantine client behavior.","inclusion":"Include work on poisoning and backdoor attacks or defenses, label flipping, free-rider detection, adversarial-update robustness, robust aggregation under Byzantine clients, malicious-majority tolerance, and convergence guarantees under arbitrary faulty updates.","exclusion":"If the adversary aims to infer private data rather than corrupt learning, use Privacy and secure computation; if integrity is enforced mainly by incentives, ledgers, or governance mechanisms, use Governance and decentralized coordination.","nearest_boundary":"Privacy and secure computation","representative_topics":["federated learning free-rider defense via model weight evolution frequency analysis","durable backdoor attacks in federated learning via low-magnitude parameter targeting","zero-knowledge proof-based byzantine robustness checks using rank statistics in federated learning"]},{"direction_index":2,"label":"D2","name":"Efficient operation and infrastructure","description":"Federated learning research whose primary objective is efficient, scalable, measurable, or deployable operation under communication, computation, energy, latency, wireless, scheduling, benchmarking, or systems constraints.","inclusion":"Include compression, pruning, quantization, low-bandwidth optimization, client scheduling, asynchronous training, straggler handling, wireless/edge/AirComp resource allocation, satellite/UAV/mobile-edge FL, benchmarks, platforms, simulators, debugging, and general evaluation infrastructure.","exclusion":"If the dominant goal is statistical personalization or heterogeneity mitigation rather than operational efficiency, use Heterogeneity and adaptive learning; if client participation is governed by payments or strategic incentives, use Governance and decentralized coordination.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated learning in low earth orbit satellite constellations with intermittent connectivity","communication-efficient federated gradient descent ascent for minimax optimization with linear convergence","interference management for over-the-air federated learning in multi-cell wireless networks"]},{"direction_index":3,"label":"D3","name":"Heterogeneity and adaptive learning","description":"Federated learning research whose primary objective is learning effectively across heterogeneous, non-IID, client-specific, weakly supervised, temporally shifting, or architecturally diverse participants.","inclusion":"Include non-IID convergence, personalization, clustered FL, domain adaptation/generalization, label-scarce or weakly supervised FL, continual/class-incremental FL, model-heterogeneity knowledge transfer, distillation, supernets, and adaptive client-specific models when the main goal is statistical adaptation.","exclusion":"If adaptation is specifically for large pretrained or foundation models, use Foundation-model and knowledge lifecycle; if the primary issue is operational bandwidth, latency, or wireless resources, use Efficient operation and infrastructure.","nearest_boundary":"Foundation-model and knowledge lifecycle","representative_topics":["federated learning adaptation under staggered distributed concept drift via hierarchical clustering","supernet training for federated learning under simultaneous data and system heterogeneity","straggler-resilient personalized federated learning under data and system heterogeneity"]},{"direction_index":4,"label":"D4","name":"Governance and decentralized coordination","description":"Federated learning research whose primary objective is organizing participation, trust, collaboration, or coordination among parties through decentralized topology, peer-to-peer optimization, incentives, fairness mechanisms, game theory, reputation, blockchain, or ledgers.","inclusion":"Include serverless and gossip FL, topology design, collaborator selection, blockchain-integrated FL, contribution valuation, auctions, contracts, token rewards, smart contracts, reputation, fairness-aware participation, and game-theoretic collaboration.","exclusion":"If decentralized communication is studied mainly as a wireless resource-allocation problem, use Efficient operation and infrastructure; if peer grouping is mainly a statistical personalization method without governance or topology focus, use Heterogeneity and adaptive learning.","nearest_boundary":"Efficient operation and infrastructure","representative_topics":["blockchain-integrated federated learning with incentive-based data integrity assurance","decentralized adaptive clustering for personalized deep learning under non-iid data shifts","modular blockchain-based federated learning with contribution-based incentive and malicious client detection"]},{"direction_index":5,"label":"D5","name":"Domain and graph-specialized FL","description":"Federated learning research whose primary objective is adapting FL to particular scientific, clinical, industrial, cyber-physical, recommendation, or graph-structured data domains where domain or data-modality constraints define the contribution.","inclusion":"Include healthcare and biomedical FL, medical imaging and EHR modeling, industrial IoT, vehicular, energy, utility, smart-grid, recommendation and cyber-physical applications, and federated graph or GNN learning when graph structure is central to the learning problem.","exclusion":"If a domain dataset is only an evaluation vehicle for a general privacy, robustness, efficiency, or heterogeneity method, classify by that broader method; if graph edges describe only client communication topology, use Governance and decentralized coordination.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated spatial-temporal graph learning with dynamic topology and relevance-driven aggregation","federated learning for privacy-preserving mammographic breast density estimation across institutions","federated prototype learning for cross-site prostate cancer detection from mri"]},{"direction_index":6,"label":"D6","name":"Partitioned-data and split architectures","description":"Federated learning research whose primary objective is designing, optimizing, securing, or analyzing architectures where features, labels, entities, or model layers are partitioned across parties rather than the standard horizontal client-sample partition.","inclusion":"Include vertical federated learning, split learning, split-federated hybrids, entity alignment, feature-partitioned training, split-point or layer partitioning, VFL-specific privacy/security protocols, and communication or computation optimization specific to these architectures.","exclusion":"If the contribution is a reusable secure aggregation or DP primitive for ordinary horizontal FL, use Privacy and secure computation; if the contribution is generic compression independent of vertical/split structure, use Efficient operation and infrastructure.","nearest_boundary":"Privacy and secure computation","representative_topics":["hybrid split-federated learning architectures for scalable privacy-preserving ml in heterogeneous d2d networks","binarized split learning for privacy leakage reduction and efficient mobile deployment","privacy-preserving vertical federated learning with heterogeneous feature support and formal security guarantees"]},{"direction_index":7,"label":"D7","name":"Foundation-model and knowledge lifecycle","description":"Federated learning research whose primary objective is controlling high-level model knowledge across the model lifecycle, including adapting large pretrained or foundation models and selectively removing, editing, or erasing learned contributions after training.","inclusion":"Include federated prompt tuning, adapters, parameter-efficient fine-tuning, pretrained transformer or vision-language model adaptation, federated foundation-model transfer, federated unlearning, certified client/data removal, and knowledge editing in FL.","exclusion":"If the problem is ordinary heterogeneous client models without pretrained/foundation-model adaptation, use Heterogeneity and adaptive learning; if removal is replaced by preventive privacy guarantees during training, use Privacy and secure computation.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated collaborative training of text-to-image transformers across distributed heterogeneous devices","federated unlearning via local unlearning initialization and minimal retraining rounds","communication-efficient prompt tuning in federated learning via split aggregation and backdoor robustness"]}],"split":"dev"}
{"episode_id":"F000_T2024-07","field_index":0,"field":"federated learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Privacy and secure computation","description":"Federated learning research whose primary objective is preventing, measuring, or formally controlling disclosure of client data, updates, identities, or intermediate representations through privacy attacks, differential privacy, or cryptographic secure computation.","inclusion":"Include work centered on gradient or membership leakage, data reconstruction, privacy defenses, DP mechanisms/accounting, secure aggregation, homomorphic encryption, MPC, signatures, zero-knowledge, or privacy-preserving aggregation protocols.","exclusion":"If the main architecture is vertical or split FL rather than a reusable privacy primitive, use Partitioned-data and split architectures; if the threat is model corruption rather than data disclosure, use Adversarial integrity and robustness.","nearest_boundary":"Partitioned-data and split architectures","representative_topics":["federated learning defense via synthetic data generation to obscure client data from gradient leakage attacks","subject-level membership inference attacks in cross-silo federated learning","subject-level differential privacy in federated learning via hierarchical gradient averaging and group privacy"]},{"direction_index":1,"label":"D1","name":"Adversarial integrity and robustness","description":"Federated learning research whose primary objective is preserving model correctness, convergence, and integrity under malicious, poisoned, backdoored, free-riding, adversarial, or Byzantine client behavior.","inclusion":"Include work on poisoning and backdoor attacks or defenses, label flipping, free-rider detection, adversarial-update robustness, robust aggregation under Byzantine clients, malicious-majority tolerance, and convergence guarantees under arbitrary faulty updates.","exclusion":"If the adversary aims to infer private data rather than corrupt learning, use Privacy and secure computation; if integrity is enforced mainly by incentives, ledgers, or governance mechanisms, use Governance and decentralized coordination.","nearest_boundary":"Privacy and secure computation","representative_topics":["federated learning free-rider defense via model weight evolution frequency analysis","durable backdoor attacks in federated learning via low-magnitude parameter targeting","zero-knowledge proof-based byzantine robustness checks using rank statistics in federated learning"]},{"direction_index":2,"label":"D2","name":"Efficient operation and infrastructure","description":"Federated learning research whose primary objective is efficient, scalable, measurable, or deployable operation under communication, computation, energy, latency, wireless, scheduling, benchmarking, or systems constraints.","inclusion":"Include compression, pruning, quantization, low-bandwidth optimization, client scheduling, asynchronous training, straggler handling, wireless/edge/AirComp resource allocation, satellite/UAV/mobile-edge FL, benchmarks, platforms, simulators, debugging, and general evaluation infrastructure.","exclusion":"If the dominant goal is statistical personalization or heterogeneity mitigation rather than operational efficiency, use Heterogeneity and adaptive learning; if client participation is governed by payments or strategic incentives, use Governance and decentralized coordination.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated learning in low earth orbit satellite constellations with intermittent connectivity","communication-efficient federated gradient descent ascent for minimax optimization with linear convergence","interference management for over-the-air federated learning in multi-cell wireless networks"]},{"direction_index":3,"label":"D3","name":"Heterogeneity and adaptive learning","description":"Federated learning research whose primary objective is learning effectively across heterogeneous, non-IID, client-specific, weakly supervised, temporally shifting, or architecturally diverse participants.","inclusion":"Include non-IID convergence, personalization, clustered FL, domain adaptation/generalization, label-scarce or weakly supervised FL, continual/class-incremental FL, model-heterogeneity knowledge transfer, distillation, supernets, and adaptive client-specific models when the main goal is statistical adaptation.","exclusion":"If adaptation is specifically for large pretrained or foundation models, use Foundation-model and knowledge lifecycle; if the primary issue is operational bandwidth, latency, or wireless resources, use Efficient operation and infrastructure.","nearest_boundary":"Foundation-model and knowledge lifecycle","representative_topics":["federated learning adaptation under staggered distributed concept drift via hierarchical clustering","supernet training for federated learning under simultaneous data and system heterogeneity","straggler-resilient personalized federated learning under data and system heterogeneity"]},{"direction_index":4,"label":"D4","name":"Governance and decentralized coordination","description":"Federated learning research whose primary objective is organizing participation, trust, collaboration, or coordination among parties through decentralized topology, peer-to-peer optimization, incentives, fairness mechanisms, game theory, reputation, blockchain, or ledgers.","inclusion":"Include serverless and gossip FL, topology design, collaborator selection, blockchain-integrated FL, contribution valuation, auctions, contracts, token rewards, smart contracts, reputation, fairness-aware participation, and game-theoretic collaboration.","exclusion":"If decentralized communication is studied mainly as a wireless resource-allocation problem, use Efficient operation and infrastructure; if peer grouping is mainly a statistical personalization method without governance or topology focus, use Heterogeneity and adaptive learning.","nearest_boundary":"Efficient operation and infrastructure","representative_topics":["blockchain-integrated federated learning with incentive-based data integrity assurance","decentralized adaptive clustering for personalized deep learning under non-iid data shifts","modular blockchain-based federated learning with contribution-based incentive and malicious client detection"]},{"direction_index":5,"label":"D5","name":"Domain and graph-specialized FL","description":"Federated learning research whose primary objective is adapting FL to particular scientific, clinical, industrial, cyber-physical, recommendation, or graph-structured data domains where domain or data-modality constraints define the contribution.","inclusion":"Include healthcare and biomedical FL, medical imaging and EHR modeling, industrial IoT, vehicular, energy, utility, smart-grid, recommendation and cyber-physical applications, and federated graph or GNN learning when graph structure is central to the learning problem.","exclusion":"If a domain dataset is only an evaluation vehicle for a general privacy, robustness, efficiency, or heterogeneity method, classify by that broader method; if graph edges describe only client communication topology, use Governance and decentralized coordination.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated spatial-temporal graph learning with dynamic topology and relevance-driven aggregation","federated learning for privacy-preserving mammographic breast density estimation across institutions","federated prototype learning for cross-site prostate cancer detection from mri"]},{"direction_index":6,"label":"D6","name":"Partitioned-data and split architectures","description":"Federated learning research whose primary objective is designing, optimizing, securing, or analyzing architectures where features, labels, entities, or model layers are partitioned across parties rather than the standard horizontal client-sample partition.","inclusion":"Include vertical federated learning, split learning, split-federated hybrids, entity alignment, feature-partitioned training, split-point or layer partitioning, VFL-specific privacy/security protocols, and communication or computation optimization specific to these architectures.","exclusion":"If the contribution is a reusable secure aggregation or DP primitive for ordinary horizontal FL, use Privacy and secure computation; if the contribution is generic compression independent of vertical/split structure, use Efficient operation and infrastructure.","nearest_boundary":"Privacy and secure computation","representative_topics":["hybrid split-federated learning architectures for scalable privacy-preserving ml in heterogeneous d2d networks","binarized split learning for privacy leakage reduction and efficient mobile deployment","privacy-preserving vertical federated learning with heterogeneous feature support and formal security guarantees"]},{"direction_index":7,"label":"D7","name":"Foundation-model and knowledge lifecycle","description":"Federated learning research whose primary objective is controlling high-level model knowledge across the model lifecycle, including adapting large pretrained or foundation models and selectively removing, editing, or erasing learned contributions after training.","inclusion":"Include federated prompt tuning, adapters, parameter-efficient fine-tuning, pretrained transformer or vision-language model adaptation, federated foundation-model transfer, federated unlearning, certified client/data removal, and knowledge editing in FL.","exclusion":"If the problem is ordinary heterogeneous client models without pretrained/foundation-model adaptation, use Heterogeneity and adaptive learning; if removal is replaced by preventive privacy guarantees during training, use Privacy and secure computation.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated collaborative training of text-to-image transformers across distributed heterogeneous devices","federated unlearning via local unlearning initialization and minimal retraining rounds","communication-efficient prompt tuning in federated learning via split aggregation and backdoor robustness"]}],"split":"dev"}
{"episode_id":"F000_T2025-01","field_index":0,"field":"federated learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Privacy and secure computation","description":"Federated learning research whose primary objective is preventing, measuring, or formally controlling disclosure of client data, updates, identities, or intermediate representations through privacy attacks, differential privacy, or cryptographic secure computation.","inclusion":"Include work centered on gradient or membership leakage, data reconstruction, privacy defenses, DP mechanisms/accounting, secure aggregation, homomorphic encryption, MPC, signatures, zero-knowledge, or privacy-preserving aggregation protocols.","exclusion":"If the main architecture is vertical or split FL rather than a reusable privacy primitive, use Partitioned-data and split architectures; if the threat is model corruption rather than data disclosure, use Adversarial integrity and robustness.","nearest_boundary":"Partitioned-data and split architectures","representative_topics":["federated learning defense via synthetic data generation to obscure client data from gradient leakage attacks","subject-level membership inference attacks in cross-silo federated learning","subject-level differential privacy in federated learning via hierarchical gradient averaging and group privacy"]},{"direction_index":1,"label":"D1","name":"Adversarial integrity and robustness","description":"Federated learning research whose primary objective is preserving model correctness, convergence, and integrity under malicious, poisoned, backdoored, free-riding, adversarial, or Byzantine client behavior.","inclusion":"Include work on poisoning and backdoor attacks or defenses, label flipping, free-rider detection, adversarial-update robustness, robust aggregation under Byzantine clients, malicious-majority tolerance, and convergence guarantees under arbitrary faulty updates.","exclusion":"If the adversary aims to infer private data rather than corrupt learning, use Privacy and secure computation; if integrity is enforced mainly by incentives, ledgers, or governance mechanisms, use Governance and decentralized coordination.","nearest_boundary":"Privacy and secure computation","representative_topics":["federated learning free-rider defense via model weight evolution frequency analysis","durable backdoor attacks in federated learning via low-magnitude parameter targeting","zero-knowledge proof-based byzantine robustness checks using rank statistics in federated learning"]},{"direction_index":2,"label":"D2","name":"Efficient operation and infrastructure","description":"Federated learning research whose primary objective is efficient, scalable, measurable, or deployable operation under communication, computation, energy, latency, wireless, scheduling, benchmarking, or systems constraints.","inclusion":"Include compression, pruning, quantization, low-bandwidth optimization, client scheduling, asynchronous training, straggler handling, wireless/edge/AirComp resource allocation, satellite/UAV/mobile-edge FL, benchmarks, platforms, simulators, debugging, and general evaluation infrastructure.","exclusion":"If the dominant goal is statistical personalization or heterogeneity mitigation rather than operational efficiency, use Heterogeneity and adaptive learning; if client participation is governed by payments or strategic incentives, use Governance and decentralized coordination.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated learning in low earth orbit satellite constellations with intermittent connectivity","communication-efficient federated gradient descent ascent for minimax optimization with linear convergence","interference management for over-the-air federated learning in multi-cell wireless networks"]},{"direction_index":3,"label":"D3","name":"Heterogeneity and adaptive learning","description":"Federated learning research whose primary objective is learning effectively across heterogeneous, non-IID, client-specific, weakly supervised, temporally shifting, or architecturally diverse participants.","inclusion":"Include non-IID convergence, personalization, clustered FL, domain adaptation/generalization, label-scarce or weakly supervised FL, continual/class-incremental FL, model-heterogeneity knowledge transfer, distillation, supernets, and adaptive client-specific models when the main goal is statistical adaptation.","exclusion":"If adaptation is specifically for large pretrained or foundation models, use Foundation-model and knowledge lifecycle; if the primary issue is operational bandwidth, latency, or wireless resources, use Efficient operation and infrastructure.","nearest_boundary":"Foundation-model and knowledge lifecycle","representative_topics":["federated learning adaptation under staggered distributed concept drift via hierarchical clustering","supernet training for federated learning under simultaneous data and system heterogeneity","straggler-resilient personalized federated learning under data and system heterogeneity"]},{"direction_index":4,"label":"D4","name":"Governance and decentralized coordination","description":"Federated learning research whose primary objective is organizing participation, trust, collaboration, or coordination among parties through decentralized topology, peer-to-peer optimization, incentives, fairness mechanisms, game theory, reputation, blockchain, or ledgers.","inclusion":"Include serverless and gossip FL, topology design, collaborator selection, blockchain-integrated FL, contribution valuation, auctions, contracts, token rewards, smart contracts, reputation, fairness-aware participation, and game-theoretic collaboration.","exclusion":"If decentralized communication is studied mainly as a wireless resource-allocation problem, use Efficient operation and infrastructure; if peer grouping is mainly a statistical personalization method without governance or topology focus, use Heterogeneity and adaptive learning.","nearest_boundary":"Efficient operation and infrastructure","representative_topics":["blockchain-integrated federated learning with incentive-based data integrity assurance","decentralized adaptive clustering for personalized deep learning under non-iid data shifts","modular blockchain-based federated learning with contribution-based incentive and malicious client detection"]},{"direction_index":5,"label":"D5","name":"Domain and graph-specialized FL","description":"Federated learning research whose primary objective is adapting FL to particular scientific, clinical, industrial, cyber-physical, recommendation, or graph-structured data domains where domain or data-modality constraints define the contribution.","inclusion":"Include healthcare and biomedical FL, medical imaging and EHR modeling, industrial IoT, vehicular, energy, utility, smart-grid, recommendation and cyber-physical applications, and federated graph or GNN learning when graph structure is central to the learning problem.","exclusion":"If a domain dataset is only an evaluation vehicle for a general privacy, robustness, efficiency, or heterogeneity method, classify by that broader method; if graph edges describe only client communication topology, use Governance and decentralized coordination.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated spatial-temporal graph learning with dynamic topology and relevance-driven aggregation","federated learning for privacy-preserving mammographic breast density estimation across institutions","federated prototype learning for cross-site prostate cancer detection from mri"]},{"direction_index":6,"label":"D6","name":"Partitioned-data and split architectures","description":"Federated learning research whose primary objective is designing, optimizing, securing, or analyzing architectures where features, labels, entities, or model layers are partitioned across parties rather than the standard horizontal client-sample partition.","inclusion":"Include vertical federated learning, split learning, split-federated hybrids, entity alignment, feature-partitioned training, split-point or layer partitioning, VFL-specific privacy/security protocols, and communication or computation optimization specific to these architectures.","exclusion":"If the contribution is a reusable secure aggregation or DP primitive for ordinary horizontal FL, use Privacy and secure computation; if the contribution is generic compression independent of vertical/split structure, use Efficient operation and infrastructure.","nearest_boundary":"Privacy and secure computation","representative_topics":["hybrid split-federated learning architectures for scalable privacy-preserving ml in heterogeneous d2d networks","binarized split learning for privacy leakage reduction and efficient mobile deployment","privacy-preserving vertical federated learning with heterogeneous feature support and formal security guarantees"]},{"direction_index":7,"label":"D7","name":"Foundation-model and knowledge lifecycle","description":"Federated learning research whose primary objective is controlling high-level model knowledge across the model lifecycle, including adapting large pretrained or foundation models and selectively removing, editing, or erasing learned contributions after training.","inclusion":"Include federated prompt tuning, adapters, parameter-efficient fine-tuning, pretrained transformer or vision-language model adaptation, federated foundation-model transfer, federated unlearning, certified client/data removal, and knowledge editing in FL.","exclusion":"If the problem is ordinary heterogeneous client models without pretrained/foundation-model adaptation, use Heterogeneity and adaptive learning; if removal is replaced by preventive privacy guarantees during training, use Privacy and secure computation.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated collaborative training of text-to-image transformers across distributed heterogeneous devices","federated unlearning via local unlearning initialization and minimal retraining rounds","communication-efficient prompt tuning in federated learning via split aggregation and backdoor robustness"]}],"split":"dev"}
{"episode_id":"F000_T2025-07","field_index":0,"field":"federated learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Privacy and secure computation","description":"Federated learning research whose primary objective is preventing, measuring, or formally controlling disclosure of client data, updates, identities, or intermediate representations through privacy attacks, differential privacy, or cryptographic secure computation.","inclusion":"Include work centered on gradient or membership leakage, data reconstruction, privacy defenses, DP mechanisms/accounting, secure aggregation, homomorphic encryption, MPC, signatures, zero-knowledge, or privacy-preserving aggregation protocols.","exclusion":"If the main architecture is vertical or split FL rather than a reusable privacy primitive, use Partitioned-data and split architectures; if the threat is model corruption rather than data disclosure, use Adversarial integrity and robustness.","nearest_boundary":"Partitioned-data and split architectures","representative_topics":["federated learning defense via synthetic data generation to obscure client data from gradient leakage attacks","subject-level membership inference attacks in cross-silo federated learning","subject-level differential privacy in federated learning via hierarchical gradient averaging and group privacy"]},{"direction_index":1,"label":"D1","name":"Adversarial integrity and robustness","description":"Federated learning research whose primary objective is preserving model correctness, convergence, and integrity under malicious, poisoned, backdoored, free-riding, adversarial, or Byzantine client behavior.","inclusion":"Include work on poisoning and backdoor attacks or defenses, label flipping, free-rider detection, adversarial-update robustness, robust aggregation under Byzantine clients, malicious-majority tolerance, and convergence guarantees under arbitrary faulty updates.","exclusion":"If the adversary aims to infer private data rather than corrupt learning, use Privacy and secure computation; if integrity is enforced mainly by incentives, ledgers, or governance mechanisms, use Governance and decentralized coordination.","nearest_boundary":"Privacy and secure computation","representative_topics":["federated learning free-rider defense via model weight evolution frequency analysis","durable backdoor attacks in federated learning via low-magnitude parameter targeting","zero-knowledge proof-based byzantine robustness checks using rank statistics in federated learning"]},{"direction_index":2,"label":"D2","name":"Efficient operation and infrastructure","description":"Federated learning research whose primary objective is efficient, scalable, measurable, or deployable operation under communication, computation, energy, latency, wireless, scheduling, benchmarking, or systems constraints.","inclusion":"Include compression, pruning, quantization, low-bandwidth optimization, client scheduling, asynchronous training, straggler handling, wireless/edge/AirComp resource allocation, satellite/UAV/mobile-edge FL, benchmarks, platforms, simulators, debugging, and general evaluation infrastructure.","exclusion":"If the dominant goal is statistical personalization or heterogeneity mitigation rather than operational efficiency, use Heterogeneity and adaptive learning; if client participation is governed by payments or strategic incentives, use Governance and decentralized coordination.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated learning in low earth orbit satellite constellations with intermittent connectivity","communication-efficient federated gradient descent ascent for minimax optimization with linear convergence","interference management for over-the-air federated learning in multi-cell wireless networks"]},{"direction_index":3,"label":"D3","name":"Heterogeneity and adaptive learning","description":"Federated learning research whose primary objective is learning effectively across heterogeneous, non-IID, client-specific, weakly supervised, temporally shifting, or architecturally diverse participants.","inclusion":"Include non-IID convergence, personalization, clustered FL, domain adaptation/generalization, label-scarce or weakly supervised FL, continual/class-incremental FL, model-heterogeneity knowledge transfer, distillation, supernets, and adaptive client-specific models when the main goal is statistical adaptation.","exclusion":"If adaptation is specifically for large pretrained or foundation models, use Foundation-model and knowledge lifecycle; if the primary issue is operational bandwidth, latency, or wireless resources, use Efficient operation and infrastructure.","nearest_boundary":"Foundation-model and knowledge lifecycle","representative_topics":["federated learning adaptation under staggered distributed concept drift via hierarchical clustering","supernet training for federated learning under simultaneous data and system heterogeneity","straggler-resilient personalized federated learning under data and system heterogeneity"]},{"direction_index":4,"label":"D4","name":"Governance and decentralized coordination","description":"Federated learning research whose primary objective is organizing participation, trust, collaboration, or coordination among parties through decentralized topology, peer-to-peer optimization, incentives, fairness mechanisms, game theory, reputation, blockchain, or ledgers.","inclusion":"Include serverless and gossip FL, topology design, collaborator selection, blockchain-integrated FL, contribution valuation, auctions, contracts, token rewards, smart contracts, reputation, fairness-aware participation, and game-theoretic collaboration.","exclusion":"If decentralized communication is studied mainly as a wireless resource-allocation problem, use Efficient operation and infrastructure; if peer grouping is mainly a statistical personalization method without governance or topology focus, use Heterogeneity and adaptive learning.","nearest_boundary":"Efficient operation and infrastructure","representative_topics":["blockchain-integrated federated learning with incentive-based data integrity assurance","decentralized adaptive clustering for personalized deep learning under non-iid data shifts","modular blockchain-based federated learning with contribution-based incentive and malicious client detection"]},{"direction_index":5,"label":"D5","name":"Domain and graph-specialized FL","description":"Federated learning research whose primary objective is adapting FL to particular scientific, clinical, industrial, cyber-physical, recommendation, or graph-structured data domains where domain or data-modality constraints define the contribution.","inclusion":"Include healthcare and biomedical FL, medical imaging and EHR modeling, industrial IoT, vehicular, energy, utility, smart-grid, recommendation and cyber-physical applications, and federated graph or GNN learning when graph structure is central to the learning problem.","exclusion":"If a domain dataset is only an evaluation vehicle for a general privacy, robustness, efficiency, or heterogeneity method, classify by that broader method; if graph edges describe only client communication topology, use Governance and decentralized coordination.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated spatial-temporal graph learning with dynamic topology and relevance-driven aggregation","federated learning for privacy-preserving mammographic breast density estimation across institutions","federated prototype learning for cross-site prostate cancer detection from mri"]},{"direction_index":6,"label":"D6","name":"Partitioned-data and split architectures","description":"Federated learning research whose primary objective is designing, optimizing, securing, or analyzing architectures where features, labels, entities, or model layers are partitioned across parties rather than the standard horizontal client-sample partition.","inclusion":"Include vertical federated learning, split learning, split-federated hybrids, entity alignment, feature-partitioned training, split-point or layer partitioning, VFL-specific privacy/security protocols, and communication or computation optimization specific to these architectures.","exclusion":"If the contribution is a reusable secure aggregation or DP primitive for ordinary horizontal FL, use Privacy and secure computation; if the contribution is generic compression independent of vertical/split structure, use Efficient operation and infrastructure.","nearest_boundary":"Privacy and secure computation","representative_topics":["hybrid split-federated learning architectures for scalable privacy-preserving ml in heterogeneous d2d networks","binarized split learning for privacy leakage reduction and efficient mobile deployment","privacy-preserving vertical federated learning with heterogeneous feature support and formal security guarantees"]},{"direction_index":7,"label":"D7","name":"Foundation-model and knowledge lifecycle","description":"Federated learning research whose primary objective is controlling high-level model knowledge across the model lifecycle, including adapting large pretrained or foundation models and selectively removing, editing, or erasing learned contributions after training.","inclusion":"Include federated prompt tuning, adapters, parameter-efficient fine-tuning, pretrained transformer or vision-language model adaptation, federated foundation-model transfer, federated unlearning, certified client/data removal, and knowledge editing in FL.","exclusion":"If the problem is ordinary heterogeneous client models without pretrained/foundation-model adaptation, use Heterogeneity and adaptive learning; if removal is replaced by preventive privacy guarantees during training, use Privacy and secure computation.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated collaborative training of text-to-image transformers across distributed heterogeneous devices","federated unlearning via local unlearning initialization and minimal retraining rounds","communication-efficient prompt tuning in federated learning via split aggregation and backdoor robustness"]}],"split":"dev"}
{"episode_id":"F000_T2026-01","field_index":0,"field":"federated learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Privacy and secure computation","description":"Federated learning research whose primary objective is preventing, measuring, or formally controlling disclosure of client data, updates, identities, or intermediate representations through privacy attacks, differential privacy, or cryptographic secure computation.","inclusion":"Include work centered on gradient or membership leakage, data reconstruction, privacy defenses, DP mechanisms/accounting, secure aggregation, homomorphic encryption, MPC, signatures, zero-knowledge, or privacy-preserving aggregation protocols.","exclusion":"If the main architecture is vertical or split FL rather than a reusable privacy primitive, use Partitioned-data and split architectures; if the threat is model corruption rather than data disclosure, use Adversarial integrity and robustness.","nearest_boundary":"Partitioned-data and split architectures","representative_topics":["federated learning defense via synthetic data generation to obscure client data from gradient leakage attacks","subject-level membership inference attacks in cross-silo federated learning","subject-level differential privacy in federated learning via hierarchical gradient averaging and group privacy"]},{"direction_index":1,"label":"D1","name":"Adversarial integrity and robustness","description":"Federated learning research whose primary objective is preserving model correctness, convergence, and integrity under malicious, poisoned, backdoored, free-riding, adversarial, or Byzantine client behavior.","inclusion":"Include work on poisoning and backdoor attacks or defenses, label flipping, free-rider detection, adversarial-update robustness, robust aggregation under Byzantine clients, malicious-majority tolerance, and convergence guarantees under arbitrary faulty updates.","exclusion":"If the adversary aims to infer private data rather than corrupt learning, use Privacy and secure computation; if integrity is enforced mainly by incentives, ledgers, or governance mechanisms, use Governance and decentralized coordination.","nearest_boundary":"Privacy and secure computation","representative_topics":["federated learning free-rider defense via model weight evolution frequency analysis","durable backdoor attacks in federated learning via low-magnitude parameter targeting","zero-knowledge proof-based byzantine robustness checks using rank statistics in federated learning"]},{"direction_index":2,"label":"D2","name":"Efficient operation and infrastructure","description":"Federated learning research whose primary objective is efficient, scalable, measurable, or deployable operation under communication, computation, energy, latency, wireless, scheduling, benchmarking, or systems constraints.","inclusion":"Include compression, pruning, quantization, low-bandwidth optimization, client scheduling, asynchronous training, straggler handling, wireless/edge/AirComp resource allocation, satellite/UAV/mobile-edge FL, benchmarks, platforms, simulators, debugging, and general evaluation infrastructure.","exclusion":"If the dominant goal is statistical personalization or heterogeneity mitigation rather than operational efficiency, use Heterogeneity and adaptive learning; if client participation is governed by payments or strategic incentives, use Governance and decentralized coordination.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated learning in low earth orbit satellite constellations with intermittent connectivity","communication-efficient federated gradient descent ascent for minimax optimization with linear convergence","interference management for over-the-air federated learning in multi-cell wireless networks"]},{"direction_index":3,"label":"D3","name":"Heterogeneity and adaptive learning","description":"Federated learning research whose primary objective is learning effectively across heterogeneous, non-IID, client-specific, weakly supervised, temporally shifting, or architecturally diverse participants.","inclusion":"Include non-IID convergence, personalization, clustered FL, domain adaptation/generalization, label-scarce or weakly supervised FL, continual/class-incremental FL, model-heterogeneity knowledge transfer, distillation, supernets, and adaptive client-specific models when the main goal is statistical adaptation.","exclusion":"If adaptation is specifically for large pretrained or foundation models, use Foundation-model and knowledge lifecycle; if the primary issue is operational bandwidth, latency, or wireless resources, use Efficient operation and infrastructure.","nearest_boundary":"Foundation-model and knowledge lifecycle","representative_topics":["federated learning adaptation under staggered distributed concept drift via hierarchical clustering","supernet training for federated learning under simultaneous data and system heterogeneity","straggler-resilient personalized federated learning under data and system heterogeneity"]},{"direction_index":4,"label":"D4","name":"Governance and decentralized coordination","description":"Federated learning research whose primary objective is organizing participation, trust, collaboration, or coordination among parties through decentralized topology, peer-to-peer optimization, incentives, fairness mechanisms, game theory, reputation, blockchain, or ledgers.","inclusion":"Include serverless and gossip FL, topology design, collaborator selection, blockchain-integrated FL, contribution valuation, auctions, contracts, token rewards, smart contracts, reputation, fairness-aware participation, and game-theoretic collaboration.","exclusion":"If decentralized communication is studied mainly as a wireless resource-allocation problem, use Efficient operation and infrastructure; if peer grouping is mainly a statistical personalization method without governance or topology focus, use Heterogeneity and adaptive learning.","nearest_boundary":"Efficient operation and infrastructure","representative_topics":["blockchain-integrated federated learning with incentive-based data integrity assurance","decentralized adaptive clustering for personalized deep learning under non-iid data shifts","modular blockchain-based federated learning with contribution-based incentive and malicious client detection"]},{"direction_index":5,"label":"D5","name":"Domain and graph-specialized FL","description":"Federated learning research whose primary objective is adapting FL to particular scientific, clinical, industrial, cyber-physical, recommendation, or graph-structured data domains where domain or data-modality constraints define the contribution.","inclusion":"Include healthcare and biomedical FL, medical imaging and EHR modeling, industrial IoT, vehicular, energy, utility, smart-grid, recommendation and cyber-physical applications, and federated graph or GNN learning when graph structure is central to the learning problem.","exclusion":"If a domain dataset is only an evaluation vehicle for a general privacy, robustness, efficiency, or heterogeneity method, classify by that broader method; if graph edges describe only client communication topology, use Governance and decentralized coordination.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated spatial-temporal graph learning with dynamic topology and relevance-driven aggregation","federated learning for privacy-preserving mammographic breast density estimation across institutions","federated prototype learning for cross-site prostate cancer detection from mri"]},{"direction_index":6,"label":"D6","name":"Partitioned-data and split architectures","description":"Federated learning research whose primary objective is designing, optimizing, securing, or analyzing architectures where features, labels, entities, or model layers are partitioned across parties rather than the standard horizontal client-sample partition.","inclusion":"Include vertical federated learning, split learning, split-federated hybrids, entity alignment, feature-partitioned training, split-point or layer partitioning, VFL-specific privacy/security protocols, and communication or computation optimization specific to these architectures.","exclusion":"If the contribution is a reusable secure aggregation or DP primitive for ordinary horizontal FL, use Privacy and secure computation; if the contribution is generic compression independent of vertical/split structure, use Efficient operation and infrastructure.","nearest_boundary":"Privacy and secure computation","representative_topics":["hybrid split-federated learning architectures for scalable privacy-preserving ml in heterogeneous d2d networks","binarized split learning for privacy leakage reduction and efficient mobile deployment","privacy-preserving vertical federated learning with heterogeneous feature support and formal security guarantees"]},{"direction_index":7,"label":"D7","name":"Foundation-model and knowledge lifecycle","description":"Federated learning research whose primary objective is controlling high-level model knowledge across the model lifecycle, including adapting large pretrained or foundation models and selectively removing, editing, or erasing learned contributions after training.","inclusion":"Include federated prompt tuning, adapters, parameter-efficient fine-tuning, pretrained transformer or vision-language model adaptation, federated foundation-model transfer, federated unlearning, certified client/data removal, and knowledge editing in FL.","exclusion":"If the problem is ordinary heterogeneous client models without pretrained/foundation-model adaptation, use Heterogeneity and adaptive learning; if removal is replaced by preventive privacy guarantees during training, use Privacy and secure computation.","nearest_boundary":"Heterogeneity and adaptive learning","representative_topics":["federated collaborative training of text-to-image transformers across distributed heterogeneous devices","federated unlearning via local unlearning initialization and minimal retraining rounds","communication-efficient prompt tuning in federated learning via split aggregation and backdoor robustness"]}],"split":"dev"}
{"episode_id":"F005_T2024-01","field_index":5,"field":"graph neural networks","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Temporal and behavioral prediction","description":"Modeling time-evolving graph signals, structures, and interaction behavior for forecasting, dynamic representation, link prediction, and recommendation objectives.","inclusion":"Include work whose primary objective is prediction over temporal, spatiotemporal, dynamic, or user-item/behavioral graph processes, including traffic and environmental forecasting, continuous-time dynamic graphs, temporal link prediction, and graph-based recommendation.","exclusion":"General architectural mechanisms for handling structure, heterophily, attention, or expressivity belong to Core graph representation architectures when prediction is not the dominant objective.","nearest_boundary":"Core graph representation architectures","representative_topics":["decoupled diffusion-inherent signal separation for dynamic spatial-temporal traffic forecasting","pre-training enhanced spatial-temporal graph neural networks for long-term multivariate time series forecasting","spatiotemporal graph neural networks for oceanic variable prediction with minimal domain knowledge"]},{"direction_index":1,"label":"D1","name":"Core graph representation architectures","description":"Designing, analyzing, and adapting general graph representation mechanisms, including message passing, expressivity, depth, heterophily handling, transformers, heterogeneous schemas, link-centric structure learning, pooling, and architecture search.","inclusion":"Include work whose primary objective is a general-purpose representation mechanism or architectural capability for graph learning, including formal expressivity, oversmoothing or over-squashing mitigation, heterophily adaptation, graph transformers, heterogeneous or higher-order GNNs, structural link/pooling methods, and GNN architecture search.","exclusion":"Methods primarily motivated by training efficiency, label scarcity, self-supervision, language-model integration, or deployment constraints belong to Learning regimes and generalization or Scale-efficient systems and reduction.","nearest_boundary":"Learning regimes and generalization","representative_topics":["2-dimensional weisfeiler-lehman graph neural networks for expressive link prediction","disentangling propagation and transformation depth to address model degradation in deep gnns","graph restructuring via adaptive spectral clustering to improve homophily in classical gnns"]},{"direction_index":2,"label":"D2","name":"Learning regimes and generalization","description":"Improving how GNNs learn, pretrain, adapt, and transfer under limited labels, self-supervision, language or foundation-model integration, distribution shift, causal or invariant structure, uncertainty, and formal generalization or stability constraints.","inclusion":"Include work whose primary objective is self-supervised or contrastive representation learning, graph prompting or foundation-model adaptation, language-model-enhanced graph learning, data-efficient supervised learning, few-shot or active learning, causal and invariant graph learning, OOD adaptation, training regularization, Bayesian or uncertainty-aware learning, and theoretical generalization or stability analysis.","exclusion":"Self-supervision or regularization used mainly to accelerate large-scale training belongs to Scale-efficient systems and reduction; adversarial threat modeling belongs to Trustworthy and secure graph learning.","nearest_boundary":"Scale-efficient systems and reduction","representative_topics":["efficient graph contrastive learning via group discrimination and binary cross-entropy","metric-based few-shot graph classification with task-conditioned embeddings and latent augmentation","diversity-aware active label sampling for graph neural network training efficiency"]},{"direction_index":3,"label":"D3","name":"Scale-efficient systems and reduction","description":"Making GNN training, inference, storage, and deployment feasible at scale through systems, sampling, distributed execution, hardware acceleration, compression, pruning, distillation, condensation, and compact representations.","inclusion":"Include work whose primary objective is reducing runtime, memory, communication, energy, model size, graph size, or data volume while preserving usable graph-learning performance.","exclusion":"Architectural changes whose main purpose is expressivity, heterophily handling, graph-level pooling, or transformer-style modeling belong to Core graph representation architectures.","nearest_boundary":"Core graph representation architectures","representative_topics":["graph neural network training for large-scale node classification via feature momentum","tensor-train compressed gnn embeddings for large graphs without node features","graph condensation by matching receptive field distributions in synthetic graph synthesis"]},{"direction_index":4,"label":"D4","name":"Trustworthy and secure graph learning","description":"Ensuring GNN reliability, accountability, safety, and governance through adversarial robustness, security analysis, fairness, privacy, federated learning, unlearning, explainability, and anomaly or fraud detection.","inclusion":"Include work whose primary objective is making graph learning trustworthy or security-relevant, including attacks and defenses, backdoors, membership inference, fairness constraints, privacy-preserving or federated training, certified unlearning, explanations of predictions or bias, and detection of anomalous or malicious graph entities.","exclusion":"Causal representation methods whose main goal is OOD generalization rather than accountability or explanation belong to Learning regimes and generalization.","nearest_boundary":"Learning regimes and generalization","representative_topics":["unsupervised graph structure refinement for robust graph neural network defense","transferable backdoor attacks on gnns via surrogate-guided perturbation-based dataset poisoning","fair graph neural networks mitigating sensitive attribute leakage through propagation-aware feature masking"]},{"direction_index":5,"label":"D5","name":"Geometric and domain-grounded modeling","description":"Embedding geometric, physical, scientific, engineering, biomedical, software, vision, or other domain structure into GNN formulations when the main objective is faithful modeling of a domain system or geometry-governed phenomenon.","inclusion":"Include work whose primary objective is E(n)/SE(3)-equivariant or geometry-aware modeling, physics-informed simulation or inverse problems, molecular and physical-system modeling, or domain-specific graph construction and knowledge integration for scientific, engineering, healthcare, vision, bioinformatics, software, or industrial applications.","exclusion":"Graph-structured decision policies or neural heuristics for optimization belong to Graph decision optimization and algorithmic reasoning.","nearest_boundary":"Graph decision optimization","representative_topics":["graph neural networks with autodecoder priors for pde-constrained inverse problem recovery","physics-aware multiplex gnns for efficient 3d molecular property prediction","equivariant transformer architecture for 3d atomistic graphs using irreducible representations"]},{"direction_index":6,"label":"D6","name":"Graph decision optimization","description":"Using GNNs to learn policies, heuristics, solvers, or algorithmic reasoning procedures for graph-structured optimization, control, allocation, decoding, clustering, and combinatorial decision problems.","inclusion":"Include work whose primary objective is solving or approximating graph-structured decisions, including NP-hard graph problems, neural combinatorial optimization, bandits, beamforming, channel decoding, optimal control, inverse control-like estimation, analogical reasoning, and matrix-completion-style algorithmic graph reasoning.","exclusion":"Predictive modeling of temporal, behavioral, or scientific graph systems without an explicit decision or solver objective belongs to Temporal and behavioral prediction or Geometric and domain-grounded modeling.","nearest_boundary":"Geometric and domain-grounded modeling","representative_topics":["neural improvement heuristics for graph combinatorial optimization with edge-structured information encoding","scattering-based graph neural networks for maximum clique approximation with rule-based decoding","benchmarking graph neural network optimizers against classical greedy algorithms for combinatorial optimization"]},{"direction_index":7,"label":"D7","name":"Evaluation and diagnostics","description":"Building benchmarks, surveys, datasets, metrics, diagnostic probes, and evaluation infrastructure for measuring GNN capabilities, costs, graph-information use, and field-level progress.","inclusion":"Include work whose primary objective is evaluation methodology rather than a new task solver, including benchmark taxonomies, controlled empirical diagnostics, long-range evaluation protocols, computation-time prediction, graph-information-use metrics, synthetic benchmark generation, benchmark platforms, and broad surveys.","exclusion":"Benchmarks or diagnostics introduced mainly to validate a specific architecture, optimizer, security method, or application belong to that primary research program.","nearest_boundary":"Core graph representation architectures","representative_topics":["taxonomy of graph benchmarks via perturbation-based sensitivity profiling for gnn evaluation","benchmarking graph transformers and message passing networks for long-range interaction reasoning","graph metric-based prediction of gnn training time for acceleration and design selection"]}],"split":"dev"}
{"episode_id":"F005_T2024-07","field_index":5,"field":"graph neural networks","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Temporal and behavioral prediction","description":"Modeling time-evolving graph signals, structures, and interaction behavior for forecasting, dynamic representation, link prediction, and recommendation objectives.","inclusion":"Include work whose primary objective is prediction over temporal, spatiotemporal, dynamic, or user-item/behavioral graph processes, including traffic and environmental forecasting, continuous-time dynamic graphs, temporal link prediction, and graph-based recommendation.","exclusion":"General architectural mechanisms for handling structure, heterophily, attention, or expressivity belong to Core graph representation architectures when prediction is not the dominant objective.","nearest_boundary":"Core graph representation architectures","representative_topics":["decoupled diffusion-inherent signal separation for dynamic spatial-temporal traffic forecasting","pre-training enhanced spatial-temporal graph neural networks for long-term multivariate time series forecasting","spatiotemporal graph neural networks for oceanic variable prediction with minimal domain knowledge"]},{"direction_index":1,"label":"D1","name":"Core graph representation architectures","description":"Designing, analyzing, and adapting general graph representation mechanisms, including message passing, expressivity, depth, heterophily handling, transformers, heterogeneous schemas, link-centric structure learning, pooling, and architecture search.","inclusion":"Include work whose primary objective is a general-purpose representation mechanism or architectural capability for graph learning, including formal expressivity, oversmoothing or over-squashing mitigation, heterophily adaptation, graph transformers, heterogeneous or higher-order GNNs, structural link/pooling methods, and GNN architecture search.","exclusion":"Methods primarily motivated by training efficiency, label scarcity, self-supervision, language-model integration, or deployment constraints belong to Learning regimes and generalization or Scale-efficient systems and reduction.","nearest_boundary":"Learning regimes and generalization","representative_topics":["2-dimensional weisfeiler-lehman graph neural networks for expressive link prediction","disentangling propagation and transformation depth to address model degradation in deep gnns","graph restructuring via adaptive spectral clustering to improve homophily in classical gnns"]},{"direction_index":2,"label":"D2","name":"Learning regimes and generalization","description":"Improving how GNNs learn, pretrain, adapt, and transfer under limited labels, self-supervision, language or foundation-model integration, distribution shift, causal or invariant structure, uncertainty, and formal generalization or stability constraints.","inclusion":"Include work whose primary objective is self-supervised or contrastive representation learning, graph prompting or foundation-model adaptation, language-model-enhanced graph learning, data-efficient supervised learning, few-shot or active learning, causal and invariant graph learning, OOD adaptation, training regularization, Bayesian or uncertainty-aware learning, and theoretical generalization or stability analysis.","exclusion":"Self-supervision or regularization used mainly to accelerate large-scale training belongs to Scale-efficient systems and reduction; adversarial threat modeling belongs to Trustworthy and secure graph learning.","nearest_boundary":"Scale-efficient systems and reduction","representative_topics":["efficient graph contrastive learning via group discrimination and binary cross-entropy","metric-based few-shot graph classification with task-conditioned embeddings and latent augmentation","diversity-aware active label sampling for graph neural network training efficiency"]},{"direction_index":3,"label":"D3","name":"Scale-efficient systems and reduction","description":"Making GNN training, inference, storage, and deployment feasible at scale through systems, sampling, distributed execution, hardware acceleration, compression, pruning, distillation, condensation, and compact representations.","inclusion":"Include work whose primary objective is reducing runtime, memory, communication, energy, model size, graph size, or data volume while preserving usable graph-learning performance.","exclusion":"Architectural changes whose main purpose is expressivity, heterophily handling, graph-level pooling, or transformer-style modeling belong to Core graph representation architectures.","nearest_boundary":"Core graph representation architectures","representative_topics":["graph neural network training for large-scale node classification via feature momentum","tensor-train compressed gnn embeddings for large graphs without node features","graph condensation by matching receptive field distributions in synthetic graph synthesis"]},{"direction_index":4,"label":"D4","name":"Trustworthy and secure graph learning","description":"Ensuring GNN reliability, accountability, safety, and governance through adversarial robustness, security analysis, fairness, privacy, federated learning, unlearning, explainability, and anomaly or fraud detection.","inclusion":"Include work whose primary objective is making graph learning trustworthy or security-relevant, including attacks and defenses, backdoors, membership inference, fairness constraints, privacy-preserving or federated training, certified unlearning, explanations of predictions or bias, and detection of anomalous or malicious graph entities.","exclusion":"Causal representation methods whose main goal is OOD generalization rather than accountability or explanation belong to Learning regimes and generalization.","nearest_boundary":"Learning regimes and generalization","representative_topics":["unsupervised graph structure refinement for robust graph neural network defense","transferable backdoor attacks on gnns via surrogate-guided perturbation-based dataset poisoning","fair graph neural networks mitigating sensitive attribute leakage through propagation-aware feature masking"]},{"direction_index":5,"label":"D5","name":"Geometric and domain-grounded modeling","description":"Embedding geometric, physical, scientific, engineering, biomedical, software, vision, or other domain structure into GNN formulations when the main objective is faithful modeling of a domain system or geometry-governed phenomenon.","inclusion":"Include work whose primary objective is E(n)/SE(3)-equivariant or geometry-aware modeling, physics-informed simulation or inverse problems, molecular and physical-system modeling, or domain-specific graph construction and knowledge integration for scientific, engineering, healthcare, vision, bioinformatics, software, or industrial applications.","exclusion":"Graph-structured decision policies or neural heuristics for optimization belong to Graph decision optimization and algorithmic reasoning.","nearest_boundary":"Graph decision optimization","representative_topics":["graph neural networks with autodecoder priors for pde-constrained inverse problem recovery","physics-aware multiplex gnns for efficient 3d molecular property prediction","equivariant transformer architecture for 3d atomistic graphs using irreducible representations"]},{"direction_index":6,"label":"D6","name":"Graph decision optimization","description":"Using GNNs to learn policies, heuristics, solvers, or algorithmic reasoning procedures for graph-structured optimization, control, allocation, decoding, clustering, and combinatorial decision problems.","inclusion":"Include work whose primary objective is solving or approximating graph-structured decisions, including NP-hard graph problems, neural combinatorial optimization, bandits, beamforming, channel decoding, optimal control, inverse control-like estimation, analogical reasoning, and matrix-completion-style algorithmic graph reasoning.","exclusion":"Predictive modeling of temporal, behavioral, or scientific graph systems without an explicit decision or solver objective belongs to Temporal and behavioral prediction or Geometric and domain-grounded modeling.","nearest_boundary":"Geometric and domain-grounded modeling","representative_topics":["neural improvement heuristics for graph combinatorial optimization with edge-structured information encoding","scattering-based graph neural networks for maximum clique approximation with rule-based decoding","benchmarking graph neural network optimizers against classical greedy algorithms for combinatorial optimization"]},{"direction_index":7,"label":"D7","name":"Evaluation and diagnostics","description":"Building benchmarks, surveys, datasets, metrics, diagnostic probes, and evaluation infrastructure for measuring GNN capabilities, costs, graph-information use, and field-level progress.","inclusion":"Include work whose primary objective is evaluation methodology rather than a new task solver, including benchmark taxonomies, controlled empirical diagnostics, long-range evaluation protocols, computation-time prediction, graph-information-use metrics, synthetic benchmark generation, benchmark platforms, and broad surveys.","exclusion":"Benchmarks or diagnostics introduced mainly to validate a specific architecture, optimizer, security method, or application belong to that primary research program.","nearest_boundary":"Core graph representation architectures","representative_topics":["taxonomy of graph benchmarks via perturbation-based sensitivity profiling for gnn evaluation","benchmarking graph transformers and message passing networks for long-range interaction reasoning","graph metric-based prediction of gnn training time for acceleration and design selection"]}],"split":"dev"}
{"episode_id":"F005_T2025-01","field_index":5,"field":"graph neural networks","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Temporal and behavioral prediction","description":"Modeling time-evolving graph signals, structures, and interaction behavior for forecasting, dynamic representation, link prediction, and recommendation objectives.","inclusion":"Include work whose primary objective is prediction over temporal, spatiotemporal, dynamic, or user-item/behavioral graph processes, including traffic and environmental forecasting, continuous-time dynamic graphs, temporal link prediction, and graph-based recommendation.","exclusion":"General architectural mechanisms for handling structure, heterophily, attention, or expressivity belong to Core graph representation architectures when prediction is not the dominant objective.","nearest_boundary":"Core graph representation architectures","representative_topics":["decoupled diffusion-inherent signal separation for dynamic spatial-temporal traffic forecasting","pre-training enhanced spatial-temporal graph neural networks for long-term multivariate time series forecasting","spatiotemporal graph neural networks for oceanic variable prediction with minimal domain knowledge"]},{"direction_index":1,"label":"D1","name":"Core graph representation architectures","description":"Designing, analyzing, and adapting general graph representation mechanisms, including message passing, expressivity, depth, heterophily handling, transformers, heterogeneous schemas, link-centric structure learning, pooling, and architecture search.","inclusion":"Include work whose primary objective is a general-purpose representation mechanism or architectural capability for graph learning, including formal expressivity, oversmoothing or over-squashing mitigation, heterophily adaptation, graph transformers, heterogeneous or higher-order GNNs, structural link/pooling methods, and GNN architecture search.","exclusion":"Methods primarily motivated by training efficiency, label scarcity, self-supervision, language-model integration, or deployment constraints belong to Learning regimes and generalization or Scale-efficient systems and reduction.","nearest_boundary":"Learning regimes and generalization","representative_topics":["2-dimensional weisfeiler-lehman graph neural networks for expressive link prediction","disentangling propagation and transformation depth to address model degradation in deep gnns","graph restructuring via adaptive spectral clustering to improve homophily in classical gnns"]},{"direction_index":2,"label":"D2","name":"Learning regimes and generalization","description":"Improving how GNNs learn, pretrain, adapt, and transfer under limited labels, self-supervision, language or foundation-model integration, distribution shift, causal or invariant structure, uncertainty, and formal generalization or stability constraints.","inclusion":"Include work whose primary objective is self-supervised or contrastive representation learning, graph prompting or foundation-model adaptation, language-model-enhanced graph learning, data-efficient supervised learning, few-shot or active learning, causal and invariant graph learning, OOD adaptation, training regularization, Bayesian or uncertainty-aware learning, and theoretical generalization or stability analysis.","exclusion":"Self-supervision or regularization used mainly to accelerate large-scale training belongs to Scale-efficient systems and reduction; adversarial threat modeling belongs to Trustworthy and secure graph learning.","nearest_boundary":"Scale-efficient systems and reduction","representative_topics":["efficient graph contrastive learning via group discrimination and binary cross-entropy","metric-based few-shot graph classification with task-conditioned embeddings and latent augmentation","diversity-aware active label sampling for graph neural network training efficiency"]},{"direction_index":3,"label":"D3","name":"Scale-efficient systems and reduction","description":"Making GNN training, inference, storage, and deployment feasible at scale through systems, sampling, distributed execution, hardware acceleration, compression, pruning, distillation, condensation, and compact representations.","inclusion":"Include work whose primary objective is reducing runtime, memory, communication, energy, model size, graph size, or data volume while preserving usable graph-learning performance.","exclusion":"Architectural changes whose main purpose is expressivity, heterophily handling, graph-level pooling, or transformer-style modeling belong to Core graph representation architectures.","nearest_boundary":"Core graph representation architectures","representative_topics":["graph neural network training for large-scale node classification via feature momentum","tensor-train compressed gnn embeddings for large graphs without node features","graph condensation by matching receptive field distributions in synthetic graph synthesis"]},{"direction_index":4,"label":"D4","name":"Trustworthy and secure graph learning","description":"Ensuring GNN reliability, accountability, safety, and governance through adversarial robustness, security analysis, fairness, privacy, federated learning, unlearning, explainability, and anomaly or fraud detection.","inclusion":"Include work whose primary objective is making graph learning trustworthy or security-relevant, including attacks and defenses, backdoors, membership inference, fairness constraints, privacy-preserving or federated training, certified unlearning, explanations of predictions or bias, and detection of anomalous or malicious graph entities.","exclusion":"Causal representation methods whose main goal is OOD generalization rather than accountability or explanation belong to Learning regimes and generalization.","nearest_boundary":"Learning regimes and generalization","representative_topics":["unsupervised graph structure refinement for robust graph neural network defense","transferable backdoor attacks on gnns via surrogate-guided perturbation-based dataset poisoning","fair graph neural networks mitigating sensitive attribute leakage through propagation-aware feature masking"]},{"direction_index":5,"label":"D5","name":"Geometric and domain-grounded modeling","description":"Embedding geometric, physical, scientific, engineering, biomedical, software, vision, or other domain structure into GNN formulations when the main objective is faithful modeling of a domain system or geometry-governed phenomenon.","inclusion":"Include work whose primary objective is E(n)/SE(3)-equivariant or geometry-aware modeling, physics-informed simulation or inverse problems, molecular and physical-system modeling, or domain-specific graph construction and knowledge integration for scientific, engineering, healthcare, vision, bioinformatics, software, or industrial applications.","exclusion":"Graph-structured decision policies or neural heuristics for optimization belong to Graph decision optimization and algorithmic reasoning.","nearest_boundary":"Graph decision optimization","representative_topics":["graph neural networks with autodecoder priors for pde-constrained inverse problem recovery","physics-aware multiplex gnns for efficient 3d molecular property prediction","equivariant transformer architecture for 3d atomistic graphs using irreducible representations"]},{"direction_index":6,"label":"D6","name":"Graph decision optimization","description":"Using GNNs to learn policies, heuristics, solvers, or algorithmic reasoning procedures for graph-structured optimization, control, allocation, decoding, clustering, and combinatorial decision problems.","inclusion":"Include work whose primary objective is solving or approximating graph-structured decisions, including NP-hard graph problems, neural combinatorial optimization, bandits, beamforming, channel decoding, optimal control, inverse control-like estimation, analogical reasoning, and matrix-completion-style algorithmic graph reasoning.","exclusion":"Predictive modeling of temporal, behavioral, or scientific graph systems without an explicit decision or solver objective belongs to Temporal and behavioral prediction or Geometric and domain-grounded modeling.","nearest_boundary":"Geometric and domain-grounded modeling","representative_topics":["neural improvement heuristics for graph combinatorial optimization with edge-structured information encoding","scattering-based graph neural networks for maximum clique approximation with rule-based decoding","benchmarking graph neural network optimizers against classical greedy algorithms for combinatorial optimization"]},{"direction_index":7,"label":"D7","name":"Evaluation and diagnostics","description":"Building benchmarks, surveys, datasets, metrics, diagnostic probes, and evaluation infrastructure for measuring GNN capabilities, costs, graph-information use, and field-level progress.","inclusion":"Include work whose primary objective is evaluation methodology rather than a new task solver, including benchmark taxonomies, controlled empirical diagnostics, long-range evaluation protocols, computation-time prediction, graph-information-use metrics, synthetic benchmark generation, benchmark platforms, and broad surveys.","exclusion":"Benchmarks or diagnostics introduced mainly to validate a specific architecture, optimizer, security method, or application belong to that primary research program.","nearest_boundary":"Core graph representation architectures","representative_topics":["taxonomy of graph benchmarks via perturbation-based sensitivity profiling for gnn evaluation","benchmarking graph transformers and message passing networks for long-range interaction reasoning","graph metric-based prediction of gnn training time for acceleration and design selection"]}],"split":"dev"}
{"episode_id":"F005_T2025-07","field_index":5,"field":"graph neural networks","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Temporal and behavioral prediction","description":"Modeling time-evolving graph signals, structures, and interaction behavior for forecasting, dynamic representation, link prediction, and recommendation objectives.","inclusion":"Include work whose primary objective is prediction over temporal, spatiotemporal, dynamic, or user-item/behavioral graph processes, including traffic and environmental forecasting, continuous-time dynamic graphs, temporal link prediction, and graph-based recommendation.","exclusion":"General architectural mechanisms for handling structure, heterophily, attention, or expressivity belong to Core graph representation architectures when prediction is not the dominant objective.","nearest_boundary":"Core graph representation architectures","representative_topics":["decoupled diffusion-inherent signal separation for dynamic spatial-temporal traffic forecasting","pre-training enhanced spatial-temporal graph neural networks for long-term multivariate time series forecasting","spatiotemporal graph neural networks for oceanic variable prediction with minimal domain knowledge"]},{"direction_index":1,"label":"D1","name":"Core graph representation architectures","description":"Designing, analyzing, and adapting general graph representation mechanisms, including message passing, expressivity, depth, heterophily handling, transformers, heterogeneous schemas, link-centric structure learning, pooling, and architecture search.","inclusion":"Include work whose primary objective is a general-purpose representation mechanism or architectural capability for graph learning, including formal expressivity, oversmoothing or over-squashing mitigation, heterophily adaptation, graph transformers, heterogeneous or higher-order GNNs, structural link/pooling methods, and GNN architecture search.","exclusion":"Methods primarily motivated by training efficiency, label scarcity, self-supervision, language-model integration, or deployment constraints belong to Learning regimes and generalization or Scale-efficient systems and reduction.","nearest_boundary":"Learning regimes and generalization","representative_topics":["2-dimensional weisfeiler-lehman graph neural networks for expressive link prediction","disentangling propagation and transformation depth to address model degradation in deep gnns","graph restructuring via adaptive spectral clustering to improve homophily in classical gnns"]},{"direction_index":2,"label":"D2","name":"Learning regimes and generalization","description":"Improving how GNNs learn, pretrain, adapt, and transfer under limited labels, self-supervision, language or foundation-model integration, distribution shift, causal or invariant structure, uncertainty, and formal generalization or stability constraints.","inclusion":"Include work whose primary objective is self-supervised or contrastive representation learning, graph prompting or foundation-model adaptation, language-model-enhanced graph learning, data-efficient supervised learning, few-shot or active learning, causal and invariant graph learning, OOD adaptation, training regularization, Bayesian or uncertainty-aware learning, and theoretical generalization or stability analysis.","exclusion":"Self-supervision or regularization used mainly to accelerate large-scale training belongs to Scale-efficient systems and reduction; adversarial threat modeling belongs to Trustworthy and secure graph learning.","nearest_boundary":"Scale-efficient systems and reduction","representative_topics":["efficient graph contrastive learning via group discrimination and binary cross-entropy","metric-based few-shot graph classification with task-conditioned embeddings and latent augmentation","diversity-aware active label sampling for graph neural network training efficiency"]},{"direction_index":3,"label":"D3","name":"Scale-efficient systems and reduction","description":"Making GNN training, inference, storage, and deployment feasible at scale through systems, sampling, distributed execution, hardware acceleration, compression, pruning, distillation, condensation, and compact representations.","inclusion":"Include work whose primary objective is reducing runtime, memory, communication, energy, model size, graph size, or data volume while preserving usable graph-learning performance.","exclusion":"Architectural changes whose main purpose is expressivity, heterophily handling, graph-level pooling, or transformer-style modeling belong to Core graph representation architectures.","nearest_boundary":"Core graph representation architectures","representative_topics":["graph neural network training for large-scale node classification via feature momentum","tensor-train compressed gnn embeddings for large graphs without node features","graph condensation by matching receptive field distributions in synthetic graph synthesis"]},{"direction_index":4,"label":"D4","name":"Trustworthy and secure graph learning","description":"Ensuring GNN reliability, accountability, safety, and governance through adversarial robustness, security analysis, fairness, privacy, federated learning, unlearning, explainability, and anomaly or fraud detection.","inclusion":"Include work whose primary objective is making graph learning trustworthy or security-relevant, including attacks and defenses, backdoors, membership inference, fairness constraints, privacy-preserving or federated training, certified unlearning, explanations of predictions or bias, and detection of anomalous or malicious graph entities.","exclusion":"Causal representation methods whose main goal is OOD generalization rather than accountability or explanation belong to Learning regimes and generalization.","nearest_boundary":"Learning regimes and generalization","representative_topics":["unsupervised graph structure refinement for robust graph neural network defense","transferable backdoor attacks on gnns via surrogate-guided perturbation-based dataset poisoning","fair graph neural networks mitigating sensitive attribute leakage through propagation-aware feature masking"]},{"direction_index":5,"label":"D5","name":"Geometric and domain-grounded modeling","description":"Embedding geometric, physical, scientific, engineering, biomedical, software, vision, or other domain structure into GNN formulations when the main objective is faithful modeling of a domain system or geometry-governed phenomenon.","inclusion":"Include work whose primary objective is E(n)/SE(3)-equivariant or geometry-aware modeling, physics-informed simulation or inverse problems, molecular and physical-system modeling, or domain-specific graph construction and knowledge integration for scientific, engineering, healthcare, vision, bioinformatics, software, or industrial applications.","exclusion":"Graph-structured decision policies or neural heuristics for optimization belong to Graph decision optimization and algorithmic reasoning.","nearest_boundary":"Graph decision optimization","representative_topics":["graph neural networks with autodecoder priors for pde-constrained inverse problem recovery","physics-aware multiplex gnns for efficient 3d molecular property prediction","equivariant transformer architecture for 3d atomistic graphs using irreducible representations"]},{"direction_index":6,"label":"D6","name":"Graph decision optimization","description":"Using GNNs to learn policies, heuristics, solvers, or algorithmic reasoning procedures for graph-structured optimization, control, allocation, decoding, clustering, and combinatorial decision problems.","inclusion":"Include work whose primary objective is solving or approximating graph-structured decisions, including NP-hard graph problems, neural combinatorial optimization, bandits, beamforming, channel decoding, optimal control, inverse control-like estimation, analogical reasoning, and matrix-completion-style algorithmic graph reasoning.","exclusion":"Predictive modeling of temporal, behavioral, or scientific graph systems without an explicit decision or solver objective belongs to Temporal and behavioral prediction or Geometric and domain-grounded modeling.","nearest_boundary":"Geometric and domain-grounded modeling","representative_topics":["neural improvement heuristics for graph combinatorial optimization with edge-structured information encoding","scattering-based graph neural networks for maximum clique approximation with rule-based decoding","benchmarking graph neural network optimizers against classical greedy algorithms for combinatorial optimization"]},{"direction_index":7,"label":"D7","name":"Evaluation and diagnostics","description":"Building benchmarks, surveys, datasets, metrics, diagnostic probes, and evaluation infrastructure for measuring GNN capabilities, costs, graph-information use, and field-level progress.","inclusion":"Include work whose primary objective is evaluation methodology rather than a new task solver, including benchmark taxonomies, controlled empirical diagnostics, long-range evaluation protocols, computation-time prediction, graph-information-use metrics, synthetic benchmark generation, benchmark platforms, and broad surveys.","exclusion":"Benchmarks or diagnostics introduced mainly to validate a specific architecture, optimizer, security method, or application belong to that primary research program.","nearest_boundary":"Core graph representation architectures","representative_topics":["taxonomy of graph benchmarks via perturbation-based sensitivity profiling for gnn evaluation","benchmarking graph transformers and message passing networks for long-range interaction reasoning","graph metric-based prediction of gnn training time for acceleration and design selection"]}],"split":"dev"}
{"episode_id":"F005_T2026-01","field_index":5,"field":"graph neural networks","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Temporal and behavioral prediction","description":"Modeling time-evolving graph signals, structures, and interaction behavior for forecasting, dynamic representation, link prediction, and recommendation objectives.","inclusion":"Include work whose primary objective is prediction over temporal, spatiotemporal, dynamic, or user-item/behavioral graph processes, including traffic and environmental forecasting, continuous-time dynamic graphs, temporal link prediction, and graph-based recommendation.","exclusion":"General architectural mechanisms for handling structure, heterophily, attention, or expressivity belong to Core graph representation architectures when prediction is not the dominant objective.","nearest_boundary":"Core graph representation architectures","representative_topics":["decoupled diffusion-inherent signal separation for dynamic spatial-temporal traffic forecasting","pre-training enhanced spatial-temporal graph neural networks for long-term multivariate time series forecasting","spatiotemporal graph neural networks for oceanic variable prediction with minimal domain knowledge"]},{"direction_index":1,"label":"D1","name":"Core graph representation architectures","description":"Designing, analyzing, and adapting general graph representation mechanisms, including message passing, expressivity, depth, heterophily handling, transformers, heterogeneous schemas, link-centric structure learning, pooling, and architecture search.","inclusion":"Include work whose primary objective is a general-purpose representation mechanism or architectural capability for graph learning, including formal expressivity, oversmoothing or over-squashing mitigation, heterophily adaptation, graph transformers, heterogeneous or higher-order GNNs, structural link/pooling methods, and GNN architecture search.","exclusion":"Methods primarily motivated by training efficiency, label scarcity, self-supervision, language-model integration, or deployment constraints belong to Learning regimes and generalization or Scale-efficient systems and reduction.","nearest_boundary":"Learning regimes and generalization","representative_topics":["2-dimensional weisfeiler-lehman graph neural networks for expressive link prediction","disentangling propagation and transformation depth to address model degradation in deep gnns","graph restructuring via adaptive spectral clustering to improve homophily in classical gnns"]},{"direction_index":2,"label":"D2","name":"Learning regimes and generalization","description":"Improving how GNNs learn, pretrain, adapt, and transfer under limited labels, self-supervision, language or foundation-model integration, distribution shift, causal or invariant structure, uncertainty, and formal generalization or stability constraints.","inclusion":"Include work whose primary objective is self-supervised or contrastive representation learning, graph prompting or foundation-model adaptation, language-model-enhanced graph learning, data-efficient supervised learning, few-shot or active learning, causal and invariant graph learning, OOD adaptation, training regularization, Bayesian or uncertainty-aware learning, and theoretical generalization or stability analysis.","exclusion":"Self-supervision or regularization used mainly to accelerate large-scale training belongs to Scale-efficient systems and reduction; adversarial threat modeling belongs to Trustworthy and secure graph learning.","nearest_boundary":"Scale-efficient systems and reduction","representative_topics":["efficient graph contrastive learning via group discrimination and binary cross-entropy","metric-based few-shot graph classification with task-conditioned embeddings and latent augmentation","diversity-aware active label sampling for graph neural network training efficiency"]},{"direction_index":3,"label":"D3","name":"Scale-efficient systems and reduction","description":"Making GNN training, inference, storage, and deployment feasible at scale through systems, sampling, distributed execution, hardware acceleration, compression, pruning, distillation, condensation, and compact representations.","inclusion":"Include work whose primary objective is reducing runtime, memory, communication, energy, model size, graph size, or data volume while preserving usable graph-learning performance.","exclusion":"Architectural changes whose main purpose is expressivity, heterophily handling, graph-level pooling, or transformer-style modeling belong to Core graph representation architectures.","nearest_boundary":"Core graph representation architectures","representative_topics":["graph neural network training for large-scale node classification via feature momentum","tensor-train compressed gnn embeddings for large graphs without node features","graph condensation by matching receptive field distributions in synthetic graph synthesis"]},{"direction_index":4,"label":"D4","name":"Trustworthy and secure graph learning","description":"Ensuring GNN reliability, accountability, safety, and governance through adversarial robustness, security analysis, fairness, privacy, federated learning, unlearning, explainability, and anomaly or fraud detection.","inclusion":"Include work whose primary objective is making graph learning trustworthy or security-relevant, including attacks and defenses, backdoors, membership inference, fairness constraints, privacy-preserving or federated training, certified unlearning, explanations of predictions or bias, and detection of anomalous or malicious graph entities.","exclusion":"Causal representation methods whose main goal is OOD generalization rather than accountability or explanation belong to Learning regimes and generalization.","nearest_boundary":"Learning regimes and generalization","representative_topics":["unsupervised graph structure refinement for robust graph neural network defense","transferable backdoor attacks on gnns via surrogate-guided perturbation-based dataset poisoning","fair graph neural networks mitigating sensitive attribute leakage through propagation-aware feature masking"]},{"direction_index":5,"label":"D5","name":"Geometric and domain-grounded modeling","description":"Embedding geometric, physical, scientific, engineering, biomedical, software, vision, or other domain structure into GNN formulations when the main objective is faithful modeling of a domain system or geometry-governed phenomenon.","inclusion":"Include work whose primary objective is E(n)/SE(3)-equivariant or geometry-aware modeling, physics-informed simulation or inverse problems, molecular and physical-system modeling, or domain-specific graph construction and knowledge integration for scientific, engineering, healthcare, vision, bioinformatics, software, or industrial applications.","exclusion":"Graph-structured decision policies or neural heuristics for optimization belong to Graph decision optimization and algorithmic reasoning.","nearest_boundary":"Graph decision optimization","representative_topics":["graph neural networks with autodecoder priors for pde-constrained inverse problem recovery","physics-aware multiplex gnns for efficient 3d molecular property prediction","equivariant transformer architecture for 3d atomistic graphs using irreducible representations"]},{"direction_index":6,"label":"D6","name":"Graph decision optimization","description":"Using GNNs to learn policies, heuristics, solvers, or algorithmic reasoning procedures for graph-structured optimization, control, allocation, decoding, clustering, and combinatorial decision problems.","inclusion":"Include work whose primary objective is solving or approximating graph-structured decisions, including NP-hard graph problems, neural combinatorial optimization, bandits, beamforming, channel decoding, optimal control, inverse control-like estimation, analogical reasoning, and matrix-completion-style algorithmic graph reasoning.","exclusion":"Predictive modeling of temporal, behavioral, or scientific graph systems without an explicit decision or solver objective belongs to Temporal and behavioral prediction or Geometric and domain-grounded modeling.","nearest_boundary":"Geometric and domain-grounded modeling","representative_topics":["neural improvement heuristics for graph combinatorial optimization with edge-structured information encoding","scattering-based graph neural networks for maximum clique approximation with rule-based decoding","benchmarking graph neural network optimizers against classical greedy algorithms for combinatorial optimization"]},{"direction_index":7,"label":"D7","name":"Evaluation and diagnostics","description":"Building benchmarks, surveys, datasets, metrics, diagnostic probes, and evaluation infrastructure for measuring GNN capabilities, costs, graph-information use, and field-level progress.","inclusion":"Include work whose primary objective is evaluation methodology rather than a new task solver, including benchmark taxonomies, controlled empirical diagnostics, long-range evaluation protocols, computation-time prediction, graph-information-use metrics, synthetic benchmark generation, benchmark platforms, and broad surveys.","exclusion":"Benchmarks or diagnostics introduced mainly to validate a specific architecture, optimizer, security method, or application belong to that primary research program.","nearest_boundary":"Core graph representation architectures","representative_topics":["taxonomy of graph benchmarks via perturbation-based sensitivity profiling for gnn evaluation","benchmarking graph transformers and message passing networks for long-range interaction reasoning","graph metric-based prediction of gnn training time for acceleration and design selection"]}],"split":"dev"}
{"episode_id":"F012_T2024-01","field_index":12,"field":"object detection","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Geometric 3D and sensor-fusion detection","description":"Detect objects by constructing 3D, BEV, range, radar, or cross-sensor scene representations from cameras, LiDAR, radar, RGB-D, or related sensors.","inclusion":"Include work whose primary objective is geometric 3D object detection, camera-only depth/BEV reasoning, single-sensor point-cloud/radar representation learning, or inference-time fusion of complementary sensors for 2D/3D detection.","exclusion":"If the main objective is label efficiency or domain transfer for 3D detectors, assign to D1; if the main objective is generic detector architecture or compression independent of 3D/sensor geometry, assign to D3.","nearest_boundary":"D1","representative_topics":["unified voxel-space multi-modality fusion for 3d object detection and tracking","depth-supervised bev 3d object detection with camera-aware depth estimation and refinement","sparse proposal-based end-to-end 3d object detection from point clouds using transformers"]},{"direction_index":1,"label":"D1","name":"Data-efficient and transfer learning","description":"Train detectors under scarce, weak, noisy, unlabeled, synthetic, or shifted data by improving supervision, pseudo-labeling, domain adaptation, and transfer protocols.","inclusion":"Include work whose primary objective is reducing annotation cost, repairing noisy labels, learning from weak or absent labels, semi-supervised 2D/3D detection, closed-set domain adaptation/generalization, synthetic-to-real transfer, or data-efficient 3D transfer.","exclusion":"If the main objective is novel-category recognition beyond a fixed label space, assign to D2; if the main objective is deployment compression or architectural efficiency rather than data/supervision efficiency, assign to D3.","nearest_boundary":"D2","representative_topics":["cross-domain detection transformer adaptation via spatial and semantic token alignment","weakly semi-supervised object detection via point annotation matching and instance learning","semi-supervised object detection with label matching via distribution-aware and instance-level self-training"]},{"direction_index":2,"label":"D2","name":"Open and novel category detection","description":"Extend detectors to categories that are unknown, novel, rare, few-shot, zero-shot, open-set, open-world, or language-defined.","inclusion":"Include work whose primary objective is category-space expansion through few-shot transfer, open-set/open-world formulations, open-vocabulary or vision-language alignment, incremental category learning, or base-to-novel generalization.","exclusion":"If categories remain fixed and the problem is target-domain shift or missing labels, assign to D1; if the work mainly improves generic DETR/CNN detector mechanics without category expansion, assign to D3.","nearest_boundary":"D1","representative_topics":["unifying multi-dataset object detection via language embedding query and category alignment","few-shot copy-paste augmentation using occlusion relationships for rare category detection","open vocabulary object detection with proposal mining and causal backdoor adjustment for novel categories"]},{"direction_index":3,"label":"D3","name":"Detector architecture and deployment efficiency","description":"Improve general-purpose detector mechanisms, training recipes, model compression, distillation, and speed-accuracy tradeoffs for deployable detection systems.","inclusion":"Include work whose primary objective is CNN/dense/single-stage/two-stage architecture design, feature pyramids, scale handling, DETR query/matching/attention mechanisms, lightweighting, pruning, quantization, distillation, or resource-aware inference.","exclusion":"If transformer or CNN mechanisms are subordinated to 3D geometry, sensor fusion, remote sensing orientation, or domain-specific clinical/maritime goals, assign to the relevant objective-specific direction instead.","nearest_boundary":"D0","representative_topics":["lightweight convolutional neck design for real-time object detection on edge devices","unified transformer framework for joint object detection and instance, panoptic, and semantic segmentation","multi-scale feature pyramid integration for transformer-based object detection"]},{"direction_index":4,"label":"D4","name":"Trustworthiness and evaluation methodology","description":"Assess, quantify, benchmark, calibrate, or improve detector reliability, security, uncertainty, robustness, and evaluation validity.","inclusion":"Include work whose primary objective is adversarial attack or defense, corruption/OOD robustness, uncertainty estimation, confidence calibration, metric design, benchmark methodology, dataset-composition analysis, annotation protocol, or diagnostic evaluation.","exclusion":"If the primary challenge is natural sensing degradation such as low light, infrared small targets, underwater optics, or adverse weather, assign to D5; if the work mainly proposes a detector architecture, assign to D3.","nearest_boundary":"D5","representative_topics":["efficient mimo uncertainty estimation for lidar-based 3d object detection under runtime constraints","longitudinal error-tolerant evaluation metrics for camera-only 3d object detection","parallelized mean average precision computation using vectorized operations for object detection"]},{"direction_index":5,"label":"D5","name":"Degraded and ambiguous appearance detection","description":"Detect objects whose visual evidence is weak, concealed, degraded, low-contrast, infrared, thermal, underwater, maritime, or otherwise difficult to separate from background.","inclusion":"Include work whose primary objective is detection under adverse visibility, low light, snow/rain/fog, infrared small-target conditions, RGB-thermal or RGB-D robustness, underwater/maritime optical degradation, camouflage, saliency, transparent/glass ambiguity, or unlabeled objectness discovery driven by foreground-background ambiguity.","exclusion":"If the primary objective is formal adversarial robustness or evaluation methodology, assign to D4; if the primary objective is overhead small/oriented remote-sensing geometry rather than degraded appearance, assign to D7.","nearest_boundary":"D4","representative_topics":["dark vessel detection in sar imagery using large-scale annotated maritime object datasets","lightweight yolo-based object detection with cross fusion for real-world snow conditions","axial-attention and multi-scale patch fusion for infrared small target detection"]},{"direction_index":6,"label":"D6","name":"Temporal and streaming detection","description":"Exploit time, motion, sequence structure, frame propagation, streaming constraints, or event streams to improve detection beyond independent single-image inference.","inclusion":"Include work whose primary objective is video object detection, temporal feature aggregation, motion-aware refinement, bounding-box propagation, streaming perception, event-camera detection, frame selection, temporal consistency, or latency-aware detection over sequences.","exclusion":"If temporal cues are used mainly for camera-only 3D geometry, assign to D0; if the application domain, such as clinical video or maritime surveillance, is the dominant research objective, assign to D7 or D5 respectively.","nearest_boundary":"D0","representative_topics":["spatio-temporal aggregation for insubstantial object detection in video","bounding box propagation for fast vehicle detection and tracking in fisheye traffic video","coarse-to-fine video-based smoky vehicle detection with 3d spatial-temporal refinement"]},{"direction_index":7,"label":"D7","name":"Domain-structured target detection","description":"Solve detection problems where specialized image domains impose distinctive target geometry, scale, orientation, morphology, operational constraints, or domain-specific target definitions.","inclusion":"Include work whose primary objective is detection in overhead/aerial/UAV/remote-sensing imagery, oriented-box remote sensing, high-resolution small-object overhead detection, clinical and biomedical object localization, or other specialized operational domains where the domain structure defines the detection problem.","exclusion":"If the domain challenge is primarily natural visibility degradation, underwater/maritime optics, infrared small targets, camouflage, or RGB-thermal robustness, assign to D5; if the contribution is a broadly general detector mechanism independent of the domain, assign to D3.","nearest_boundary":"D5","representative_topics":["self-supervised pretraining for label-efficient chest x-ray anomaly object detection","knowledge distillation for lightweight oriented object detection on aerial images","modulated kalman iou loss for consistent oriented bounding box regression in aerial images"]}],"split":"dev"}
{"episode_id":"F012_T2024-07","field_index":12,"field":"object detection","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Geometric 3D and sensor-fusion detection","description":"Detect objects by constructing 3D, BEV, range, radar, or cross-sensor scene representations from cameras, LiDAR, radar, RGB-D, or related sensors.","inclusion":"Include work whose primary objective is geometric 3D object detection, camera-only depth/BEV reasoning, single-sensor point-cloud/radar representation learning, or inference-time fusion of complementary sensors for 2D/3D detection.","exclusion":"If the main objective is label efficiency or domain transfer for 3D detectors, assign to D1; if the main objective is generic detector architecture or compression independent of 3D/sensor geometry, assign to D3.","nearest_boundary":"D1","representative_topics":["unified voxel-space multi-modality fusion for 3d object detection and tracking","depth-supervised bev 3d object detection with camera-aware depth estimation and refinement","sparse proposal-based end-to-end 3d object detection from point clouds using transformers"]},{"direction_index":1,"label":"D1","name":"Data-efficient and transfer learning","description":"Train detectors under scarce, weak, noisy, unlabeled, synthetic, or shifted data by improving supervision, pseudo-labeling, domain adaptation, and transfer protocols.","inclusion":"Include work whose primary objective is reducing annotation cost, repairing noisy labels, learning from weak or absent labels, semi-supervised 2D/3D detection, closed-set domain adaptation/generalization, synthetic-to-real transfer, or data-efficient 3D transfer.","exclusion":"If the main objective is novel-category recognition beyond a fixed label space, assign to D2; if the main objective is deployment compression or architectural efficiency rather than data/supervision efficiency, assign to D3.","nearest_boundary":"D2","representative_topics":["cross-domain detection transformer adaptation via spatial and semantic token alignment","weakly semi-supervised object detection via point annotation matching and instance learning","semi-supervised object detection with label matching via distribution-aware and instance-level self-training"]},{"direction_index":2,"label":"D2","name":"Open and novel category detection","description":"Extend detectors to categories that are unknown, novel, rare, few-shot, zero-shot, open-set, open-world, or language-defined.","inclusion":"Include work whose primary objective is category-space expansion through few-shot transfer, open-set/open-world formulations, open-vocabulary or vision-language alignment, incremental category learning, or base-to-novel generalization.","exclusion":"If categories remain fixed and the problem is target-domain shift or missing labels, assign to D1; if the work mainly improves generic DETR/CNN detector mechanics without category expansion, assign to D3.","nearest_boundary":"D1","representative_topics":["unifying multi-dataset object detection via language embedding query and category alignment","few-shot copy-paste augmentation using occlusion relationships for rare category detection","open vocabulary object detection with proposal mining and causal backdoor adjustment for novel categories"]},{"direction_index":3,"label":"D3","name":"Detector architecture and deployment efficiency","description":"Improve general-purpose detector mechanisms, training recipes, model compression, distillation, and speed-accuracy tradeoffs for deployable detection systems.","inclusion":"Include work whose primary objective is CNN/dense/single-stage/two-stage architecture design, feature pyramids, scale handling, DETR query/matching/attention mechanisms, lightweighting, pruning, quantization, distillation, or resource-aware inference.","exclusion":"If transformer or CNN mechanisms are subordinated to 3D geometry, sensor fusion, remote sensing orientation, or domain-specific clinical/maritime goals, assign to the relevant objective-specific direction instead.","nearest_boundary":"D0","representative_topics":["lightweight convolutional neck design for real-time object detection on edge devices","unified transformer framework for joint object detection and instance, panoptic, and semantic segmentation","multi-scale feature pyramid integration for transformer-based object detection"]},{"direction_index":4,"label":"D4","name":"Trustworthiness and evaluation methodology","description":"Assess, quantify, benchmark, calibrate, or improve detector reliability, security, uncertainty, robustness, and evaluation validity.","inclusion":"Include work whose primary objective is adversarial attack or defense, corruption/OOD robustness, uncertainty estimation, confidence calibration, metric design, benchmark methodology, dataset-composition analysis, annotation protocol, or diagnostic evaluation.","exclusion":"If the primary challenge is natural sensing degradation such as low light, infrared small targets, underwater optics, or adverse weather, assign to D5; if the work mainly proposes a detector architecture, assign to D3.","nearest_boundary":"D5","representative_topics":["efficient mimo uncertainty estimation for lidar-based 3d object detection under runtime constraints","longitudinal error-tolerant evaluation metrics for camera-only 3d object detection","parallelized mean average precision computation using vectorized operations for object detection"]},{"direction_index":5,"label":"D5","name":"Degraded and ambiguous appearance detection","description":"Detect objects whose visual evidence is weak, concealed, degraded, low-contrast, infrared, thermal, underwater, maritime, or otherwise difficult to separate from background.","inclusion":"Include work whose primary objective is detection under adverse visibility, low light, snow/rain/fog, infrared small-target conditions, RGB-thermal or RGB-D robustness, underwater/maritime optical degradation, camouflage, saliency, transparent/glass ambiguity, or unlabeled objectness discovery driven by foreground-background ambiguity.","exclusion":"If the primary objective is formal adversarial robustness or evaluation methodology, assign to D4; if the primary objective is overhead small/oriented remote-sensing geometry rather than degraded appearance, assign to D7.","nearest_boundary":"D4","representative_topics":["dark vessel detection in sar imagery using large-scale annotated maritime object datasets","lightweight yolo-based object detection with cross fusion for real-world snow conditions","axial-attention and multi-scale patch fusion for infrared small target detection"]},{"direction_index":6,"label":"D6","name":"Temporal and streaming detection","description":"Exploit time, motion, sequence structure, frame propagation, streaming constraints, or event streams to improve detection beyond independent single-image inference.","inclusion":"Include work whose primary objective is video object detection, temporal feature aggregation, motion-aware refinement, bounding-box propagation, streaming perception, event-camera detection, frame selection, temporal consistency, or latency-aware detection over sequences.","exclusion":"If temporal cues are used mainly for camera-only 3D geometry, assign to D0; if the application domain, such as clinical video or maritime surveillance, is the dominant research objective, assign to D7 or D5 respectively.","nearest_boundary":"D0","representative_topics":["spatio-temporal aggregation for insubstantial object detection in video","bounding box propagation for fast vehicle detection and tracking in fisheye traffic video","coarse-to-fine video-based smoky vehicle detection with 3d spatial-temporal refinement"]},{"direction_index":7,"label":"D7","name":"Domain-structured target detection","description":"Solve detection problems where specialized image domains impose distinctive target geometry, scale, orientation, morphology, operational constraints, or domain-specific target definitions.","inclusion":"Include work whose primary objective is detection in overhead/aerial/UAV/remote-sensing imagery, oriented-box remote sensing, high-resolution small-object overhead detection, clinical and biomedical object localization, or other specialized operational domains where the domain structure defines the detection problem.","exclusion":"If the domain challenge is primarily natural visibility degradation, underwater/maritime optics, infrared small targets, camouflage, or RGB-thermal robustness, assign to D5; if the contribution is a broadly general detector mechanism independent of the domain, assign to D3.","nearest_boundary":"D5","representative_topics":["self-supervised pretraining for label-efficient chest x-ray anomaly object detection","knowledge distillation for lightweight oriented object detection on aerial images","modulated kalman iou loss for consistent oriented bounding box regression in aerial images"]}],"split":"dev"}
{"episode_id":"F012_T2025-01","field_index":12,"field":"object detection","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Geometric 3D and sensor-fusion detection","description":"Detect objects by constructing 3D, BEV, range, radar, or cross-sensor scene representations from cameras, LiDAR, radar, RGB-D, or related sensors.","inclusion":"Include work whose primary objective is geometric 3D object detection, camera-only depth/BEV reasoning, single-sensor point-cloud/radar representation learning, or inference-time fusion of complementary sensors for 2D/3D detection.","exclusion":"If the main objective is label efficiency or domain transfer for 3D detectors, assign to D1; if the main objective is generic detector architecture or compression independent of 3D/sensor geometry, assign to D3.","nearest_boundary":"D1","representative_topics":["unified voxel-space multi-modality fusion for 3d object detection and tracking","depth-supervised bev 3d object detection with camera-aware depth estimation and refinement","sparse proposal-based end-to-end 3d object detection from point clouds using transformers"]},{"direction_index":1,"label":"D1","name":"Data-efficient and transfer learning","description":"Train detectors under scarce, weak, noisy, unlabeled, synthetic, or shifted data by improving supervision, pseudo-labeling, domain adaptation, and transfer protocols.","inclusion":"Include work whose primary objective is reducing annotation cost, repairing noisy labels, learning from weak or absent labels, semi-supervised 2D/3D detection, closed-set domain adaptation/generalization, synthetic-to-real transfer, or data-efficient 3D transfer.","exclusion":"If the main objective is novel-category recognition beyond a fixed label space, assign to D2; if the main objective is deployment compression or architectural efficiency rather than data/supervision efficiency, assign to D3.","nearest_boundary":"D2","representative_topics":["cross-domain detection transformer adaptation via spatial and semantic token alignment","weakly semi-supervised object detection via point annotation matching and instance learning","semi-supervised object detection with label matching via distribution-aware and instance-level self-training"]},{"direction_index":2,"label":"D2","name":"Open and novel category detection","description":"Extend detectors to categories that are unknown, novel, rare, few-shot, zero-shot, open-set, open-world, or language-defined.","inclusion":"Include work whose primary objective is category-space expansion through few-shot transfer, open-set/open-world formulations, open-vocabulary or vision-language alignment, incremental category learning, or base-to-novel generalization.","exclusion":"If categories remain fixed and the problem is target-domain shift or missing labels, assign to D1; if the work mainly improves generic DETR/CNN detector mechanics without category expansion, assign to D3.","nearest_boundary":"D1","representative_topics":["unifying multi-dataset object detection via language embedding query and category alignment","few-shot copy-paste augmentation using occlusion relationships for rare category detection","open vocabulary object detection with proposal mining and causal backdoor adjustment for novel categories"]},{"direction_index":3,"label":"D3","name":"Detector architecture and deployment efficiency","description":"Improve general-purpose detector mechanisms, training recipes, model compression, distillation, and speed-accuracy tradeoffs for deployable detection systems.","inclusion":"Include work whose primary objective is CNN/dense/single-stage/two-stage architecture design, feature pyramids, scale handling, DETR query/matching/attention mechanisms, lightweighting, pruning, quantization, distillation, or resource-aware inference.","exclusion":"If transformer or CNN mechanisms are subordinated to 3D geometry, sensor fusion, remote sensing orientation, or domain-specific clinical/maritime goals, assign to the relevant objective-specific direction instead.","nearest_boundary":"D0","representative_topics":["lightweight convolutional neck design for real-time object detection on edge devices","unified transformer framework for joint object detection and instance, panoptic, and semantic segmentation","multi-scale feature pyramid integration for transformer-based object detection"]},{"direction_index":4,"label":"D4","name":"Trustworthiness and evaluation methodology","description":"Assess, quantify, benchmark, calibrate, or improve detector reliability, security, uncertainty, robustness, and evaluation validity.","inclusion":"Include work whose primary objective is adversarial attack or defense, corruption/OOD robustness, uncertainty estimation, confidence calibration, metric design, benchmark methodology, dataset-composition analysis, annotation protocol, or diagnostic evaluation.","exclusion":"If the primary challenge is natural sensing degradation such as low light, infrared small targets, underwater optics, or adverse weather, assign to D5; if the work mainly proposes a detector architecture, assign to D3.","nearest_boundary":"D5","representative_topics":["efficient mimo uncertainty estimation for lidar-based 3d object detection under runtime constraints","longitudinal error-tolerant evaluation metrics for camera-only 3d object detection","parallelized mean average precision computation using vectorized operations for object detection"]},{"direction_index":5,"label":"D5","name":"Degraded and ambiguous appearance detection","description":"Detect objects whose visual evidence is weak, concealed, degraded, low-contrast, infrared, thermal, underwater, maritime, or otherwise difficult to separate from background.","inclusion":"Include work whose primary objective is detection under adverse visibility, low light, snow/rain/fog, infrared small-target conditions, RGB-thermal or RGB-D robustness, underwater/maritime optical degradation, camouflage, saliency, transparent/glass ambiguity, or unlabeled objectness discovery driven by foreground-background ambiguity.","exclusion":"If the primary objective is formal adversarial robustness or evaluation methodology, assign to D4; if the primary objective is overhead small/oriented remote-sensing geometry rather than degraded appearance, assign to D7.","nearest_boundary":"D4","representative_topics":["dark vessel detection in sar imagery using large-scale annotated maritime object datasets","lightweight yolo-based object detection with cross fusion for real-world snow conditions","axial-attention and multi-scale patch fusion for infrared small target detection"]},{"direction_index":6,"label":"D6","name":"Temporal and streaming detection","description":"Exploit time, motion, sequence structure, frame propagation, streaming constraints, or event streams to improve detection beyond independent single-image inference.","inclusion":"Include work whose primary objective is video object detection, temporal feature aggregation, motion-aware refinement, bounding-box propagation, streaming perception, event-camera detection, frame selection, temporal consistency, or latency-aware detection over sequences.","exclusion":"If temporal cues are used mainly for camera-only 3D geometry, assign to D0; if the application domain, such as clinical video or maritime surveillance, is the dominant research objective, assign to D7 or D5 respectively.","nearest_boundary":"D0","representative_topics":["spatio-temporal aggregation for insubstantial object detection in video","bounding box propagation for fast vehicle detection and tracking in fisheye traffic video","coarse-to-fine video-based smoky vehicle detection with 3d spatial-temporal refinement"]},{"direction_index":7,"label":"D7","name":"Domain-structured target detection","description":"Solve detection problems where specialized image domains impose distinctive target geometry, scale, orientation, morphology, operational constraints, or domain-specific target definitions.","inclusion":"Include work whose primary objective is detection in overhead/aerial/UAV/remote-sensing imagery, oriented-box remote sensing, high-resolution small-object overhead detection, clinical and biomedical object localization, or other specialized operational domains where the domain structure defines the detection problem.","exclusion":"If the domain challenge is primarily natural visibility degradation, underwater/maritime optics, infrared small targets, camouflage, or RGB-thermal robustness, assign to D5; if the contribution is a broadly general detector mechanism independent of the domain, assign to D3.","nearest_boundary":"D5","representative_topics":["self-supervised pretraining for label-efficient chest x-ray anomaly object detection","knowledge distillation for lightweight oriented object detection on aerial images","modulated kalman iou loss for consistent oriented bounding box regression in aerial images"]}],"split":"dev"}
{"episode_id":"F012_T2025-07","field_index":12,"field":"object detection","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Geometric 3D and sensor-fusion detection","description":"Detect objects by constructing 3D, BEV, range, radar, or cross-sensor scene representations from cameras, LiDAR, radar, RGB-D, or related sensors.","inclusion":"Include work whose primary objective is geometric 3D object detection, camera-only depth/BEV reasoning, single-sensor point-cloud/radar representation learning, or inference-time fusion of complementary sensors for 2D/3D detection.","exclusion":"If the main objective is label efficiency or domain transfer for 3D detectors, assign to D1; if the main objective is generic detector architecture or compression independent of 3D/sensor geometry, assign to D3.","nearest_boundary":"D1","representative_topics":["unified voxel-space multi-modality fusion for 3d object detection and tracking","depth-supervised bev 3d object detection with camera-aware depth estimation and refinement","sparse proposal-based end-to-end 3d object detection from point clouds using transformers"]},{"direction_index":1,"label":"D1","name":"Data-efficient and transfer learning","description":"Train detectors under scarce, weak, noisy, unlabeled, synthetic, or shifted data by improving supervision, pseudo-labeling, domain adaptation, and transfer protocols.","inclusion":"Include work whose primary objective is reducing annotation cost, repairing noisy labels, learning from weak or absent labels, semi-supervised 2D/3D detection, closed-set domain adaptation/generalization, synthetic-to-real transfer, or data-efficient 3D transfer.","exclusion":"If the main objective is novel-category recognition beyond a fixed label space, assign to D2; if the main objective is deployment compression or architectural efficiency rather than data/supervision efficiency, assign to D3.","nearest_boundary":"D2","representative_topics":["cross-domain detection transformer adaptation via spatial and semantic token alignment","weakly semi-supervised object detection via point annotation matching and instance learning","semi-supervised object detection with label matching via distribution-aware and instance-level self-training"]},{"direction_index":2,"label":"D2","name":"Open and novel category detection","description":"Extend detectors to categories that are unknown, novel, rare, few-shot, zero-shot, open-set, open-world, or language-defined.","inclusion":"Include work whose primary objective is category-space expansion through few-shot transfer, open-set/open-world formulations, open-vocabulary or vision-language alignment, incremental category learning, or base-to-novel generalization.","exclusion":"If categories remain fixed and the problem is target-domain shift or missing labels, assign to D1; if the work mainly improves generic DETR/CNN detector mechanics without category expansion, assign to D3.","nearest_boundary":"D1","representative_topics":["unifying multi-dataset object detection via language embedding query and category alignment","few-shot copy-paste augmentation using occlusion relationships for rare category detection","open vocabulary object detection with proposal mining and causal backdoor adjustment for novel categories"]},{"direction_index":3,"label":"D3","name":"Detector architecture and deployment efficiency","description":"Improve general-purpose detector mechanisms, training recipes, model compression, distillation, and speed-accuracy tradeoffs for deployable detection systems.","inclusion":"Include work whose primary objective is CNN/dense/single-stage/two-stage architecture design, feature pyramids, scale handling, DETR query/matching/attention mechanisms, lightweighting, pruning, quantization, distillation, or resource-aware inference.","exclusion":"If transformer or CNN mechanisms are subordinated to 3D geometry, sensor fusion, remote sensing orientation, or domain-specific clinical/maritime goals, assign to the relevant objective-specific direction instead.","nearest_boundary":"D0","representative_topics":["lightweight convolutional neck design for real-time object detection on edge devices","unified transformer framework for joint object detection and instance, panoptic, and semantic segmentation","multi-scale feature pyramid integration for transformer-based object detection"]},{"direction_index":4,"label":"D4","name":"Trustworthiness and evaluation methodology","description":"Assess, quantify, benchmark, calibrate, or improve detector reliability, security, uncertainty, robustness, and evaluation validity.","inclusion":"Include work whose primary objective is adversarial attack or defense, corruption/OOD robustness, uncertainty estimation, confidence calibration, metric design, benchmark methodology, dataset-composition analysis, annotation protocol, or diagnostic evaluation.","exclusion":"If the primary challenge is natural sensing degradation such as low light, infrared small targets, underwater optics, or adverse weather, assign to D5; if the work mainly proposes a detector architecture, assign to D3.","nearest_boundary":"D5","representative_topics":["efficient mimo uncertainty estimation for lidar-based 3d object detection under runtime constraints","longitudinal error-tolerant evaluation metrics for camera-only 3d object detection","parallelized mean average precision computation using vectorized operations for object detection"]},{"direction_index":5,"label":"D5","name":"Degraded and ambiguous appearance detection","description":"Detect objects whose visual evidence is weak, concealed, degraded, low-contrast, infrared, thermal, underwater, maritime, or otherwise difficult to separate from background.","inclusion":"Include work whose primary objective is detection under adverse visibility, low light, snow/rain/fog, infrared small-target conditions, RGB-thermal or RGB-D robustness, underwater/maritime optical degradation, camouflage, saliency, transparent/glass ambiguity, or unlabeled objectness discovery driven by foreground-background ambiguity.","exclusion":"If the primary objective is formal adversarial robustness or evaluation methodology, assign to D4; if the primary objective is overhead small/oriented remote-sensing geometry rather than degraded appearance, assign to D7.","nearest_boundary":"D4","representative_topics":["dark vessel detection in sar imagery using large-scale annotated maritime object datasets","lightweight yolo-based object detection with cross fusion for real-world snow conditions","axial-attention and multi-scale patch fusion for infrared small target detection"]},{"direction_index":6,"label":"D6","name":"Temporal and streaming detection","description":"Exploit time, motion, sequence structure, frame propagation, streaming constraints, or event streams to improve detection beyond independent single-image inference.","inclusion":"Include work whose primary objective is video object detection, temporal feature aggregation, motion-aware refinement, bounding-box propagation, streaming perception, event-camera detection, frame selection, temporal consistency, or latency-aware detection over sequences.","exclusion":"If temporal cues are used mainly for camera-only 3D geometry, assign to D0; if the application domain, such as clinical video or maritime surveillance, is the dominant research objective, assign to D7 or D5 respectively.","nearest_boundary":"D0","representative_topics":["spatio-temporal aggregation for insubstantial object detection in video","bounding box propagation for fast vehicle detection and tracking in fisheye traffic video","coarse-to-fine video-based smoky vehicle detection with 3d spatial-temporal refinement"]},{"direction_index":7,"label":"D7","name":"Domain-structured target detection","description":"Solve detection problems where specialized image domains impose distinctive target geometry, scale, orientation, morphology, operational constraints, or domain-specific target definitions.","inclusion":"Include work whose primary objective is detection in overhead/aerial/UAV/remote-sensing imagery, oriented-box remote sensing, high-resolution small-object overhead detection, clinical and biomedical object localization, or other specialized operational domains where the domain structure defines the detection problem.","exclusion":"If the domain challenge is primarily natural visibility degradation, underwater/maritime optics, infrared small targets, camouflage, or RGB-thermal robustness, assign to D5; if the contribution is a broadly general detector mechanism independent of the domain, assign to D3.","nearest_boundary":"D5","representative_topics":["self-supervised pretraining for label-efficient chest x-ray anomaly object detection","knowledge distillation for lightweight oriented object detection on aerial images","modulated kalman iou loss for consistent oriented bounding box regression in aerial images"]}],"split":"dev"}
{"episode_id":"F012_T2026-01","field_index":12,"field":"object detection","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Geometric 3D and sensor-fusion detection","description":"Detect objects by constructing 3D, BEV, range, radar, or cross-sensor scene representations from cameras, LiDAR, radar, RGB-D, or related sensors.","inclusion":"Include work whose primary objective is geometric 3D object detection, camera-only depth/BEV reasoning, single-sensor point-cloud/radar representation learning, or inference-time fusion of complementary sensors for 2D/3D detection.","exclusion":"If the main objective is label efficiency or domain transfer for 3D detectors, assign to D1; if the main objective is generic detector architecture or compression independent of 3D/sensor geometry, assign to D3.","nearest_boundary":"D1","representative_topics":["unified voxel-space multi-modality fusion for 3d object detection and tracking","depth-supervised bev 3d object detection with camera-aware depth estimation and refinement","sparse proposal-based end-to-end 3d object detection from point clouds using transformers"]},{"direction_index":1,"label":"D1","name":"Data-efficient and transfer learning","description":"Train detectors under scarce, weak, noisy, unlabeled, synthetic, or shifted data by improving supervision, pseudo-labeling, domain adaptation, and transfer protocols.","inclusion":"Include work whose primary objective is reducing annotation cost, repairing noisy labels, learning from weak or absent labels, semi-supervised 2D/3D detection, closed-set domain adaptation/generalization, synthetic-to-real transfer, or data-efficient 3D transfer.","exclusion":"If the main objective is novel-category recognition beyond a fixed label space, assign to D2; if the main objective is deployment compression or architectural efficiency rather than data/supervision efficiency, assign to D3.","nearest_boundary":"D2","representative_topics":["cross-domain detection transformer adaptation via spatial and semantic token alignment","weakly semi-supervised object detection via point annotation matching and instance learning","semi-supervised object detection with label matching via distribution-aware and instance-level self-training"]},{"direction_index":2,"label":"D2","name":"Open and novel category detection","description":"Extend detectors to categories that are unknown, novel, rare, few-shot, zero-shot, open-set, open-world, or language-defined.","inclusion":"Include work whose primary objective is category-space expansion through few-shot transfer, open-set/open-world formulations, open-vocabulary or vision-language alignment, incremental category learning, or base-to-novel generalization.","exclusion":"If categories remain fixed and the problem is target-domain shift or missing labels, assign to D1; if the work mainly improves generic DETR/CNN detector mechanics without category expansion, assign to D3.","nearest_boundary":"D1","representative_topics":["unifying multi-dataset object detection via language embedding query and category alignment","few-shot copy-paste augmentation using occlusion relationships for rare category detection","open vocabulary object detection with proposal mining and causal backdoor adjustment for novel categories"]},{"direction_index":3,"label":"D3","name":"Detector architecture and deployment efficiency","description":"Improve general-purpose detector mechanisms, training recipes, model compression, distillation, and speed-accuracy tradeoffs for deployable detection systems.","inclusion":"Include work whose primary objective is CNN/dense/single-stage/two-stage architecture design, feature pyramids, scale handling, DETR query/matching/attention mechanisms, lightweighting, pruning, quantization, distillation, or resource-aware inference.","exclusion":"If transformer or CNN mechanisms are subordinated to 3D geometry, sensor fusion, remote sensing orientation, or domain-specific clinical/maritime goals, assign to the relevant objective-specific direction instead.","nearest_boundary":"D0","representative_topics":["lightweight convolutional neck design for real-time object detection on edge devices","unified transformer framework for joint object detection and instance, panoptic, and semantic segmentation","multi-scale feature pyramid integration for transformer-based object detection"]},{"direction_index":4,"label":"D4","name":"Trustworthiness and evaluation methodology","description":"Assess, quantify, benchmark, calibrate, or improve detector reliability, security, uncertainty, robustness, and evaluation validity.","inclusion":"Include work whose primary objective is adversarial attack or defense, corruption/OOD robustness, uncertainty estimation, confidence calibration, metric design, benchmark methodology, dataset-composition analysis, annotation protocol, or diagnostic evaluation.","exclusion":"If the primary challenge is natural sensing degradation such as low light, infrared small targets, underwater optics, or adverse weather, assign to D5; if the work mainly proposes a detector architecture, assign to D3.","nearest_boundary":"D5","representative_topics":["efficient mimo uncertainty estimation for lidar-based 3d object detection under runtime constraints","longitudinal error-tolerant evaluation metrics for camera-only 3d object detection","parallelized mean average precision computation using vectorized operations for object detection"]},{"direction_index":5,"label":"D5","name":"Degraded and ambiguous appearance detection","description":"Detect objects whose visual evidence is weak, concealed, degraded, low-contrast, infrared, thermal, underwater, maritime, or otherwise difficult to separate from background.","inclusion":"Include work whose primary objective is detection under adverse visibility, low light, snow/rain/fog, infrared small-target conditions, RGB-thermal or RGB-D robustness, underwater/maritime optical degradation, camouflage, saliency, transparent/glass ambiguity, or unlabeled objectness discovery driven by foreground-background ambiguity.","exclusion":"If the primary objective is formal adversarial robustness or evaluation methodology, assign to D4; if the primary objective is overhead small/oriented remote-sensing geometry rather than degraded appearance, assign to D7.","nearest_boundary":"D4","representative_topics":["dark vessel detection in sar imagery using large-scale annotated maritime object datasets","lightweight yolo-based object detection with cross fusion for real-world snow conditions","axial-attention and multi-scale patch fusion for infrared small target detection"]},{"direction_index":6,"label":"D6","name":"Temporal and streaming detection","description":"Exploit time, motion, sequence structure, frame propagation, streaming constraints, or event streams to improve detection beyond independent single-image inference.","inclusion":"Include work whose primary objective is video object detection, temporal feature aggregation, motion-aware refinement, bounding-box propagation, streaming perception, event-camera detection, frame selection, temporal consistency, or latency-aware detection over sequences.","exclusion":"If temporal cues are used mainly for camera-only 3D geometry, assign to D0; if the application domain, such as clinical video or maritime surveillance, is the dominant research objective, assign to D7 or D5 respectively.","nearest_boundary":"D0","representative_topics":["spatio-temporal aggregation for insubstantial object detection in video","bounding box propagation for fast vehicle detection and tracking in fisheye traffic video","coarse-to-fine video-based smoky vehicle detection with 3d spatial-temporal refinement"]},{"direction_index":7,"label":"D7","name":"Domain-structured target detection","description":"Solve detection problems where specialized image domains impose distinctive target geometry, scale, orientation, morphology, operational constraints, or domain-specific target definitions.","inclusion":"Include work whose primary objective is detection in overhead/aerial/UAV/remote-sensing imagery, oriented-box remote sensing, high-resolution small-object overhead detection, clinical and biomedical object localization, or other specialized operational domains where the domain structure defines the detection problem.","exclusion":"If the domain challenge is primarily natural visibility degradation, underwater/maritime optics, infrared small targets, camouflage, or RGB-thermal robustness, assign to D5; if the contribution is a broadly general detector mechanism independent of the domain, assign to D3.","nearest_boundary":"D5","representative_topics":["self-supervised pretraining for label-efficient chest x-ray anomaly object detection","knowledge distillation for lightweight oriented object detection on aerial images","modulated kalman iou loss for consistent oriented bounding box regression in aerial images"]}],"split":"dev"}
{"episode_id":"F014_T2024-01","field_index":14,"field":"adversarial robustness","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Physical-world evasion","description":"Constructing and evaluating adversarial attacks intended to remain effective under real-world sensing, printing, illumination, geometry, optics, motion, capture, or deployment constraints.","inclusion":"Include atoms whose primary objective is a physically realizable adversarial artifact or signal, such as patches, masks, camouflage, printed patterns, optical or light manipulation, 3D objects, or transformation-robust perturbations deployed against real sensors.","exclusion":"Digital-only evasion methods, query-limited black-box attacks, and defenses against physical attacks belong to the corresponding attack-method or defense program instead.","nearest_boundary":"Digital evasion attack generation","representative_topics":["light-based physical adversarial attacks on vision dnn models using laser spot optimization","physical adversarial attacks on dnns via optical zoom lens image scaling","black-box physical impersonation attacks on face recognition via transferable adversarial masks"]},{"direction_index":1,"label":"D1","name":"Digital evasion attack generation","description":"Developing digital test-time adversarial example generation methods, including white-box attacks, transfer-based black-box attacks, query-efficient attacks, decision-based attacks, and empirical analyses of digital evasion mechanisms.","inclusion":"Include atoms whose primary objective is to craft or improve digital evasion attacks through optimization, surrogate transfer, query reduction, gradient estimation, transformations, ensembles, feature manipulation, or model-mechanism exploitation.","exclusion":"Physically deployed attacks belong to physical-world evasion; training-time data or model corruption belongs to poisoning and backdoors; modality-specific work belongs to domain-constrained robustness when domain structure rather than attack methodology is primary.","nearest_boundary":"Physical-world evasion","representative_topics":["real-time transferable adversarial attacks on face recognition via feed-forward transformation networks","dark knowledge-augmented surrogate models for transferable adversarial example generation","saliency-guided black-box adversarial attacks for imperceptible perturbations"]},{"direction_index":2,"label":"D2","name":"Training-time poisoning and backdoors","description":"Studying adversarial manipulation of training data, labels, user profiles, graph training samples, federated updates, or model parameters, and defenses that preserve model integrity under such training-time compromise.","inclusion":"Include atoms whose primary objective is poisoning, label flipping, data insertion, profile pollution, backdoor or trojan injection, training-set sanitization, poisoning-robust learning, or certification against training-time corruption.","exclusion":"Test-time evasion without training compromise belongs to evasion attack generation; graph attacks belong to domain-constrained robustness when graph topology and message passing rather than the training-time threat surface is central.","nearest_boundary":"Domain-constrained robustness","representative_topics":["adversarial label noise attacks on support vector machines via optimal label-flipping strategies","defending sequential recommenders against substitution-based profile pollution via neighborhood sampling and adversarial training","training-time poisoning attacks on k-nearest neighbor classification via data insertion"]},{"direction_index":3,"label":"D3","name":"Robust model learning and design","description":"Improving robustness by changing how models are trained, adapted, distilled, regularized, architected, ensembled, represented, or robustified from pretrained models.","inclusion":"Include atoms whose primary objective is adversarial training, robust optimization, regularization, fast or stable training, robust distillation, pretrained-model adaptation, robust fine-tuning, robust subnetworks, intrinsic robust architectures, robust representations, ensembles, sparsity, quantization, curvature control, or architectural design principles.","exclusion":"Test-time-only detection or purification belongs to inference-time mitigation; formal guarantees belong to certification; benchmarks or theory without a new robust learning/design method belong to evaluation and foundations.","nearest_boundary":"Inference-time mitigation","representative_topics":["confidence threshold reduction for simultaneous adversarial robustness and generalization improvement","federated adversarial training for vision transformers with similarity-weighted aggregation under non-iid data","feature distillation for accuracy-preserving adversarial training with predictive feature retention"]},{"direction_index":4,"label":"D4","name":"Inference-time mitigation","description":"Defending deployed models at test time by detecting, rejecting, localizing, purifying, reconstructing, denoising, transforming, correcting, or dynamically routing adversarial inputs.","inclusion":"Include atoms whose primary objective is adversarial example detection, OOD-style detection for adversarial inputs, patch localization or removal, denoising and reconstruction purification, gradient or feature anomaly detection, moving-target defenses, plug-in preprocessing, or test-time correction without full robust retraining.","exclusion":"Training-procedure defenses belong to robust model learning and design; defenses with formal robustness guarantees belong to certification; attack papers against detectors belong to the relevant attack-generation program.","nearest_boundary":"Robust model learning and design","representative_topics":["attack-agnostic adversarial detection via statistical and curvature-based anomaly quantification","adversarial correction via intermediate layer feature analysis and conditional distribution matching","moving target defense against adversarial evasion using out-of-distribution detection and model pool rotation"]},{"direction_index":5,"label":"D5","name":"Certified and verifiable robustness","description":"Producing formal deterministic or probabilistic guarantees that a model is robust within specified perturbation sets, threat models, or certified regions.","inclusion":"Include atoms whose primary objective is randomized smoothing, deterministic verification, interval or polyhedral bounds, certified training, formal verification solvers, Lipschitz or bound-propagation certification, certified purification, or analysis of certificate soundness and limits.","exclusion":"Empirical defenses without guarantees belong to robust learning or inference-time mitigation; broad mathematical theory without model-level certificates belongs to evaluation and foundations.","nearest_boundary":"Robustness evaluation and foundations","representative_topics":["adversarial attacks on cascading certified robust ensembles with unsound robustness guarantees","classifier-specific certification of randomized smoothing to overcome dimensionality-induced robustness limits","relaxed neuron pruning for scalable certifiable robustness via linearity grafting"]},{"direction_index":6,"label":"D6","name":"Domain-constrained robustness","description":"Adapting attacks, defenses, training, evaluation, or analysis to domains and modalities whose structure fundamentally changes the robustness problem.","inclusion":"Include atoms whose primary objective is robustness for graph learning, NLP, large language or multimodal models, structured vision, medical imaging, object detection, segmentation, point clouds, cybersecurity, wireless and infrastructure systems, reinforcement learning, control, time series, recommender systems, audio, biometrics, or other specialized domains where domain constraints drive the contribution.","exclusion":"General-purpose attacks or defenses merely evaluated on a domain dataset belong to the relevant attack or defense program; physical deployment belongs to physical-world evasion when real-world artifact construction is primary.","nearest_boundary":"Digital evasion attack generation","representative_topics":["unsupervised graph structure refinement for robust graph neural network defense","camouflaged node injection attacks on gnns via ego-network distribution similarity","character-level adversarial defense via text normalization for hate speech and nli models"]},{"direction_index":7,"label":"D7","name":"Robustness evaluation and foundations","description":"Measuring, benchmarking, formalizing, explaining, and empirically or theoretically understanding adversarial robustness, including broader robustness to corruptions, semantic variation, perceptual shifts, and distribution shifts.","inclusion":"Include atoms whose primary objective is a benchmark, metric, dataset, protocol, survey, diagnostic tool, comparative evaluation, interpretability study, empirical understanding, theoretical bound, minimax or learnability analysis, geometric characterization, surrogate-risk theory, or reframing of robustness beyond standard norm-bounded perturbations.","exclusion":"New attack algorithms belong to evasion attack generation; new defenses belong to robust learning, inference-time mitigation, or certification; domain-specific adaptation belongs to domain-constrained robustness when the domain constraints are the main scientific objective.","nearest_boundary":"Certified and verifiable robustness","representative_topics":["joint adversarial and corruption robustness training with generalization preservation","revising adversarial robustness formulation to address perceptible and geometric adversarial vulnerability","theoretical foundations of adversarial reprogramming in relu networks via gradient flow analysis"]}],"split":"dev"}
{"episode_id":"F014_T2024-07","field_index":14,"field":"adversarial robustness","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Physical-world evasion","description":"Constructing and evaluating adversarial attacks intended to remain effective under real-world sensing, printing, illumination, geometry, optics, motion, capture, or deployment constraints.","inclusion":"Include atoms whose primary objective is a physically realizable adversarial artifact or signal, such as patches, masks, camouflage, printed patterns, optical or light manipulation, 3D objects, or transformation-robust perturbations deployed against real sensors.","exclusion":"Digital-only evasion methods, query-limited black-box attacks, and defenses against physical attacks belong to the corresponding attack-method or defense program instead.","nearest_boundary":"Digital evasion attack generation","representative_topics":["light-based physical adversarial attacks on vision dnn models using laser spot optimization","physical adversarial attacks on dnns via optical zoom lens image scaling","black-box physical impersonation attacks on face recognition via transferable adversarial masks"]},{"direction_index":1,"label":"D1","name":"Digital evasion attack generation","description":"Developing digital test-time adversarial example generation methods, including white-box attacks, transfer-based black-box attacks, query-efficient attacks, decision-based attacks, and empirical analyses of digital evasion mechanisms.","inclusion":"Include atoms whose primary objective is to craft or improve digital evasion attacks through optimization, surrogate transfer, query reduction, gradient estimation, transformations, ensembles, feature manipulation, or model-mechanism exploitation.","exclusion":"Physically deployed attacks belong to physical-world evasion; training-time data or model corruption belongs to poisoning and backdoors; modality-specific work belongs to domain-constrained robustness when domain structure rather than attack methodology is primary.","nearest_boundary":"Physical-world evasion","representative_topics":["real-time transferable adversarial attacks on face recognition via feed-forward transformation networks","dark knowledge-augmented surrogate models for transferable adversarial example generation","saliency-guided black-box adversarial attacks for imperceptible perturbations"]},{"direction_index":2,"label":"D2","name":"Training-time poisoning and backdoors","description":"Studying adversarial manipulation of training data, labels, user profiles, graph training samples, federated updates, or model parameters, and defenses that preserve model integrity under such training-time compromise.","inclusion":"Include atoms whose primary objective is poisoning, label flipping, data insertion, profile pollution, backdoor or trojan injection, training-set sanitization, poisoning-robust learning, or certification against training-time corruption.","exclusion":"Test-time evasion without training compromise belongs to evasion attack generation; graph attacks belong to domain-constrained robustness when graph topology and message passing rather than the training-time threat surface is central.","nearest_boundary":"Domain-constrained robustness","representative_topics":["adversarial label noise attacks on support vector machines via optimal label-flipping strategies","defending sequential recommenders against substitution-based profile pollution via neighborhood sampling and adversarial training","training-time poisoning attacks on k-nearest neighbor classification via data insertion"]},{"direction_index":3,"label":"D3","name":"Robust model learning and design","description":"Improving robustness by changing how models are trained, adapted, distilled, regularized, architected, ensembled, represented, or robustified from pretrained models.","inclusion":"Include atoms whose primary objective is adversarial training, robust optimization, regularization, fast or stable training, robust distillation, pretrained-model adaptation, robust fine-tuning, robust subnetworks, intrinsic robust architectures, robust representations, ensembles, sparsity, quantization, curvature control, or architectural design principles.","exclusion":"Test-time-only detection or purification belongs to inference-time mitigation; formal guarantees belong to certification; benchmarks or theory without a new robust learning/design method belong to evaluation and foundations.","nearest_boundary":"Inference-time mitigation","representative_topics":["confidence threshold reduction for simultaneous adversarial robustness and generalization improvement","federated adversarial training for vision transformers with similarity-weighted aggregation under non-iid data","feature distillation for accuracy-preserving adversarial training with predictive feature retention"]},{"direction_index":4,"label":"D4","name":"Inference-time mitigation","description":"Defending deployed models at test time by detecting, rejecting, localizing, purifying, reconstructing, denoising, transforming, correcting, or dynamically routing adversarial inputs.","inclusion":"Include atoms whose primary objective is adversarial example detection, OOD-style detection for adversarial inputs, patch localization or removal, denoising and reconstruction purification, gradient or feature anomaly detection, moving-target defenses, plug-in preprocessing, or test-time correction without full robust retraining.","exclusion":"Training-procedure defenses belong to robust model learning and design; defenses with formal robustness guarantees belong to certification; attack papers against detectors belong to the relevant attack-generation program.","nearest_boundary":"Robust model learning and design","representative_topics":["attack-agnostic adversarial detection via statistical and curvature-based anomaly quantification","adversarial correction via intermediate layer feature analysis and conditional distribution matching","moving target defense against adversarial evasion using out-of-distribution detection and model pool rotation"]},{"direction_index":5,"label":"D5","name":"Certified and verifiable robustness","description":"Producing formal deterministic or probabilistic guarantees that a model is robust within specified perturbation sets, threat models, or certified regions.","inclusion":"Include atoms whose primary objective is randomized smoothing, deterministic verification, interval or polyhedral bounds, certified training, formal verification solvers, Lipschitz or bound-propagation certification, certified purification, or analysis of certificate soundness and limits.","exclusion":"Empirical defenses without guarantees belong to robust learning or inference-time mitigation; broad mathematical theory without model-level certificates belongs to evaluation and foundations.","nearest_boundary":"Robustness evaluation and foundations","representative_topics":["adversarial attacks on cascading certified robust ensembles with unsound robustness guarantees","classifier-specific certification of randomized smoothing to overcome dimensionality-induced robustness limits","relaxed neuron pruning for scalable certifiable robustness via linearity grafting"]},{"direction_index":6,"label":"D6","name":"Domain-constrained robustness","description":"Adapting attacks, defenses, training, evaluation, or analysis to domains and modalities whose structure fundamentally changes the robustness problem.","inclusion":"Include atoms whose primary objective is robustness for graph learning, NLP, large language or multimodal models, structured vision, medical imaging, object detection, segmentation, point clouds, cybersecurity, wireless and infrastructure systems, reinforcement learning, control, time series, recommender systems, audio, biometrics, or other specialized domains where domain constraints drive the contribution.","exclusion":"General-purpose attacks or defenses merely evaluated on a domain dataset belong to the relevant attack or defense program; physical deployment belongs to physical-world evasion when real-world artifact construction is primary.","nearest_boundary":"Digital evasion attack generation","representative_topics":["unsupervised graph structure refinement for robust graph neural network defense","camouflaged node injection attacks on gnns via ego-network distribution similarity","character-level adversarial defense via text normalization for hate speech and nli models"]},{"direction_index":7,"label":"D7","name":"Robustness evaluation and foundations","description":"Measuring, benchmarking, formalizing, explaining, and empirically or theoretically understanding adversarial robustness, including broader robustness to corruptions, semantic variation, perceptual shifts, and distribution shifts.","inclusion":"Include atoms whose primary objective is a benchmark, metric, dataset, protocol, survey, diagnostic tool, comparative evaluation, interpretability study, empirical understanding, theoretical bound, minimax or learnability analysis, geometric characterization, surrogate-risk theory, or reframing of robustness beyond standard norm-bounded perturbations.","exclusion":"New attack algorithms belong to evasion attack generation; new defenses belong to robust learning, inference-time mitigation, or certification; domain-specific adaptation belongs to domain-constrained robustness when the domain constraints are the main scientific objective.","nearest_boundary":"Certified and verifiable robustness","representative_topics":["joint adversarial and corruption robustness training with generalization preservation","revising adversarial robustness formulation to address perceptible and geometric adversarial vulnerability","theoretical foundations of adversarial reprogramming in relu networks via gradient flow analysis"]}],"split":"dev"}
{"episode_id":"F014_T2025-01","field_index":14,"field":"adversarial robustness","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Physical-world evasion","description":"Constructing and evaluating adversarial attacks intended to remain effective under real-world sensing, printing, illumination, geometry, optics, motion, capture, or deployment constraints.","inclusion":"Include atoms whose primary objective is a physically realizable adversarial artifact or signal, such as patches, masks, camouflage, printed patterns, optical or light manipulation, 3D objects, or transformation-robust perturbations deployed against real sensors.","exclusion":"Digital-only evasion methods, query-limited black-box attacks, and defenses against physical attacks belong to the corresponding attack-method or defense program instead.","nearest_boundary":"Digital evasion attack generation","representative_topics":["light-based physical adversarial attacks on vision dnn models using laser spot optimization","physical adversarial attacks on dnns via optical zoom lens image scaling","black-box physical impersonation attacks on face recognition via transferable adversarial masks"]},{"direction_index":1,"label":"D1","name":"Digital evasion attack generation","description":"Developing digital test-time adversarial example generation methods, including white-box attacks, transfer-based black-box attacks, query-efficient attacks, decision-based attacks, and empirical analyses of digital evasion mechanisms.","inclusion":"Include atoms whose primary objective is to craft or improve digital evasion attacks through optimization, surrogate transfer, query reduction, gradient estimation, transformations, ensembles, feature manipulation, or model-mechanism exploitation.","exclusion":"Physically deployed attacks belong to physical-world evasion; training-time data or model corruption belongs to poisoning and backdoors; modality-specific work belongs to domain-constrained robustness when domain structure rather than attack methodology is primary.","nearest_boundary":"Physical-world evasion","representative_topics":["real-time transferable adversarial attacks on face recognition via feed-forward transformation networks","dark knowledge-augmented surrogate models for transferable adversarial example generation","saliency-guided black-box adversarial attacks for imperceptible perturbations"]},{"direction_index":2,"label":"D2","name":"Training-time poisoning and backdoors","description":"Studying adversarial manipulation of training data, labels, user profiles, graph training samples, federated updates, or model parameters, and defenses that preserve model integrity under such training-time compromise.","inclusion":"Include atoms whose primary objective is poisoning, label flipping, data insertion, profile pollution, backdoor or trojan injection, training-set sanitization, poisoning-robust learning, or certification against training-time corruption.","exclusion":"Test-time evasion without training compromise belongs to evasion attack generation; graph attacks belong to domain-constrained robustness when graph topology and message passing rather than the training-time threat surface is central.","nearest_boundary":"Domain-constrained robustness","representative_topics":["adversarial label noise attacks on support vector machines via optimal label-flipping strategies","defending sequential recommenders against substitution-based profile pollution via neighborhood sampling and adversarial training","training-time poisoning attacks on k-nearest neighbor classification via data insertion"]},{"direction_index":3,"label":"D3","name":"Robust model learning and design","description":"Improving robustness by changing how models are trained, adapted, distilled, regularized, architected, ensembled, represented, or robustified from pretrained models.","inclusion":"Include atoms whose primary objective is adversarial training, robust optimization, regularization, fast or stable training, robust distillation, pretrained-model adaptation, robust fine-tuning, robust subnetworks, intrinsic robust architectures, robust representations, ensembles, sparsity, quantization, curvature control, or architectural design principles.","exclusion":"Test-time-only detection or purification belongs to inference-time mitigation; formal guarantees belong to certification; benchmarks or theory without a new robust learning/design method belong to evaluation and foundations.","nearest_boundary":"Inference-time mitigation","representative_topics":["confidence threshold reduction for simultaneous adversarial robustness and generalization improvement","federated adversarial training for vision transformers with similarity-weighted aggregation under non-iid data","feature distillation for accuracy-preserving adversarial training with predictive feature retention"]},{"direction_index":4,"label":"D4","name":"Inference-time mitigation","description":"Defending deployed models at test time by detecting, rejecting, localizing, purifying, reconstructing, denoising, transforming, correcting, or dynamically routing adversarial inputs.","inclusion":"Include atoms whose primary objective is adversarial example detection, OOD-style detection for adversarial inputs, patch localization or removal, denoising and reconstruction purification, gradient or feature anomaly detection, moving-target defenses, plug-in preprocessing, or test-time correction without full robust retraining.","exclusion":"Training-procedure defenses belong to robust model learning and design; defenses with formal robustness guarantees belong to certification; attack papers against detectors belong to the relevant attack-generation program.","nearest_boundary":"Robust model learning and design","representative_topics":["attack-agnostic adversarial detection via statistical and curvature-based anomaly quantification","adversarial correction via intermediate layer feature analysis and conditional distribution matching","moving target defense against adversarial evasion using out-of-distribution detection and model pool rotation"]},{"direction_index":5,"label":"D5","name":"Certified and verifiable robustness","description":"Producing formal deterministic or probabilistic guarantees that a model is robust within specified perturbation sets, threat models, or certified regions.","inclusion":"Include atoms whose primary objective is randomized smoothing, deterministic verification, interval or polyhedral bounds, certified training, formal verification solvers, Lipschitz or bound-propagation certification, certified purification, or analysis of certificate soundness and limits.","exclusion":"Empirical defenses without guarantees belong to robust learning or inference-time mitigation; broad mathematical theory without model-level certificates belongs to evaluation and foundations.","nearest_boundary":"Robustness evaluation and foundations","representative_topics":["adversarial attacks on cascading certified robust ensembles with unsound robustness guarantees","classifier-specific certification of randomized smoothing to overcome dimensionality-induced robustness limits","relaxed neuron pruning for scalable certifiable robustness via linearity grafting"]},{"direction_index":6,"label":"D6","name":"Domain-constrained robustness","description":"Adapting attacks, defenses, training, evaluation, or analysis to domains and modalities whose structure fundamentally changes the robustness problem.","inclusion":"Include atoms whose primary objective is robustness for graph learning, NLP, large language or multimodal models, structured vision, medical imaging, object detection, segmentation, point clouds, cybersecurity, wireless and infrastructure systems, reinforcement learning, control, time series, recommender systems, audio, biometrics, or other specialized domains where domain constraints drive the contribution.","exclusion":"General-purpose attacks or defenses merely evaluated on a domain dataset belong to the relevant attack or defense program; physical deployment belongs to physical-world evasion when real-world artifact construction is primary.","nearest_boundary":"Digital evasion attack generation","representative_topics":["unsupervised graph structure refinement for robust graph neural network defense","camouflaged node injection attacks on gnns via ego-network distribution similarity","character-level adversarial defense via text normalization for hate speech and nli models"]},{"direction_index":7,"label":"D7","name":"Robustness evaluation and foundations","description":"Measuring, benchmarking, formalizing, explaining, and empirically or theoretically understanding adversarial robustness, including broader robustness to corruptions, semantic variation, perceptual shifts, and distribution shifts.","inclusion":"Include atoms whose primary objective is a benchmark, metric, dataset, protocol, survey, diagnostic tool, comparative evaluation, interpretability study, empirical understanding, theoretical bound, minimax or learnability analysis, geometric characterization, surrogate-risk theory, or reframing of robustness beyond standard norm-bounded perturbations.","exclusion":"New attack algorithms belong to evasion attack generation; new defenses belong to robust learning, inference-time mitigation, or certification; domain-specific adaptation belongs to domain-constrained robustness when the domain constraints are the main scientific objective.","nearest_boundary":"Certified and verifiable robustness","representative_topics":["joint adversarial and corruption robustness training with generalization preservation","revising adversarial robustness formulation to address perceptible and geometric adversarial vulnerability","theoretical foundations of adversarial reprogramming in relu networks via gradient flow analysis"]}],"split":"dev"}
{"episode_id":"F014_T2025-07","field_index":14,"field":"adversarial robustness","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Physical-world evasion","description":"Constructing and evaluating adversarial attacks intended to remain effective under real-world sensing, printing, illumination, geometry, optics, motion, capture, or deployment constraints.","inclusion":"Include atoms whose primary objective is a physically realizable adversarial artifact or signal, such as patches, masks, camouflage, printed patterns, optical or light manipulation, 3D objects, or transformation-robust perturbations deployed against real sensors.","exclusion":"Digital-only evasion methods, query-limited black-box attacks, and defenses against physical attacks belong to the corresponding attack-method or defense program instead.","nearest_boundary":"Digital evasion attack generation","representative_topics":["light-based physical adversarial attacks on vision dnn models using laser spot optimization","physical adversarial attacks on dnns via optical zoom lens image scaling","black-box physical impersonation attacks on face recognition via transferable adversarial masks"]},{"direction_index":1,"label":"D1","name":"Digital evasion attack generation","description":"Developing digital test-time adversarial example generation methods, including white-box attacks, transfer-based black-box attacks, query-efficient attacks, decision-based attacks, and empirical analyses of digital evasion mechanisms.","inclusion":"Include atoms whose primary objective is to craft or improve digital evasion attacks through optimization, surrogate transfer, query reduction, gradient estimation, transformations, ensembles, feature manipulation, or model-mechanism exploitation.","exclusion":"Physically deployed attacks belong to physical-world evasion; training-time data or model corruption belongs to poisoning and backdoors; modality-specific work belongs to domain-constrained robustness when domain structure rather than attack methodology is primary.","nearest_boundary":"Physical-world evasion","representative_topics":["real-time transferable adversarial attacks on face recognition via feed-forward transformation networks","dark knowledge-augmented surrogate models for transferable adversarial example generation","saliency-guided black-box adversarial attacks for imperceptible perturbations"]},{"direction_index":2,"label":"D2","name":"Training-time poisoning and backdoors","description":"Studying adversarial manipulation of training data, labels, user profiles, graph training samples, federated updates, or model parameters, and defenses that preserve model integrity under such training-time compromise.","inclusion":"Include atoms whose primary objective is poisoning, label flipping, data insertion, profile pollution, backdoor or trojan injection, training-set sanitization, poisoning-robust learning, or certification against training-time corruption.","exclusion":"Test-time evasion without training compromise belongs to evasion attack generation; graph attacks belong to domain-constrained robustness when graph topology and message passing rather than the training-time threat surface is central.","nearest_boundary":"Domain-constrained robustness","representative_topics":["adversarial label noise attacks on support vector machines via optimal label-flipping strategies","defending sequential recommenders against substitution-based profile pollution via neighborhood sampling and adversarial training","training-time poisoning attacks on k-nearest neighbor classification via data insertion"]},{"direction_index":3,"label":"D3","name":"Robust model learning and design","description":"Improving robustness by changing how models are trained, adapted, distilled, regularized, architected, ensembled, represented, or robustified from pretrained models.","inclusion":"Include atoms whose primary objective is adversarial training, robust optimization, regularization, fast or stable training, robust distillation, pretrained-model adaptation, robust fine-tuning, robust subnetworks, intrinsic robust architectures, robust representations, ensembles, sparsity, quantization, curvature control, or architectural design principles.","exclusion":"Test-time-only detection or purification belongs to inference-time mitigation; formal guarantees belong to certification; benchmarks or theory without a new robust learning/design method belong to evaluation and foundations.","nearest_boundary":"Inference-time mitigation","representative_topics":["confidence threshold reduction for simultaneous adversarial robustness and generalization improvement","federated adversarial training for vision transformers with similarity-weighted aggregation under non-iid data","feature distillation for accuracy-preserving adversarial training with predictive feature retention"]},{"direction_index":4,"label":"D4","name":"Inference-time mitigation","description":"Defending deployed models at test time by detecting, rejecting, localizing, purifying, reconstructing, denoising, transforming, correcting, or dynamically routing adversarial inputs.","inclusion":"Include atoms whose primary objective is adversarial example detection, OOD-style detection for adversarial inputs, patch localization or removal, denoising and reconstruction purification, gradient or feature anomaly detection, moving-target defenses, plug-in preprocessing, or test-time correction without full robust retraining.","exclusion":"Training-procedure defenses belong to robust model learning and design; defenses with formal robustness guarantees belong to certification; attack papers against detectors belong to the relevant attack-generation program.","nearest_boundary":"Robust model learning and design","representative_topics":["attack-agnostic adversarial detection via statistical and curvature-based anomaly quantification","adversarial correction via intermediate layer feature analysis and conditional distribution matching","moving target defense against adversarial evasion using out-of-distribution detection and model pool rotation"]},{"direction_index":5,"label":"D5","name":"Certified and verifiable robustness","description":"Producing formal deterministic or probabilistic guarantees that a model is robust within specified perturbation sets, threat models, or certified regions.","inclusion":"Include atoms whose primary objective is randomized smoothing, deterministic verification, interval or polyhedral bounds, certified training, formal verification solvers, Lipschitz or bound-propagation certification, certified purification, or analysis of certificate soundness and limits.","exclusion":"Empirical defenses without guarantees belong to robust learning or inference-time mitigation; broad mathematical theory without model-level certificates belongs to evaluation and foundations.","nearest_boundary":"Robustness evaluation and foundations","representative_topics":["adversarial attacks on cascading certified robust ensembles with unsound robustness guarantees","classifier-specific certification of randomized smoothing to overcome dimensionality-induced robustness limits","relaxed neuron pruning for scalable certifiable robustness via linearity grafting"]},{"direction_index":6,"label":"D6","name":"Domain-constrained robustness","description":"Adapting attacks, defenses, training, evaluation, or analysis to domains and modalities whose structure fundamentally changes the robustness problem.","inclusion":"Include atoms whose primary objective is robustness for graph learning, NLP, large language or multimodal models, structured vision, medical imaging, object detection, segmentation, point clouds, cybersecurity, wireless and infrastructure systems, reinforcement learning, control, time series, recommender systems, audio, biometrics, or other specialized domains where domain constraints drive the contribution.","exclusion":"General-purpose attacks or defenses merely evaluated on a domain dataset belong to the relevant attack or defense program; physical deployment belongs to physical-world evasion when real-world artifact construction is primary.","nearest_boundary":"Digital evasion attack generation","representative_topics":["unsupervised graph structure refinement for robust graph neural network defense","camouflaged node injection attacks on gnns via ego-network distribution similarity","character-level adversarial defense via text normalization for hate speech and nli models"]},{"direction_index":7,"label":"D7","name":"Robustness evaluation and foundations","description":"Measuring, benchmarking, formalizing, explaining, and empirically or theoretically understanding adversarial robustness, including broader robustness to corruptions, semantic variation, perceptual shifts, and distribution shifts.","inclusion":"Include atoms whose primary objective is a benchmark, metric, dataset, protocol, survey, diagnostic tool, comparative evaluation, interpretability study, empirical understanding, theoretical bound, minimax or learnability analysis, geometric characterization, surrogate-risk theory, or reframing of robustness beyond standard norm-bounded perturbations.","exclusion":"New attack algorithms belong to evasion attack generation; new defenses belong to robust learning, inference-time mitigation, or certification; domain-specific adaptation belongs to domain-constrained robustness when the domain constraints are the main scientific objective.","nearest_boundary":"Certified and verifiable robustness","representative_topics":["joint adversarial and corruption robustness training with generalization preservation","revising adversarial robustness formulation to address perceptible and geometric adversarial vulnerability","theoretical foundations of adversarial reprogramming in relu networks via gradient flow analysis"]}],"split":"dev"}
{"episode_id":"F014_T2026-01","field_index":14,"field":"adversarial robustness","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Physical-world evasion","description":"Constructing and evaluating adversarial attacks intended to remain effective under real-world sensing, printing, illumination, geometry, optics, motion, capture, or deployment constraints.","inclusion":"Include atoms whose primary objective is a physically realizable adversarial artifact or signal, such as patches, masks, camouflage, printed patterns, optical or light manipulation, 3D objects, or transformation-robust perturbations deployed against real sensors.","exclusion":"Digital-only evasion methods, query-limited black-box attacks, and defenses against physical attacks belong to the corresponding attack-method or defense program instead.","nearest_boundary":"Digital evasion attack generation","representative_topics":["light-based physical adversarial attacks on vision dnn models using laser spot optimization","physical adversarial attacks on dnns via optical zoom lens image scaling","black-box physical impersonation attacks on face recognition via transferable adversarial masks"]},{"direction_index":1,"label":"D1","name":"Digital evasion attack generation","description":"Developing digital test-time adversarial example generation methods, including white-box attacks, transfer-based black-box attacks, query-efficient attacks, decision-based attacks, and empirical analyses of digital evasion mechanisms.","inclusion":"Include atoms whose primary objective is to craft or improve digital evasion attacks through optimization, surrogate transfer, query reduction, gradient estimation, transformations, ensembles, feature manipulation, or model-mechanism exploitation.","exclusion":"Physically deployed attacks belong to physical-world evasion; training-time data or model corruption belongs to poisoning and backdoors; modality-specific work belongs to domain-constrained robustness when domain structure rather than attack methodology is primary.","nearest_boundary":"Physical-world evasion","representative_topics":["real-time transferable adversarial attacks on face recognition via feed-forward transformation networks","dark knowledge-augmented surrogate models for transferable adversarial example generation","saliency-guided black-box adversarial attacks for imperceptible perturbations"]},{"direction_index":2,"label":"D2","name":"Training-time poisoning and backdoors","description":"Studying adversarial manipulation of training data, labels, user profiles, graph training samples, federated updates, or model parameters, and defenses that preserve model integrity under such training-time compromise.","inclusion":"Include atoms whose primary objective is poisoning, label flipping, data insertion, profile pollution, backdoor or trojan injection, training-set sanitization, poisoning-robust learning, or certification against training-time corruption.","exclusion":"Test-time evasion without training compromise belongs to evasion attack generation; graph attacks belong to domain-constrained robustness when graph topology and message passing rather than the training-time threat surface is central.","nearest_boundary":"Domain-constrained robustness","representative_topics":["adversarial label noise attacks on support vector machines via optimal label-flipping strategies","defending sequential recommenders against substitution-based profile pollution via neighborhood sampling and adversarial training","training-time poisoning attacks on k-nearest neighbor classification via data insertion"]},{"direction_index":3,"label":"D3","name":"Robust model learning and design","description":"Improving robustness by changing how models are trained, adapted, distilled, regularized, architected, ensembled, represented, or robustified from pretrained models.","inclusion":"Include atoms whose primary objective is adversarial training, robust optimization, regularization, fast or stable training, robust distillation, pretrained-model adaptation, robust fine-tuning, robust subnetworks, intrinsic robust architectures, robust representations, ensembles, sparsity, quantization, curvature control, or architectural design principles.","exclusion":"Test-time-only detection or purification belongs to inference-time mitigation; formal guarantees belong to certification; benchmarks or theory without a new robust learning/design method belong to evaluation and foundations.","nearest_boundary":"Inference-time mitigation","representative_topics":["confidence threshold reduction for simultaneous adversarial robustness and generalization improvement","federated adversarial training for vision transformers with similarity-weighted aggregation under non-iid data","feature distillation for accuracy-preserving adversarial training with predictive feature retention"]},{"direction_index":4,"label":"D4","name":"Inference-time mitigation","description":"Defending deployed models at test time by detecting, rejecting, localizing, purifying, reconstructing, denoising, transforming, correcting, or dynamically routing adversarial inputs.","inclusion":"Include atoms whose primary objective is adversarial example detection, OOD-style detection for adversarial inputs, patch localization or removal, denoising and reconstruction purification, gradient or feature anomaly detection, moving-target defenses, plug-in preprocessing, or test-time correction without full robust retraining.","exclusion":"Training-procedure defenses belong to robust model learning and design; defenses with formal robustness guarantees belong to certification; attack papers against detectors belong to the relevant attack-generation program.","nearest_boundary":"Robust model learning and design","representative_topics":["attack-agnostic adversarial detection via statistical and curvature-based anomaly quantification","adversarial correction via intermediate layer feature analysis and conditional distribution matching","moving target defense against adversarial evasion using out-of-distribution detection and model pool rotation"]},{"direction_index":5,"label":"D5","name":"Certified and verifiable robustness","description":"Producing formal deterministic or probabilistic guarantees that a model is robust within specified perturbation sets, threat models, or certified regions.","inclusion":"Include atoms whose primary objective is randomized smoothing, deterministic verification, interval or polyhedral bounds, certified training, formal verification solvers, Lipschitz or bound-propagation certification, certified purification, or analysis of certificate soundness and limits.","exclusion":"Empirical defenses without guarantees belong to robust learning or inference-time mitigation; broad mathematical theory without model-level certificates belongs to evaluation and foundations.","nearest_boundary":"Robustness evaluation and foundations","representative_topics":["adversarial attacks on cascading certified robust ensembles with unsound robustness guarantees","classifier-specific certification of randomized smoothing to overcome dimensionality-induced robustness limits","relaxed neuron pruning for scalable certifiable robustness via linearity grafting"]},{"direction_index":6,"label":"D6","name":"Domain-constrained robustness","description":"Adapting attacks, defenses, training, evaluation, or analysis to domains and modalities whose structure fundamentally changes the robustness problem.","inclusion":"Include atoms whose primary objective is robustness for graph learning, NLP, large language or multimodal models, structured vision, medical imaging, object detection, segmentation, point clouds, cybersecurity, wireless and infrastructure systems, reinforcement learning, control, time series, recommender systems, audio, biometrics, or other specialized domains where domain constraints drive the contribution.","exclusion":"General-purpose attacks or defenses merely evaluated on a domain dataset belong to the relevant attack or defense program; physical deployment belongs to physical-world evasion when real-world artifact construction is primary.","nearest_boundary":"Digital evasion attack generation","representative_topics":["unsupervised graph structure refinement for robust graph neural network defense","camouflaged node injection attacks on gnns via ego-network distribution similarity","character-level adversarial defense via text normalization for hate speech and nli models"]},{"direction_index":7,"label":"D7","name":"Robustness evaluation and foundations","description":"Measuring, benchmarking, formalizing, explaining, and empirically or theoretically understanding adversarial robustness, including broader robustness to corruptions, semantic variation, perceptual shifts, and distribution shifts.","inclusion":"Include atoms whose primary objective is a benchmark, metric, dataset, protocol, survey, diagnostic tool, comparative evaluation, interpretability study, empirical understanding, theoretical bound, minimax or learnability analysis, geometric characterization, surrogate-risk theory, or reframing of robustness beyond standard norm-bounded perturbations.","exclusion":"New attack algorithms belong to evasion attack generation; new defenses belong to robust learning, inference-time mitigation, or certification; domain-specific adaptation belongs to domain-constrained robustness when the domain constraints are the main scientific objective.","nearest_boundary":"Certified and verifiable robustness","representative_topics":["joint adversarial and corruption robustness training with generalization preservation","revising adversarial robustness formulation to address perceptible and geometric adversarial vulnerability","theoretical foundations of adversarial reprogramming in relu networks via gradient flow analysis"]}],"split":"dev"}
{"episode_id":"F017_T2024-01","field_index":17,"field":"multi-agent reinforcement learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cooperative coordination mechanisms","description":"Learning scalable cooperative joint behavior through communication, credit assignment, relational or graph structure, CTDE/decentralized architectures, role or heterogeneity handling, and cooperative policy optimization.","inclusion":"Include source atoms whose primary objective is to make cooperative agents coordinate better under shared or largely aligned objectives by designing messages, value decompositions, centralized-training/decentralized-execution architectures, decentralized policy updates, graph/attention interaction models, role discovery, parameter sharing, or heterogeneous-agent cooperation mechanisms.","exclusion":"If the main objective is a game-theoretic solution concept, opponent/partner mental-state modeling, or incentives among self-interested agents, assign to strategic interaction; if the main objective is sparse-reward exploration, offline learning, transfer, or bandit regret, assign to data-efficient adaptation.","nearest_boundary":"D1","representative_topics":["decentralized multi-agent reinforcement learning via distribution matching to expert visitation distributions","universally expressive communication protocols in multi-agent reinforcement learning via gnn augmentation","policy gradient methods for cooperative marl beyond value decomposition and parameter sharing"]},{"direction_index":1,"label":"D1","name":"Strategic interaction and incentives","description":"Learning, analyzing, or shaping behavior in strategic multi-agent settings with distinct preferences, adaptive partners, equilibrium concepts, opponent models, theory of mind, social dilemmas, or incentive mechanisms.","inclusion":"Include source atoms whose primary objective is equilibrium computation or convergence in games, no-regret/self-play dynamics, opponent or partner modeling, intention or belief inference, ad hoc strategic adaptation, cooperation emergence, social dilemmas, contracts, markets, fairness, or incentives for mixed-motive agents.","exclusion":"If agents are treated primarily as a cooperative team with known aligned objectives and the contribution is a coordination architecture or policy optimizer, assign to cooperative coordination mechanisms; if the issue is malicious attacks or formal safety constraints, assign to trustworthy MARL.","nearest_boundary":"D0","representative_topics":["sample-efficient equilibrium learning in partially observable markov games via optimistic mle","convergence guarantees and price of anarchy bounds for softmax policy gradient in markov potential games","hierarchical reinforcement learning with opponent modeling for cooperative multi-agent control"]},{"direction_index":2,"label":"D2","name":"Data-efficient adaptation","description":"Improving learning efficiency and adaptability under sparse rewards, limited interaction, fixed data, regret constraints, changing tasks, unseen partners, or changing team composition.","inclusion":"Include source atoms whose primary objective is exploration, intrinsic motivation, curriculum or hierarchy for sparse feedback, offline MARL, model-based data-limited learning, transfer, task/team adaptation, zero-shot or ad hoc teamwork generalization, and multi-agent bandit regret or best-arm identification.","exclusion":"If the main contribution is a general cooperative coordination architecture trained online in a fixed team, assign to cooperative coordination mechanisms; if the main contribution is equilibrium theory rather than data limitation or adaptation, assign to strategic interaction and incentives.","nearest_boundary":"D0","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","offline nash equilibrium learning in zero-sum markov games via pessimistic model-based confidence bounds","adaptive curriculum goal generation via stein variational sampling for multi-goal reinforcement learning"]},{"direction_index":3,"label":"D3","name":"Trustworthy MARL","description":"Making MARL systems safe, robust, secure, privacy-preserving, interpretable, diagnosable, or compliant with formal/risk constraints.","inclusion":"Include source atoms whose primary objective is adversarial attack or defense, certified robustness, constrained or risk-aware policy learning, temporal-logic or formal safety, privacy-preserving or federated training, secure experience sharing, interpretability, explainability, symbolic reasoning, diagnostics, or novelty handling for trustworthiness.","exclusion":"If federation or safety is incidental to a telecom, energy, robotics, or mobility deployment, assign to the corresponding application-objective direction; if reward shaping is mainly for exploration rather than transparency or constraints, assign to data-efficient adaptation.","nearest_boundary":"D6","representative_topics":["reward poisoning attacks on offline multi-agent reinforcement learning via markov perfect equilibrium","near-optimal multi-agent coverage learning under unknown safety constraints and partial observability","primal-dual penalty reformulation for probabilistic constraints in constrained multiagent reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Large-population abstractions","description":"Scaling MARL to very large or infinite populations by replacing explicit many-agent interaction with population-level, mean-field, graphon, or collective-dynamics abstractions.","inclusion":"Include source atoms whose primary objective is tractability or analysis for large-population MARL through mean-field reinforcement learning, mean-field games, graphon games, major-minor population models, aggregate distribution dynamics, or population-level convergence and safety methods.","exclusion":"If scalability is pursued through finite-team communication, CTDE, graph neural networks, hierarchy, or parameter sharing, assign to cooperative coordination mechanisms; if the main objective is a domain system such as load balancing rather than the population abstraction itself, assign to networked infrastructure control.","nearest_boundary":"D0","representative_topics":["mean-field control for scalable load balancing under synchronization delay in parallel queuing systems","oracle-free online mean-field game learning via single-sample-path mean-field equilibrium approximation","scalable multi-agent reinforcement learning via mean-field and collective intelligence methods for large populations"]},{"direction_index":5,"label":"D5","name":"Evaluation infrastructure","description":"Building, standardizing, or synthesizing benchmarks, simulators, platforms, evaluation protocols, scaling studies, surveys, replications, and reproducibility resources for MARL.","inclusion":"Include source atoms whose primary objective is assessment infrastructure or empirical synthesis rather than a new MARL learning mechanism: benchmark environments, scalable simulators, evaluation suites, community challenges, reproducibility tools, scaling laws, surveys, and diagnostic taxonomies.","exclusion":"If an environment or simulator is introduced mainly to validate a new algorithmic mechanism, assign to that mechanism direction; if a benchmark is primarily a domain deployment objective such as transportation or robotics, assign to the relevant domain direction.","nearest_boundary":"D7","representative_topics":["scalable training and evaluation platform for multi-robot reinforcement learning","scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research","distributed local simulation with influence-aware models for scalable multi-agent reinforcement learning"]},{"direction_index":6,"label":"D6","name":"Networked infrastructure control","description":"Applying or adapting MARL to optimize resource allocation and operational control in communication, computing, energy, power-grid, queuing, IoT, and other networked infrastructure systems.","inclusion":"Include source atoms whose primary objective is a domain-defined infrastructure-control problem such as wireless channel access, spectrum or telecom resource management, edge association, power control, microgrid or energy management, voltage stabilization, network load balancing, IoT data collection, or large queuing-system control.","exclusion":"If the primary objective is privacy-preserving/federated training rather than infrastructure performance, assign to trustworthy MARL; if the primary objective is physical movement, fleet routing, robotics, or UAV coordination, assign to embodied mobility systems.","nearest_boundary":"D7","representative_topics":["distributed marl for fair network load balancing via markov potential game nash equilibrium","transformer-based multi-agent actor-critic for active voltage control in renewable-integrated power grids","scalable multi-agent reinforcement learning for joint channel access policy and signaling in wireless multiple-access networks"]},{"direction_index":7,"label":"D7","name":"Embodied mobility systems","description":"Applying or adapting MARL to coordinate physically or logistically moving agents such as vehicles, robots, UAVs, pursuers, fleets, and embodied swarms.","inclusion":"Include source atoms whose primary objective is multi-agent mobility or embodied coordination: vehicle routing, autonomous driving, fleet dispatch, multi-robot coordination, UAV trajectory planning, formation, coverage, pursuit, pedestrian-aware planning, manipulation, motion planning, and robotic or transportation system integration.","exclusion":"If the physical domain is only a benchmark for a general algorithm, assign to the relevant algorithmic direction or evaluation infrastructure; if the main objective is network/energy resource allocation rather than movement or embodied coordination, assign to networked infrastructure control.","nearest_boundary":"D6","representative_topics":["multi-agent neural rewriting for vehicle routing under partial observability and limited cost disclosure","scalable training and evaluation platform for multi-robot reinforcement learning","scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research"]}],"split":"dev"}
{"episode_id":"F017_T2024-07","field_index":17,"field":"multi-agent reinforcement learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cooperative coordination mechanisms","description":"Learning scalable cooperative joint behavior through communication, credit assignment, relational or graph structure, CTDE/decentralized architectures, role or heterogeneity handling, and cooperative policy optimization.","inclusion":"Include source atoms whose primary objective is to make cooperative agents coordinate better under shared or largely aligned objectives by designing messages, value decompositions, centralized-training/decentralized-execution architectures, decentralized policy updates, graph/attention interaction models, role discovery, parameter sharing, or heterogeneous-agent cooperation mechanisms.","exclusion":"If the main objective is a game-theoretic solution concept, opponent/partner mental-state modeling, or incentives among self-interested agents, assign to strategic interaction; if the main objective is sparse-reward exploration, offline learning, transfer, or bandit regret, assign to data-efficient adaptation.","nearest_boundary":"D1","representative_topics":["decentralized multi-agent reinforcement learning via distribution matching to expert visitation distributions","universally expressive communication protocols in multi-agent reinforcement learning via gnn augmentation","policy gradient methods for cooperative marl beyond value decomposition and parameter sharing"]},{"direction_index":1,"label":"D1","name":"Strategic interaction and incentives","description":"Learning, analyzing, or shaping behavior in strategic multi-agent settings with distinct preferences, adaptive partners, equilibrium concepts, opponent models, theory of mind, social dilemmas, or incentive mechanisms.","inclusion":"Include source atoms whose primary objective is equilibrium computation or convergence in games, no-regret/self-play dynamics, opponent or partner modeling, intention or belief inference, ad hoc strategic adaptation, cooperation emergence, social dilemmas, contracts, markets, fairness, or incentives for mixed-motive agents.","exclusion":"If agents are treated primarily as a cooperative team with known aligned objectives and the contribution is a coordination architecture or policy optimizer, assign to cooperative coordination mechanisms; if the issue is malicious attacks or formal safety constraints, assign to trustworthy MARL.","nearest_boundary":"D0","representative_topics":["sample-efficient equilibrium learning in partially observable markov games via optimistic mle","convergence guarantees and price of anarchy bounds for softmax policy gradient in markov potential games","hierarchical reinforcement learning with opponent modeling for cooperative multi-agent control"]},{"direction_index":2,"label":"D2","name":"Data-efficient adaptation","description":"Improving learning efficiency and adaptability under sparse rewards, limited interaction, fixed data, regret constraints, changing tasks, unseen partners, or changing team composition.","inclusion":"Include source atoms whose primary objective is exploration, intrinsic motivation, curriculum or hierarchy for sparse feedback, offline MARL, model-based data-limited learning, transfer, task/team adaptation, zero-shot or ad hoc teamwork generalization, and multi-agent bandit regret or best-arm identification.","exclusion":"If the main contribution is a general cooperative coordination architecture trained online in a fixed team, assign to cooperative coordination mechanisms; if the main contribution is equilibrium theory rather than data limitation or adaptation, assign to strategic interaction and incentives.","nearest_boundary":"D0","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","offline nash equilibrium learning in zero-sum markov games via pessimistic model-based confidence bounds","adaptive curriculum goal generation via stein variational sampling for multi-goal reinforcement learning"]},{"direction_index":3,"label":"D3","name":"Trustworthy MARL","description":"Making MARL systems safe, robust, secure, privacy-preserving, interpretable, diagnosable, or compliant with formal/risk constraints.","inclusion":"Include source atoms whose primary objective is adversarial attack or defense, certified robustness, constrained or risk-aware policy learning, temporal-logic or formal safety, privacy-preserving or federated training, secure experience sharing, interpretability, explainability, symbolic reasoning, diagnostics, or novelty handling for trustworthiness.","exclusion":"If federation or safety is incidental to a telecom, energy, robotics, or mobility deployment, assign to the corresponding application-objective direction; if reward shaping is mainly for exploration rather than transparency or constraints, assign to data-efficient adaptation.","nearest_boundary":"D6","representative_topics":["reward poisoning attacks on offline multi-agent reinforcement learning via markov perfect equilibrium","near-optimal multi-agent coverage learning under unknown safety constraints and partial observability","primal-dual penalty reformulation for probabilistic constraints in constrained multiagent reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Large-population abstractions","description":"Scaling MARL to very large or infinite populations by replacing explicit many-agent interaction with population-level, mean-field, graphon, or collective-dynamics abstractions.","inclusion":"Include source atoms whose primary objective is tractability or analysis for large-population MARL through mean-field reinforcement learning, mean-field games, graphon games, major-minor population models, aggregate distribution dynamics, or population-level convergence and safety methods.","exclusion":"If scalability is pursued through finite-team communication, CTDE, graph neural networks, hierarchy, or parameter sharing, assign to cooperative coordination mechanisms; if the main objective is a domain system such as load balancing rather than the population abstraction itself, assign to networked infrastructure control.","nearest_boundary":"D0","representative_topics":["mean-field control for scalable load balancing under synchronization delay in parallel queuing systems","oracle-free online mean-field game learning via single-sample-path mean-field equilibrium approximation","scalable multi-agent reinforcement learning via mean-field and collective intelligence methods for large populations"]},{"direction_index":5,"label":"D5","name":"Evaluation infrastructure","description":"Building, standardizing, or synthesizing benchmarks, simulators, platforms, evaluation protocols, scaling studies, surveys, replications, and reproducibility resources for MARL.","inclusion":"Include source atoms whose primary objective is assessment infrastructure or empirical synthesis rather than a new MARL learning mechanism: benchmark environments, scalable simulators, evaluation suites, community challenges, reproducibility tools, scaling laws, surveys, and diagnostic taxonomies.","exclusion":"If an environment or simulator is introduced mainly to validate a new algorithmic mechanism, assign to that mechanism direction; if a benchmark is primarily a domain deployment objective such as transportation or robotics, assign to the relevant domain direction.","nearest_boundary":"D7","representative_topics":["scalable training and evaluation platform for multi-robot reinforcement learning","scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research","distributed local simulation with influence-aware models for scalable multi-agent reinforcement learning"]},{"direction_index":6,"label":"D6","name":"Networked infrastructure control","description":"Applying or adapting MARL to optimize resource allocation and operational control in communication, computing, energy, power-grid, queuing, IoT, and other networked infrastructure systems.","inclusion":"Include source atoms whose primary objective is a domain-defined infrastructure-control problem such as wireless channel access, spectrum or telecom resource management, edge association, power control, microgrid or energy management, voltage stabilization, network load balancing, IoT data collection, or large queuing-system control.","exclusion":"If the primary objective is privacy-preserving/federated training rather than infrastructure performance, assign to trustworthy MARL; if the primary objective is physical movement, fleet routing, robotics, or UAV coordination, assign to embodied mobility systems.","nearest_boundary":"D7","representative_topics":["distributed marl for fair network load balancing via markov potential game nash equilibrium","transformer-based multi-agent actor-critic for active voltage control in renewable-integrated power grids","scalable multi-agent reinforcement learning for joint channel access policy and signaling in wireless multiple-access networks"]},{"direction_index":7,"label":"D7","name":"Embodied mobility systems","description":"Applying or adapting MARL to coordinate physically or logistically moving agents such as vehicles, robots, UAVs, pursuers, fleets, and embodied swarms.","inclusion":"Include source atoms whose primary objective is multi-agent mobility or embodied coordination: vehicle routing, autonomous driving, fleet dispatch, multi-robot coordination, UAV trajectory planning, formation, coverage, pursuit, pedestrian-aware planning, manipulation, motion planning, and robotic or transportation system integration.","exclusion":"If the physical domain is only a benchmark for a general algorithm, assign to the relevant algorithmic direction or evaluation infrastructure; if the main objective is network/energy resource allocation rather than movement or embodied coordination, assign to networked infrastructure control.","nearest_boundary":"D6","representative_topics":["multi-agent neural rewriting for vehicle routing under partial observability and limited cost disclosure","scalable training and evaluation platform for multi-robot reinforcement learning","scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research"]}],"split":"dev"}
{"episode_id":"F017_T2025-01","field_index":17,"field":"multi-agent reinforcement learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cooperative coordination mechanisms","description":"Learning scalable cooperative joint behavior through communication, credit assignment, relational or graph structure, CTDE/decentralized architectures, role or heterogeneity handling, and cooperative policy optimization.","inclusion":"Include source atoms whose primary objective is to make cooperative agents coordinate better under shared or largely aligned objectives by designing messages, value decompositions, centralized-training/decentralized-execution architectures, decentralized policy updates, graph/attention interaction models, role discovery, parameter sharing, or heterogeneous-agent cooperation mechanisms.","exclusion":"If the main objective is a game-theoretic solution concept, opponent/partner mental-state modeling, or incentives among self-interested agents, assign to strategic interaction; if the main objective is sparse-reward exploration, offline learning, transfer, or bandit regret, assign to data-efficient adaptation.","nearest_boundary":"D1","representative_topics":["decentralized multi-agent reinforcement learning via distribution matching to expert visitation distributions","universally expressive communication protocols in multi-agent reinforcement learning via gnn augmentation","policy gradient methods for cooperative marl beyond value decomposition and parameter sharing"]},{"direction_index":1,"label":"D1","name":"Strategic interaction and incentives","description":"Learning, analyzing, or shaping behavior in strategic multi-agent settings with distinct preferences, adaptive partners, equilibrium concepts, opponent models, theory of mind, social dilemmas, or incentive mechanisms.","inclusion":"Include source atoms whose primary objective is equilibrium computation or convergence in games, no-regret/self-play dynamics, opponent or partner modeling, intention or belief inference, ad hoc strategic adaptation, cooperation emergence, social dilemmas, contracts, markets, fairness, or incentives for mixed-motive agents.","exclusion":"If agents are treated primarily as a cooperative team with known aligned objectives and the contribution is a coordination architecture or policy optimizer, assign to cooperative coordination mechanisms; if the issue is malicious attacks or formal safety constraints, assign to trustworthy MARL.","nearest_boundary":"D0","representative_topics":["sample-efficient equilibrium learning in partially observable markov games via optimistic mle","convergence guarantees and price of anarchy bounds for softmax policy gradient in markov potential games","hierarchical reinforcement learning with opponent modeling for cooperative multi-agent control"]},{"direction_index":2,"label":"D2","name":"Data-efficient adaptation","description":"Improving learning efficiency and adaptability under sparse rewards, limited interaction, fixed data, regret constraints, changing tasks, unseen partners, or changing team composition.","inclusion":"Include source atoms whose primary objective is exploration, intrinsic motivation, curriculum or hierarchy for sparse feedback, offline MARL, model-based data-limited learning, transfer, task/team adaptation, zero-shot or ad hoc teamwork generalization, and multi-agent bandit regret or best-arm identification.","exclusion":"If the main contribution is a general cooperative coordination architecture trained online in a fixed team, assign to cooperative coordination mechanisms; if the main contribution is equilibrium theory rather than data limitation or adaptation, assign to strategic interaction and incentives.","nearest_boundary":"D0","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","offline nash equilibrium learning in zero-sum markov games via pessimistic model-based confidence bounds","adaptive curriculum goal generation via stein variational sampling for multi-goal reinforcement learning"]},{"direction_index":3,"label":"D3","name":"Trustworthy MARL","description":"Making MARL systems safe, robust, secure, privacy-preserving, interpretable, diagnosable, or compliant with formal/risk constraints.","inclusion":"Include source atoms whose primary objective is adversarial attack or defense, certified robustness, constrained or risk-aware policy learning, temporal-logic or formal safety, privacy-preserving or federated training, secure experience sharing, interpretability, explainability, symbolic reasoning, diagnostics, or novelty handling for trustworthiness.","exclusion":"If federation or safety is incidental to a telecom, energy, robotics, or mobility deployment, assign to the corresponding application-objective direction; if reward shaping is mainly for exploration rather than transparency or constraints, assign to data-efficient adaptation.","nearest_boundary":"D6","representative_topics":["reward poisoning attacks on offline multi-agent reinforcement learning via markov perfect equilibrium","near-optimal multi-agent coverage learning under unknown safety constraints and partial observability","primal-dual penalty reformulation for probabilistic constraints in constrained multiagent reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Large-population abstractions","description":"Scaling MARL to very large or infinite populations by replacing explicit many-agent interaction with population-level, mean-field, graphon, or collective-dynamics abstractions.","inclusion":"Include source atoms whose primary objective is tractability or analysis for large-population MARL through mean-field reinforcement learning, mean-field games, graphon games, major-minor population models, aggregate distribution dynamics, or population-level convergence and safety methods.","exclusion":"If scalability is pursued through finite-team communication, CTDE, graph neural networks, hierarchy, or parameter sharing, assign to cooperative coordination mechanisms; if the main objective is a domain system such as load balancing rather than the population abstraction itself, assign to networked infrastructure control.","nearest_boundary":"D0","representative_topics":["mean-field control for scalable load balancing under synchronization delay in parallel queuing systems","oracle-free online mean-field game learning via single-sample-path mean-field equilibrium approximation","scalable multi-agent reinforcement learning via mean-field and collective intelligence methods for large populations"]},{"direction_index":5,"label":"D5","name":"Evaluation infrastructure","description":"Building, standardizing, or synthesizing benchmarks, simulators, platforms, evaluation protocols, scaling studies, surveys, replications, and reproducibility resources for MARL.","inclusion":"Include source atoms whose primary objective is assessment infrastructure or empirical synthesis rather than a new MARL learning mechanism: benchmark environments, scalable simulators, evaluation suites, community challenges, reproducibility tools, scaling laws, surveys, and diagnostic taxonomies.","exclusion":"If an environment or simulator is introduced mainly to validate a new algorithmic mechanism, assign to that mechanism direction; if a benchmark is primarily a domain deployment objective such as transportation or robotics, assign to the relevant domain direction.","nearest_boundary":"D7","representative_topics":["scalable training and evaluation platform for multi-robot reinforcement learning","scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research","distributed local simulation with influence-aware models for scalable multi-agent reinforcement learning"]},{"direction_index":6,"label":"D6","name":"Networked infrastructure control","description":"Applying or adapting MARL to optimize resource allocation and operational control in communication, computing, energy, power-grid, queuing, IoT, and other networked infrastructure systems.","inclusion":"Include source atoms whose primary objective is a domain-defined infrastructure-control problem such as wireless channel access, spectrum or telecom resource management, edge association, power control, microgrid or energy management, voltage stabilization, network load balancing, IoT data collection, or large queuing-system control.","exclusion":"If the primary objective is privacy-preserving/federated training rather than infrastructure performance, assign to trustworthy MARL; if the primary objective is physical movement, fleet routing, robotics, or UAV coordination, assign to embodied mobility systems.","nearest_boundary":"D7","representative_topics":["distributed marl for fair network load balancing via markov potential game nash equilibrium","transformer-based multi-agent actor-critic for active voltage control in renewable-integrated power grids","scalable multi-agent reinforcement learning for joint channel access policy and signaling in wireless multiple-access networks"]},{"direction_index":7,"label":"D7","name":"Embodied mobility systems","description":"Applying or adapting MARL to coordinate physically or logistically moving agents such as vehicles, robots, UAVs, pursuers, fleets, and embodied swarms.","inclusion":"Include source atoms whose primary objective is multi-agent mobility or embodied coordination: vehicle routing, autonomous driving, fleet dispatch, multi-robot coordination, UAV trajectory planning, formation, coverage, pursuit, pedestrian-aware planning, manipulation, motion planning, and robotic or transportation system integration.","exclusion":"If the physical domain is only a benchmark for a general algorithm, assign to the relevant algorithmic direction or evaluation infrastructure; if the main objective is network/energy resource allocation rather than movement or embodied coordination, assign to networked infrastructure control.","nearest_boundary":"D6","representative_topics":["multi-agent neural rewriting for vehicle routing under partial observability and limited cost disclosure","scalable training and evaluation platform for multi-robot reinforcement learning","scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research"]}],"split":"dev"}
{"episode_id":"F017_T2025-07","field_index":17,"field":"multi-agent reinforcement learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cooperative coordination mechanisms","description":"Learning scalable cooperative joint behavior through communication, credit assignment, relational or graph structure, CTDE/decentralized architectures, role or heterogeneity handling, and cooperative policy optimization.","inclusion":"Include source atoms whose primary objective is to make cooperative agents coordinate better under shared or largely aligned objectives by designing messages, value decompositions, centralized-training/decentralized-execution architectures, decentralized policy updates, graph/attention interaction models, role discovery, parameter sharing, or heterogeneous-agent cooperation mechanisms.","exclusion":"If the main objective is a game-theoretic solution concept, opponent/partner mental-state modeling, or incentives among self-interested agents, assign to strategic interaction; if the main objective is sparse-reward exploration, offline learning, transfer, or bandit regret, assign to data-efficient adaptation.","nearest_boundary":"D1","representative_topics":["decentralized multi-agent reinforcement learning via distribution matching to expert visitation distributions","universally expressive communication protocols in multi-agent reinforcement learning via gnn augmentation","policy gradient methods for cooperative marl beyond value decomposition and parameter sharing"]},{"direction_index":1,"label":"D1","name":"Strategic interaction and incentives","description":"Learning, analyzing, or shaping behavior in strategic multi-agent settings with distinct preferences, adaptive partners, equilibrium concepts, opponent models, theory of mind, social dilemmas, or incentive mechanisms.","inclusion":"Include source atoms whose primary objective is equilibrium computation or convergence in games, no-regret/self-play dynamics, opponent or partner modeling, intention or belief inference, ad hoc strategic adaptation, cooperation emergence, social dilemmas, contracts, markets, fairness, or incentives for mixed-motive agents.","exclusion":"If agents are treated primarily as a cooperative team with known aligned objectives and the contribution is a coordination architecture or policy optimizer, assign to cooperative coordination mechanisms; if the issue is malicious attacks or formal safety constraints, assign to trustworthy MARL.","nearest_boundary":"D0","representative_topics":["sample-efficient equilibrium learning in partially observable markov games via optimistic mle","convergence guarantees and price of anarchy bounds for softmax policy gradient in markov potential games","hierarchical reinforcement learning with opponent modeling for cooperative multi-agent control"]},{"direction_index":2,"label":"D2","name":"Data-efficient adaptation","description":"Improving learning efficiency and adaptability under sparse rewards, limited interaction, fixed data, regret constraints, changing tasks, unseen partners, or changing team composition.","inclusion":"Include source atoms whose primary objective is exploration, intrinsic motivation, curriculum or hierarchy for sparse feedback, offline MARL, model-based data-limited learning, transfer, task/team adaptation, zero-shot or ad hoc teamwork generalization, and multi-agent bandit regret or best-arm identification.","exclusion":"If the main contribution is a general cooperative coordination architecture trained online in a fixed team, assign to cooperative coordination mechanisms; if the main contribution is equilibrium theory rather than data limitation or adaptation, assign to strategic interaction and incentives.","nearest_boundary":"D0","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","offline nash equilibrium learning in zero-sum markov games via pessimistic model-based confidence bounds","adaptive curriculum goal generation via stein variational sampling for multi-goal reinforcement learning"]},{"direction_index":3,"label":"D3","name":"Trustworthy MARL","description":"Making MARL systems safe, robust, secure, privacy-preserving, interpretable, diagnosable, or compliant with formal/risk constraints.","inclusion":"Include source atoms whose primary objective is adversarial attack or defense, certified robustness, constrained or risk-aware policy learning, temporal-logic or formal safety, privacy-preserving or federated training, secure experience sharing, interpretability, explainability, symbolic reasoning, diagnostics, or novelty handling for trustworthiness.","exclusion":"If federation or safety is incidental to a telecom, energy, robotics, or mobility deployment, assign to the corresponding application-objective direction; if reward shaping is mainly for exploration rather than transparency or constraints, assign to data-efficient adaptation.","nearest_boundary":"D6","representative_topics":["reward poisoning attacks on offline multi-agent reinforcement learning via markov perfect equilibrium","near-optimal multi-agent coverage learning under unknown safety constraints and partial observability","primal-dual penalty reformulation for probabilistic constraints in constrained multiagent reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Large-population abstractions","description":"Scaling MARL to very large or infinite populations by replacing explicit many-agent interaction with population-level, mean-field, graphon, or collective-dynamics abstractions.","inclusion":"Include source atoms whose primary objective is tractability or analysis for large-population MARL through mean-field reinforcement learning, mean-field games, graphon games, major-minor population models, aggregate distribution dynamics, or population-level convergence and safety methods.","exclusion":"If scalability is pursued through finite-team communication, CTDE, graph neural networks, hierarchy, or parameter sharing, assign to cooperative coordination mechanisms; if the main objective is a domain system such as load balancing rather than the population abstraction itself, assign to networked infrastructure control.","nearest_boundary":"D0","representative_topics":["mean-field control for scalable load balancing under synchronization delay in parallel queuing systems","oracle-free online mean-field game learning via single-sample-path mean-field equilibrium approximation","scalable multi-agent reinforcement learning via mean-field and collective intelligence methods for large populations"]},{"direction_index":5,"label":"D5","name":"Evaluation infrastructure","description":"Building, standardizing, or synthesizing benchmarks, simulators, platforms, evaluation protocols, scaling studies, surveys, replications, and reproducibility resources for MARL.","inclusion":"Include source atoms whose primary objective is assessment infrastructure or empirical synthesis rather than a new MARL learning mechanism: benchmark environments, scalable simulators, evaluation suites, community challenges, reproducibility tools, scaling laws, surveys, and diagnostic taxonomies.","exclusion":"If an environment or simulator is introduced mainly to validate a new algorithmic mechanism, assign to that mechanism direction; if a benchmark is primarily a domain deployment objective such as transportation or robotics, assign to the relevant domain direction.","nearest_boundary":"D7","representative_topics":["scalable training and evaluation platform for multi-robot reinforcement learning","scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research","distributed local simulation with influence-aware models for scalable multi-agent reinforcement learning"]},{"direction_index":6,"label":"D6","name":"Networked infrastructure control","description":"Applying or adapting MARL to optimize resource allocation and operational control in communication, computing, energy, power-grid, queuing, IoT, and other networked infrastructure systems.","inclusion":"Include source atoms whose primary objective is a domain-defined infrastructure-control problem such as wireless channel access, spectrum or telecom resource management, edge association, power control, microgrid or energy management, voltage stabilization, network load balancing, IoT data collection, or large queuing-system control.","exclusion":"If the primary objective is privacy-preserving/federated training rather than infrastructure performance, assign to trustworthy MARL; if the primary objective is physical movement, fleet routing, robotics, or UAV coordination, assign to embodied mobility systems.","nearest_boundary":"D7","representative_topics":["distributed marl for fair network load balancing via markov potential game nash equilibrium","transformer-based multi-agent actor-critic for active voltage control in renewable-integrated power grids","scalable multi-agent reinforcement learning for joint channel access policy and signaling in wireless multiple-access networks"]},{"direction_index":7,"label":"D7","name":"Embodied mobility systems","description":"Applying or adapting MARL to coordinate physically or logistically moving agents such as vehicles, robots, UAVs, pursuers, fleets, and embodied swarms.","inclusion":"Include source atoms whose primary objective is multi-agent mobility or embodied coordination: vehicle routing, autonomous driving, fleet dispatch, multi-robot coordination, UAV trajectory planning, formation, coverage, pursuit, pedestrian-aware planning, manipulation, motion planning, and robotic or transportation system integration.","exclusion":"If the physical domain is only a benchmark for a general algorithm, assign to the relevant algorithmic direction or evaluation infrastructure; if the main objective is network/energy resource allocation rather than movement or embodied coordination, assign to networked infrastructure control.","nearest_boundary":"D6","representative_topics":["multi-agent neural rewriting for vehicle routing under partial observability and limited cost disclosure","scalable training and evaluation platform for multi-robot reinforcement learning","scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research"]}],"split":"dev"}
{"episode_id":"F017_T2026-01","field_index":17,"field":"multi-agent reinforcement learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cooperative coordination mechanisms","description":"Learning scalable cooperative joint behavior through communication, credit assignment, relational or graph structure, CTDE/decentralized architectures, role or heterogeneity handling, and cooperative policy optimization.","inclusion":"Include source atoms whose primary objective is to make cooperative agents coordinate better under shared or largely aligned objectives by designing messages, value decompositions, centralized-training/decentralized-execution architectures, decentralized policy updates, graph/attention interaction models, role discovery, parameter sharing, or heterogeneous-agent cooperation mechanisms.","exclusion":"If the main objective is a game-theoretic solution concept, opponent/partner mental-state modeling, or incentives among self-interested agents, assign to strategic interaction; if the main objective is sparse-reward exploration, offline learning, transfer, or bandit regret, assign to data-efficient adaptation.","nearest_boundary":"D1","representative_topics":["decentralized multi-agent reinforcement learning via distribution matching to expert visitation distributions","universally expressive communication protocols in multi-agent reinforcement learning via gnn augmentation","policy gradient methods for cooperative marl beyond value decomposition and parameter sharing"]},{"direction_index":1,"label":"D1","name":"Strategic interaction and incentives","description":"Learning, analyzing, or shaping behavior in strategic multi-agent settings with distinct preferences, adaptive partners, equilibrium concepts, opponent models, theory of mind, social dilemmas, or incentive mechanisms.","inclusion":"Include source atoms whose primary objective is equilibrium computation or convergence in games, no-regret/self-play dynamics, opponent or partner modeling, intention or belief inference, ad hoc strategic adaptation, cooperation emergence, social dilemmas, contracts, markets, fairness, or incentives for mixed-motive agents.","exclusion":"If agents are treated primarily as a cooperative team with known aligned objectives and the contribution is a coordination architecture or policy optimizer, assign to cooperative coordination mechanisms; if the issue is malicious attacks or formal safety constraints, assign to trustworthy MARL.","nearest_boundary":"D0","representative_topics":["sample-efficient equilibrium learning in partially observable markov games via optimistic mle","convergence guarantees and price of anarchy bounds for softmax policy gradient in markov potential games","hierarchical reinforcement learning with opponent modeling for cooperative multi-agent control"]},{"direction_index":2,"label":"D2","name":"Data-efficient adaptation","description":"Improving learning efficiency and adaptability under sparse rewards, limited interaction, fixed data, regret constraints, changing tasks, unseen partners, or changing team composition.","inclusion":"Include source atoms whose primary objective is exploration, intrinsic motivation, curriculum or hierarchy for sparse feedback, offline MARL, model-based data-limited learning, transfer, task/team adaptation, zero-shot or ad hoc teamwork generalization, and multi-agent bandit regret or best-arm identification.","exclusion":"If the main contribution is a general cooperative coordination architecture trained online in a fixed team, assign to cooperative coordination mechanisms; if the main contribution is equilibrium theory rather than data limitation or adaptation, assign to strategic interaction and incentives.","nearest_boundary":"D0","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","offline nash equilibrium learning in zero-sum markov games via pessimistic model-based confidence bounds","adaptive curriculum goal generation via stein variational sampling for multi-goal reinforcement learning"]},{"direction_index":3,"label":"D3","name":"Trustworthy MARL","description":"Making MARL systems safe, robust, secure, privacy-preserving, interpretable, diagnosable, or compliant with formal/risk constraints.","inclusion":"Include source atoms whose primary objective is adversarial attack or defense, certified robustness, constrained or risk-aware policy learning, temporal-logic or formal safety, privacy-preserving or federated training, secure experience sharing, interpretability, explainability, symbolic reasoning, diagnostics, or novelty handling for trustworthiness.","exclusion":"If federation or safety is incidental to a telecom, energy, robotics, or mobility deployment, assign to the corresponding application-objective direction; if reward shaping is mainly for exploration rather than transparency or constraints, assign to data-efficient adaptation.","nearest_boundary":"D6","representative_topics":["reward poisoning attacks on offline multi-agent reinforcement learning via markov perfect equilibrium","near-optimal multi-agent coverage learning under unknown safety constraints and partial observability","primal-dual penalty reformulation for probabilistic constraints in constrained multiagent reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Large-population abstractions","description":"Scaling MARL to very large or infinite populations by replacing explicit many-agent interaction with population-level, mean-field, graphon, or collective-dynamics abstractions.","inclusion":"Include source atoms whose primary objective is tractability or analysis for large-population MARL through mean-field reinforcement learning, mean-field games, graphon games, major-minor population models, aggregate distribution dynamics, or population-level convergence and safety methods.","exclusion":"If scalability is pursued through finite-team communication, CTDE, graph neural networks, hierarchy, or parameter sharing, assign to cooperative coordination mechanisms; if the main objective is a domain system such as load balancing rather than the population abstraction itself, assign to networked infrastructure control.","nearest_boundary":"D0","representative_topics":["mean-field control for scalable load balancing under synchronization delay in parallel queuing systems","oracle-free online mean-field game learning via single-sample-path mean-field equilibrium approximation","scalable multi-agent reinforcement learning via mean-field and collective intelligence methods for large populations"]},{"direction_index":5,"label":"D5","name":"Evaluation infrastructure","description":"Building, standardizing, or synthesizing benchmarks, simulators, platforms, evaluation protocols, scaling studies, surveys, replications, and reproducibility resources for MARL.","inclusion":"Include source atoms whose primary objective is assessment infrastructure or empirical synthesis rather than a new MARL learning mechanism: benchmark environments, scalable simulators, evaluation suites, community challenges, reproducibility tools, scaling laws, surveys, and diagnostic taxonomies.","exclusion":"If an environment or simulator is introduced mainly to validate a new algorithmic mechanism, assign to that mechanism direction; if a benchmark is primarily a domain deployment objective such as transportation or robotics, assign to the relevant domain direction.","nearest_boundary":"D7","representative_topics":["scalable training and evaluation platform for multi-robot reinforcement learning","scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research","distributed local simulation with influence-aware models for scalable multi-agent reinforcement learning"]},{"direction_index":6,"label":"D6","name":"Networked infrastructure control","description":"Applying or adapting MARL to optimize resource allocation and operational control in communication, computing, energy, power-grid, queuing, IoT, and other networked infrastructure systems.","inclusion":"Include source atoms whose primary objective is a domain-defined infrastructure-control problem such as wireless channel access, spectrum or telecom resource management, edge association, power control, microgrid or energy management, voltage stabilization, network load balancing, IoT data collection, or large queuing-system control.","exclusion":"If the primary objective is privacy-preserving/federated training rather than infrastructure performance, assign to trustworthy MARL; if the primary objective is physical movement, fleet routing, robotics, or UAV coordination, assign to embodied mobility systems.","nearest_boundary":"D7","representative_topics":["distributed marl for fair network load balancing via markov potential game nash equilibrium","transformer-based multi-agent actor-critic for active voltage control in renewable-integrated power grids","scalable multi-agent reinforcement learning for joint channel access policy and signaling in wireless multiple-access networks"]},{"direction_index":7,"label":"D7","name":"Embodied mobility systems","description":"Applying or adapting MARL to coordinate physically or logistically moving agents such as vehicles, robots, UAVs, pursuers, fleets, and embodied swarms.","inclusion":"Include source atoms whose primary objective is multi-agent mobility or embodied coordination: vehicle routing, autonomous driving, fleet dispatch, multi-robot coordination, UAV trajectory planning, formation, coverage, pursuit, pedestrian-aware planning, manipulation, motion planning, and robotic or transportation system integration.","exclusion":"If the physical domain is only a benchmark for a general algorithm, assign to the relevant algorithmic direction or evaluation infrastructure; if the main objective is network/energy resource allocation rather than movement or embodied coordination, assign to networked infrastructure control.","nearest_boundary":"D6","representative_topics":["multi-agent neural rewriting for vehicle routing under partial observability and limited cost disclosure","scalable training and evaluation platform for multi-robot reinforcement learning","scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research"]}],"split":"dev"}
{"episode_id":"F022_T2024-01","field_index":22,"field":"reinforcement learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Logged-Data Policy Learning","description":"Learning, evaluating, correcting, or stabilizing RL policies when the available experience is off-policy, logged, replayed, or fixed, with emphasis on distribution mismatch, estimator bias/variance, and extrapolation control.","inclusion":"Include atoms whose primary objective is offline policy optimization, off-policy evaluation, off-policy actor-critic or TD correction, conservative or pessimistic learning from fixed datasets, replay-based stabilization, or value estimation from behavior-policy data.","exclusion":"If the primary object is generic policy-gradient or Q-learning convergence without an off-policy/logged-data bottleneck, place it in Core RL Optimization Theory; if a learned dynamics model is the central contribution, place it in Model-Based, Representation, and Transfer Learning.","nearest_boundary":"Core RL Optimization Theory","representative_topics":["gap-dependent sample complexity bounds for offline reinforcement learning under coverage assumptions","model generation with provable coverability for offline dynamics-aware policy learning","differentially private offline reinforcement learning with instance-dependent learning bounds"]},{"direction_index":1,"label":"D1","name":"Exploration and Temporal Abstraction","description":"Designing mechanisms that help agents discover useful states, goals, options, skills, and long-horizon action structures under sparse rewards or hard exploration.","inclusion":"Include atoms whose primary objective is intrinsic motivation, curiosity, novelty, uncertainty-driven exploration, adaptive goal generation, option discovery, skill reuse, hierarchical control, subgoal learning, or temporally extended decision structures.","exclusion":"If goals or skills are mainly a representation-transfer device without temporal decomposition or exploration focus, place it in Model-Based, Representation, and Transfer Learning; if reward modification is mainly task-specification learning from humans or demonstrations, place it in Reward Specification and Human Alignment.","nearest_boundary":"Model-Based, Representation, and Transfer Learning","representative_topics":["policy churn as implicit exploration in deep value-based reinforcement learning","action noise scaling schedules and types in off-policy deep reinforcement learning exploration","heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations"]},{"direction_index":2,"label":"D2","name":"Reward Specification and Human Alignment","description":"Inferring, shaping, diagnosing, or aligning reward and constraint specifications from demonstrations, preferences, language, logic, domain knowledge, or human feedback.","inclusion":"Include atoms whose primary objective is inverse RL, inverse constrained RL, preference-based reward learning, RLHF-style reward modeling, reward hacking or overoptimization analysis, human-in-the-loop feedback, language/logic reward construction, or transferable reward specification.","exclusion":"If reward signals are introduced mainly to drive exploration rather than specify a task, place it in Exploration and Temporal Abstraction; if fixed-dataset policy optimization is the main contribution, place it in Logged-Data Policy Learning.","nearest_boundary":"Exploration and Temporal Abstraction","representative_topics":["learning soft constraints from expert demonstrations via iterative constrained optimization in inverse reinforcement learning","regret-based reward function learning from human preferences in rlhf","case-based inverse reinforcement learning via temporal coherence for strategy-level imitation"]},{"direction_index":3,"label":"D3","name":"Safety and Robustness Guarantees","description":"Training and analyzing policies that satisfy safety, stability, risk, feasibility, or robustness requirements under constraints, perturbations, uncertainty, attacks, or corrupted data.","inclusion":"Include atoms whose primary objective is constrained MDP optimization, safe exploration, zero-violation learning, barrier/Lyapunov/stability certification, risk-sensitive safety, adversarial robustness, distributionally robust RL, poisoning or Byzantine robustness, and defenses against perturbations.","exclusion":"If robustness is only dataset-support conservatism for offline learning, place it in Logged-Data Policy Learning; if constraints are inferred as reward/task specifications rather than enforced during policy learning or deployment, place it in Reward Specification and Human Alignment.","nearest_boundary":"Logged-Data Policy Learning","representative_topics":["byzantine-robust distributed reinforcement learning via robust mean estimation and pessimistic/optimistic value iteration","krasovskii-constrained reinforcement learning with provable stability guarantees for nonlinear systems","safe exploration via safety state augmentation and budget scheduling in model-free reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Model-Based, Representation, and Transfer Learning","description":"Learning dynamics models, latent state/task representations, abstractions, world models, and transferable knowledge structures that improve planning, generalization, adaptation, or sample efficiency.","inclusion":"Include atoms whose primary objective is learned dynamics modeling, world models, model-predictive planning, value-equivalent or causal models, latent representation learning, visual/entity/history encodings, task embeddings, sim-to-real transfer, lifelong or continual adaptation, meta-RL, and pretraining for transfer.","exclusion":"If a model is used mainly to solve a fixed offline-data policy optimization problem, place it in Logged-Data Policy Learning; if explicit option or skill hierarchy is the central object, place it in Exploration and Temporal Abstraction.","nearest_boundary":"Exploration and Temporal Abstraction","representative_topics":["lifelong reinforcement learning with provably efficient multi-task policy learning via value distillation","adaptive scheduling of data augmentation distillation for deep reinforcement learning generalization","transformer-based q-networks for stable history encoding in partially observable reinforcement learning"]},{"direction_index":5,"label":"D5","name":"Multi-Agent Coordination and Games","description":"Developing RL methods and theory for multiple learning agents, including cooperative coordination, communication, social mechanisms, decentralized execution, and equilibrium learning in Markov games.","inclusion":"Include atoms whose primary objective is cooperative MARL, CTDE, value decomposition, credit assignment, ad-hoc teamwork, learned communication, social dilemmas, incentive mechanisms, Markov-game equilibrium computation, no-regret dynamics, mean-field or competitive multi-agent learning, and multi-agent coordination theory.","exclusion":"If the multi-agent setting is incidental to a domain deployment problem such as traffic, energy, or robotics, place it in Operational RL Deployment and Evaluation; if adversarial interaction is primarily a single-agent robustness formulation, place it in Safety and Robustness Guarantees.","nearest_boundary":"Operational RL Deployment and Evaluation","representative_topics":["relation-aware value function factorization for ad-hoc cooperation in multi-agent reinforcement learning","sample-efficient equilibrium learning in partially observable markov games via optimistic mle","decentralized optimistic hyperpolicy mirror descent for no-regret learning in markov games"]},{"direction_index":6,"label":"D6","name":"Core RL Optimization Theory","description":"Establishing foundational algorithms, convergence guarantees, regret bounds, sample complexity, and optimization properties for single-agent RL and MDP variants.","inclusion":"Include atoms whose primary objective is policy-gradient, natural-policy-gradient, actor-critic, Q-learning, TD, value-iteration, posterior-sampling, regularization, structured-MDP, average-cost, LQR, risk-sensitive, or quantum policy-optimization theory when not primarily offline, multi-agent, or safety-constrained.","exclusion":"If the guarantee is mainly about coverage and extrapolation from fixed datasets, place it in Logged-Data Policy Learning; if the theory concerns equilibrium or no-regret dynamics among multiple agents, place it in Multi-Agent Coordination and Games.","nearest_boundary":"Logged-Data Policy Learning","representative_topics":["predecessor feature credit assignment via bootstrapped expected trace in reinforcement learning","stabilizing q-learning via modular algorithmic tools for provably efficient linear mdp learning","finite-time complexity analysis of entropy-regularized neural natural actor-critic algorithms"]},{"direction_index":7,"label":"D7","name":"Operational RL Deployment and Evaluation","description":"Adapting, testing, benchmarking, explaining, and deploying RL for concrete operational decision systems, including physical infrastructure, embodied agents, markets, combinatorial design, and RL software/evaluation infrastructure.","inclusion":"Include atoms whose primary objective is RL for infrastructure and network control, energy systems, traffic and autonomous vehicles, robotics and embodied control, finance and economic policy, combinatorial optimization or generative design, domain-specific RL formulations, benchmark/simulator construction, scalable RL systems, agent testing, diagnostics, and explainability.","exclusion":"If a paper’s central contribution is a general RL algorithmic mechanism independent of the operational domain, place it in the corresponding methodological direction such as Model-Based, Safety, Multi-Agent, or Logged-Data Policy Learning; if the benchmark is only an evaluation venue for a new algorithm, do not place it here.","nearest_boundary":"Model-Based, Representation, and Transfer Learning","representative_topics":["reinforcement learning with greedy search for sequential scene object placement","transformer-based multi-agent actor-critic for active voltage control in renewable-integrated power grids","optimization-assisted ensemble deep reinforcement learning for scalable unit commitment"]}],"split":"dev"}
{"episode_id":"F022_T2024-07","field_index":22,"field":"reinforcement learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Logged-Data Policy Learning","description":"Learning, evaluating, correcting, or stabilizing RL policies when the available experience is off-policy, logged, replayed, or fixed, with emphasis on distribution mismatch, estimator bias/variance, and extrapolation control.","inclusion":"Include atoms whose primary objective is offline policy optimization, off-policy evaluation, off-policy actor-critic or TD correction, conservative or pessimistic learning from fixed datasets, replay-based stabilization, or value estimation from behavior-policy data.","exclusion":"If the primary object is generic policy-gradient or Q-learning convergence without an off-policy/logged-data bottleneck, place it in Core RL Optimization Theory; if a learned dynamics model is the central contribution, place it in Model-Based, Representation, and Transfer Learning.","nearest_boundary":"Core RL Optimization Theory","representative_topics":["gap-dependent sample complexity bounds for offline reinforcement learning under coverage assumptions","model generation with provable coverability for offline dynamics-aware policy learning","differentially private offline reinforcement learning with instance-dependent learning bounds"]},{"direction_index":1,"label":"D1","name":"Exploration and Temporal Abstraction","description":"Designing mechanisms that help agents discover useful states, goals, options, skills, and long-horizon action structures under sparse rewards or hard exploration.","inclusion":"Include atoms whose primary objective is intrinsic motivation, curiosity, novelty, uncertainty-driven exploration, adaptive goal generation, option discovery, skill reuse, hierarchical control, subgoal learning, or temporally extended decision structures.","exclusion":"If goals or skills are mainly a representation-transfer device without temporal decomposition or exploration focus, place it in Model-Based, Representation, and Transfer Learning; if reward modification is mainly task-specification learning from humans or demonstrations, place it in Reward Specification and Human Alignment.","nearest_boundary":"Model-Based, Representation, and Transfer Learning","representative_topics":["policy churn as implicit exploration in deep value-based reinforcement learning","action noise scaling schedules and types in off-policy deep reinforcement learning exploration","heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations"]},{"direction_index":2,"label":"D2","name":"Reward Specification and Human Alignment","description":"Inferring, shaping, diagnosing, or aligning reward and constraint specifications from demonstrations, preferences, language, logic, domain knowledge, or human feedback.","inclusion":"Include atoms whose primary objective is inverse RL, inverse constrained RL, preference-based reward learning, RLHF-style reward modeling, reward hacking or overoptimization analysis, human-in-the-loop feedback, language/logic reward construction, or transferable reward specification.","exclusion":"If reward signals are introduced mainly to drive exploration rather than specify a task, place it in Exploration and Temporal Abstraction; if fixed-dataset policy optimization is the main contribution, place it in Logged-Data Policy Learning.","nearest_boundary":"Exploration and Temporal Abstraction","representative_topics":["learning soft constraints from expert demonstrations via iterative constrained optimization in inverse reinforcement learning","regret-based reward function learning from human preferences in rlhf","case-based inverse reinforcement learning via temporal coherence for strategy-level imitation"]},{"direction_index":3,"label":"D3","name":"Safety and Robustness Guarantees","description":"Training and analyzing policies that satisfy safety, stability, risk, feasibility, or robustness requirements under constraints, perturbations, uncertainty, attacks, or corrupted data.","inclusion":"Include atoms whose primary objective is constrained MDP optimization, safe exploration, zero-violation learning, barrier/Lyapunov/stability certification, risk-sensitive safety, adversarial robustness, distributionally robust RL, poisoning or Byzantine robustness, and defenses against perturbations.","exclusion":"If robustness is only dataset-support conservatism for offline learning, place it in Logged-Data Policy Learning; if constraints are inferred as reward/task specifications rather than enforced during policy learning or deployment, place it in Reward Specification and Human Alignment.","nearest_boundary":"Logged-Data Policy Learning","representative_topics":["byzantine-robust distributed reinforcement learning via robust mean estimation and pessimistic/optimistic value iteration","krasovskii-constrained reinforcement learning with provable stability guarantees for nonlinear systems","safe exploration via safety state augmentation and budget scheduling in model-free reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Model-Based, Representation, and Transfer Learning","description":"Learning dynamics models, latent state/task representations, abstractions, world models, and transferable knowledge structures that improve planning, generalization, adaptation, or sample efficiency.","inclusion":"Include atoms whose primary objective is learned dynamics modeling, world models, model-predictive planning, value-equivalent or causal models, latent representation learning, visual/entity/history encodings, task embeddings, sim-to-real transfer, lifelong or continual adaptation, meta-RL, and pretraining for transfer.","exclusion":"If a model is used mainly to solve a fixed offline-data policy optimization problem, place it in Logged-Data Policy Learning; if explicit option or skill hierarchy is the central object, place it in Exploration and Temporal Abstraction.","nearest_boundary":"Exploration and Temporal Abstraction","representative_topics":["lifelong reinforcement learning with provably efficient multi-task policy learning via value distillation","adaptive scheduling of data augmentation distillation for deep reinforcement learning generalization","transformer-based q-networks for stable history encoding in partially observable reinforcement learning"]},{"direction_index":5,"label":"D5","name":"Multi-Agent Coordination and Games","description":"Developing RL methods and theory for multiple learning agents, including cooperative coordination, communication, social mechanisms, decentralized execution, and equilibrium learning in Markov games.","inclusion":"Include atoms whose primary objective is cooperative MARL, CTDE, value decomposition, credit assignment, ad-hoc teamwork, learned communication, social dilemmas, incentive mechanisms, Markov-game equilibrium computation, no-regret dynamics, mean-field or competitive multi-agent learning, and multi-agent coordination theory.","exclusion":"If the multi-agent setting is incidental to a domain deployment problem such as traffic, energy, or robotics, place it in Operational RL Deployment and Evaluation; if adversarial interaction is primarily a single-agent robustness formulation, place it in Safety and Robustness Guarantees.","nearest_boundary":"Operational RL Deployment and Evaluation","representative_topics":["relation-aware value function factorization for ad-hoc cooperation in multi-agent reinforcement learning","sample-efficient equilibrium learning in partially observable markov games via optimistic mle","decentralized optimistic hyperpolicy mirror descent for no-regret learning in markov games"]},{"direction_index":6,"label":"D6","name":"Core RL Optimization Theory","description":"Establishing foundational algorithms, convergence guarantees, regret bounds, sample complexity, and optimization properties for single-agent RL and MDP variants.","inclusion":"Include atoms whose primary objective is policy-gradient, natural-policy-gradient, actor-critic, Q-learning, TD, value-iteration, posterior-sampling, regularization, structured-MDP, average-cost, LQR, risk-sensitive, or quantum policy-optimization theory when not primarily offline, multi-agent, or safety-constrained.","exclusion":"If the guarantee is mainly about coverage and extrapolation from fixed datasets, place it in Logged-Data Policy Learning; if the theory concerns equilibrium or no-regret dynamics among multiple agents, place it in Multi-Agent Coordination and Games.","nearest_boundary":"Logged-Data Policy Learning","representative_topics":["predecessor feature credit assignment via bootstrapped expected trace in reinforcement learning","stabilizing q-learning via modular algorithmic tools for provably efficient linear mdp learning","finite-time complexity analysis of entropy-regularized neural natural actor-critic algorithms"]},{"direction_index":7,"label":"D7","name":"Operational RL Deployment and Evaluation","description":"Adapting, testing, benchmarking, explaining, and deploying RL for concrete operational decision systems, including physical infrastructure, embodied agents, markets, combinatorial design, and RL software/evaluation infrastructure.","inclusion":"Include atoms whose primary objective is RL for infrastructure and network control, energy systems, traffic and autonomous vehicles, robotics and embodied control, finance and economic policy, combinatorial optimization or generative design, domain-specific RL formulations, benchmark/simulator construction, scalable RL systems, agent testing, diagnostics, and explainability.","exclusion":"If a paper’s central contribution is a general RL algorithmic mechanism independent of the operational domain, place it in the corresponding methodological direction such as Model-Based, Safety, Multi-Agent, or Logged-Data Policy Learning; if the benchmark is only an evaluation venue for a new algorithm, do not place it here.","nearest_boundary":"Model-Based, Representation, and Transfer Learning","representative_topics":["reinforcement learning with greedy search for sequential scene object placement","transformer-based multi-agent actor-critic for active voltage control in renewable-integrated power grids","optimization-assisted ensemble deep reinforcement learning for scalable unit commitment"]}],"split":"dev"}
{"episode_id":"F022_T2025-01","field_index":22,"field":"reinforcement learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Logged-Data Policy Learning","description":"Learning, evaluating, correcting, or stabilizing RL policies when the available experience is off-policy, logged, replayed, or fixed, with emphasis on distribution mismatch, estimator bias/variance, and extrapolation control.","inclusion":"Include atoms whose primary objective is offline policy optimization, off-policy evaluation, off-policy actor-critic or TD correction, conservative or pessimistic learning from fixed datasets, replay-based stabilization, or value estimation from behavior-policy data.","exclusion":"If the primary object is generic policy-gradient or Q-learning convergence without an off-policy/logged-data bottleneck, place it in Core RL Optimization Theory; if a learned dynamics model is the central contribution, place it in Model-Based, Representation, and Transfer Learning.","nearest_boundary":"Core RL Optimization Theory","representative_topics":["gap-dependent sample complexity bounds for offline reinforcement learning under coverage assumptions","model generation with provable coverability for offline dynamics-aware policy learning","differentially private offline reinforcement learning with instance-dependent learning bounds"]},{"direction_index":1,"label":"D1","name":"Exploration and Temporal Abstraction","description":"Designing mechanisms that help agents discover useful states, goals, options, skills, and long-horizon action structures under sparse rewards or hard exploration.","inclusion":"Include atoms whose primary objective is intrinsic motivation, curiosity, novelty, uncertainty-driven exploration, adaptive goal generation, option discovery, skill reuse, hierarchical control, subgoal learning, or temporally extended decision structures.","exclusion":"If goals or skills are mainly a representation-transfer device without temporal decomposition or exploration focus, place it in Model-Based, Representation, and Transfer Learning; if reward modification is mainly task-specification learning from humans or demonstrations, place it in Reward Specification and Human Alignment.","nearest_boundary":"Model-Based, Representation, and Transfer Learning","representative_topics":["policy churn as implicit exploration in deep value-based reinforcement learning","action noise scaling schedules and types in off-policy deep reinforcement learning exploration","heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations"]},{"direction_index":2,"label":"D2","name":"Reward Specification and Human Alignment","description":"Inferring, shaping, diagnosing, or aligning reward and constraint specifications from demonstrations, preferences, language, logic, domain knowledge, or human feedback.","inclusion":"Include atoms whose primary objective is inverse RL, inverse constrained RL, preference-based reward learning, RLHF-style reward modeling, reward hacking or overoptimization analysis, human-in-the-loop feedback, language/logic reward construction, or transferable reward specification.","exclusion":"If reward signals are introduced mainly to drive exploration rather than specify a task, place it in Exploration and Temporal Abstraction; if fixed-dataset policy optimization is the main contribution, place it in Logged-Data Policy Learning.","nearest_boundary":"Exploration and Temporal Abstraction","representative_topics":["learning soft constraints from expert demonstrations via iterative constrained optimization in inverse reinforcement learning","regret-based reward function learning from human preferences in rlhf","case-based inverse reinforcement learning via temporal coherence for strategy-level imitation"]},{"direction_index":3,"label":"D3","name":"Safety and Robustness Guarantees","description":"Training and analyzing policies that satisfy safety, stability, risk, feasibility, or robustness requirements under constraints, perturbations, uncertainty, attacks, or corrupted data.","inclusion":"Include atoms whose primary objective is constrained MDP optimization, safe exploration, zero-violation learning, barrier/Lyapunov/stability certification, risk-sensitive safety, adversarial robustness, distributionally robust RL, poisoning or Byzantine robustness, and defenses against perturbations.","exclusion":"If robustness is only dataset-support conservatism for offline learning, place it in Logged-Data Policy Learning; if constraints are inferred as reward/task specifications rather than enforced during policy learning or deployment, place it in Reward Specification and Human Alignment.","nearest_boundary":"Logged-Data Policy Learning","representative_topics":["byzantine-robust distributed reinforcement learning via robust mean estimation and pessimistic/optimistic value iteration","krasovskii-constrained reinforcement learning with provable stability guarantees for nonlinear systems","safe exploration via safety state augmentation and budget scheduling in model-free reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Model-Based, Representation, and Transfer Learning","description":"Learning dynamics models, latent state/task representations, abstractions, world models, and transferable knowledge structures that improve planning, generalization, adaptation, or sample efficiency.","inclusion":"Include atoms whose primary objective is learned dynamics modeling, world models, model-predictive planning, value-equivalent or causal models, latent representation learning, visual/entity/history encodings, task embeddings, sim-to-real transfer, lifelong or continual adaptation, meta-RL, and pretraining for transfer.","exclusion":"If a model is used mainly to solve a fixed offline-data policy optimization problem, place it in Logged-Data Policy Learning; if explicit option or skill hierarchy is the central object, place it in Exploration and Temporal Abstraction.","nearest_boundary":"Exploration and Temporal Abstraction","representative_topics":["lifelong reinforcement learning with provably efficient multi-task policy learning via value distillation","adaptive scheduling of data augmentation distillation for deep reinforcement learning generalization","transformer-based q-networks for stable history encoding in partially observable reinforcement learning"]},{"direction_index":5,"label":"D5","name":"Multi-Agent Coordination and Games","description":"Developing RL methods and theory for multiple learning agents, including cooperative coordination, communication, social mechanisms, decentralized execution, and equilibrium learning in Markov games.","inclusion":"Include atoms whose primary objective is cooperative MARL, CTDE, value decomposition, credit assignment, ad-hoc teamwork, learned communication, social dilemmas, incentive mechanisms, Markov-game equilibrium computation, no-regret dynamics, mean-field or competitive multi-agent learning, and multi-agent coordination theory.","exclusion":"If the multi-agent setting is incidental to a domain deployment problem such as traffic, energy, or robotics, place it in Operational RL Deployment and Evaluation; if adversarial interaction is primarily a single-agent robustness formulation, place it in Safety and Robustness Guarantees.","nearest_boundary":"Operational RL Deployment and Evaluation","representative_topics":["relation-aware value function factorization for ad-hoc cooperation in multi-agent reinforcement learning","sample-efficient equilibrium learning in partially observable markov games via optimistic mle","decentralized optimistic hyperpolicy mirror descent for no-regret learning in markov games"]},{"direction_index":6,"label":"D6","name":"Core RL Optimization Theory","description":"Establishing foundational algorithms, convergence guarantees, regret bounds, sample complexity, and optimization properties for single-agent RL and MDP variants.","inclusion":"Include atoms whose primary objective is policy-gradient, natural-policy-gradient, actor-critic, Q-learning, TD, value-iteration, posterior-sampling, regularization, structured-MDP, average-cost, LQR, risk-sensitive, or quantum policy-optimization theory when not primarily offline, multi-agent, or safety-constrained.","exclusion":"If the guarantee is mainly about coverage and extrapolation from fixed datasets, place it in Logged-Data Policy Learning; if the theory concerns equilibrium or no-regret dynamics among multiple agents, place it in Multi-Agent Coordination and Games.","nearest_boundary":"Logged-Data Policy Learning","representative_topics":["predecessor feature credit assignment via bootstrapped expected trace in reinforcement learning","stabilizing q-learning via modular algorithmic tools for provably efficient linear mdp learning","finite-time complexity analysis of entropy-regularized neural natural actor-critic algorithms"]},{"direction_index":7,"label":"D7","name":"Operational RL Deployment and Evaluation","description":"Adapting, testing, benchmarking, explaining, and deploying RL for concrete operational decision systems, including physical infrastructure, embodied agents, markets, combinatorial design, and RL software/evaluation infrastructure.","inclusion":"Include atoms whose primary objective is RL for infrastructure and network control, energy systems, traffic and autonomous vehicles, robotics and embodied control, finance and economic policy, combinatorial optimization or generative design, domain-specific RL formulations, benchmark/simulator construction, scalable RL systems, agent testing, diagnostics, and explainability.","exclusion":"If a paper’s central contribution is a general RL algorithmic mechanism independent of the operational domain, place it in the corresponding methodological direction such as Model-Based, Safety, Multi-Agent, or Logged-Data Policy Learning; if the benchmark is only an evaluation venue for a new algorithm, do not place it here.","nearest_boundary":"Model-Based, Representation, and Transfer Learning","representative_topics":["reinforcement learning with greedy search for sequential scene object placement","transformer-based multi-agent actor-critic for active voltage control in renewable-integrated power grids","optimization-assisted ensemble deep reinforcement learning for scalable unit commitment"]}],"split":"dev"}
{"episode_id":"F022_T2025-07","field_index":22,"field":"reinforcement learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Logged-Data Policy Learning","description":"Learning, evaluating, correcting, or stabilizing RL policies when the available experience is off-policy, logged, replayed, or fixed, with emphasis on distribution mismatch, estimator bias/variance, and extrapolation control.","inclusion":"Include atoms whose primary objective is offline policy optimization, off-policy evaluation, off-policy actor-critic or TD correction, conservative or pessimistic learning from fixed datasets, replay-based stabilization, or value estimation from behavior-policy data.","exclusion":"If the primary object is generic policy-gradient or Q-learning convergence without an off-policy/logged-data bottleneck, place it in Core RL Optimization Theory; if a learned dynamics model is the central contribution, place it in Model-Based, Representation, and Transfer Learning.","nearest_boundary":"Core RL Optimization Theory","representative_topics":["gap-dependent sample complexity bounds for offline reinforcement learning under coverage assumptions","model generation with provable coverability for offline dynamics-aware policy learning","differentially private offline reinforcement learning with instance-dependent learning bounds"]},{"direction_index":1,"label":"D1","name":"Exploration and Temporal Abstraction","description":"Designing mechanisms that help agents discover useful states, goals, options, skills, and long-horizon action structures under sparse rewards or hard exploration.","inclusion":"Include atoms whose primary objective is intrinsic motivation, curiosity, novelty, uncertainty-driven exploration, adaptive goal generation, option discovery, skill reuse, hierarchical control, subgoal learning, or temporally extended decision structures.","exclusion":"If goals or skills are mainly a representation-transfer device without temporal decomposition or exploration focus, place it in Model-Based, Representation, and Transfer Learning; if reward modification is mainly task-specification learning from humans or demonstrations, place it in Reward Specification and Human Alignment.","nearest_boundary":"Model-Based, Representation, and Transfer Learning","representative_topics":["policy churn as implicit exploration in deep value-based reinforcement learning","action noise scaling schedules and types in off-policy deep reinforcement learning exploration","heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations"]},{"direction_index":2,"label":"D2","name":"Reward Specification and Human Alignment","description":"Inferring, shaping, diagnosing, or aligning reward and constraint specifications from demonstrations, preferences, language, logic, domain knowledge, or human feedback.","inclusion":"Include atoms whose primary objective is inverse RL, inverse constrained RL, preference-based reward learning, RLHF-style reward modeling, reward hacking or overoptimization analysis, human-in-the-loop feedback, language/logic reward construction, or transferable reward specification.","exclusion":"If reward signals are introduced mainly to drive exploration rather than specify a task, place it in Exploration and Temporal Abstraction; if fixed-dataset policy optimization is the main contribution, place it in Logged-Data Policy Learning.","nearest_boundary":"Exploration and Temporal Abstraction","representative_topics":["learning soft constraints from expert demonstrations via iterative constrained optimization in inverse reinforcement learning","regret-based reward function learning from human preferences in rlhf","case-based inverse reinforcement learning via temporal coherence for strategy-level imitation"]},{"direction_index":3,"label":"D3","name":"Safety and Robustness Guarantees","description":"Training and analyzing policies that satisfy safety, stability, risk, feasibility, or robustness requirements under constraints, perturbations, uncertainty, attacks, or corrupted data.","inclusion":"Include atoms whose primary objective is constrained MDP optimization, safe exploration, zero-violation learning, barrier/Lyapunov/stability certification, risk-sensitive safety, adversarial robustness, distributionally robust RL, poisoning or Byzantine robustness, and defenses against perturbations.","exclusion":"If robustness is only dataset-support conservatism for offline learning, place it in Logged-Data Policy Learning; if constraints are inferred as reward/task specifications rather than enforced during policy learning or deployment, place it in Reward Specification and Human Alignment.","nearest_boundary":"Logged-Data Policy Learning","representative_topics":["byzantine-robust distributed reinforcement learning via robust mean estimation and pessimistic/optimistic value iteration","krasovskii-constrained reinforcement learning with provable stability guarantees for nonlinear systems","safe exploration via safety state augmentation and budget scheduling in model-free reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Model-Based, Representation, and Transfer Learning","description":"Learning dynamics models, latent state/task representations, abstractions, world models, and transferable knowledge structures that improve planning, generalization, adaptation, or sample efficiency.","inclusion":"Include atoms whose primary objective is learned dynamics modeling, world models, model-predictive planning, value-equivalent or causal models, latent representation learning, visual/entity/history encodings, task embeddings, sim-to-real transfer, lifelong or continual adaptation, meta-RL, and pretraining for transfer.","exclusion":"If a model is used mainly to solve a fixed offline-data policy optimization problem, place it in Logged-Data Policy Learning; if explicit option or skill hierarchy is the central object, place it in Exploration and Temporal Abstraction.","nearest_boundary":"Exploration and Temporal Abstraction","representative_topics":["lifelong reinforcement learning with provably efficient multi-task policy learning via value distillation","adaptive scheduling of data augmentation distillation for deep reinforcement learning generalization","transformer-based q-networks for stable history encoding in partially observable reinforcement learning"]},{"direction_index":5,"label":"D5","name":"Multi-Agent Coordination and Games","description":"Developing RL methods and theory for multiple learning agents, including cooperative coordination, communication, social mechanisms, decentralized execution, and equilibrium learning in Markov games.","inclusion":"Include atoms whose primary objective is cooperative MARL, CTDE, value decomposition, credit assignment, ad-hoc teamwork, learned communication, social dilemmas, incentive mechanisms, Markov-game equilibrium computation, no-regret dynamics, mean-field or competitive multi-agent learning, and multi-agent coordination theory.","exclusion":"If the multi-agent setting is incidental to a domain deployment problem such as traffic, energy, or robotics, place it in Operational RL Deployment and Evaluation; if adversarial interaction is primarily a single-agent robustness formulation, place it in Safety and Robustness Guarantees.","nearest_boundary":"Operational RL Deployment and Evaluation","representative_topics":["relation-aware value function factorization for ad-hoc cooperation in multi-agent reinforcement learning","sample-efficient equilibrium learning in partially observable markov games via optimistic mle","decentralized optimistic hyperpolicy mirror descent for no-regret learning in markov games"]},{"direction_index":6,"label":"D6","name":"Core RL Optimization Theory","description":"Establishing foundational algorithms, convergence guarantees, regret bounds, sample complexity, and optimization properties for single-agent RL and MDP variants.","inclusion":"Include atoms whose primary objective is policy-gradient, natural-policy-gradient, actor-critic, Q-learning, TD, value-iteration, posterior-sampling, regularization, structured-MDP, average-cost, LQR, risk-sensitive, or quantum policy-optimization theory when not primarily offline, multi-agent, or safety-constrained.","exclusion":"If the guarantee is mainly about coverage and extrapolation from fixed datasets, place it in Logged-Data Policy Learning; if the theory concerns equilibrium or no-regret dynamics among multiple agents, place it in Multi-Agent Coordination and Games.","nearest_boundary":"Logged-Data Policy Learning","representative_topics":["predecessor feature credit assignment via bootstrapped expected trace in reinforcement learning","stabilizing q-learning via modular algorithmic tools for provably efficient linear mdp learning","finite-time complexity analysis of entropy-regularized neural natural actor-critic algorithms"]},{"direction_index":7,"label":"D7","name":"Operational RL Deployment and Evaluation","description":"Adapting, testing, benchmarking, explaining, and deploying RL for concrete operational decision systems, including physical infrastructure, embodied agents, markets, combinatorial design, and RL software/evaluation infrastructure.","inclusion":"Include atoms whose primary objective is RL for infrastructure and network control, energy systems, traffic and autonomous vehicles, robotics and embodied control, finance and economic policy, combinatorial optimization or generative design, domain-specific RL formulations, benchmark/simulator construction, scalable RL systems, agent testing, diagnostics, and explainability.","exclusion":"If a paper’s central contribution is a general RL algorithmic mechanism independent of the operational domain, place it in the corresponding methodological direction such as Model-Based, Safety, Multi-Agent, or Logged-Data Policy Learning; if the benchmark is only an evaluation venue for a new algorithm, do not place it here.","nearest_boundary":"Model-Based, Representation, and Transfer Learning","representative_topics":["reinforcement learning with greedy search for sequential scene object placement","transformer-based multi-agent actor-critic for active voltage control in renewable-integrated power grids","optimization-assisted ensemble deep reinforcement learning for scalable unit commitment"]}],"split":"dev"}
{"episode_id":"F022_T2026-01","field_index":22,"field":"reinforcement learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Logged-Data Policy Learning","description":"Learning, evaluating, correcting, or stabilizing RL policies when the available experience is off-policy, logged, replayed, or fixed, with emphasis on distribution mismatch, estimator bias/variance, and extrapolation control.","inclusion":"Include atoms whose primary objective is offline policy optimization, off-policy evaluation, off-policy actor-critic or TD correction, conservative or pessimistic learning from fixed datasets, replay-based stabilization, or value estimation from behavior-policy data.","exclusion":"If the primary object is generic policy-gradient or Q-learning convergence without an off-policy/logged-data bottleneck, place it in Core RL Optimization Theory; if a learned dynamics model is the central contribution, place it in Model-Based, Representation, and Transfer Learning.","nearest_boundary":"Core RL Optimization Theory","representative_topics":["gap-dependent sample complexity bounds for offline reinforcement learning under coverage assumptions","model generation with provable coverability for offline dynamics-aware policy learning","differentially private offline reinforcement learning with instance-dependent learning bounds"]},{"direction_index":1,"label":"D1","name":"Exploration and Temporal Abstraction","description":"Designing mechanisms that help agents discover useful states, goals, options, skills, and long-horizon action structures under sparse rewards or hard exploration.","inclusion":"Include atoms whose primary objective is intrinsic motivation, curiosity, novelty, uncertainty-driven exploration, adaptive goal generation, option discovery, skill reuse, hierarchical control, subgoal learning, or temporally extended decision structures.","exclusion":"If goals or skills are mainly a representation-transfer device without temporal decomposition or exploration focus, place it in Model-Based, Representation, and Transfer Learning; if reward modification is mainly task-specification learning from humans or demonstrations, place it in Reward Specification and Human Alignment.","nearest_boundary":"Model-Based, Representation, and Transfer Learning","representative_topics":["policy churn as implicit exploration in deep value-based reinforcement learning","action noise scaling schedules and types in off-policy deep reinforcement learning exploration","heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations"]},{"direction_index":2,"label":"D2","name":"Reward Specification and Human Alignment","description":"Inferring, shaping, diagnosing, or aligning reward and constraint specifications from demonstrations, preferences, language, logic, domain knowledge, or human feedback.","inclusion":"Include atoms whose primary objective is inverse RL, inverse constrained RL, preference-based reward learning, RLHF-style reward modeling, reward hacking or overoptimization analysis, human-in-the-loop feedback, language/logic reward construction, or transferable reward specification.","exclusion":"If reward signals are introduced mainly to drive exploration rather than specify a task, place it in Exploration and Temporal Abstraction; if fixed-dataset policy optimization is the main contribution, place it in Logged-Data Policy Learning.","nearest_boundary":"Exploration and Temporal Abstraction","representative_topics":["learning soft constraints from expert demonstrations via iterative constrained optimization in inverse reinforcement learning","regret-based reward function learning from human preferences in rlhf","case-based inverse reinforcement learning via temporal coherence for strategy-level imitation"]},{"direction_index":3,"label":"D3","name":"Safety and Robustness Guarantees","description":"Training and analyzing policies that satisfy safety, stability, risk, feasibility, or robustness requirements under constraints, perturbations, uncertainty, attacks, or corrupted data.","inclusion":"Include atoms whose primary objective is constrained MDP optimization, safe exploration, zero-violation learning, barrier/Lyapunov/stability certification, risk-sensitive safety, adversarial robustness, distributionally robust RL, poisoning or Byzantine robustness, and defenses against perturbations.","exclusion":"If robustness is only dataset-support conservatism for offline learning, place it in Logged-Data Policy Learning; if constraints are inferred as reward/task specifications rather than enforced during policy learning or deployment, place it in Reward Specification and Human Alignment.","nearest_boundary":"Logged-Data Policy Learning","representative_topics":["byzantine-robust distributed reinforcement learning via robust mean estimation and pessimistic/optimistic value iteration","krasovskii-constrained reinforcement learning with provable stability guarantees for nonlinear systems","safe exploration via safety state augmentation and budget scheduling in model-free reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Model-Based, Representation, and Transfer Learning","description":"Learning dynamics models, latent state/task representations, abstractions, world models, and transferable knowledge structures that improve planning, generalization, adaptation, or sample efficiency.","inclusion":"Include atoms whose primary objective is learned dynamics modeling, world models, model-predictive planning, value-equivalent or causal models, latent representation learning, visual/entity/history encodings, task embeddings, sim-to-real transfer, lifelong or continual adaptation, meta-RL, and pretraining for transfer.","exclusion":"If a model is used mainly to solve a fixed offline-data policy optimization problem, place it in Logged-Data Policy Learning; if explicit option or skill hierarchy is the central object, place it in Exploration and Temporal Abstraction.","nearest_boundary":"Exploration and Temporal Abstraction","representative_topics":["lifelong reinforcement learning with provably efficient multi-task policy learning via value distillation","adaptive scheduling of data augmentation distillation for deep reinforcement learning generalization","transformer-based q-networks for stable history encoding in partially observable reinforcement learning"]},{"direction_index":5,"label":"D5","name":"Multi-Agent Coordination and Games","description":"Developing RL methods and theory for multiple learning agents, including cooperative coordination, communication, social mechanisms, decentralized execution, and equilibrium learning in Markov games.","inclusion":"Include atoms whose primary objective is cooperative MARL, CTDE, value decomposition, credit assignment, ad-hoc teamwork, learned communication, social dilemmas, incentive mechanisms, Markov-game equilibrium computation, no-regret dynamics, mean-field or competitive multi-agent learning, and multi-agent coordination theory.","exclusion":"If the multi-agent setting is incidental to a domain deployment problem such as traffic, energy, or robotics, place it in Operational RL Deployment and Evaluation; if adversarial interaction is primarily a single-agent robustness formulation, place it in Safety and Robustness Guarantees.","nearest_boundary":"Operational RL Deployment and Evaluation","representative_topics":["relation-aware value function factorization for ad-hoc cooperation in multi-agent reinforcement learning","sample-efficient equilibrium learning in partially observable markov games via optimistic mle","decentralized optimistic hyperpolicy mirror descent for no-regret learning in markov games"]},{"direction_index":6,"label":"D6","name":"Core RL Optimization Theory","description":"Establishing foundational algorithms, convergence guarantees, regret bounds, sample complexity, and optimization properties for single-agent RL and MDP variants.","inclusion":"Include atoms whose primary objective is policy-gradient, natural-policy-gradient, actor-critic, Q-learning, TD, value-iteration, posterior-sampling, regularization, structured-MDP, average-cost, LQR, risk-sensitive, or quantum policy-optimization theory when not primarily offline, multi-agent, or safety-constrained.","exclusion":"If the guarantee is mainly about coverage and extrapolation from fixed datasets, place it in Logged-Data Policy Learning; if the theory concerns equilibrium or no-regret dynamics among multiple agents, place it in Multi-Agent Coordination and Games.","nearest_boundary":"Logged-Data Policy Learning","representative_topics":["predecessor feature credit assignment via bootstrapped expected trace in reinforcement learning","stabilizing q-learning via modular algorithmic tools for provably efficient linear mdp learning","finite-time complexity analysis of entropy-regularized neural natural actor-critic algorithms"]},{"direction_index":7,"label":"D7","name":"Operational RL Deployment and Evaluation","description":"Adapting, testing, benchmarking, explaining, and deploying RL for concrete operational decision systems, including physical infrastructure, embodied agents, markets, combinatorial design, and RL software/evaluation infrastructure.","inclusion":"Include atoms whose primary objective is RL for infrastructure and network control, energy systems, traffic and autonomous vehicles, robotics and embodied control, finance and economic policy, combinatorial optimization or generative design, domain-specific RL formulations, benchmark/simulator construction, scalable RL systems, agent testing, diagnostics, and explainability.","exclusion":"If a paper’s central contribution is a general RL algorithmic mechanism independent of the operational domain, place it in the corresponding methodological direction such as Model-Based, Safety, Multi-Agent, or Logged-Data Policy Learning; if the benchmark is only an evaluation venue for a new algorithm, do not place it here.","nearest_boundary":"Model-Based, Representation, and Transfer Learning","representative_topics":["reinforcement learning with greedy search for sequential scene object placement","transformer-based multi-agent actor-critic for active voltage control in renewable-integrated power grids","optimization-assisted ensemble deep reinforcement learning for scalable unit commitment"]}],"split":"dev"}
{"episode_id":"F024_T2024-01","field_index":24,"field":"multimodal large language models","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Capability evaluation and diagnostics","description":"Research whose primary objective is to measure, benchmark, diagnose, or compare multimodal large language model capabilities, weaknesses, robustness profiles, perception, grounding, and reasoning behavior.","inclusion":"Include papers proposing benchmarks, evaluation suites, metrics, diagnostic protocols, controlled capability tests, evaluator frameworks, or systematic empirical assessments where evaluation is the main contribution.","exclusion":"If evaluation is used mainly to validate a new model, training method, reliability intervention, or domain adaptation, assign to that primary modeling or mitigation direction instead.","nearest_boundary":"Trustworthy reliability and factuality","representative_topics":["comprehensive evaluation benchmark for multimodal large language model perception and cognition","benchmarking generative multimodal llm comprehension across spatial and temporal evaluation dimensions","bridge-architecture evaluation for complex visual reasoning via object-level and pre-training analysis"]},{"direction_index":1,"label":"D1","name":"General MLLM construction","description":"Research whose primary objective is to build, align, architect, pretrain, instruction-tune, or synthesize foundations for general-purpose multimodal LLMs, including vision-language assistants, modular representation design, unified multimodal generation, and broad field-level architectural synthesis.","inclusion":"Include work centered on visual or multimodal instruction tuning, modality-to-LLM alignment, projection or adapter bridges when used for general assistant construction, unified any-to-any or autoregressive multimodal architectures, mixed-modal pretraining, core representation design, and broad surveys of MLLM foundations.","exclusion":"If the primary goal is explicit spatial outputs, efficiency, domain specialization, temporal/audio modeling, safety mitigation, or benchmark construction, assign to that more specific objective instead.","nearest_boundary":"Efficient, scalable, and cross-lingual adaptation","representative_topics":["language model as universal interface for modular multimodal encoder integration via semi-causal pretraining","unified transformer model via discrete token homogenization for diverse vision and language tasks","modularized multimodal pretraining with disentangled and collaborative modality modules"]},{"direction_index":2,"label":"D2","name":"Trustworthy reliability and factuality","description":"Research whose primary objective is to understand, detect, reduce, or control failures of trustworthiness in MLLMs, including hallucination, factual inconsistency, adversarial vulnerability, jailbreaks, robustness failures, catastrophic forgetting, interpretability, model editing, and safety-relevant alignment weaknesses.","inclusion":"Include papers primarily proposing hallucination detection or mitigation, factuality alignment, robust instruction tuning, reward or preference methods for visual factuality, adversarial attack or defense analysis, jailbreak studies, model editing, neuron-level interpretation, or continual-learning reliability analysis.","exclusion":"General capability benchmarks without a reliability-specific intervention belong to capability evaluation; ordinary instruction tuning without a factuality, safety, or robustness objective belongs to general MLLM construction.","nearest_boundary":"Capability evaluation and diagnostics","representative_topics":["mitigating multimodal hallucination via robust instruction tuning with adversarial negative samples","multimodal reward models and preference optimization for hallucination detection and prevention in vision language models","contrastive instruction tuning for hallucination reduction in vision-language models"]},{"direction_index":3,"label":"D3","name":"Fine-grained visual grounding","description":"Research whose primary objective is to make MLLMs understand, refer to, localize, output, or manipulate explicit visual entities and spatial structures such as regions, coordinates, bounding boxes, masks, OCR-visible text, segmentation, geometric relations, and pixel- or panoptic-level representations.","inclusion":"Include work on referring dialogue, coordinate input or output, region-level alignment, spatially grounded instruction tuning, segmentation-linked MLLMs, OCR-aware fine-grained perception, open-world regional recognition, and pixel-level or mask-based interaction.","exclusion":"Broad image-chat or VQA without explicit spatial, region, OCR, or pixel-level grounding belongs to general MLLM construction or reasoning; domain-specific grounding whose main objective is a vertical application belongs to domain-specialized adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["multimodal llm for spatial referential dialogue using natural language coordinate inputs and outputs","visual instruction tuning for text-rich image understanding via ocr and caption-guided data augmentation","visual grounding in multimodal llms via segmentation-based cross-modal entity correspondence"]},{"direction_index":4,"label":"D4","name":"Reasoning, prompting, and agency","description":"Research whose primary objective is to improve multimodal task solving through inference-time reasoning strategies, prompting, in-context learning, retrieval or expert-assisted reasoning, tool orchestration, memory, multi-agent collaboration, embodied interaction, navigation, or action loops.","inclusion":"Include papers on multimodal chain-of-thought, visual prompting, zero- or few-shot prompting, in-context learning, prompt-based expert calls, retrieval-augmented reasoning, multimodal agents, tool use, plan-execute-inspect loops, web or GUI navigation, multi-agent systems, embodied agents, and perception-action reasoning.","exclusion":"If the main contribution is supervised model alignment or architecture rather than inference procedure or agency, assign to general MLLM construction; if a vertical domain rather than general agency is central, assign to domain-specialized adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["zero-shot vqa with frozen llms via image-to-prompt bridging without end-to-end training","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","embodied multimodal language models grounding continuous sensor modalities for embodied reasoning tasks"]},{"direction_index":5,"label":"D5","name":"Domain-specialized adaptation","description":"Research whose primary objective is to adapt, train, evaluate, or construct data for MLLMs in specialized professional, scientific, clinical, technical, document-intensive, or other vertical domains where domain knowledge and domain conventions are central.","inclusion":"Include work on healthcare, dentistry, ophthalmology, radiology, science instruction data, document understanding, finance, autonomous driving, robotics-as-domain, marine or other professional domains when domain data, domain outputs, or domain-specific evaluation drive the contribution.","exclusion":"General assistant methods merely demonstrated on domain examples belong to general MLLM construction; general agentic control or navigation belongs to reasoning, prompting, and agency; general region/OCR grounding belongs to fine-grained visual grounding.","nearest_boundary":"Reasoning, prompting, and agency","representative_topics":["multimodal large language models for automated and cross-modal dental diagnosis","instruction-tuned vision-language llm for chest x-ray understanding and bidirectional generation","vision-language pre-training for embodied planning and control via chain-of-thought ego-centric data"]},{"direction_index":6,"label":"D6","name":"Efficient, scalable, and cross-lingual adaptation","description":"Research whose primary objective is to make MLLMs cheaper, smaller, more deployable, more scalable, or transferable across languages through parameter-efficient tuning, compression, lightweight grounding, retrieval-assisted adaptation, sparse experts, scaling-law analysis, visual tokenizer design, or multilingual/cross-lingual extension.","inclusion":"Include papers focused on LoRA or adapter tuning, lightweight projections, pruning, distillation, quantization, retrieval-augmented parameter-efficient adaptation, sparse MoE scaling, scaling laws, deployability, multilingual vision-language scaling, bilingual instruction tuning, and cross-lingual multimodal transfer when efficiency, scale, or language transfer is central.","exclusion":"Full general instruction tuning without an efficiency or cross-lingual objective belongs to general MLLM construction; modality-specific speech or video modeling belongs to temporal, audio, and speech multimodality.","nearest_boundary":"General MLLM construction","representative_topics":["jointly scaled multilingual vision-language model with large-capacity vision transformer and modular encoder-decoder interface","scaling laws for generative mixed-modal language models with inter-modal synergy and competition","efficient visual grounding of frozen language models via linear layer finetuning"]},{"direction_index":7,"label":"D7","name":"Temporal, audio, and speech multimodality","description":"Research whose primary objective is to extend MLLMs to temporally extended, video, audio, speech, music, sound, or audio-visual inputs and outputs, including temporal reasoning, video conversation, audio-visual comprehension, speech-language alignment, and audio generation or translation.","inclusion":"Include papers on video-language models, long-video understanding, temporal adapters, audio-visual video assistants, speech-language instruction following, audio and music understanding or generation, speech translation through LLMs, and multimodal systems where audio, speech, or temporal sequence modeling is the main contribution.","exclusion":"Unified any-to-any architectures whose main contribution is modality-general output unification belong to general MLLM construction; prompt-only temporal reasoning belongs to reasoning, prompting, and agency; cross-lingual transfer without speech/audio emphasis belongs to efficient, scalable, and cross-lingual adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["multimodal llm augmentation for audio understanding and generation across diverse sound modalities","unified video sequence understanding via large language model token sequence reasoning","audio-visual language model for temporal video comprehension via cross-modal q-former alignment"]}],"split":"dev"}
{"episode_id":"F024_T2024-07","field_index":24,"field":"multimodal large language models","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Capability evaluation and diagnostics","description":"Research whose primary objective is to measure, benchmark, diagnose, or compare multimodal large language model capabilities, weaknesses, robustness profiles, perception, grounding, and reasoning behavior.","inclusion":"Include papers proposing benchmarks, evaluation suites, metrics, diagnostic protocols, controlled capability tests, evaluator frameworks, or systematic empirical assessments where evaluation is the main contribution.","exclusion":"If evaluation is used mainly to validate a new model, training method, reliability intervention, or domain adaptation, assign to that primary modeling or mitigation direction instead.","nearest_boundary":"Trustworthy reliability and factuality","representative_topics":["comprehensive evaluation benchmark for multimodal large language model perception and cognition","benchmarking generative multimodal llm comprehension across spatial and temporal evaluation dimensions","bridge-architecture evaluation for complex visual reasoning via object-level and pre-training analysis"]},{"direction_index":1,"label":"D1","name":"General MLLM construction","description":"Research whose primary objective is to build, align, architect, pretrain, instruction-tune, or synthesize foundations for general-purpose multimodal LLMs, including vision-language assistants, modular representation design, unified multimodal generation, and broad field-level architectural synthesis.","inclusion":"Include work centered on visual or multimodal instruction tuning, modality-to-LLM alignment, projection or adapter bridges when used for general assistant construction, unified any-to-any or autoregressive multimodal architectures, mixed-modal pretraining, core representation design, and broad surveys of MLLM foundations.","exclusion":"If the primary goal is explicit spatial outputs, efficiency, domain specialization, temporal/audio modeling, safety mitigation, or benchmark construction, assign to that more specific objective instead.","nearest_boundary":"Efficient, scalable, and cross-lingual adaptation","representative_topics":["language model as universal interface for modular multimodal encoder integration via semi-causal pretraining","unified transformer model via discrete token homogenization for diverse vision and language tasks","modularized multimodal pretraining with disentangled and collaborative modality modules"]},{"direction_index":2,"label":"D2","name":"Trustworthy reliability and factuality","description":"Research whose primary objective is to understand, detect, reduce, or control failures of trustworthiness in MLLMs, including hallucination, factual inconsistency, adversarial vulnerability, jailbreaks, robustness failures, catastrophic forgetting, interpretability, model editing, and safety-relevant alignment weaknesses.","inclusion":"Include papers primarily proposing hallucination detection or mitigation, factuality alignment, robust instruction tuning, reward or preference methods for visual factuality, adversarial attack or defense analysis, jailbreak studies, model editing, neuron-level interpretation, or continual-learning reliability analysis.","exclusion":"General capability benchmarks without a reliability-specific intervention belong to capability evaluation; ordinary instruction tuning without a factuality, safety, or robustness objective belongs to general MLLM construction.","nearest_boundary":"Capability evaluation and diagnostics","representative_topics":["mitigating multimodal hallucination via robust instruction tuning with adversarial negative samples","multimodal reward models and preference optimization for hallucination detection and prevention in vision language models","contrastive instruction tuning for hallucination reduction in vision-language models"]},{"direction_index":3,"label":"D3","name":"Fine-grained visual grounding","description":"Research whose primary objective is to make MLLMs understand, refer to, localize, output, or manipulate explicit visual entities and spatial structures such as regions, coordinates, bounding boxes, masks, OCR-visible text, segmentation, geometric relations, and pixel- or panoptic-level representations.","inclusion":"Include work on referring dialogue, coordinate input or output, region-level alignment, spatially grounded instruction tuning, segmentation-linked MLLMs, OCR-aware fine-grained perception, open-world regional recognition, and pixel-level or mask-based interaction.","exclusion":"Broad image-chat or VQA without explicit spatial, region, OCR, or pixel-level grounding belongs to general MLLM construction or reasoning; domain-specific grounding whose main objective is a vertical application belongs to domain-specialized adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["multimodal llm for spatial referential dialogue using natural language coordinate inputs and outputs","visual instruction tuning for text-rich image understanding via ocr and caption-guided data augmentation","visual grounding in multimodal llms via segmentation-based cross-modal entity correspondence"]},{"direction_index":4,"label":"D4","name":"Reasoning, prompting, and agency","description":"Research whose primary objective is to improve multimodal task solving through inference-time reasoning strategies, prompting, in-context learning, retrieval or expert-assisted reasoning, tool orchestration, memory, multi-agent collaboration, embodied interaction, navigation, or action loops.","inclusion":"Include papers on multimodal chain-of-thought, visual prompting, zero- or few-shot prompting, in-context learning, prompt-based expert calls, retrieval-augmented reasoning, multimodal agents, tool use, plan-execute-inspect loops, web or GUI navigation, multi-agent systems, embodied agents, and perception-action reasoning.","exclusion":"If the main contribution is supervised model alignment or architecture rather than inference procedure or agency, assign to general MLLM construction; if a vertical domain rather than general agency is central, assign to domain-specialized adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["zero-shot vqa with frozen llms via image-to-prompt bridging without end-to-end training","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","embodied multimodal language models grounding continuous sensor modalities for embodied reasoning tasks"]},{"direction_index":5,"label":"D5","name":"Domain-specialized adaptation","description":"Research whose primary objective is to adapt, train, evaluate, or construct data for MLLMs in specialized professional, scientific, clinical, technical, document-intensive, or other vertical domains where domain knowledge and domain conventions are central.","inclusion":"Include work on healthcare, dentistry, ophthalmology, radiology, science instruction data, document understanding, finance, autonomous driving, robotics-as-domain, marine or other professional domains when domain data, domain outputs, or domain-specific evaluation drive the contribution.","exclusion":"General assistant methods merely demonstrated on domain examples belong to general MLLM construction; general agentic control or navigation belongs to reasoning, prompting, and agency; general region/OCR grounding belongs to fine-grained visual grounding.","nearest_boundary":"Reasoning, prompting, and agency","representative_topics":["multimodal large language models for automated and cross-modal dental diagnosis","instruction-tuned vision-language llm for chest x-ray understanding and bidirectional generation","vision-language pre-training for embodied planning and control via chain-of-thought ego-centric data"]},{"direction_index":6,"label":"D6","name":"Efficient, scalable, and cross-lingual adaptation","description":"Research whose primary objective is to make MLLMs cheaper, smaller, more deployable, more scalable, or transferable across languages through parameter-efficient tuning, compression, lightweight grounding, retrieval-assisted adaptation, sparse experts, scaling-law analysis, visual tokenizer design, or multilingual/cross-lingual extension.","inclusion":"Include papers focused on LoRA or adapter tuning, lightweight projections, pruning, distillation, quantization, retrieval-augmented parameter-efficient adaptation, sparse MoE scaling, scaling laws, deployability, multilingual vision-language scaling, bilingual instruction tuning, and cross-lingual multimodal transfer when efficiency, scale, or language transfer is central.","exclusion":"Full general instruction tuning without an efficiency or cross-lingual objective belongs to general MLLM construction; modality-specific speech or video modeling belongs to temporal, audio, and speech multimodality.","nearest_boundary":"General MLLM construction","representative_topics":["jointly scaled multilingual vision-language model with large-capacity vision transformer and modular encoder-decoder interface","scaling laws for generative mixed-modal language models with inter-modal synergy and competition","efficient visual grounding of frozen language models via linear layer finetuning"]},{"direction_index":7,"label":"D7","name":"Temporal, audio, and speech multimodality","description":"Research whose primary objective is to extend MLLMs to temporally extended, video, audio, speech, music, sound, or audio-visual inputs and outputs, including temporal reasoning, video conversation, audio-visual comprehension, speech-language alignment, and audio generation or translation.","inclusion":"Include papers on video-language models, long-video understanding, temporal adapters, audio-visual video assistants, speech-language instruction following, audio and music understanding or generation, speech translation through LLMs, and multimodal systems where audio, speech, or temporal sequence modeling is the main contribution.","exclusion":"Unified any-to-any architectures whose main contribution is modality-general output unification belong to general MLLM construction; prompt-only temporal reasoning belongs to reasoning, prompting, and agency; cross-lingual transfer without speech/audio emphasis belongs to efficient, scalable, and cross-lingual adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["multimodal llm augmentation for audio understanding and generation across diverse sound modalities","unified video sequence understanding via large language model token sequence reasoning","audio-visual language model for temporal video comprehension via cross-modal q-former alignment"]}],"split":"dev"}
{"episode_id":"F024_T2025-01","field_index":24,"field":"multimodal large language models","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Capability evaluation and diagnostics","description":"Research whose primary objective is to measure, benchmark, diagnose, or compare multimodal large language model capabilities, weaknesses, robustness profiles, perception, grounding, and reasoning behavior.","inclusion":"Include papers proposing benchmarks, evaluation suites, metrics, diagnostic protocols, controlled capability tests, evaluator frameworks, or systematic empirical assessments where evaluation is the main contribution.","exclusion":"If evaluation is used mainly to validate a new model, training method, reliability intervention, or domain adaptation, assign to that primary modeling or mitigation direction instead.","nearest_boundary":"Trustworthy reliability and factuality","representative_topics":["comprehensive evaluation benchmark for multimodal large language model perception and cognition","benchmarking generative multimodal llm comprehension across spatial and temporal evaluation dimensions","bridge-architecture evaluation for complex visual reasoning via object-level and pre-training analysis"]},{"direction_index":1,"label":"D1","name":"General MLLM construction","description":"Research whose primary objective is to build, align, architect, pretrain, instruction-tune, or synthesize foundations for general-purpose multimodal LLMs, including vision-language assistants, modular representation design, unified multimodal generation, and broad field-level architectural synthesis.","inclusion":"Include work centered on visual or multimodal instruction tuning, modality-to-LLM alignment, projection or adapter bridges when used for general assistant construction, unified any-to-any or autoregressive multimodal architectures, mixed-modal pretraining, core representation design, and broad surveys of MLLM foundations.","exclusion":"If the primary goal is explicit spatial outputs, efficiency, domain specialization, temporal/audio modeling, safety mitigation, or benchmark construction, assign to that more specific objective instead.","nearest_boundary":"Efficient, scalable, and cross-lingual adaptation","representative_topics":["language model as universal interface for modular multimodal encoder integration via semi-causal pretraining","unified transformer model via discrete token homogenization for diverse vision and language tasks","modularized multimodal pretraining with disentangled and collaborative modality modules"]},{"direction_index":2,"label":"D2","name":"Trustworthy reliability and factuality","description":"Research whose primary objective is to understand, detect, reduce, or control failures of trustworthiness in MLLMs, including hallucination, factual inconsistency, adversarial vulnerability, jailbreaks, robustness failures, catastrophic forgetting, interpretability, model editing, and safety-relevant alignment weaknesses.","inclusion":"Include papers primarily proposing hallucination detection or mitigation, factuality alignment, robust instruction tuning, reward or preference methods for visual factuality, adversarial attack or defense analysis, jailbreak studies, model editing, neuron-level interpretation, or continual-learning reliability analysis.","exclusion":"General capability benchmarks without a reliability-specific intervention belong to capability evaluation; ordinary instruction tuning without a factuality, safety, or robustness objective belongs to general MLLM construction.","nearest_boundary":"Capability evaluation and diagnostics","representative_topics":["mitigating multimodal hallucination via robust instruction tuning with adversarial negative samples","multimodal reward models and preference optimization for hallucination detection and prevention in vision language models","contrastive instruction tuning for hallucination reduction in vision-language models"]},{"direction_index":3,"label":"D3","name":"Fine-grained visual grounding","description":"Research whose primary objective is to make MLLMs understand, refer to, localize, output, or manipulate explicit visual entities and spatial structures such as regions, coordinates, bounding boxes, masks, OCR-visible text, segmentation, geometric relations, and pixel- or panoptic-level representations.","inclusion":"Include work on referring dialogue, coordinate input or output, region-level alignment, spatially grounded instruction tuning, segmentation-linked MLLMs, OCR-aware fine-grained perception, open-world regional recognition, and pixel-level or mask-based interaction.","exclusion":"Broad image-chat or VQA without explicit spatial, region, OCR, or pixel-level grounding belongs to general MLLM construction or reasoning; domain-specific grounding whose main objective is a vertical application belongs to domain-specialized adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["multimodal llm for spatial referential dialogue using natural language coordinate inputs and outputs","visual instruction tuning for text-rich image understanding via ocr and caption-guided data augmentation","visual grounding in multimodal llms via segmentation-based cross-modal entity correspondence"]},{"direction_index":4,"label":"D4","name":"Reasoning, prompting, and agency","description":"Research whose primary objective is to improve multimodal task solving through inference-time reasoning strategies, prompting, in-context learning, retrieval or expert-assisted reasoning, tool orchestration, memory, multi-agent collaboration, embodied interaction, navigation, or action loops.","inclusion":"Include papers on multimodal chain-of-thought, visual prompting, zero- or few-shot prompting, in-context learning, prompt-based expert calls, retrieval-augmented reasoning, multimodal agents, tool use, plan-execute-inspect loops, web or GUI navigation, multi-agent systems, embodied agents, and perception-action reasoning.","exclusion":"If the main contribution is supervised model alignment or architecture rather than inference procedure or agency, assign to general MLLM construction; if a vertical domain rather than general agency is central, assign to domain-specialized adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["zero-shot vqa with frozen llms via image-to-prompt bridging without end-to-end training","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","embodied multimodal language models grounding continuous sensor modalities for embodied reasoning tasks"]},{"direction_index":5,"label":"D5","name":"Domain-specialized adaptation","description":"Research whose primary objective is to adapt, train, evaluate, or construct data for MLLMs in specialized professional, scientific, clinical, technical, document-intensive, or other vertical domains where domain knowledge and domain conventions are central.","inclusion":"Include work on healthcare, dentistry, ophthalmology, radiology, science instruction data, document understanding, finance, autonomous driving, robotics-as-domain, marine or other professional domains when domain data, domain outputs, or domain-specific evaluation drive the contribution.","exclusion":"General assistant methods merely demonstrated on domain examples belong to general MLLM construction; general agentic control or navigation belongs to reasoning, prompting, and agency; general region/OCR grounding belongs to fine-grained visual grounding.","nearest_boundary":"Reasoning, prompting, and agency","representative_topics":["multimodal large language models for automated and cross-modal dental diagnosis","instruction-tuned vision-language llm for chest x-ray understanding and bidirectional generation","vision-language pre-training for embodied planning and control via chain-of-thought ego-centric data"]},{"direction_index":6,"label":"D6","name":"Efficient, scalable, and cross-lingual adaptation","description":"Research whose primary objective is to make MLLMs cheaper, smaller, more deployable, more scalable, or transferable across languages through parameter-efficient tuning, compression, lightweight grounding, retrieval-assisted adaptation, sparse experts, scaling-law analysis, visual tokenizer design, or multilingual/cross-lingual extension.","inclusion":"Include papers focused on LoRA or adapter tuning, lightweight projections, pruning, distillation, quantization, retrieval-augmented parameter-efficient adaptation, sparse MoE scaling, scaling laws, deployability, multilingual vision-language scaling, bilingual instruction tuning, and cross-lingual multimodal transfer when efficiency, scale, or language transfer is central.","exclusion":"Full general instruction tuning without an efficiency or cross-lingual objective belongs to general MLLM construction; modality-specific speech or video modeling belongs to temporal, audio, and speech multimodality.","nearest_boundary":"General MLLM construction","representative_topics":["jointly scaled multilingual vision-language model with large-capacity vision transformer and modular encoder-decoder interface","scaling laws for generative mixed-modal language models with inter-modal synergy and competition","efficient visual grounding of frozen language models via linear layer finetuning"]},{"direction_index":7,"label":"D7","name":"Temporal, audio, and speech multimodality","description":"Research whose primary objective is to extend MLLMs to temporally extended, video, audio, speech, music, sound, or audio-visual inputs and outputs, including temporal reasoning, video conversation, audio-visual comprehension, speech-language alignment, and audio generation or translation.","inclusion":"Include papers on video-language models, long-video understanding, temporal adapters, audio-visual video assistants, speech-language instruction following, audio and music understanding or generation, speech translation through LLMs, and multimodal systems where audio, speech, or temporal sequence modeling is the main contribution.","exclusion":"Unified any-to-any architectures whose main contribution is modality-general output unification belong to general MLLM construction; prompt-only temporal reasoning belongs to reasoning, prompting, and agency; cross-lingual transfer without speech/audio emphasis belongs to efficient, scalable, and cross-lingual adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["multimodal llm augmentation for audio understanding and generation across diverse sound modalities","unified video sequence understanding via large language model token sequence reasoning","audio-visual language model for temporal video comprehension via cross-modal q-former alignment"]}],"split":"dev"}
{"episode_id":"F024_T2025-07","field_index":24,"field":"multimodal large language models","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Capability evaluation and diagnostics","description":"Research whose primary objective is to measure, benchmark, diagnose, or compare multimodal large language model capabilities, weaknesses, robustness profiles, perception, grounding, and reasoning behavior.","inclusion":"Include papers proposing benchmarks, evaluation suites, metrics, diagnostic protocols, controlled capability tests, evaluator frameworks, or systematic empirical assessments where evaluation is the main contribution.","exclusion":"If evaluation is used mainly to validate a new model, training method, reliability intervention, or domain adaptation, assign to that primary modeling or mitigation direction instead.","nearest_boundary":"Trustworthy reliability and factuality","representative_topics":["comprehensive evaluation benchmark for multimodal large language model perception and cognition","benchmarking generative multimodal llm comprehension across spatial and temporal evaluation dimensions","bridge-architecture evaluation for complex visual reasoning via object-level and pre-training analysis"]},{"direction_index":1,"label":"D1","name":"General MLLM construction","description":"Research whose primary objective is to build, align, architect, pretrain, instruction-tune, or synthesize foundations for general-purpose multimodal LLMs, including vision-language assistants, modular representation design, unified multimodal generation, and broad field-level architectural synthesis.","inclusion":"Include work centered on visual or multimodal instruction tuning, modality-to-LLM alignment, projection or adapter bridges when used for general assistant construction, unified any-to-any or autoregressive multimodal architectures, mixed-modal pretraining, core representation design, and broad surveys of MLLM foundations.","exclusion":"If the primary goal is explicit spatial outputs, efficiency, domain specialization, temporal/audio modeling, safety mitigation, or benchmark construction, assign to that more specific objective instead.","nearest_boundary":"Efficient, scalable, and cross-lingual adaptation","representative_topics":["language model as universal interface for modular multimodal encoder integration via semi-causal pretraining","unified transformer model via discrete token homogenization for diverse vision and language tasks","modularized multimodal pretraining with disentangled and collaborative modality modules"]},{"direction_index":2,"label":"D2","name":"Trustworthy reliability and factuality","description":"Research whose primary objective is to understand, detect, reduce, or control failures of trustworthiness in MLLMs, including hallucination, factual inconsistency, adversarial vulnerability, jailbreaks, robustness failures, catastrophic forgetting, interpretability, model editing, and safety-relevant alignment weaknesses.","inclusion":"Include papers primarily proposing hallucination detection or mitigation, factuality alignment, robust instruction tuning, reward or preference methods for visual factuality, adversarial attack or defense analysis, jailbreak studies, model editing, neuron-level interpretation, or continual-learning reliability analysis.","exclusion":"General capability benchmarks without a reliability-specific intervention belong to capability evaluation; ordinary instruction tuning without a factuality, safety, or robustness objective belongs to general MLLM construction.","nearest_boundary":"Capability evaluation and diagnostics","representative_topics":["mitigating multimodal hallucination via robust instruction tuning with adversarial negative samples","multimodal reward models and preference optimization for hallucination detection and prevention in vision language models","contrastive instruction tuning for hallucination reduction in vision-language models"]},{"direction_index":3,"label":"D3","name":"Fine-grained visual grounding","description":"Research whose primary objective is to make MLLMs understand, refer to, localize, output, or manipulate explicit visual entities and spatial structures such as regions, coordinates, bounding boxes, masks, OCR-visible text, segmentation, geometric relations, and pixel- or panoptic-level representations.","inclusion":"Include work on referring dialogue, coordinate input or output, region-level alignment, spatially grounded instruction tuning, segmentation-linked MLLMs, OCR-aware fine-grained perception, open-world regional recognition, and pixel-level or mask-based interaction.","exclusion":"Broad image-chat or VQA without explicit spatial, region, OCR, or pixel-level grounding belongs to general MLLM construction or reasoning; domain-specific grounding whose main objective is a vertical application belongs to domain-specialized adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["multimodal llm for spatial referential dialogue using natural language coordinate inputs and outputs","visual instruction tuning for text-rich image understanding via ocr and caption-guided data augmentation","visual grounding in multimodal llms via segmentation-based cross-modal entity correspondence"]},{"direction_index":4,"label":"D4","name":"Reasoning, prompting, and agency","description":"Research whose primary objective is to improve multimodal task solving through inference-time reasoning strategies, prompting, in-context learning, retrieval or expert-assisted reasoning, tool orchestration, memory, multi-agent collaboration, embodied interaction, navigation, or action loops.","inclusion":"Include papers on multimodal chain-of-thought, visual prompting, zero- or few-shot prompting, in-context learning, prompt-based expert calls, retrieval-augmented reasoning, multimodal agents, tool use, plan-execute-inspect loops, web or GUI navigation, multi-agent systems, embodied agents, and perception-action reasoning.","exclusion":"If the main contribution is supervised model alignment or architecture rather than inference procedure or agency, assign to general MLLM construction; if a vertical domain rather than general agency is central, assign to domain-specialized adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["zero-shot vqa with frozen llms via image-to-prompt bridging without end-to-end training","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","embodied multimodal language models grounding continuous sensor modalities for embodied reasoning tasks"]},{"direction_index":5,"label":"D5","name":"Domain-specialized adaptation","description":"Research whose primary objective is to adapt, train, evaluate, or construct data for MLLMs in specialized professional, scientific, clinical, technical, document-intensive, or other vertical domains where domain knowledge and domain conventions are central.","inclusion":"Include work on healthcare, dentistry, ophthalmology, radiology, science instruction data, document understanding, finance, autonomous driving, robotics-as-domain, marine or other professional domains when domain data, domain outputs, or domain-specific evaluation drive the contribution.","exclusion":"General assistant methods merely demonstrated on domain examples belong to general MLLM construction; general agentic control or navigation belongs to reasoning, prompting, and agency; general region/OCR grounding belongs to fine-grained visual grounding.","nearest_boundary":"Reasoning, prompting, and agency","representative_topics":["multimodal large language models for automated and cross-modal dental diagnosis","instruction-tuned vision-language llm for chest x-ray understanding and bidirectional generation","vision-language pre-training for embodied planning and control via chain-of-thought ego-centric data"]},{"direction_index":6,"label":"D6","name":"Efficient, scalable, and cross-lingual adaptation","description":"Research whose primary objective is to make MLLMs cheaper, smaller, more deployable, more scalable, or transferable across languages through parameter-efficient tuning, compression, lightweight grounding, retrieval-assisted adaptation, sparse experts, scaling-law analysis, visual tokenizer design, or multilingual/cross-lingual extension.","inclusion":"Include papers focused on LoRA or adapter tuning, lightweight projections, pruning, distillation, quantization, retrieval-augmented parameter-efficient adaptation, sparse MoE scaling, scaling laws, deployability, multilingual vision-language scaling, bilingual instruction tuning, and cross-lingual multimodal transfer when efficiency, scale, or language transfer is central.","exclusion":"Full general instruction tuning without an efficiency or cross-lingual objective belongs to general MLLM construction; modality-specific speech or video modeling belongs to temporal, audio, and speech multimodality.","nearest_boundary":"General MLLM construction","representative_topics":["jointly scaled multilingual vision-language model with large-capacity vision transformer and modular encoder-decoder interface","scaling laws for generative mixed-modal language models with inter-modal synergy and competition","efficient visual grounding of frozen language models via linear layer finetuning"]},{"direction_index":7,"label":"D7","name":"Temporal, audio, and speech multimodality","description":"Research whose primary objective is to extend MLLMs to temporally extended, video, audio, speech, music, sound, or audio-visual inputs and outputs, including temporal reasoning, video conversation, audio-visual comprehension, speech-language alignment, and audio generation or translation.","inclusion":"Include papers on video-language models, long-video understanding, temporal adapters, audio-visual video assistants, speech-language instruction following, audio and music understanding or generation, speech translation through LLMs, and multimodal systems where audio, speech, or temporal sequence modeling is the main contribution.","exclusion":"Unified any-to-any architectures whose main contribution is modality-general output unification belong to general MLLM construction; prompt-only temporal reasoning belongs to reasoning, prompting, and agency; cross-lingual transfer without speech/audio emphasis belongs to efficient, scalable, and cross-lingual adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["multimodal llm augmentation for audio understanding and generation across diverse sound modalities","unified video sequence understanding via large language model token sequence reasoning","audio-visual language model for temporal video comprehension via cross-modal q-former alignment"]}],"split":"dev"}
{"episode_id":"F024_T2026-01","field_index":24,"field":"multimodal large language models","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Capability evaluation and diagnostics","description":"Research whose primary objective is to measure, benchmark, diagnose, or compare multimodal large language model capabilities, weaknesses, robustness profiles, perception, grounding, and reasoning behavior.","inclusion":"Include papers proposing benchmarks, evaluation suites, metrics, diagnostic protocols, controlled capability tests, evaluator frameworks, or systematic empirical assessments where evaluation is the main contribution.","exclusion":"If evaluation is used mainly to validate a new model, training method, reliability intervention, or domain adaptation, assign to that primary modeling or mitigation direction instead.","nearest_boundary":"Trustworthy reliability and factuality","representative_topics":["comprehensive evaluation benchmark for multimodal large language model perception and cognition","benchmarking generative multimodal llm comprehension across spatial and temporal evaluation dimensions","bridge-architecture evaluation for complex visual reasoning via object-level and pre-training analysis"]},{"direction_index":1,"label":"D1","name":"General MLLM construction","description":"Research whose primary objective is to build, align, architect, pretrain, instruction-tune, or synthesize foundations for general-purpose multimodal LLMs, including vision-language assistants, modular representation design, unified multimodal generation, and broad field-level architectural synthesis.","inclusion":"Include work centered on visual or multimodal instruction tuning, modality-to-LLM alignment, projection or adapter bridges when used for general assistant construction, unified any-to-any or autoregressive multimodal architectures, mixed-modal pretraining, core representation design, and broad surveys of MLLM foundations.","exclusion":"If the primary goal is explicit spatial outputs, efficiency, domain specialization, temporal/audio modeling, safety mitigation, or benchmark construction, assign to that more specific objective instead.","nearest_boundary":"Efficient, scalable, and cross-lingual adaptation","representative_topics":["language model as universal interface for modular multimodal encoder integration via semi-causal pretraining","unified transformer model via discrete token homogenization for diverse vision and language tasks","modularized multimodal pretraining with disentangled and collaborative modality modules"]},{"direction_index":2,"label":"D2","name":"Trustworthy reliability and factuality","description":"Research whose primary objective is to understand, detect, reduce, or control failures of trustworthiness in MLLMs, including hallucination, factual inconsistency, adversarial vulnerability, jailbreaks, robustness failures, catastrophic forgetting, interpretability, model editing, and safety-relevant alignment weaknesses.","inclusion":"Include papers primarily proposing hallucination detection or mitigation, factuality alignment, robust instruction tuning, reward or preference methods for visual factuality, adversarial attack or defense analysis, jailbreak studies, model editing, neuron-level interpretation, or continual-learning reliability analysis.","exclusion":"General capability benchmarks without a reliability-specific intervention belong to capability evaluation; ordinary instruction tuning without a factuality, safety, or robustness objective belongs to general MLLM construction.","nearest_boundary":"Capability evaluation and diagnostics","representative_topics":["mitigating multimodal hallucination via robust instruction tuning with adversarial negative samples","multimodal reward models and preference optimization for hallucination detection and prevention in vision language models","contrastive instruction tuning for hallucination reduction in vision-language models"]},{"direction_index":3,"label":"D3","name":"Fine-grained visual grounding","description":"Research whose primary objective is to make MLLMs understand, refer to, localize, output, or manipulate explicit visual entities and spatial structures such as regions, coordinates, bounding boxes, masks, OCR-visible text, segmentation, geometric relations, and pixel- or panoptic-level representations.","inclusion":"Include work on referring dialogue, coordinate input or output, region-level alignment, spatially grounded instruction tuning, segmentation-linked MLLMs, OCR-aware fine-grained perception, open-world regional recognition, and pixel-level or mask-based interaction.","exclusion":"Broad image-chat or VQA without explicit spatial, region, OCR, or pixel-level grounding belongs to general MLLM construction or reasoning; domain-specific grounding whose main objective is a vertical application belongs to domain-specialized adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["multimodal llm for spatial referential dialogue using natural language coordinate inputs and outputs","visual instruction tuning for text-rich image understanding via ocr and caption-guided data augmentation","visual grounding in multimodal llms via segmentation-based cross-modal entity correspondence"]},{"direction_index":4,"label":"D4","name":"Reasoning, prompting, and agency","description":"Research whose primary objective is to improve multimodal task solving through inference-time reasoning strategies, prompting, in-context learning, retrieval or expert-assisted reasoning, tool orchestration, memory, multi-agent collaboration, embodied interaction, navigation, or action loops.","inclusion":"Include papers on multimodal chain-of-thought, visual prompting, zero- or few-shot prompting, in-context learning, prompt-based expert calls, retrieval-augmented reasoning, multimodal agents, tool use, plan-execute-inspect loops, web or GUI navigation, multi-agent systems, embodied agents, and perception-action reasoning.","exclusion":"If the main contribution is supervised model alignment or architecture rather than inference procedure or agency, assign to general MLLM construction; if a vertical domain rather than general agency is central, assign to domain-specialized adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["zero-shot vqa with frozen llms via image-to-prompt bridging without end-to-end training","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","embodied multimodal language models grounding continuous sensor modalities for embodied reasoning tasks"]},{"direction_index":5,"label":"D5","name":"Domain-specialized adaptation","description":"Research whose primary objective is to adapt, train, evaluate, or construct data for MLLMs in specialized professional, scientific, clinical, technical, document-intensive, or other vertical domains where domain knowledge and domain conventions are central.","inclusion":"Include work on healthcare, dentistry, ophthalmology, radiology, science instruction data, document understanding, finance, autonomous driving, robotics-as-domain, marine or other professional domains when domain data, domain outputs, or domain-specific evaluation drive the contribution.","exclusion":"General assistant methods merely demonstrated on domain examples belong to general MLLM construction; general agentic control or navigation belongs to reasoning, prompting, and agency; general region/OCR grounding belongs to fine-grained visual grounding.","nearest_boundary":"Reasoning, prompting, and agency","representative_topics":["multimodal large language models for automated and cross-modal dental diagnosis","instruction-tuned vision-language llm for chest x-ray understanding and bidirectional generation","vision-language pre-training for embodied planning and control via chain-of-thought ego-centric data"]},{"direction_index":6,"label":"D6","name":"Efficient, scalable, and cross-lingual adaptation","description":"Research whose primary objective is to make MLLMs cheaper, smaller, more deployable, more scalable, or transferable across languages through parameter-efficient tuning, compression, lightweight grounding, retrieval-assisted adaptation, sparse experts, scaling-law analysis, visual tokenizer design, or multilingual/cross-lingual extension.","inclusion":"Include papers focused on LoRA or adapter tuning, lightweight projections, pruning, distillation, quantization, retrieval-augmented parameter-efficient adaptation, sparse MoE scaling, scaling laws, deployability, multilingual vision-language scaling, bilingual instruction tuning, and cross-lingual multimodal transfer when efficiency, scale, or language transfer is central.","exclusion":"Full general instruction tuning without an efficiency or cross-lingual objective belongs to general MLLM construction; modality-specific speech or video modeling belongs to temporal, audio, and speech multimodality.","nearest_boundary":"General MLLM construction","representative_topics":["jointly scaled multilingual vision-language model with large-capacity vision transformer and modular encoder-decoder interface","scaling laws for generative mixed-modal language models with inter-modal synergy and competition","efficient visual grounding of frozen language models via linear layer finetuning"]},{"direction_index":7,"label":"D7","name":"Temporal, audio, and speech multimodality","description":"Research whose primary objective is to extend MLLMs to temporally extended, video, audio, speech, music, sound, or audio-visual inputs and outputs, including temporal reasoning, video conversation, audio-visual comprehension, speech-language alignment, and audio generation or translation.","inclusion":"Include papers on video-language models, long-video understanding, temporal adapters, audio-visual video assistants, speech-language instruction following, audio and music understanding or generation, speech translation through LLMs, and multimodal systems where audio, speech, or temporal sequence modeling is the main contribution.","exclusion":"Unified any-to-any architectures whose main contribution is modality-general output unification belong to general MLLM construction; prompt-only temporal reasoning belongs to reasoning, prompting, and agency; cross-lingual transfer without speech/audio emphasis belongs to efficient, scalable, and cross-lingual adaptation.","nearest_boundary":"General MLLM construction","representative_topics":["multimodal llm augmentation for audio understanding and generation across diverse sound modalities","unified video sequence understanding via large language model token sequence reasoning","audio-visual language model for temporal video comprehension via cross-modal q-former alignment"]}],"split":"dev"}
{"episode_id":"F025_T2024-01","field_index":25,"field":"clinical natural language processing","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Clinical fact and code structuring","description":"Methods that convert clinical or health-related free text into structured clinical facts, entities, attributes, relations, phenotypes, social-context variables, medication-safety signals, or standardized administrative and diagnostic codes.","inclusion":"Include work whose primary output is a structured representation derived from text, such as clinical entities, relations, events, SDOH variables, medication changes, adverse drug events, ontology-normalized facts, or ICD/DRG/CPT-like code sets.","exclusion":"Work whose primary output is a patient-level prognosis, trial match, narrative summary, dialogue response, or reusable model belongs to the corresponding decision-support, summarization, communication, or model-development direction.","nearest_boundary":"Decision, prediction, and trial workflow support","representative_topics":["weakly supervised icd code classification of clinical notes using label descriptions and language models","pretrained language model fine-tuning for multilabel icd coding from clinical text","bert-based extraction of medication change attributes from clinical narrative text"]},{"direction_index":1,"label":"D1","name":"Clinical summarization and communication","description":"Systems that generate, simplify, restructure, or converse over clinical content for professional documentation, patient understanding, or consultation-style interaction.","inclusion":"Include work producing narrative clinical summaries, discharge summaries, note generation, dialogue-to-note outputs, plain-language explanations, lay summaries, report simplifications, patient-facing communication, or clinical consultation dialogue systems.","exclusion":"Image-grounded report generation belongs to image-grounded medical language; clinician-facing diagnostic decisions without a communication-centered artifact belong to decision support.","nearest_boundary":"Image-grounded medical language","representative_topics":["granularity optimization for extractive discharge summary generation from unstructured clinical records","ontology-guided multi-document abstractive summarisation of inpatient clinical records","survey of transformer and deep learning methods for medical document summarization"]},{"direction_index":2,"label":"D2","name":"Image-grounded medical language","description":"Vision-language clinical NLP where medical images are primary inputs and language is generated, summarized, or used interactively around imaging findings.","inclusion":"Include radiology report generation, image-grounded impression generation, medical image captioning, image-based diagnostic dialogue, and visual question answering in medical or surgical imaging contexts.","exclusion":"Text-only radiology summarization belongs to clinical summarization and communication; non-image diagnostic reasoning from notes or structured text belongs to decision support.","nearest_boundary":"Clinical summarization and communication","representative_topics":["self-guided radiology report generation via unsupervised domain knowledge and self-improvement comparison","vision-language encoder-decoder models for automated radiology report generation","chest x-ray captioning via multimodal transformer fusion with pathology localization heatmaps"]},{"direction_index":3,"label":"D3","name":"Reusable models and resource adaptation","description":"Research whose main objective is to create, adapt, prompt, transfer, or supply reusable clinical NLP models, resources, data, toolkits, or sample-efficient capabilities across tasks, languages, or low-label settings.","inclusion":"Include domain-adaptive pretraining, instruction or prompt methods as general mechanisms, cross-lingual transfer, low-resource and few-shot learning as the central challenge, synthetic clinical data resources, vocabularies, embeddings, model-ready datasets, and general-purpose clinical NLP toolkits.","exclusion":"If adaptation or prompting is only a means to solve a specific extraction, coding, summarization, trial, diagnosis, or patient-communication workflow, classify by that workflow; if the main contribution is model auditing or privacy protection, use trustworthiness and privacy.","nearest_boundary":"Trustworthy and privacy-preserving NLP","representative_topics":["cross-lingual consumer health vocabulary construction via word-embedding alignment of layperson-generated corpora","low-resource german medical named entity recognition via pretrained language model transfer learning","data-centric meta-dataset curation for biomedical language model pretraining and evaluation"]},{"direction_index":4,"label":"D4","name":"Trustworthy and privacy-preserving NLP","description":"Evaluation, benchmarking, safety analysis, bias auditing, robustness assessment, de-identification, anonymization, and privacy-preserving operation of clinical NLP systems.","inclusion":"Include work primarily measuring or improving trust, safety, fairness, reliability, hallucination behavior, failure modes, benchmark validity, de-identification, pseudonymization, anonymization effects, differential privacy, or federated privacy-preserving clinical NLP evaluation/training.","exclusion":"Work that releases a general model or resource without trust/privacy as the central objective belongs to reusable model adaptation; task-specific evaluations belong to the task direction.","nearest_boundary":"Reusable models and resource adaptation","representative_topics":["federated model-to-data sandbox for unbiased multi-site clinical nlp evaluation","prompt engineering and reasoning strategies for llms on medical question answering benchmarks","gender bias analysis in clinical word embeddings for stereotypical medical categories"]},{"direction_index":5,"label":"D5","name":"Decision, prediction, and trial workflow support","description":"NLP systems that support clinical action by predicting patient outcomes, retrieving or representing EHR evidence, performing diagnostic or treatment reasoning, or processing clinical trials and recruitment workflows.","inclusion":"Include patient-level mortality/readmission/risk prediction, EHR representation for downstream clinical endpoints, cohort or evidence retrieval for care, diagnostic reasoning, treatment or management recommendation, clinical trial retrieval, eligibility processing, patient-trial matching, and trial outcome prediction.","exclusion":"Discrete entity/code extraction without a decision or endpoint belongs to clinical fact and code structuring; evaluation-only studies of reasoning ability belong to trustworthiness and privacy; patient-facing explanations belong to summarization and communication.","nearest_boundary":"Clinical fact and code structuring","representative_topics":["reinforcement learning for scalable online disease diagnosis via multi-model actor-critic fusion","language-based contextual embeddings for tabular medical record representation in clinical prediction","zero-shot clinical trial retrieval via self-supervised meta-structure and ontology-guided contrastive learning"]},{"direction_index":6,"label":"D6","name":"Mental and neurocognitive language assessment","description":"NLP for assessing, monitoring, explaining, or supporting mental health, psychotherapy, counseling, cognitive decline, dementia, Alzheimer’s disease, aphasia, and related language-disorder conditions.","inclusion":"Include depression, anxiety, suicide-risk, psychotherapy dialogue modeling, counseling summarization or strategy analysis, working-alliance prediction, dementia or Alzheimer’s detection from language, ASR effects on dementia classification, and aphasia-oriented language assistance.","exclusion":"Population-level health social media surveillance without individual mental-health or counseling assessment belongs to public health evidence surveillance; general patient-facing health support belongs to summarization and communication.","nearest_boundary":"Public health and biomedical evidence surveillance","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","behavioral interpretability tests for generalizable bert-based depression detection from text","machine learning classifiers for automated dementia language impairment assessment"]},{"direction_index":7,"label":"D7","name":"Public health and biomedical evidence surveillance","description":"NLP for monitoring, retrieving, classifying, and synthesizing health evidence from public, social, regulatory, or biomedical document collections rather than individual clinical records.","inclusion":"Include social media health surveillance, misinformation or conflicting health-claim detection, disease mention monitoring in public posts, public-health signal tracking, biomedical literature screening, systematic review automation, regulatory medical-document classification, and knowledge discovery over biomedical document collections.","exclusion":"Medication or SDOH extraction from clinical documents belongs to clinical fact and code structuring; mental-health assessment of individuals from social media belongs to mental and neurocognitive assessment; trial-specific processing belongs to decision, prediction, and trial workflow support.","nearest_boundary":"Mental and neurocognitive language assessment","representative_topics":["low-resource disease named entity recognition in spanish social media using medical gazetteers","transformer-based premise classification in covid-19 social media health text","pre-trained language model fine-tuning for contextualized health conflict detection with synthetic data augmentation"]}],"split":"dev"}
{"episode_id":"F025_T2024-07","field_index":25,"field":"clinical natural language processing","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Clinical fact and code structuring","description":"Methods that convert clinical or health-related free text into structured clinical facts, entities, attributes, relations, phenotypes, social-context variables, medication-safety signals, or standardized administrative and diagnostic codes.","inclusion":"Include work whose primary output is a structured representation derived from text, such as clinical entities, relations, events, SDOH variables, medication changes, adverse drug events, ontology-normalized facts, or ICD/DRG/CPT-like code sets.","exclusion":"Work whose primary output is a patient-level prognosis, trial match, narrative summary, dialogue response, or reusable model belongs to the corresponding decision-support, summarization, communication, or model-development direction.","nearest_boundary":"Decision, prediction, and trial workflow support","representative_topics":["weakly supervised icd code classification of clinical notes using label descriptions and language models","pretrained language model fine-tuning for multilabel icd coding from clinical text","bert-based extraction of medication change attributes from clinical narrative text"]},{"direction_index":1,"label":"D1","name":"Clinical summarization and communication","description":"Systems that generate, simplify, restructure, or converse over clinical content for professional documentation, patient understanding, or consultation-style interaction.","inclusion":"Include work producing narrative clinical summaries, discharge summaries, note generation, dialogue-to-note outputs, plain-language explanations, lay summaries, report simplifications, patient-facing communication, or clinical consultation dialogue systems.","exclusion":"Image-grounded report generation belongs to image-grounded medical language; clinician-facing diagnostic decisions without a communication-centered artifact belong to decision support.","nearest_boundary":"Image-grounded medical language","representative_topics":["granularity optimization for extractive discharge summary generation from unstructured clinical records","ontology-guided multi-document abstractive summarisation of inpatient clinical records","survey of transformer and deep learning methods for medical document summarization"]},{"direction_index":2,"label":"D2","name":"Image-grounded medical language","description":"Vision-language clinical NLP where medical images are primary inputs and language is generated, summarized, or used interactively around imaging findings.","inclusion":"Include radiology report generation, image-grounded impression generation, medical image captioning, image-based diagnostic dialogue, and visual question answering in medical or surgical imaging contexts.","exclusion":"Text-only radiology summarization belongs to clinical summarization and communication; non-image diagnostic reasoning from notes or structured text belongs to decision support.","nearest_boundary":"Clinical summarization and communication","representative_topics":["self-guided radiology report generation via unsupervised domain knowledge and self-improvement comparison","vision-language encoder-decoder models for automated radiology report generation","chest x-ray captioning via multimodal transformer fusion with pathology localization heatmaps"]},{"direction_index":3,"label":"D3","name":"Reusable models and resource adaptation","description":"Research whose main objective is to create, adapt, prompt, transfer, or supply reusable clinical NLP models, resources, data, toolkits, or sample-efficient capabilities across tasks, languages, or low-label settings.","inclusion":"Include domain-adaptive pretraining, instruction or prompt methods as general mechanisms, cross-lingual transfer, low-resource and few-shot learning as the central challenge, synthetic clinical data resources, vocabularies, embeddings, model-ready datasets, and general-purpose clinical NLP toolkits.","exclusion":"If adaptation or prompting is only a means to solve a specific extraction, coding, summarization, trial, diagnosis, or patient-communication workflow, classify by that workflow; if the main contribution is model auditing or privacy protection, use trustworthiness and privacy.","nearest_boundary":"Trustworthy and privacy-preserving NLP","representative_topics":["cross-lingual consumer health vocabulary construction via word-embedding alignment of layperson-generated corpora","low-resource german medical named entity recognition via pretrained language model transfer learning","data-centric meta-dataset curation for biomedical language model pretraining and evaluation"]},{"direction_index":4,"label":"D4","name":"Trustworthy and privacy-preserving NLP","description":"Evaluation, benchmarking, safety analysis, bias auditing, robustness assessment, de-identification, anonymization, and privacy-preserving operation of clinical NLP systems.","inclusion":"Include work primarily measuring or improving trust, safety, fairness, reliability, hallucination behavior, failure modes, benchmark validity, de-identification, pseudonymization, anonymization effects, differential privacy, or federated privacy-preserving clinical NLP evaluation/training.","exclusion":"Work that releases a general model or resource without trust/privacy as the central objective belongs to reusable model adaptation; task-specific evaluations belong to the task direction.","nearest_boundary":"Reusable models and resource adaptation","representative_topics":["federated model-to-data sandbox for unbiased multi-site clinical nlp evaluation","prompt engineering and reasoning strategies for llms on medical question answering benchmarks","gender bias analysis in clinical word embeddings for stereotypical medical categories"]},{"direction_index":5,"label":"D5","name":"Decision, prediction, and trial workflow support","description":"NLP systems that support clinical action by predicting patient outcomes, retrieving or representing EHR evidence, performing diagnostic or treatment reasoning, or processing clinical trials and recruitment workflows.","inclusion":"Include patient-level mortality/readmission/risk prediction, EHR representation for downstream clinical endpoints, cohort or evidence retrieval for care, diagnostic reasoning, treatment or management recommendation, clinical trial retrieval, eligibility processing, patient-trial matching, and trial outcome prediction.","exclusion":"Discrete entity/code extraction without a decision or endpoint belongs to clinical fact and code structuring; evaluation-only studies of reasoning ability belong to trustworthiness and privacy; patient-facing explanations belong to summarization and communication.","nearest_boundary":"Clinical fact and code structuring","representative_topics":["reinforcement learning for scalable online disease diagnosis via multi-model actor-critic fusion","language-based contextual embeddings for tabular medical record representation in clinical prediction","zero-shot clinical trial retrieval via self-supervised meta-structure and ontology-guided contrastive learning"]},{"direction_index":6,"label":"D6","name":"Mental and neurocognitive language assessment","description":"NLP for assessing, monitoring, explaining, or supporting mental health, psychotherapy, counseling, cognitive decline, dementia, Alzheimer’s disease, aphasia, and related language-disorder conditions.","inclusion":"Include depression, anxiety, suicide-risk, psychotherapy dialogue modeling, counseling summarization or strategy analysis, working-alliance prediction, dementia or Alzheimer’s detection from language, ASR effects on dementia classification, and aphasia-oriented language assistance.","exclusion":"Population-level health social media surveillance without individual mental-health or counseling assessment belongs to public health evidence surveillance; general patient-facing health support belongs to summarization and communication.","nearest_boundary":"Public health and biomedical evidence surveillance","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","behavioral interpretability tests for generalizable bert-based depression detection from text","machine learning classifiers for automated dementia language impairment assessment"]},{"direction_index":7,"label":"D7","name":"Public health and biomedical evidence surveillance","description":"NLP for monitoring, retrieving, classifying, and synthesizing health evidence from public, social, regulatory, or biomedical document collections rather than individual clinical records.","inclusion":"Include social media health surveillance, misinformation or conflicting health-claim detection, disease mention monitoring in public posts, public-health signal tracking, biomedical literature screening, systematic review automation, regulatory medical-document classification, and knowledge discovery over biomedical document collections.","exclusion":"Medication or SDOH extraction from clinical documents belongs to clinical fact and code structuring; mental-health assessment of individuals from social media belongs to mental and neurocognitive assessment; trial-specific processing belongs to decision, prediction, and trial workflow support.","nearest_boundary":"Mental and neurocognitive language assessment","representative_topics":["low-resource disease named entity recognition in spanish social media using medical gazetteers","transformer-based premise classification in covid-19 social media health text","pre-trained language model fine-tuning for contextualized health conflict detection with synthetic data augmentation"]}],"split":"dev"}
{"episode_id":"F025_T2025-01","field_index":25,"field":"clinical natural language processing","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Clinical fact and code structuring","description":"Methods that convert clinical or health-related free text into structured clinical facts, entities, attributes, relations, phenotypes, social-context variables, medication-safety signals, or standardized administrative and diagnostic codes.","inclusion":"Include work whose primary output is a structured representation derived from text, such as clinical entities, relations, events, SDOH variables, medication changes, adverse drug events, ontology-normalized facts, or ICD/DRG/CPT-like code sets.","exclusion":"Work whose primary output is a patient-level prognosis, trial match, narrative summary, dialogue response, or reusable model belongs to the corresponding decision-support, summarization, communication, or model-development direction.","nearest_boundary":"Decision, prediction, and trial workflow support","representative_topics":["weakly supervised icd code classification of clinical notes using label descriptions and language models","pretrained language model fine-tuning for multilabel icd coding from clinical text","bert-based extraction of medication change attributes from clinical narrative text"]},{"direction_index":1,"label":"D1","name":"Clinical summarization and communication","description":"Systems that generate, simplify, restructure, or converse over clinical content for professional documentation, patient understanding, or consultation-style interaction.","inclusion":"Include work producing narrative clinical summaries, discharge summaries, note generation, dialogue-to-note outputs, plain-language explanations, lay summaries, report simplifications, patient-facing communication, or clinical consultation dialogue systems.","exclusion":"Image-grounded report generation belongs to image-grounded medical language; clinician-facing diagnostic decisions without a communication-centered artifact belong to decision support.","nearest_boundary":"Image-grounded medical language","representative_topics":["granularity optimization for extractive discharge summary generation from unstructured clinical records","ontology-guided multi-document abstractive summarisation of inpatient clinical records","survey of transformer and deep learning methods for medical document summarization"]},{"direction_index":2,"label":"D2","name":"Image-grounded medical language","description":"Vision-language clinical NLP where medical images are primary inputs and language is generated, summarized, or used interactively around imaging findings.","inclusion":"Include radiology report generation, image-grounded impression generation, medical image captioning, image-based diagnostic dialogue, and visual question answering in medical or surgical imaging contexts.","exclusion":"Text-only radiology summarization belongs to clinical summarization and communication; non-image diagnostic reasoning from notes or structured text belongs to decision support.","nearest_boundary":"Clinical summarization and communication","representative_topics":["self-guided radiology report generation via unsupervised domain knowledge and self-improvement comparison","vision-language encoder-decoder models for automated radiology report generation","chest x-ray captioning via multimodal transformer fusion with pathology localization heatmaps"]},{"direction_index":3,"label":"D3","name":"Reusable models and resource adaptation","description":"Research whose main objective is to create, adapt, prompt, transfer, or supply reusable clinical NLP models, resources, data, toolkits, or sample-efficient capabilities across tasks, languages, or low-label settings.","inclusion":"Include domain-adaptive pretraining, instruction or prompt methods as general mechanisms, cross-lingual transfer, low-resource and few-shot learning as the central challenge, synthetic clinical data resources, vocabularies, embeddings, model-ready datasets, and general-purpose clinical NLP toolkits.","exclusion":"If adaptation or prompting is only a means to solve a specific extraction, coding, summarization, trial, diagnosis, or patient-communication workflow, classify by that workflow; if the main contribution is model auditing or privacy protection, use trustworthiness and privacy.","nearest_boundary":"Trustworthy and privacy-preserving NLP","representative_topics":["cross-lingual consumer health vocabulary construction via word-embedding alignment of layperson-generated corpora","low-resource german medical named entity recognition via pretrained language model transfer learning","data-centric meta-dataset curation for biomedical language model pretraining and evaluation"]},{"direction_index":4,"label":"D4","name":"Trustworthy and privacy-preserving NLP","description":"Evaluation, benchmarking, safety analysis, bias auditing, robustness assessment, de-identification, anonymization, and privacy-preserving operation of clinical NLP systems.","inclusion":"Include work primarily measuring or improving trust, safety, fairness, reliability, hallucination behavior, failure modes, benchmark validity, de-identification, pseudonymization, anonymization effects, differential privacy, or federated privacy-preserving clinical NLP evaluation/training.","exclusion":"Work that releases a general model or resource without trust/privacy as the central objective belongs to reusable model adaptation; task-specific evaluations belong to the task direction.","nearest_boundary":"Reusable models and resource adaptation","representative_topics":["federated model-to-data sandbox for unbiased multi-site clinical nlp evaluation","prompt engineering and reasoning strategies for llms on medical question answering benchmarks","gender bias analysis in clinical word embeddings for stereotypical medical categories"]},{"direction_index":5,"label":"D5","name":"Decision, prediction, and trial workflow support","description":"NLP systems that support clinical action by predicting patient outcomes, retrieving or representing EHR evidence, performing diagnostic or treatment reasoning, or processing clinical trials and recruitment workflows.","inclusion":"Include patient-level mortality/readmission/risk prediction, EHR representation for downstream clinical endpoints, cohort or evidence retrieval for care, diagnostic reasoning, treatment or management recommendation, clinical trial retrieval, eligibility processing, patient-trial matching, and trial outcome prediction.","exclusion":"Discrete entity/code extraction without a decision or endpoint belongs to clinical fact and code structuring; evaluation-only studies of reasoning ability belong to trustworthiness and privacy; patient-facing explanations belong to summarization and communication.","nearest_boundary":"Clinical fact and code structuring","representative_topics":["reinforcement learning for scalable online disease diagnosis via multi-model actor-critic fusion","language-based contextual embeddings for tabular medical record representation in clinical prediction","zero-shot clinical trial retrieval via self-supervised meta-structure and ontology-guided contrastive learning"]},{"direction_index":6,"label":"D6","name":"Mental and neurocognitive language assessment","description":"NLP for assessing, monitoring, explaining, or supporting mental health, psychotherapy, counseling, cognitive decline, dementia, Alzheimer’s disease, aphasia, and related language-disorder conditions.","inclusion":"Include depression, anxiety, suicide-risk, psychotherapy dialogue modeling, counseling summarization or strategy analysis, working-alliance prediction, dementia or Alzheimer’s detection from language, ASR effects on dementia classification, and aphasia-oriented language assistance.","exclusion":"Population-level health social media surveillance without individual mental-health or counseling assessment belongs to public health evidence surveillance; general patient-facing health support belongs to summarization and communication.","nearest_boundary":"Public health and biomedical evidence surveillance","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","behavioral interpretability tests for generalizable bert-based depression detection from text","machine learning classifiers for automated dementia language impairment assessment"]},{"direction_index":7,"label":"D7","name":"Public health and biomedical evidence surveillance","description":"NLP for monitoring, retrieving, classifying, and synthesizing health evidence from public, social, regulatory, or biomedical document collections rather than individual clinical records.","inclusion":"Include social media health surveillance, misinformation or conflicting health-claim detection, disease mention monitoring in public posts, public-health signal tracking, biomedical literature screening, systematic review automation, regulatory medical-document classification, and knowledge discovery over biomedical document collections.","exclusion":"Medication or SDOH extraction from clinical documents belongs to clinical fact and code structuring; mental-health assessment of individuals from social media belongs to mental and neurocognitive assessment; trial-specific processing belongs to decision, prediction, and trial workflow support.","nearest_boundary":"Mental and neurocognitive language assessment","representative_topics":["low-resource disease named entity recognition in spanish social media using medical gazetteers","transformer-based premise classification in covid-19 social media health text","pre-trained language model fine-tuning for contextualized health conflict detection with synthetic data augmentation"]}],"split":"dev"}
{"episode_id":"F025_T2025-07","field_index":25,"field":"clinical natural language processing","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Clinical fact and code structuring","description":"Methods that convert clinical or health-related free text into structured clinical facts, entities, attributes, relations, phenotypes, social-context variables, medication-safety signals, or standardized administrative and diagnostic codes.","inclusion":"Include work whose primary output is a structured representation derived from text, such as clinical entities, relations, events, SDOH variables, medication changes, adverse drug events, ontology-normalized facts, or ICD/DRG/CPT-like code sets.","exclusion":"Work whose primary output is a patient-level prognosis, trial match, narrative summary, dialogue response, or reusable model belongs to the corresponding decision-support, summarization, communication, or model-development direction.","nearest_boundary":"Decision, prediction, and trial workflow support","representative_topics":["weakly supervised icd code classification of clinical notes using label descriptions and language models","pretrained language model fine-tuning for multilabel icd coding from clinical text","bert-based extraction of medication change attributes from clinical narrative text"]},{"direction_index":1,"label":"D1","name":"Clinical summarization and communication","description":"Systems that generate, simplify, restructure, or converse over clinical content for professional documentation, patient understanding, or consultation-style interaction.","inclusion":"Include work producing narrative clinical summaries, discharge summaries, note generation, dialogue-to-note outputs, plain-language explanations, lay summaries, report simplifications, patient-facing communication, or clinical consultation dialogue systems.","exclusion":"Image-grounded report generation belongs to image-grounded medical language; clinician-facing diagnostic decisions without a communication-centered artifact belong to decision support.","nearest_boundary":"Image-grounded medical language","representative_topics":["granularity optimization for extractive discharge summary generation from unstructured clinical records","ontology-guided multi-document abstractive summarisation of inpatient clinical records","survey of transformer and deep learning methods for medical document summarization"]},{"direction_index":2,"label":"D2","name":"Image-grounded medical language","description":"Vision-language clinical NLP where medical images are primary inputs and language is generated, summarized, or used interactively around imaging findings.","inclusion":"Include radiology report generation, image-grounded impression generation, medical image captioning, image-based diagnostic dialogue, and visual question answering in medical or surgical imaging contexts.","exclusion":"Text-only radiology summarization belongs to clinical summarization and communication; non-image diagnostic reasoning from notes or structured text belongs to decision support.","nearest_boundary":"Clinical summarization and communication","representative_topics":["self-guided radiology report generation via unsupervised domain knowledge and self-improvement comparison","vision-language encoder-decoder models for automated radiology report generation","chest x-ray captioning via multimodal transformer fusion with pathology localization heatmaps"]},{"direction_index":3,"label":"D3","name":"Reusable models and resource adaptation","description":"Research whose main objective is to create, adapt, prompt, transfer, or supply reusable clinical NLP models, resources, data, toolkits, or sample-efficient capabilities across tasks, languages, or low-label settings.","inclusion":"Include domain-adaptive pretraining, instruction or prompt methods as general mechanisms, cross-lingual transfer, low-resource and few-shot learning as the central challenge, synthetic clinical data resources, vocabularies, embeddings, model-ready datasets, and general-purpose clinical NLP toolkits.","exclusion":"If adaptation or prompting is only a means to solve a specific extraction, coding, summarization, trial, diagnosis, or patient-communication workflow, classify by that workflow; if the main contribution is model auditing or privacy protection, use trustworthiness and privacy.","nearest_boundary":"Trustworthy and privacy-preserving NLP","representative_topics":["cross-lingual consumer health vocabulary construction via word-embedding alignment of layperson-generated corpora","low-resource german medical named entity recognition via pretrained language model transfer learning","data-centric meta-dataset curation for biomedical language model pretraining and evaluation"]},{"direction_index":4,"label":"D4","name":"Trustworthy and privacy-preserving NLP","description":"Evaluation, benchmarking, safety analysis, bias auditing, robustness assessment, de-identification, anonymization, and privacy-preserving operation of clinical NLP systems.","inclusion":"Include work primarily measuring or improving trust, safety, fairness, reliability, hallucination behavior, failure modes, benchmark validity, de-identification, pseudonymization, anonymization effects, differential privacy, or federated privacy-preserving clinical NLP evaluation/training.","exclusion":"Work that releases a general model or resource without trust/privacy as the central objective belongs to reusable model adaptation; task-specific evaluations belong to the task direction.","nearest_boundary":"Reusable models and resource adaptation","representative_topics":["federated model-to-data sandbox for unbiased multi-site clinical nlp evaluation","prompt engineering and reasoning strategies for llms on medical question answering benchmarks","gender bias analysis in clinical word embeddings for stereotypical medical categories"]},{"direction_index":5,"label":"D5","name":"Decision, prediction, and trial workflow support","description":"NLP systems that support clinical action by predicting patient outcomes, retrieving or representing EHR evidence, performing diagnostic or treatment reasoning, or processing clinical trials and recruitment workflows.","inclusion":"Include patient-level mortality/readmission/risk prediction, EHR representation for downstream clinical endpoints, cohort or evidence retrieval for care, diagnostic reasoning, treatment or management recommendation, clinical trial retrieval, eligibility processing, patient-trial matching, and trial outcome prediction.","exclusion":"Discrete entity/code extraction without a decision or endpoint belongs to clinical fact and code structuring; evaluation-only studies of reasoning ability belong to trustworthiness and privacy; patient-facing explanations belong to summarization and communication.","nearest_boundary":"Clinical fact and code structuring","representative_topics":["reinforcement learning for scalable online disease diagnosis via multi-model actor-critic fusion","language-based contextual embeddings for tabular medical record representation in clinical prediction","zero-shot clinical trial retrieval via self-supervised meta-structure and ontology-guided contrastive learning"]},{"direction_index":6,"label":"D6","name":"Mental and neurocognitive language assessment","description":"NLP for assessing, monitoring, explaining, or supporting mental health, psychotherapy, counseling, cognitive decline, dementia, Alzheimer’s disease, aphasia, and related language-disorder conditions.","inclusion":"Include depression, anxiety, suicide-risk, psychotherapy dialogue modeling, counseling summarization or strategy analysis, working-alliance prediction, dementia or Alzheimer’s detection from language, ASR effects on dementia classification, and aphasia-oriented language assistance.","exclusion":"Population-level health social media surveillance without individual mental-health or counseling assessment belongs to public health evidence surveillance; general patient-facing health support belongs to summarization and communication.","nearest_boundary":"Public health and biomedical evidence surveillance","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","behavioral interpretability tests for generalizable bert-based depression detection from text","machine learning classifiers for automated dementia language impairment assessment"]},{"direction_index":7,"label":"D7","name":"Public health and biomedical evidence surveillance","description":"NLP for monitoring, retrieving, classifying, and synthesizing health evidence from public, social, regulatory, or biomedical document collections rather than individual clinical records.","inclusion":"Include social media health surveillance, misinformation or conflicting health-claim detection, disease mention monitoring in public posts, public-health signal tracking, biomedical literature screening, systematic review automation, regulatory medical-document classification, and knowledge discovery over biomedical document collections.","exclusion":"Medication or SDOH extraction from clinical documents belongs to clinical fact and code structuring; mental-health assessment of individuals from social media belongs to mental and neurocognitive assessment; trial-specific processing belongs to decision, prediction, and trial workflow support.","nearest_boundary":"Mental and neurocognitive language assessment","representative_topics":["low-resource disease named entity recognition in spanish social media using medical gazetteers","transformer-based premise classification in covid-19 social media health text","pre-trained language model fine-tuning for contextualized health conflict detection with synthetic data augmentation"]}],"split":"dev"}
{"episode_id":"F025_T2026-01","field_index":25,"field":"clinical natural language processing","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Clinical fact and code structuring","description":"Methods that convert clinical or health-related free text into structured clinical facts, entities, attributes, relations, phenotypes, social-context variables, medication-safety signals, or standardized administrative and diagnostic codes.","inclusion":"Include work whose primary output is a structured representation derived from text, such as clinical entities, relations, events, SDOH variables, medication changes, adverse drug events, ontology-normalized facts, or ICD/DRG/CPT-like code sets.","exclusion":"Work whose primary output is a patient-level prognosis, trial match, narrative summary, dialogue response, or reusable model belongs to the corresponding decision-support, summarization, communication, or model-development direction.","nearest_boundary":"Decision, prediction, and trial workflow support","representative_topics":["weakly supervised icd code classification of clinical notes using label descriptions and language models","pretrained language model fine-tuning for multilabel icd coding from clinical text","bert-based extraction of medication change attributes from clinical narrative text"]},{"direction_index":1,"label":"D1","name":"Clinical summarization and communication","description":"Systems that generate, simplify, restructure, or converse over clinical content for professional documentation, patient understanding, or consultation-style interaction.","inclusion":"Include work producing narrative clinical summaries, discharge summaries, note generation, dialogue-to-note outputs, plain-language explanations, lay summaries, report simplifications, patient-facing communication, or clinical consultation dialogue systems.","exclusion":"Image-grounded report generation belongs to image-grounded medical language; clinician-facing diagnostic decisions without a communication-centered artifact belong to decision support.","nearest_boundary":"Image-grounded medical language","representative_topics":["granularity optimization for extractive discharge summary generation from unstructured clinical records","ontology-guided multi-document abstractive summarisation of inpatient clinical records","survey of transformer and deep learning methods for medical document summarization"]},{"direction_index":2,"label":"D2","name":"Image-grounded medical language","description":"Vision-language clinical NLP where medical images are primary inputs and language is generated, summarized, or used interactively around imaging findings.","inclusion":"Include radiology report generation, image-grounded impression generation, medical image captioning, image-based diagnostic dialogue, and visual question answering in medical or surgical imaging contexts.","exclusion":"Text-only radiology summarization belongs to clinical summarization and communication; non-image diagnostic reasoning from notes or structured text belongs to decision support.","nearest_boundary":"Clinical summarization and communication","representative_topics":["self-guided radiology report generation via unsupervised domain knowledge and self-improvement comparison","vision-language encoder-decoder models for automated radiology report generation","chest x-ray captioning via multimodal transformer fusion with pathology localization heatmaps"]},{"direction_index":3,"label":"D3","name":"Reusable models and resource adaptation","description":"Research whose main objective is to create, adapt, prompt, transfer, or supply reusable clinical NLP models, resources, data, toolkits, or sample-efficient capabilities across tasks, languages, or low-label settings.","inclusion":"Include domain-adaptive pretraining, instruction or prompt methods as general mechanisms, cross-lingual transfer, low-resource and few-shot learning as the central challenge, synthetic clinical data resources, vocabularies, embeddings, model-ready datasets, and general-purpose clinical NLP toolkits.","exclusion":"If adaptation or prompting is only a means to solve a specific extraction, coding, summarization, trial, diagnosis, or patient-communication workflow, classify by that workflow; if the main contribution is model auditing or privacy protection, use trustworthiness and privacy.","nearest_boundary":"Trustworthy and privacy-preserving NLP","representative_topics":["cross-lingual consumer health vocabulary construction via word-embedding alignment of layperson-generated corpora","low-resource german medical named entity recognition via pretrained language model transfer learning","data-centric meta-dataset curation for biomedical language model pretraining and evaluation"]},{"direction_index":4,"label":"D4","name":"Trustworthy and privacy-preserving NLP","description":"Evaluation, benchmarking, safety analysis, bias auditing, robustness assessment, de-identification, anonymization, and privacy-preserving operation of clinical NLP systems.","inclusion":"Include work primarily measuring or improving trust, safety, fairness, reliability, hallucination behavior, failure modes, benchmark validity, de-identification, pseudonymization, anonymization effects, differential privacy, or federated privacy-preserving clinical NLP evaluation/training.","exclusion":"Work that releases a general model or resource without trust/privacy as the central objective belongs to reusable model adaptation; task-specific evaluations belong to the task direction.","nearest_boundary":"Reusable models and resource adaptation","representative_topics":["federated model-to-data sandbox for unbiased multi-site clinical nlp evaluation","prompt engineering and reasoning strategies for llms on medical question answering benchmarks","gender bias analysis in clinical word embeddings for stereotypical medical categories"]},{"direction_index":5,"label":"D5","name":"Decision, prediction, and trial workflow support","description":"NLP systems that support clinical action by predicting patient outcomes, retrieving or representing EHR evidence, performing diagnostic or treatment reasoning, or processing clinical trials and recruitment workflows.","inclusion":"Include patient-level mortality/readmission/risk prediction, EHR representation for downstream clinical endpoints, cohort or evidence retrieval for care, diagnostic reasoning, treatment or management recommendation, clinical trial retrieval, eligibility processing, patient-trial matching, and trial outcome prediction.","exclusion":"Discrete entity/code extraction without a decision or endpoint belongs to clinical fact and code structuring; evaluation-only studies of reasoning ability belong to trustworthiness and privacy; patient-facing explanations belong to summarization and communication.","nearest_boundary":"Clinical fact and code structuring","representative_topics":["reinforcement learning for scalable online disease diagnosis via multi-model actor-critic fusion","language-based contextual embeddings for tabular medical record representation in clinical prediction","zero-shot clinical trial retrieval via self-supervised meta-structure and ontology-guided contrastive learning"]},{"direction_index":6,"label":"D6","name":"Mental and neurocognitive language assessment","description":"NLP for assessing, monitoring, explaining, or supporting mental health, psychotherapy, counseling, cognitive decline, dementia, Alzheimer’s disease, aphasia, and related language-disorder conditions.","inclusion":"Include depression, anxiety, suicide-risk, psychotherapy dialogue modeling, counseling summarization or strategy analysis, working-alliance prediction, dementia or Alzheimer’s detection from language, ASR effects on dementia classification, and aphasia-oriented language assistance.","exclusion":"Population-level health social media surveillance without individual mental-health or counseling assessment belongs to public health evidence surveillance; general patient-facing health support belongs to summarization and communication.","nearest_boundary":"Public health and biomedical evidence surveillance","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","behavioral interpretability tests for generalizable bert-based depression detection from text","machine learning classifiers for automated dementia language impairment assessment"]},{"direction_index":7,"label":"D7","name":"Public health and biomedical evidence surveillance","description":"NLP for monitoring, retrieving, classifying, and synthesizing health evidence from public, social, regulatory, or biomedical document collections rather than individual clinical records.","inclusion":"Include social media health surveillance, misinformation or conflicting health-claim detection, disease mention monitoring in public posts, public-health signal tracking, biomedical literature screening, systematic review automation, regulatory medical-document classification, and knowledge discovery over biomedical document collections.","exclusion":"Medication or SDOH extraction from clinical documents belongs to clinical fact and code structuring; mental-health assessment of individuals from social media belongs to mental and neurocognitive assessment; trial-specific processing belongs to decision, prediction, and trial workflow support.","nearest_boundary":"Mental and neurocognitive language assessment","representative_topics":["low-resource disease named entity recognition in spanish social media using medical gazetteers","transformer-based premise classification in covid-19 social media health text","pre-trained language model fine-tuning for contextualized health conflict detection with synthetic data augmentation"]}],"split":"dev"}
{"episode_id":"F028_T2024-01","field_index":28,"field":"quantum machine learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Trainable variational circuits","description":"Improve the optimization behavior, trainability, architecture selection, initialization, compression, and parameter efficiency of parameterized quantum circuits used as learning models.","inclusion":"Include work whose primary objective is diagnosing or mitigating barren plateaus, laziness, vanishing gradients, optimizer pathologies, shot-efficient gradients, pruning, compression, overparameterization, hyperparameter effects, or architecture search for VQCs/QNNs.","exclusion":"If the main objective is a supervised predictor, kernel method, generator, distributed protocol, trust guarantee, physical-system learner, or structural prior, assign to that objective even if trainability is evaluated.","nearest_boundary":"Foundations and structural priors","representative_topics":["theoretical analysis of laziness and barren plateau in overparameterized variational quantum circuits","hyperparameter importance analysis of quantum neural networks using functional anova on small-scale datasets","quantum circuit architecture search using gate-level distance metrics and bayesian optimization"]},{"direction_index":1,"label":"D1","name":"Foundations and structural priors","description":"Establish formal capacity, learnability, generalization, advantage, limitation, and inductive-bias principles for quantum learning models, including symmetry-, geometry-, representation-, and tensor-network-based architectural priors.","inclusion":"Include work whose primary contribution is a theorem, bound, analytic framework, quantum-versus-classical comparison, or principled structural prior explaining what quantum learning models can represent, generalize, or learn.","exclusion":"If the main objective is empirical optimization of trainability, a task-specific supervised predictor, kernel construction, or learning an unknown physical quantum object, assign to that nearer program.","nearest_boundary":"Trainable variational circuits","representative_topics":["theoretical error bounds for variational quantum circuits with tensor-train dimensionality reduction for functional regression","theoretical bounds on data encoding concentration in parameterized quantum circuits for quantum classifiers","classical surrogate analysis of variational quantum learning models and their inductive biases"]},{"direction_index":2,"label":"D2","name":"Classical-data prediction models","description":"Develop quantum or hybrid models for supervised prediction on classical data, including neural classifiers/regressors, quantum kernels, feature maps, graph-data models, financial predictors, and other structured classical-data prediction pipelines.","inclusion":"Include work whose primary objective is classification or regression performance, feature extraction, kernel construction, domain prediction, graph learning, or financial/economic prediction using quantum, hybrid, or quantum-inspired models.","exclusion":"If the central object is a generative model, reinforcement-learning agent, language/sequence representation, trainability method, general theorem, or annealing/linear-algebra primitive, assign to that program instead.","nearest_boundary":"Generative distribution learning","representative_topics":["variational quantum approximate svm with inference transfer for scalable hyper-linear classification","scalability analysis of quantum convolutional neural networks via measurement step reduction","quantum feature extraction for thz multi-layer depth and content recognition"]},{"direction_index":3,"label":"D3","name":"Generative distribution learning","description":"Create, train, analyze, and evaluate quantum or hybrid generative models for sampling, distribution learning, data generation, molecular generation, image generation, and Born-machine-style modeling.","inclusion":"Include work whose primary output is a generative model or generative-learning method, including quantum GANs, Born machines, VAEs, energy-based models, tensor-network generative models, and training or evaluation methods for generators.","exclusion":"If the generator is mainly a tool for tomography, physical-state characterization, supervised prediction, or generic VQC trainability analysis, assign to that primary objective.","nearest_boundary":"Quantum physical-system learning","representative_topics":["survey of quantum generative learning models for near-term quantum advantage","tensor network generative models with entanglement entropy variations for unsupervised informative feature recognition","sample-efficient quantum born machine training via coding rate reduction to prevent mode collapse"]},{"direction_index":4,"label":"D4","name":"Quantum physical-system learning","description":"Use learning methods to reconstruct, characterize, predict, simulate, or discover properties of quantum states, processes, Hamiltonians, circuits, many-body systems, dynamics, and quantum hardware.","inclusion":"Include tomography, shadow methods, Hamiltonian/channel/unitary learning, neural quantum-state representations, learning quantum symmetries, predicting random-circuit outputs, many-body property prediction, open-system dynamics learning, and quantum-system simulation where the quantum system is the object of study.","exclusion":"If quantum hardware or quantum circuits are used mainly as computational substrates for classical-data ML, supervised prediction, generative modeling, or scientific PDE solving, assign to those objectives.","nearest_boundary":"Foundations and structural priors","representative_topics":["adaptive online shadow tomography with dynamic regret bounds for evolving quantum states","adaptive versus nonadaptive quantum state tomography with incoherent measurements and memory constraints","neural quantum states for feynman-kitaev ground state encoding of quantum dynamics"]},{"direction_index":5,"label":"D5","name":"Sequential, language, and decision learning","description":"Develop quantum models for temporally ordered data, language structure, recurrent or reservoir computation, attention, reinforcement learning, bandits, control, and sequential decision-making.","inclusion":"Include quantum RL, policy iteration, actor-critic or Q-learning, bandits, multi-agent RL, quantum recurrent networks, QLSTMs, reservoir computing, temporal prediction, quantum attention for sequences, and quantum NLP or compositional linguistic models.","exclusion":"If the primary objective is ordinary supervised prediction on non-sequential data, distributed/federated protocol design, or general VQC optimization, assign to the corresponding program.","nearest_boundary":"Classical-data prediction models","representative_topics":["quantum circuit implementations of word classification, bigram modeling, and ambiguity resolution","quantum policy iteration using amplitude estimation and grover search for sample-efficient reinforcement learning","continuous-variable quantum recurrent networks for efficient time series learning"]},{"direction_index":6,"label":"D6","name":"Dependable and deployed QML systems","description":"Build, integrate, distribute, secure, verify, explain, benchmark, or operationalize QML systems through platforms, workflows, federated protocols, privacy mechanisms, robustness methods, fairness checks, uncertainty guarantees, and interpretability tools.","inclusion":"Include work whose primary objective is system-level QML infrastructure, reusable frameworks, AutoML/workflow automation, cross-platform deployment, distributed or federated learning, privacy-preserving learning, adversarial robustness, formal verification, fairness, explainability, or reliability guarantees.","exclusion":"If a paper primarily proposes a single predictive architecture, trainability method, kernel, generator, or physical-system learner and only uses infrastructure or robustness as evaluation context, assign to that model objective.","nearest_boundary":"Trainable variational circuits","representative_topics":["software framework for quantum kernel machine learning with integrated pipeline and visualization","slimmable quantum federated learning with dynamic angle and pole parameter exploitation under communication constraints","fairness verification of noisy quantum machine learning models via tensor network bias kernel detection"]},{"direction_index":7,"label":"D7","name":"Quantum algorithmic ML primitives","description":"Design quantum, quantum-inspired, annealing, Ising/QUBO, linear-algebra, clustering, and scientific-computing primitives that reformulate or accelerate machine-learning and computational-science subproblems.","inclusion":"Include work whose primary objective is a quantum algorithmic primitive or optimization formulation for clustering, feature selection, matrix factorization, decision trees, Bayesian inference, neural-network training, splines, PDEs, physics-informed networks, scientific simulation, QUBO/Ising encodings, adiabatic methods, or annealer-based ML optimization.","exclusion":"If the main objective is a gate-based VQC training method, kernel predictor, supervised architecture, generative model, or physical quantum-system learner, assign to that more direct program.","nearest_boundary":"Classical-data prediction models","representative_topics":["coreset selection strategies for noise-robust quantum clustering on near-term devices","variational quantum clustering via non-orthogonal states and tensor network simulation","quantum algorithms for bayesian inference including mcmc, gaussian processes, and deep learning"]}],"split":"dev"}
{"episode_id":"F028_T2024-07","field_index":28,"field":"quantum machine learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Trainable variational circuits","description":"Improve the optimization behavior, trainability, architecture selection, initialization, compression, and parameter efficiency of parameterized quantum circuits used as learning models.","inclusion":"Include work whose primary objective is diagnosing or mitigating barren plateaus, laziness, vanishing gradients, optimizer pathologies, shot-efficient gradients, pruning, compression, overparameterization, hyperparameter effects, or architecture search for VQCs/QNNs.","exclusion":"If the main objective is a supervised predictor, kernel method, generator, distributed protocol, trust guarantee, physical-system learner, or structural prior, assign to that objective even if trainability is evaluated.","nearest_boundary":"Foundations and structural priors","representative_topics":["theoretical analysis of laziness and barren plateau in overparameterized variational quantum circuits","hyperparameter importance analysis of quantum neural networks using functional anova on small-scale datasets","quantum circuit architecture search using gate-level distance metrics and bayesian optimization"]},{"direction_index":1,"label":"D1","name":"Foundations and structural priors","description":"Establish formal capacity, learnability, generalization, advantage, limitation, and inductive-bias principles for quantum learning models, including symmetry-, geometry-, representation-, and tensor-network-based architectural priors.","inclusion":"Include work whose primary contribution is a theorem, bound, analytic framework, quantum-versus-classical comparison, or principled structural prior explaining what quantum learning models can represent, generalize, or learn.","exclusion":"If the main objective is empirical optimization of trainability, a task-specific supervised predictor, kernel construction, or learning an unknown physical quantum object, assign to that nearer program.","nearest_boundary":"Trainable variational circuits","representative_topics":["theoretical error bounds for variational quantum circuits with tensor-train dimensionality reduction for functional regression","theoretical bounds on data encoding concentration in parameterized quantum circuits for quantum classifiers","classical surrogate analysis of variational quantum learning models and their inductive biases"]},{"direction_index":2,"label":"D2","name":"Classical-data prediction models","description":"Develop quantum or hybrid models for supervised prediction on classical data, including neural classifiers/regressors, quantum kernels, feature maps, graph-data models, financial predictors, and other structured classical-data prediction pipelines.","inclusion":"Include work whose primary objective is classification or regression performance, feature extraction, kernel construction, domain prediction, graph learning, or financial/economic prediction using quantum, hybrid, or quantum-inspired models.","exclusion":"If the central object is a generative model, reinforcement-learning agent, language/sequence representation, trainability method, general theorem, or annealing/linear-algebra primitive, assign to that program instead.","nearest_boundary":"Generative distribution learning","representative_topics":["variational quantum approximate svm with inference transfer for scalable hyper-linear classification","scalability analysis of quantum convolutional neural networks via measurement step reduction","quantum feature extraction for thz multi-layer depth and content recognition"]},{"direction_index":3,"label":"D3","name":"Generative distribution learning","description":"Create, train, analyze, and evaluate quantum or hybrid generative models for sampling, distribution learning, data generation, molecular generation, image generation, and Born-machine-style modeling.","inclusion":"Include work whose primary output is a generative model or generative-learning method, including quantum GANs, Born machines, VAEs, energy-based models, tensor-network generative models, and training or evaluation methods for generators.","exclusion":"If the generator is mainly a tool for tomography, physical-state characterization, supervised prediction, or generic VQC trainability analysis, assign to that primary objective.","nearest_boundary":"Quantum physical-system learning","representative_topics":["survey of quantum generative learning models for near-term quantum advantage","tensor network generative models with entanglement entropy variations for unsupervised informative feature recognition","sample-efficient quantum born machine training via coding rate reduction to prevent mode collapse"]},{"direction_index":4,"label":"D4","name":"Quantum physical-system learning","description":"Use learning methods to reconstruct, characterize, predict, simulate, or discover properties of quantum states, processes, Hamiltonians, circuits, many-body systems, dynamics, and quantum hardware.","inclusion":"Include tomography, shadow methods, Hamiltonian/channel/unitary learning, neural quantum-state representations, learning quantum symmetries, predicting random-circuit outputs, many-body property prediction, open-system dynamics learning, and quantum-system simulation where the quantum system is the object of study.","exclusion":"If quantum hardware or quantum circuits are used mainly as computational substrates for classical-data ML, supervised prediction, generative modeling, or scientific PDE solving, assign to those objectives.","nearest_boundary":"Foundations and structural priors","representative_topics":["adaptive online shadow tomography with dynamic regret bounds for evolving quantum states","adaptive versus nonadaptive quantum state tomography with incoherent measurements and memory constraints","neural quantum states for feynman-kitaev ground state encoding of quantum dynamics"]},{"direction_index":5,"label":"D5","name":"Sequential, language, and decision learning","description":"Develop quantum models for temporally ordered data, language structure, recurrent or reservoir computation, attention, reinforcement learning, bandits, control, and sequential decision-making.","inclusion":"Include quantum RL, policy iteration, actor-critic or Q-learning, bandits, multi-agent RL, quantum recurrent networks, QLSTMs, reservoir computing, temporal prediction, quantum attention for sequences, and quantum NLP or compositional linguistic models.","exclusion":"If the primary objective is ordinary supervised prediction on non-sequential data, distributed/federated protocol design, or general VQC optimization, assign to the corresponding program.","nearest_boundary":"Classical-data prediction models","representative_topics":["quantum circuit implementations of word classification, bigram modeling, and ambiguity resolution","quantum policy iteration using amplitude estimation and grover search for sample-efficient reinforcement learning","continuous-variable quantum recurrent networks for efficient time series learning"]},{"direction_index":6,"label":"D6","name":"Dependable and deployed QML systems","description":"Build, integrate, distribute, secure, verify, explain, benchmark, or operationalize QML systems through platforms, workflows, federated protocols, privacy mechanisms, robustness methods, fairness checks, uncertainty guarantees, and interpretability tools.","inclusion":"Include work whose primary objective is system-level QML infrastructure, reusable frameworks, AutoML/workflow automation, cross-platform deployment, distributed or federated learning, privacy-preserving learning, adversarial robustness, formal verification, fairness, explainability, or reliability guarantees.","exclusion":"If a paper primarily proposes a single predictive architecture, trainability method, kernel, generator, or physical-system learner and only uses infrastructure or robustness as evaluation context, assign to that model objective.","nearest_boundary":"Trainable variational circuits","representative_topics":["software framework for quantum kernel machine learning with integrated pipeline and visualization","slimmable quantum federated learning with dynamic angle and pole parameter exploitation under communication constraints","fairness verification of noisy quantum machine learning models via tensor network bias kernel detection"]},{"direction_index":7,"label":"D7","name":"Quantum algorithmic ML primitives","description":"Design quantum, quantum-inspired, annealing, Ising/QUBO, linear-algebra, clustering, and scientific-computing primitives that reformulate or accelerate machine-learning and computational-science subproblems.","inclusion":"Include work whose primary objective is a quantum algorithmic primitive or optimization formulation for clustering, feature selection, matrix factorization, decision trees, Bayesian inference, neural-network training, splines, PDEs, physics-informed networks, scientific simulation, QUBO/Ising encodings, adiabatic methods, or annealer-based ML optimization.","exclusion":"If the main objective is a gate-based VQC training method, kernel predictor, supervised architecture, generative model, or physical quantum-system learner, assign to that more direct program.","nearest_boundary":"Classical-data prediction models","representative_topics":["coreset selection strategies for noise-robust quantum clustering on near-term devices","variational quantum clustering via non-orthogonal states and tensor network simulation","quantum algorithms for bayesian inference including mcmc, gaussian processes, and deep learning"]}],"split":"dev"}
{"episode_id":"F028_T2025-01","field_index":28,"field":"quantum machine learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Trainable variational circuits","description":"Improve the optimization behavior, trainability, architecture selection, initialization, compression, and parameter efficiency of parameterized quantum circuits used as learning models.","inclusion":"Include work whose primary objective is diagnosing or mitigating barren plateaus, laziness, vanishing gradients, optimizer pathologies, shot-efficient gradients, pruning, compression, overparameterization, hyperparameter effects, or architecture search for VQCs/QNNs.","exclusion":"If the main objective is a supervised predictor, kernel method, generator, distributed protocol, trust guarantee, physical-system learner, or structural prior, assign to that objective even if trainability is evaluated.","nearest_boundary":"Foundations and structural priors","representative_topics":["theoretical analysis of laziness and barren plateau in overparameterized variational quantum circuits","hyperparameter importance analysis of quantum neural networks using functional anova on small-scale datasets","quantum circuit architecture search using gate-level distance metrics and bayesian optimization"]},{"direction_index":1,"label":"D1","name":"Foundations and structural priors","description":"Establish formal capacity, learnability, generalization, advantage, limitation, and inductive-bias principles for quantum learning models, including symmetry-, geometry-, representation-, and tensor-network-based architectural priors.","inclusion":"Include work whose primary contribution is a theorem, bound, analytic framework, quantum-versus-classical comparison, or principled structural prior explaining what quantum learning models can represent, generalize, or learn.","exclusion":"If the main objective is empirical optimization of trainability, a task-specific supervised predictor, kernel construction, or learning an unknown physical quantum object, assign to that nearer program.","nearest_boundary":"Trainable variational circuits","representative_topics":["theoretical error bounds for variational quantum circuits with tensor-train dimensionality reduction for functional regression","theoretical bounds on data encoding concentration in parameterized quantum circuits for quantum classifiers","classical surrogate analysis of variational quantum learning models and their inductive biases"]},{"direction_index":2,"label":"D2","name":"Classical-data prediction models","description":"Develop quantum or hybrid models for supervised prediction on classical data, including neural classifiers/regressors, quantum kernels, feature maps, graph-data models, financial predictors, and other structured classical-data prediction pipelines.","inclusion":"Include work whose primary objective is classification or regression performance, feature extraction, kernel construction, domain prediction, graph learning, or financial/economic prediction using quantum, hybrid, or quantum-inspired models.","exclusion":"If the central object is a generative model, reinforcement-learning agent, language/sequence representation, trainability method, general theorem, or annealing/linear-algebra primitive, assign to that program instead.","nearest_boundary":"Generative distribution learning","representative_topics":["variational quantum approximate svm with inference transfer for scalable hyper-linear classification","scalability analysis of quantum convolutional neural networks via measurement step reduction","quantum feature extraction for thz multi-layer depth and content recognition"]},{"direction_index":3,"label":"D3","name":"Generative distribution learning","description":"Create, train, analyze, and evaluate quantum or hybrid generative models for sampling, distribution learning, data generation, molecular generation, image generation, and Born-machine-style modeling.","inclusion":"Include work whose primary output is a generative model or generative-learning method, including quantum GANs, Born machines, VAEs, energy-based models, tensor-network generative models, and training or evaluation methods for generators.","exclusion":"If the generator is mainly a tool for tomography, physical-state characterization, supervised prediction, or generic VQC trainability analysis, assign to that primary objective.","nearest_boundary":"Quantum physical-system learning","representative_topics":["survey of quantum generative learning models for near-term quantum advantage","tensor network generative models with entanglement entropy variations for unsupervised informative feature recognition","sample-efficient quantum born machine training via coding rate reduction to prevent mode collapse"]},{"direction_index":4,"label":"D4","name":"Quantum physical-system learning","description":"Use learning methods to reconstruct, characterize, predict, simulate, or discover properties of quantum states, processes, Hamiltonians, circuits, many-body systems, dynamics, and quantum hardware.","inclusion":"Include tomography, shadow methods, Hamiltonian/channel/unitary learning, neural quantum-state representations, learning quantum symmetries, predicting random-circuit outputs, many-body property prediction, open-system dynamics learning, and quantum-system simulation where the quantum system is the object of study.","exclusion":"If quantum hardware or quantum circuits are used mainly as computational substrates for classical-data ML, supervised prediction, generative modeling, or scientific PDE solving, assign to those objectives.","nearest_boundary":"Foundations and structural priors","representative_topics":["adaptive online shadow tomography with dynamic regret bounds for evolving quantum states","adaptive versus nonadaptive quantum state tomography with incoherent measurements and memory constraints","neural quantum states for feynman-kitaev ground state encoding of quantum dynamics"]},{"direction_index":5,"label":"D5","name":"Sequential, language, and decision learning","description":"Develop quantum models for temporally ordered data, language structure, recurrent or reservoir computation, attention, reinforcement learning, bandits, control, and sequential decision-making.","inclusion":"Include quantum RL, policy iteration, actor-critic or Q-learning, bandits, multi-agent RL, quantum recurrent networks, QLSTMs, reservoir computing, temporal prediction, quantum attention for sequences, and quantum NLP or compositional linguistic models.","exclusion":"If the primary objective is ordinary supervised prediction on non-sequential data, distributed/federated protocol design, or general VQC optimization, assign to the corresponding program.","nearest_boundary":"Classical-data prediction models","representative_topics":["quantum circuit implementations of word classification, bigram modeling, and ambiguity resolution","quantum policy iteration using amplitude estimation and grover search for sample-efficient reinforcement learning","continuous-variable quantum recurrent networks for efficient time series learning"]},{"direction_index":6,"label":"D6","name":"Dependable and deployed QML systems","description":"Build, integrate, distribute, secure, verify, explain, benchmark, or operationalize QML systems through platforms, workflows, federated protocols, privacy mechanisms, robustness methods, fairness checks, uncertainty guarantees, and interpretability tools.","inclusion":"Include work whose primary objective is system-level QML infrastructure, reusable frameworks, AutoML/workflow automation, cross-platform deployment, distributed or federated learning, privacy-preserving learning, adversarial robustness, formal verification, fairness, explainability, or reliability guarantees.","exclusion":"If a paper primarily proposes a single predictive architecture, trainability method, kernel, generator, or physical-system learner and only uses infrastructure or robustness as evaluation context, assign to that model objective.","nearest_boundary":"Trainable variational circuits","representative_topics":["software framework for quantum kernel machine learning with integrated pipeline and visualization","slimmable quantum federated learning with dynamic angle and pole parameter exploitation under communication constraints","fairness verification of noisy quantum machine learning models via tensor network bias kernel detection"]},{"direction_index":7,"label":"D7","name":"Quantum algorithmic ML primitives","description":"Design quantum, quantum-inspired, annealing, Ising/QUBO, linear-algebra, clustering, and scientific-computing primitives that reformulate or accelerate machine-learning and computational-science subproblems.","inclusion":"Include work whose primary objective is a quantum algorithmic primitive or optimization formulation for clustering, feature selection, matrix factorization, decision trees, Bayesian inference, neural-network training, splines, PDEs, physics-informed networks, scientific simulation, QUBO/Ising encodings, adiabatic methods, or annealer-based ML optimization.","exclusion":"If the main objective is a gate-based VQC training method, kernel predictor, supervised architecture, generative model, or physical quantum-system learner, assign to that more direct program.","nearest_boundary":"Classical-data prediction models","representative_topics":["coreset selection strategies for noise-robust quantum clustering on near-term devices","variational quantum clustering via non-orthogonal states and tensor network simulation","quantum algorithms for bayesian inference including mcmc, gaussian processes, and deep learning"]}],"split":"dev"}
{"episode_id":"F028_T2025-07","field_index":28,"field":"quantum machine learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Trainable variational circuits","description":"Improve the optimization behavior, trainability, architecture selection, initialization, compression, and parameter efficiency of parameterized quantum circuits used as learning models.","inclusion":"Include work whose primary objective is diagnosing or mitigating barren plateaus, laziness, vanishing gradients, optimizer pathologies, shot-efficient gradients, pruning, compression, overparameterization, hyperparameter effects, or architecture search for VQCs/QNNs.","exclusion":"If the main objective is a supervised predictor, kernel method, generator, distributed protocol, trust guarantee, physical-system learner, or structural prior, assign to that objective even if trainability is evaluated.","nearest_boundary":"Foundations and structural priors","representative_topics":["theoretical analysis of laziness and barren plateau in overparameterized variational quantum circuits","hyperparameter importance analysis of quantum neural networks using functional anova on small-scale datasets","quantum circuit architecture search using gate-level distance metrics and bayesian optimization"]},{"direction_index":1,"label":"D1","name":"Foundations and structural priors","description":"Establish formal capacity, learnability, generalization, advantage, limitation, and inductive-bias principles for quantum learning models, including symmetry-, geometry-, representation-, and tensor-network-based architectural priors.","inclusion":"Include work whose primary contribution is a theorem, bound, analytic framework, quantum-versus-classical comparison, or principled structural prior explaining what quantum learning models can represent, generalize, or learn.","exclusion":"If the main objective is empirical optimization of trainability, a task-specific supervised predictor, kernel construction, or learning an unknown physical quantum object, assign to that nearer program.","nearest_boundary":"Trainable variational circuits","representative_topics":["theoretical error bounds for variational quantum circuits with tensor-train dimensionality reduction for functional regression","theoretical bounds on data encoding concentration in parameterized quantum circuits for quantum classifiers","classical surrogate analysis of variational quantum learning models and their inductive biases"]},{"direction_index":2,"label":"D2","name":"Classical-data prediction models","description":"Develop quantum or hybrid models for supervised prediction on classical data, including neural classifiers/regressors, quantum kernels, feature maps, graph-data models, financial predictors, and other structured classical-data prediction pipelines.","inclusion":"Include work whose primary objective is classification or regression performance, feature extraction, kernel construction, domain prediction, graph learning, or financial/economic prediction using quantum, hybrid, or quantum-inspired models.","exclusion":"If the central object is a generative model, reinforcement-learning agent, language/sequence representation, trainability method, general theorem, or annealing/linear-algebra primitive, assign to that program instead.","nearest_boundary":"Generative distribution learning","representative_topics":["variational quantum approximate svm with inference transfer for scalable hyper-linear classification","scalability analysis of quantum convolutional neural networks via measurement step reduction","quantum feature extraction for thz multi-layer depth and content recognition"]},{"direction_index":3,"label":"D3","name":"Generative distribution learning","description":"Create, train, analyze, and evaluate quantum or hybrid generative models for sampling, distribution learning, data generation, molecular generation, image generation, and Born-machine-style modeling.","inclusion":"Include work whose primary output is a generative model or generative-learning method, including quantum GANs, Born machines, VAEs, energy-based models, tensor-network generative models, and training or evaluation methods for generators.","exclusion":"If the generator is mainly a tool for tomography, physical-state characterization, supervised prediction, or generic VQC trainability analysis, assign to that primary objective.","nearest_boundary":"Quantum physical-system learning","representative_topics":["survey of quantum generative learning models for near-term quantum advantage","tensor network generative models with entanglement entropy variations for unsupervised informative feature recognition","sample-efficient quantum born machine training via coding rate reduction to prevent mode collapse"]},{"direction_index":4,"label":"D4","name":"Quantum physical-system learning","description":"Use learning methods to reconstruct, characterize, predict, simulate, or discover properties of quantum states, processes, Hamiltonians, circuits, many-body systems, dynamics, and quantum hardware.","inclusion":"Include tomography, shadow methods, Hamiltonian/channel/unitary learning, neural quantum-state representations, learning quantum symmetries, predicting random-circuit outputs, many-body property prediction, open-system dynamics learning, and quantum-system simulation where the quantum system is the object of study.","exclusion":"If quantum hardware or quantum circuits are used mainly as computational substrates for classical-data ML, supervised prediction, generative modeling, or scientific PDE solving, assign to those objectives.","nearest_boundary":"Foundations and structural priors","representative_topics":["adaptive online shadow tomography with dynamic regret bounds for evolving quantum states","adaptive versus nonadaptive quantum state tomography with incoherent measurements and memory constraints","neural quantum states for feynman-kitaev ground state encoding of quantum dynamics"]},{"direction_index":5,"label":"D5","name":"Sequential, language, and decision learning","description":"Develop quantum models for temporally ordered data, language structure, recurrent or reservoir computation, attention, reinforcement learning, bandits, control, and sequential decision-making.","inclusion":"Include quantum RL, policy iteration, actor-critic or Q-learning, bandits, multi-agent RL, quantum recurrent networks, QLSTMs, reservoir computing, temporal prediction, quantum attention for sequences, and quantum NLP or compositional linguistic models.","exclusion":"If the primary objective is ordinary supervised prediction on non-sequential data, distributed/federated protocol design, or general VQC optimization, assign to the corresponding program.","nearest_boundary":"Classical-data prediction models","representative_topics":["quantum circuit implementations of word classification, bigram modeling, and ambiguity resolution","quantum policy iteration using amplitude estimation and grover search for sample-efficient reinforcement learning","continuous-variable quantum recurrent networks for efficient time series learning"]},{"direction_index":6,"label":"D6","name":"Dependable and deployed QML systems","description":"Build, integrate, distribute, secure, verify, explain, benchmark, or operationalize QML systems through platforms, workflows, federated protocols, privacy mechanisms, robustness methods, fairness checks, uncertainty guarantees, and interpretability tools.","inclusion":"Include work whose primary objective is system-level QML infrastructure, reusable frameworks, AutoML/workflow automation, cross-platform deployment, distributed or federated learning, privacy-preserving learning, adversarial robustness, formal verification, fairness, explainability, or reliability guarantees.","exclusion":"If a paper primarily proposes a single predictive architecture, trainability method, kernel, generator, or physical-system learner and only uses infrastructure or robustness as evaluation context, assign to that model objective.","nearest_boundary":"Trainable variational circuits","representative_topics":["software framework for quantum kernel machine learning with integrated pipeline and visualization","slimmable quantum federated learning with dynamic angle and pole parameter exploitation under communication constraints","fairness verification of noisy quantum machine learning models via tensor network bias kernel detection"]},{"direction_index":7,"label":"D7","name":"Quantum algorithmic ML primitives","description":"Design quantum, quantum-inspired, annealing, Ising/QUBO, linear-algebra, clustering, and scientific-computing primitives that reformulate or accelerate machine-learning and computational-science subproblems.","inclusion":"Include work whose primary objective is a quantum algorithmic primitive or optimization formulation for clustering, feature selection, matrix factorization, decision trees, Bayesian inference, neural-network training, splines, PDEs, physics-informed networks, scientific simulation, QUBO/Ising encodings, adiabatic methods, or annealer-based ML optimization.","exclusion":"If the main objective is a gate-based VQC training method, kernel predictor, supervised architecture, generative model, or physical quantum-system learner, assign to that more direct program.","nearest_boundary":"Classical-data prediction models","representative_topics":["coreset selection strategies for noise-robust quantum clustering on near-term devices","variational quantum clustering via non-orthogonal states and tensor network simulation","quantum algorithms for bayesian inference including mcmc, gaussian processes, and deep learning"]}],"split":"dev"}
{"episode_id":"F028_T2026-01","field_index":28,"field":"quantum machine learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Trainable variational circuits","description":"Improve the optimization behavior, trainability, architecture selection, initialization, compression, and parameter efficiency of parameterized quantum circuits used as learning models.","inclusion":"Include work whose primary objective is diagnosing or mitigating barren plateaus, laziness, vanishing gradients, optimizer pathologies, shot-efficient gradients, pruning, compression, overparameterization, hyperparameter effects, or architecture search for VQCs/QNNs.","exclusion":"If the main objective is a supervised predictor, kernel method, generator, distributed protocol, trust guarantee, physical-system learner, or structural prior, assign to that objective even if trainability is evaluated.","nearest_boundary":"Foundations and structural priors","representative_topics":["theoretical analysis of laziness and barren plateau in overparameterized variational quantum circuits","hyperparameter importance analysis of quantum neural networks using functional anova on small-scale datasets","quantum circuit architecture search using gate-level distance metrics and bayesian optimization"]},{"direction_index":1,"label":"D1","name":"Foundations and structural priors","description":"Establish formal capacity, learnability, generalization, advantage, limitation, and inductive-bias principles for quantum learning models, including symmetry-, geometry-, representation-, and tensor-network-based architectural priors.","inclusion":"Include work whose primary contribution is a theorem, bound, analytic framework, quantum-versus-classical comparison, or principled structural prior explaining what quantum learning models can represent, generalize, or learn.","exclusion":"If the main objective is empirical optimization of trainability, a task-specific supervised predictor, kernel construction, or learning an unknown physical quantum object, assign to that nearer program.","nearest_boundary":"Trainable variational circuits","representative_topics":["theoretical error bounds for variational quantum circuits with tensor-train dimensionality reduction for functional regression","theoretical bounds on data encoding concentration in parameterized quantum circuits for quantum classifiers","classical surrogate analysis of variational quantum learning models and their inductive biases"]},{"direction_index":2,"label":"D2","name":"Classical-data prediction models","description":"Develop quantum or hybrid models for supervised prediction on classical data, including neural classifiers/regressors, quantum kernels, feature maps, graph-data models, financial predictors, and other structured classical-data prediction pipelines.","inclusion":"Include work whose primary objective is classification or regression performance, feature extraction, kernel construction, domain prediction, graph learning, or financial/economic prediction using quantum, hybrid, or quantum-inspired models.","exclusion":"If the central object is a generative model, reinforcement-learning agent, language/sequence representation, trainability method, general theorem, or annealing/linear-algebra primitive, assign to that program instead.","nearest_boundary":"Generative distribution learning","representative_topics":["variational quantum approximate svm with inference transfer for scalable hyper-linear classification","scalability analysis of quantum convolutional neural networks via measurement step reduction","quantum feature extraction for thz multi-layer depth and content recognition"]},{"direction_index":3,"label":"D3","name":"Generative distribution learning","description":"Create, train, analyze, and evaluate quantum or hybrid generative models for sampling, distribution learning, data generation, molecular generation, image generation, and Born-machine-style modeling.","inclusion":"Include work whose primary output is a generative model or generative-learning method, including quantum GANs, Born machines, VAEs, energy-based models, tensor-network generative models, and training or evaluation methods for generators.","exclusion":"If the generator is mainly a tool for tomography, physical-state characterization, supervised prediction, or generic VQC trainability analysis, assign to that primary objective.","nearest_boundary":"Quantum physical-system learning","representative_topics":["survey of quantum generative learning models for near-term quantum advantage","tensor network generative models with entanglement entropy variations for unsupervised informative feature recognition","sample-efficient quantum born machine training via coding rate reduction to prevent mode collapse"]},{"direction_index":4,"label":"D4","name":"Quantum physical-system learning","description":"Use learning methods to reconstruct, characterize, predict, simulate, or discover properties of quantum states, processes, Hamiltonians, circuits, many-body systems, dynamics, and quantum hardware.","inclusion":"Include tomography, shadow methods, Hamiltonian/channel/unitary learning, neural quantum-state representations, learning quantum symmetries, predicting random-circuit outputs, many-body property prediction, open-system dynamics learning, and quantum-system simulation where the quantum system is the object of study.","exclusion":"If quantum hardware or quantum circuits are used mainly as computational substrates for classical-data ML, supervised prediction, generative modeling, or scientific PDE solving, assign to those objectives.","nearest_boundary":"Foundations and structural priors","representative_topics":["adaptive online shadow tomography with dynamic regret bounds for evolving quantum states","adaptive versus nonadaptive quantum state tomography with incoherent measurements and memory constraints","neural quantum states for feynman-kitaev ground state encoding of quantum dynamics"]},{"direction_index":5,"label":"D5","name":"Sequential, language, and decision learning","description":"Develop quantum models for temporally ordered data, language structure, recurrent or reservoir computation, attention, reinforcement learning, bandits, control, and sequential decision-making.","inclusion":"Include quantum RL, policy iteration, actor-critic or Q-learning, bandits, multi-agent RL, quantum recurrent networks, QLSTMs, reservoir computing, temporal prediction, quantum attention for sequences, and quantum NLP or compositional linguistic models.","exclusion":"If the primary objective is ordinary supervised prediction on non-sequential data, distributed/federated protocol design, or general VQC optimization, assign to the corresponding program.","nearest_boundary":"Classical-data prediction models","representative_topics":["quantum circuit implementations of word classification, bigram modeling, and ambiguity resolution","quantum policy iteration using amplitude estimation and grover search for sample-efficient reinforcement learning","continuous-variable quantum recurrent networks for efficient time series learning"]},{"direction_index":6,"label":"D6","name":"Dependable and deployed QML systems","description":"Build, integrate, distribute, secure, verify, explain, benchmark, or operationalize QML systems through platforms, workflows, federated protocols, privacy mechanisms, robustness methods, fairness checks, uncertainty guarantees, and interpretability tools.","inclusion":"Include work whose primary objective is system-level QML infrastructure, reusable frameworks, AutoML/workflow automation, cross-platform deployment, distributed or federated learning, privacy-preserving learning, adversarial robustness, formal verification, fairness, explainability, or reliability guarantees.","exclusion":"If a paper primarily proposes a single predictive architecture, trainability method, kernel, generator, or physical-system learner and only uses infrastructure or robustness as evaluation context, assign to that model objective.","nearest_boundary":"Trainable variational circuits","representative_topics":["software framework for quantum kernel machine learning with integrated pipeline and visualization","slimmable quantum federated learning with dynamic angle and pole parameter exploitation under communication constraints","fairness verification of noisy quantum machine learning models via tensor network bias kernel detection"]},{"direction_index":7,"label":"D7","name":"Quantum algorithmic ML primitives","description":"Design quantum, quantum-inspired, annealing, Ising/QUBO, linear-algebra, clustering, and scientific-computing primitives that reformulate or accelerate machine-learning and computational-science subproblems.","inclusion":"Include work whose primary objective is a quantum algorithmic primitive or optimization formulation for clustering, feature selection, matrix factorization, decision trees, Bayesian inference, neural-network training, splines, PDEs, physics-informed networks, scientific simulation, QUBO/Ising encodings, adiabatic methods, or annealer-based ML optimization.","exclusion":"If the main objective is a gate-based VQC training method, kernel predictor, supervised architecture, generative model, or physical quantum-system learner, assign to that more direct program.","nearest_boundary":"Classical-data prediction models","representative_topics":["coreset selection strategies for noise-robust quantum clustering on near-term devices","variational quantum clustering via non-orthogonal states and tensor network simulation","quantum algorithms for bayesian inference including mcmc, gaussian processes, and deep learning"]}],"split":"dev"}
{"episode_id":"F030_T2024-01","field_index":30,"field":"neural network theory","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Representation and function-space capacity","description":"Characterize what neural networks can represent or approximate, and how architecture, activation choice, geometry, operator targets, and function-space structure determine expressive capacity.","inclusion":"Include atoms whose primary objective is approximation theory, expressivity, exact representation, activation-function effects on representable functions, ReLU/polyhedral function geometry, operator/PDE/manifold approximation, or functional-analytic foundations of network function classes.","exclusion":"If the primary objective is statistical risk, optimizer-selected solutions, training convergence, or graph/Transformer/symmetry-specific computational distinction, assign to the corresponding generalization, training, or structured-architecture direction instead.","nearest_boundary":"D7","representative_topics":["l^p norm approximation lower bounds for feed-forward neural networks via fat-shattering dimension","exact neural network representation of convex piecewise quadratic functions via max-out topology","relu neural network approximation of parametric hyperbolic conservation laws with generalization error bounds"]},{"direction_index":1,"label":"D1","name":"Generalization, robustness, and risk regimes","description":"Explain out-of-sample behavior of neural networks, including sample complexity, excess risk, consistency, robustness under perturbations, and overparameterized interpolation or overfitting regimes.","inclusion":"Include atoms whose primary objective is statistical generalization bounds, sample complexity, PAC-Bayes or complexity measures, consistency, label complexity, adversarial or Lipschitz robustness, benign/harmful overfitting, double descent as a risk phenomenon, memorization, or broad inductive-bias explanations of generalization.","exclusion":"If the central object is the optimizer's selected solution, loss-landscape dynamics, kernel-limit asymptotics, or Bayesian/statistical-physics ensemble formalism, assign to those directions instead.","nearest_boundary":"D4","representative_topics":["decision boundary variability as a generalization measure in deep learning theory","hessian-based pac-bayesian generalization bounds for robust fine-tuning under label noise","consistency theory for regularized neural networks via sieve and minimal network methods"]},{"direction_index":2,"label":"D2","name":"Kernel and infinite-width limits","description":"Analyze neural networks through NTK, NNGP, random-feature, large-width, infinite-depth/width, lazy-training, and finite-width-correction limits.","inclusion":"Include atoms whose primary objective is deriving or using kernel-limit objects, infinite-width Gaussian-process correspondences, NTK spectra, lazy-regime training/generalization, random-feature asymptotics, or quantitative deviations from such limits.","exclusion":"If the work primarily studies feature learning beyond fixed kernels, ordinary optimization convergence, or functional-analysis foundations not centered on NTK/NNGP limits, assign to the relevant neighboring direction.","nearest_boundary":"D4","representative_topics":["generalization bounds for overparameterized 3-layer neural tangent kernel models with hidden-layer interactions","neural covariance sde for shaped infinite-depth-and-width feedforward network initialization","spectral bias bounds for deep networks via fisher information low effective rank in kernel regime"]},{"direction_index":3,"label":"D3","name":"Terminal representation geometry","description":"Study the limiting geometry of learned class representations and classifiers at late training, especially neural collapse and related terminal feature-collapse phenomena.","inclusion":"Include atoms whose primary objective is simplex/equiangular class geometry, class-mean collapse, classifier-feature duality, neural collapse under imbalance, label noise, normalization, robustness, transfer, or perturbation of terminal class geometry.","exclusion":"If the objective is general representation learning over training time, feature emergence, robustness certification, or broad generalization theory rather than terminal class geometry, assign to the neighboring representation or generalization direction.","nearest_boundary":"D4","representative_topics":["neural collapse as a terminal training state for generalization and transfer learning","theoretical analysis of neural collapse under label noise via memorization-dilation","invariant neural collapse geometry characterization under class imbalance via simplex-encoded label interpolation"]},{"direction_index":4,"label":"D4","name":"Training dynamics and optimizer mechanisms","description":"Characterize how neural-network training proceeds and what mechanisms govern convergence, representation formation, implicit regularization, loss geometry, sharpness, and phase behavior during optimization.","inclusion":"Include atoms whose primary objective is gradient-flow or SGD/GD dynamics, global/local convergence, loss landscapes, saddle or stationary-point structure, flatness/sharpness, edge-of-stability, implicit bias toward selected solutions, dropout or weight-decay regularization effects, feature-learning dynamics, lazy-to-rich transitions, spectral/simple-to-complex learning order, or mean-field training convergence.","exclusion":"If the main result is a kernel-limit asymptotic, a statistical generalization bound, terminal neural-collapse geometry, or a robustness property independent of training mechanism, assign to those directions instead.","nearest_boundary":"D2","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","stochastic gradient descent and weight decay bias toward low-rank weight matrix learning","theoretical analysis of feature learning emergence in neural networks via input structure exploitation"]},{"direction_index":5,"label":"D5","name":"Efficiency, sparsity, and learnability limits","description":"Establish theoretical limits and guarantees for efficient neural-network learning or representation, including sparse/pruned parameterizations and computational hardness of learning network classes.","inclusion":"Include atoms whose primary objective is sparse-network expressivity, pruning or lottery-ticket guarantees, low-rank or structured parameter efficiency, compact parameterizations, SQ or NP-hardness lower bounds, finite-sample identification, stationarity-testing hardness, or efficient algorithms for restricted neural classes.","exclusion":"If sparsity, low rank, or compactness is only the optimizer-selected outcome, assign to training dynamics; if parameter count is merely part of approximation-rate tradeoffs, assign to representation capacity.","nearest_boundary":"D4","representative_topics":["strong lottery ticket hypothesis generalization to arbitrary group-equivariant neural networks via pruning","theoretical analysis of random mask pruning via overparameterization and sparse network containment","strong lottery ticket hypothesis with epsilon-perturbation to relax over-parameterization requirements"]},{"direction_index":6,"label":"D6","name":"Probabilistic and physics-based network theory","description":"Explain neural networks using Bayesian inference, variational inference, ensemble theory, statistical mechanics, field theory, energy landscapes, entropy, and physical analogies.","inclusion":"Include atoms whose primary objective is posterior behavior, Bayesian neural-network asymptotics or sampling, variational Bayesian theory, statistical-mechanical learning, entropy or volume explanations, replica/field-theoretic finite-width analysis, Boltzmann/Hopfield energy models, or physics analogies for signal propagation and emergent network phenomena.","exclusion":"If probability appears only as a statistical learning bound, SGD noise model for flat minima, or mean-field convergence proof, assign to generalization or training dynamics instead.","nearest_boundary":"D1","representative_topics":["bayesian neural network posterior convergence rates in besov space with adaptive smoothness","repriorisation for accelerated mcmc sampling in wide bayesian neural networks","statistical mechanical theory of weight-distribution-constrained learning via optimal transport and information geometry"]},{"direction_index":7,"label":"D7","name":"Structured and nonstandard architectures","description":"Analyze neural-network theories whose primary objective is tied to architectural structure beyond standard feedforward networks, including graph, attention, sequence, recurrent, equivariant, invariant, reservoir, energy-based, biological, and specialized dynamical architectures.","inclusion":"Include atoms whose primary objective is Transformer or attention expressivity, GNN/WL/logical graph expressivity, recurrent or sequence computational power, memory and temporal dynamics, group-equivariant or invariant layer construction and universality, graphon or geometric graph theory, reservoir/spiking/energy-based/neural-ODE or biologically motivated architectures, and other specialized architectural computation theories.","exclusion":"If the same architecture is used only as an example for ordinary approximation, generalization, kernel limits, or optimizer convergence, assign to those broader objective directions instead.","nearest_boundary":"D0","representative_topics":["linear region bounds for graph convolutional network expressivity across depth","theoretical analysis of rank collapse and gradient propagation in transformer self-attention layers","complexity-theoretic characterization of log-precision transformers via parallelism tradeoff"]}],"split":"dev"}
{"episode_id":"F030_T2024-07","field_index":30,"field":"neural network theory","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Representation and function-space capacity","description":"Characterize what neural networks can represent or approximate, and how architecture, activation choice, geometry, operator targets, and function-space structure determine expressive capacity.","inclusion":"Include atoms whose primary objective is approximation theory, expressivity, exact representation, activation-function effects on representable functions, ReLU/polyhedral function geometry, operator/PDE/manifold approximation, or functional-analytic foundations of network function classes.","exclusion":"If the primary objective is statistical risk, optimizer-selected solutions, training convergence, or graph/Transformer/symmetry-specific computational distinction, assign to the corresponding generalization, training, or structured-architecture direction instead.","nearest_boundary":"D7","representative_topics":["l^p norm approximation lower bounds for feed-forward neural networks via fat-shattering dimension","exact neural network representation of convex piecewise quadratic functions via max-out topology","relu neural network approximation of parametric hyperbolic conservation laws with generalization error bounds"]},{"direction_index":1,"label":"D1","name":"Generalization, robustness, and risk regimes","description":"Explain out-of-sample behavior of neural networks, including sample complexity, excess risk, consistency, robustness under perturbations, and overparameterized interpolation or overfitting regimes.","inclusion":"Include atoms whose primary objective is statistical generalization bounds, sample complexity, PAC-Bayes or complexity measures, consistency, label complexity, adversarial or Lipschitz robustness, benign/harmful overfitting, double descent as a risk phenomenon, memorization, or broad inductive-bias explanations of generalization.","exclusion":"If the central object is the optimizer's selected solution, loss-landscape dynamics, kernel-limit asymptotics, or Bayesian/statistical-physics ensemble formalism, assign to those directions instead.","nearest_boundary":"D4","representative_topics":["decision boundary variability as a generalization measure in deep learning theory","hessian-based pac-bayesian generalization bounds for robust fine-tuning under label noise","consistency theory for regularized neural networks via sieve and minimal network methods"]},{"direction_index":2,"label":"D2","name":"Kernel and infinite-width limits","description":"Analyze neural networks through NTK, NNGP, random-feature, large-width, infinite-depth/width, lazy-training, and finite-width-correction limits.","inclusion":"Include atoms whose primary objective is deriving or using kernel-limit objects, infinite-width Gaussian-process correspondences, NTK spectra, lazy-regime training/generalization, random-feature asymptotics, or quantitative deviations from such limits.","exclusion":"If the work primarily studies feature learning beyond fixed kernels, ordinary optimization convergence, or functional-analysis foundations not centered on NTK/NNGP limits, assign to the relevant neighboring direction.","nearest_boundary":"D4","representative_topics":["generalization bounds for overparameterized 3-layer neural tangent kernel models with hidden-layer interactions","neural covariance sde for shaped infinite-depth-and-width feedforward network initialization","spectral bias bounds for deep networks via fisher information low effective rank in kernel regime"]},{"direction_index":3,"label":"D3","name":"Terminal representation geometry","description":"Study the limiting geometry of learned class representations and classifiers at late training, especially neural collapse and related terminal feature-collapse phenomena.","inclusion":"Include atoms whose primary objective is simplex/equiangular class geometry, class-mean collapse, classifier-feature duality, neural collapse under imbalance, label noise, normalization, robustness, transfer, or perturbation of terminal class geometry.","exclusion":"If the objective is general representation learning over training time, feature emergence, robustness certification, or broad generalization theory rather than terminal class geometry, assign to the neighboring representation or generalization direction.","nearest_boundary":"D4","representative_topics":["neural collapse as a terminal training state for generalization and transfer learning","theoretical analysis of neural collapse under label noise via memorization-dilation","invariant neural collapse geometry characterization under class imbalance via simplex-encoded label interpolation"]},{"direction_index":4,"label":"D4","name":"Training dynamics and optimizer mechanisms","description":"Characterize how neural-network training proceeds and what mechanisms govern convergence, representation formation, implicit regularization, loss geometry, sharpness, and phase behavior during optimization.","inclusion":"Include atoms whose primary objective is gradient-flow or SGD/GD dynamics, global/local convergence, loss landscapes, saddle or stationary-point structure, flatness/sharpness, edge-of-stability, implicit bias toward selected solutions, dropout or weight-decay regularization effects, feature-learning dynamics, lazy-to-rich transitions, spectral/simple-to-complex learning order, or mean-field training convergence.","exclusion":"If the main result is a kernel-limit asymptotic, a statistical generalization bound, terminal neural-collapse geometry, or a robustness property independent of training mechanism, assign to those directions instead.","nearest_boundary":"D2","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","stochastic gradient descent and weight decay bias toward low-rank weight matrix learning","theoretical analysis of feature learning emergence in neural networks via input structure exploitation"]},{"direction_index":5,"label":"D5","name":"Efficiency, sparsity, and learnability limits","description":"Establish theoretical limits and guarantees for efficient neural-network learning or representation, including sparse/pruned parameterizations and computational hardness of learning network classes.","inclusion":"Include atoms whose primary objective is sparse-network expressivity, pruning or lottery-ticket guarantees, low-rank or structured parameter efficiency, compact parameterizations, SQ or NP-hardness lower bounds, finite-sample identification, stationarity-testing hardness, or efficient algorithms for restricted neural classes.","exclusion":"If sparsity, low rank, or compactness is only the optimizer-selected outcome, assign to training dynamics; if parameter count is merely part of approximation-rate tradeoffs, assign to representation capacity.","nearest_boundary":"D4","representative_topics":["strong lottery ticket hypothesis generalization to arbitrary group-equivariant neural networks via pruning","theoretical analysis of random mask pruning via overparameterization and sparse network containment","strong lottery ticket hypothesis with epsilon-perturbation to relax over-parameterization requirements"]},{"direction_index":6,"label":"D6","name":"Probabilistic and physics-based network theory","description":"Explain neural networks using Bayesian inference, variational inference, ensemble theory, statistical mechanics, field theory, energy landscapes, entropy, and physical analogies.","inclusion":"Include atoms whose primary objective is posterior behavior, Bayesian neural-network asymptotics or sampling, variational Bayesian theory, statistical-mechanical learning, entropy or volume explanations, replica/field-theoretic finite-width analysis, Boltzmann/Hopfield energy models, or physics analogies for signal propagation and emergent network phenomena.","exclusion":"If probability appears only as a statistical learning bound, SGD noise model for flat minima, or mean-field convergence proof, assign to generalization or training dynamics instead.","nearest_boundary":"D1","representative_topics":["bayesian neural network posterior convergence rates in besov space with adaptive smoothness","repriorisation for accelerated mcmc sampling in wide bayesian neural networks","statistical mechanical theory of weight-distribution-constrained learning via optimal transport and information geometry"]},{"direction_index":7,"label":"D7","name":"Structured and nonstandard architectures","description":"Analyze neural-network theories whose primary objective is tied to architectural structure beyond standard feedforward networks, including graph, attention, sequence, recurrent, equivariant, invariant, reservoir, energy-based, biological, and specialized dynamical architectures.","inclusion":"Include atoms whose primary objective is Transformer or attention expressivity, GNN/WL/logical graph expressivity, recurrent or sequence computational power, memory and temporal dynamics, group-equivariant or invariant layer construction and universality, graphon or geometric graph theory, reservoir/spiking/energy-based/neural-ODE or biologically motivated architectures, and other specialized architectural computation theories.","exclusion":"If the same architecture is used only as an example for ordinary approximation, generalization, kernel limits, or optimizer convergence, assign to those broader objective directions instead.","nearest_boundary":"D0","representative_topics":["linear region bounds for graph convolutional network expressivity across depth","theoretical analysis of rank collapse and gradient propagation in transformer self-attention layers","complexity-theoretic characterization of log-precision transformers via parallelism tradeoff"]}],"split":"dev"}
{"episode_id":"F030_T2025-01","field_index":30,"field":"neural network theory","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Representation and function-space capacity","description":"Characterize what neural networks can represent or approximate, and how architecture, activation choice, geometry, operator targets, and function-space structure determine expressive capacity.","inclusion":"Include atoms whose primary objective is approximation theory, expressivity, exact representation, activation-function effects on representable functions, ReLU/polyhedral function geometry, operator/PDE/manifold approximation, or functional-analytic foundations of network function classes.","exclusion":"If the primary objective is statistical risk, optimizer-selected solutions, training convergence, or graph/Transformer/symmetry-specific computational distinction, assign to the corresponding generalization, training, or structured-architecture direction instead.","nearest_boundary":"D7","representative_topics":["l^p norm approximation lower bounds for feed-forward neural networks via fat-shattering dimension","exact neural network representation of convex piecewise quadratic functions via max-out topology","relu neural network approximation of parametric hyperbolic conservation laws with generalization error bounds"]},{"direction_index":1,"label":"D1","name":"Generalization, robustness, and risk regimes","description":"Explain out-of-sample behavior of neural networks, including sample complexity, excess risk, consistency, robustness under perturbations, and overparameterized interpolation or overfitting regimes.","inclusion":"Include atoms whose primary objective is statistical generalization bounds, sample complexity, PAC-Bayes or complexity measures, consistency, label complexity, adversarial or Lipschitz robustness, benign/harmful overfitting, double descent as a risk phenomenon, memorization, or broad inductive-bias explanations of generalization.","exclusion":"If the central object is the optimizer's selected solution, loss-landscape dynamics, kernel-limit asymptotics, or Bayesian/statistical-physics ensemble formalism, assign to those directions instead.","nearest_boundary":"D4","representative_topics":["decision boundary variability as a generalization measure in deep learning theory","hessian-based pac-bayesian generalization bounds for robust fine-tuning under label noise","consistency theory for regularized neural networks via sieve and minimal network methods"]},{"direction_index":2,"label":"D2","name":"Kernel and infinite-width limits","description":"Analyze neural networks through NTK, NNGP, random-feature, large-width, infinite-depth/width, lazy-training, and finite-width-correction limits.","inclusion":"Include atoms whose primary objective is deriving or using kernel-limit objects, infinite-width Gaussian-process correspondences, NTK spectra, lazy-regime training/generalization, random-feature asymptotics, or quantitative deviations from such limits.","exclusion":"If the work primarily studies feature learning beyond fixed kernels, ordinary optimization convergence, or functional-analysis foundations not centered on NTK/NNGP limits, assign to the relevant neighboring direction.","nearest_boundary":"D4","representative_topics":["generalization bounds for overparameterized 3-layer neural tangent kernel models with hidden-layer interactions","neural covariance sde for shaped infinite-depth-and-width feedforward network initialization","spectral bias bounds for deep networks via fisher information low effective rank in kernel regime"]},{"direction_index":3,"label":"D3","name":"Terminal representation geometry","description":"Study the limiting geometry of learned class representations and classifiers at late training, especially neural collapse and related terminal feature-collapse phenomena.","inclusion":"Include atoms whose primary objective is simplex/equiangular class geometry, class-mean collapse, classifier-feature duality, neural collapse under imbalance, label noise, normalization, robustness, transfer, or perturbation of terminal class geometry.","exclusion":"If the objective is general representation learning over training time, feature emergence, robustness certification, or broad generalization theory rather than terminal class geometry, assign to the neighboring representation or generalization direction.","nearest_boundary":"D4","representative_topics":["neural collapse as a terminal training state for generalization and transfer learning","theoretical analysis of neural collapse under label noise via memorization-dilation","invariant neural collapse geometry characterization under class imbalance via simplex-encoded label interpolation"]},{"direction_index":4,"label":"D4","name":"Training dynamics and optimizer mechanisms","description":"Characterize how neural-network training proceeds and what mechanisms govern convergence, representation formation, implicit regularization, loss geometry, sharpness, and phase behavior during optimization.","inclusion":"Include atoms whose primary objective is gradient-flow or SGD/GD dynamics, global/local convergence, loss landscapes, saddle or stationary-point structure, flatness/sharpness, edge-of-stability, implicit bias toward selected solutions, dropout or weight-decay regularization effects, feature-learning dynamics, lazy-to-rich transitions, spectral/simple-to-complex learning order, or mean-field training convergence.","exclusion":"If the main result is a kernel-limit asymptotic, a statistical generalization bound, terminal neural-collapse geometry, or a robustness property independent of training mechanism, assign to those directions instead.","nearest_boundary":"D2","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","stochastic gradient descent and weight decay bias toward low-rank weight matrix learning","theoretical analysis of feature learning emergence in neural networks via input structure exploitation"]},{"direction_index":5,"label":"D5","name":"Efficiency, sparsity, and learnability limits","description":"Establish theoretical limits and guarantees for efficient neural-network learning or representation, including sparse/pruned parameterizations and computational hardness of learning network classes.","inclusion":"Include atoms whose primary objective is sparse-network expressivity, pruning or lottery-ticket guarantees, low-rank or structured parameter efficiency, compact parameterizations, SQ or NP-hardness lower bounds, finite-sample identification, stationarity-testing hardness, or efficient algorithms for restricted neural classes.","exclusion":"If sparsity, low rank, or compactness is only the optimizer-selected outcome, assign to training dynamics; if parameter count is merely part of approximation-rate tradeoffs, assign to representation capacity.","nearest_boundary":"D4","representative_topics":["strong lottery ticket hypothesis generalization to arbitrary group-equivariant neural networks via pruning","theoretical analysis of random mask pruning via overparameterization and sparse network containment","strong lottery ticket hypothesis with epsilon-perturbation to relax over-parameterization requirements"]},{"direction_index":6,"label":"D6","name":"Probabilistic and physics-based network theory","description":"Explain neural networks using Bayesian inference, variational inference, ensemble theory, statistical mechanics, field theory, energy landscapes, entropy, and physical analogies.","inclusion":"Include atoms whose primary objective is posterior behavior, Bayesian neural-network asymptotics or sampling, variational Bayesian theory, statistical-mechanical learning, entropy or volume explanations, replica/field-theoretic finite-width analysis, Boltzmann/Hopfield energy models, or physics analogies for signal propagation and emergent network phenomena.","exclusion":"If probability appears only as a statistical learning bound, SGD noise model for flat minima, or mean-field convergence proof, assign to generalization or training dynamics instead.","nearest_boundary":"D1","representative_topics":["bayesian neural network posterior convergence rates in besov space with adaptive smoothness","repriorisation for accelerated mcmc sampling in wide bayesian neural networks","statistical mechanical theory of weight-distribution-constrained learning via optimal transport and information geometry"]},{"direction_index":7,"label":"D7","name":"Structured and nonstandard architectures","description":"Analyze neural-network theories whose primary objective is tied to architectural structure beyond standard feedforward networks, including graph, attention, sequence, recurrent, equivariant, invariant, reservoir, energy-based, biological, and specialized dynamical architectures.","inclusion":"Include atoms whose primary objective is Transformer or attention expressivity, GNN/WL/logical graph expressivity, recurrent or sequence computational power, memory and temporal dynamics, group-equivariant or invariant layer construction and universality, graphon or geometric graph theory, reservoir/spiking/energy-based/neural-ODE or biologically motivated architectures, and other specialized architectural computation theories.","exclusion":"If the same architecture is used only as an example for ordinary approximation, generalization, kernel limits, or optimizer convergence, assign to those broader objective directions instead.","nearest_boundary":"D0","representative_topics":["linear region bounds for graph convolutional network expressivity across depth","theoretical analysis of rank collapse and gradient propagation in transformer self-attention layers","complexity-theoretic characterization of log-precision transformers via parallelism tradeoff"]}],"split":"dev"}
{"episode_id":"F030_T2025-07","field_index":30,"field":"neural network theory","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Representation and function-space capacity","description":"Characterize what neural networks can represent or approximate, and how architecture, activation choice, geometry, operator targets, and function-space structure determine expressive capacity.","inclusion":"Include atoms whose primary objective is approximation theory, expressivity, exact representation, activation-function effects on representable functions, ReLU/polyhedral function geometry, operator/PDE/manifold approximation, or functional-analytic foundations of network function classes.","exclusion":"If the primary objective is statistical risk, optimizer-selected solutions, training convergence, or graph/Transformer/symmetry-specific computational distinction, assign to the corresponding generalization, training, or structured-architecture direction instead.","nearest_boundary":"D7","representative_topics":["l^p norm approximation lower bounds for feed-forward neural networks via fat-shattering dimension","exact neural network representation of convex piecewise quadratic functions via max-out topology","relu neural network approximation of parametric hyperbolic conservation laws with generalization error bounds"]},{"direction_index":1,"label":"D1","name":"Generalization, robustness, and risk regimes","description":"Explain out-of-sample behavior of neural networks, including sample complexity, excess risk, consistency, robustness under perturbations, and overparameterized interpolation or overfitting regimes.","inclusion":"Include atoms whose primary objective is statistical generalization bounds, sample complexity, PAC-Bayes or complexity measures, consistency, label complexity, adversarial or Lipschitz robustness, benign/harmful overfitting, double descent as a risk phenomenon, memorization, or broad inductive-bias explanations of generalization.","exclusion":"If the central object is the optimizer's selected solution, loss-landscape dynamics, kernel-limit asymptotics, or Bayesian/statistical-physics ensemble formalism, assign to those directions instead.","nearest_boundary":"D4","representative_topics":["decision boundary variability as a generalization measure in deep learning theory","hessian-based pac-bayesian generalization bounds for robust fine-tuning under label noise","consistency theory for regularized neural networks via sieve and minimal network methods"]},{"direction_index":2,"label":"D2","name":"Kernel and infinite-width limits","description":"Analyze neural networks through NTK, NNGP, random-feature, large-width, infinite-depth/width, lazy-training, and finite-width-correction limits.","inclusion":"Include atoms whose primary objective is deriving or using kernel-limit objects, infinite-width Gaussian-process correspondences, NTK spectra, lazy-regime training/generalization, random-feature asymptotics, or quantitative deviations from such limits.","exclusion":"If the work primarily studies feature learning beyond fixed kernels, ordinary optimization convergence, or functional-analysis foundations not centered on NTK/NNGP limits, assign to the relevant neighboring direction.","nearest_boundary":"D4","representative_topics":["generalization bounds for overparameterized 3-layer neural tangent kernel models with hidden-layer interactions","neural covariance sde for shaped infinite-depth-and-width feedforward network initialization","spectral bias bounds for deep networks via fisher information low effective rank in kernel regime"]},{"direction_index":3,"label":"D3","name":"Terminal representation geometry","description":"Study the limiting geometry of learned class representations and classifiers at late training, especially neural collapse and related terminal feature-collapse phenomena.","inclusion":"Include atoms whose primary objective is simplex/equiangular class geometry, class-mean collapse, classifier-feature duality, neural collapse under imbalance, label noise, normalization, robustness, transfer, or perturbation of terminal class geometry.","exclusion":"If the objective is general representation learning over training time, feature emergence, robustness certification, or broad generalization theory rather than terminal class geometry, assign to the neighboring representation or generalization direction.","nearest_boundary":"D4","representative_topics":["neural collapse as a terminal training state for generalization and transfer learning","theoretical analysis of neural collapse under label noise via memorization-dilation","invariant neural collapse geometry characterization under class imbalance via simplex-encoded label interpolation"]},{"direction_index":4,"label":"D4","name":"Training dynamics and optimizer mechanisms","description":"Characterize how neural-network training proceeds and what mechanisms govern convergence, representation formation, implicit regularization, loss geometry, sharpness, and phase behavior during optimization.","inclusion":"Include atoms whose primary objective is gradient-flow or SGD/GD dynamics, global/local convergence, loss landscapes, saddle or stationary-point structure, flatness/sharpness, edge-of-stability, implicit bias toward selected solutions, dropout or weight-decay regularization effects, feature-learning dynamics, lazy-to-rich transitions, spectral/simple-to-complex learning order, or mean-field training convergence.","exclusion":"If the main result is a kernel-limit asymptotic, a statistical generalization bound, terminal neural-collapse geometry, or a robustness property independent of training mechanism, assign to those directions instead.","nearest_boundary":"D2","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","stochastic gradient descent and weight decay bias toward low-rank weight matrix learning","theoretical analysis of feature learning emergence in neural networks via input structure exploitation"]},{"direction_index":5,"label":"D5","name":"Efficiency, sparsity, and learnability limits","description":"Establish theoretical limits and guarantees for efficient neural-network learning or representation, including sparse/pruned parameterizations and computational hardness of learning network classes.","inclusion":"Include atoms whose primary objective is sparse-network expressivity, pruning or lottery-ticket guarantees, low-rank or structured parameter efficiency, compact parameterizations, SQ or NP-hardness lower bounds, finite-sample identification, stationarity-testing hardness, or efficient algorithms for restricted neural classes.","exclusion":"If sparsity, low rank, or compactness is only the optimizer-selected outcome, assign to training dynamics; if parameter count is merely part of approximation-rate tradeoffs, assign to representation capacity.","nearest_boundary":"D4","representative_topics":["strong lottery ticket hypothesis generalization to arbitrary group-equivariant neural networks via pruning","theoretical analysis of random mask pruning via overparameterization and sparse network containment","strong lottery ticket hypothesis with epsilon-perturbation to relax over-parameterization requirements"]},{"direction_index":6,"label":"D6","name":"Probabilistic and physics-based network theory","description":"Explain neural networks using Bayesian inference, variational inference, ensemble theory, statistical mechanics, field theory, energy landscapes, entropy, and physical analogies.","inclusion":"Include atoms whose primary objective is posterior behavior, Bayesian neural-network asymptotics or sampling, variational Bayesian theory, statistical-mechanical learning, entropy or volume explanations, replica/field-theoretic finite-width analysis, Boltzmann/Hopfield energy models, or physics analogies for signal propagation and emergent network phenomena.","exclusion":"If probability appears only as a statistical learning bound, SGD noise model for flat minima, or mean-field convergence proof, assign to generalization or training dynamics instead.","nearest_boundary":"D1","representative_topics":["bayesian neural network posterior convergence rates in besov space with adaptive smoothness","repriorisation for accelerated mcmc sampling in wide bayesian neural networks","statistical mechanical theory of weight-distribution-constrained learning via optimal transport and information geometry"]},{"direction_index":7,"label":"D7","name":"Structured and nonstandard architectures","description":"Analyze neural-network theories whose primary objective is tied to architectural structure beyond standard feedforward networks, including graph, attention, sequence, recurrent, equivariant, invariant, reservoir, energy-based, biological, and specialized dynamical architectures.","inclusion":"Include atoms whose primary objective is Transformer or attention expressivity, GNN/WL/logical graph expressivity, recurrent or sequence computational power, memory and temporal dynamics, group-equivariant or invariant layer construction and universality, graphon or geometric graph theory, reservoir/spiking/energy-based/neural-ODE or biologically motivated architectures, and other specialized architectural computation theories.","exclusion":"If the same architecture is used only as an example for ordinary approximation, generalization, kernel limits, or optimizer convergence, assign to those broader objective directions instead.","nearest_boundary":"D0","representative_topics":["linear region bounds for graph convolutional network expressivity across depth","theoretical analysis of rank collapse and gradient propagation in transformer self-attention layers","complexity-theoretic characterization of log-precision transformers via parallelism tradeoff"]}],"split":"dev"}
{"episode_id":"F030_T2026-01","field_index":30,"field":"neural network theory","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Representation and function-space capacity","description":"Characterize what neural networks can represent or approximate, and how architecture, activation choice, geometry, operator targets, and function-space structure determine expressive capacity.","inclusion":"Include atoms whose primary objective is approximation theory, expressivity, exact representation, activation-function effects on representable functions, ReLU/polyhedral function geometry, operator/PDE/manifold approximation, or functional-analytic foundations of network function classes.","exclusion":"If the primary objective is statistical risk, optimizer-selected solutions, training convergence, or graph/Transformer/symmetry-specific computational distinction, assign to the corresponding generalization, training, or structured-architecture direction instead.","nearest_boundary":"D7","representative_topics":["l^p norm approximation lower bounds for feed-forward neural networks via fat-shattering dimension","exact neural network representation of convex piecewise quadratic functions via max-out topology","relu neural network approximation of parametric hyperbolic conservation laws with generalization error bounds"]},{"direction_index":1,"label":"D1","name":"Generalization, robustness, and risk regimes","description":"Explain out-of-sample behavior of neural networks, including sample complexity, excess risk, consistency, robustness under perturbations, and overparameterized interpolation or overfitting regimes.","inclusion":"Include atoms whose primary objective is statistical generalization bounds, sample complexity, PAC-Bayes or complexity measures, consistency, label complexity, adversarial or Lipschitz robustness, benign/harmful overfitting, double descent as a risk phenomenon, memorization, or broad inductive-bias explanations of generalization.","exclusion":"If the central object is the optimizer's selected solution, loss-landscape dynamics, kernel-limit asymptotics, or Bayesian/statistical-physics ensemble formalism, assign to those directions instead.","nearest_boundary":"D4","representative_topics":["decision boundary variability as a generalization measure in deep learning theory","hessian-based pac-bayesian generalization bounds for robust fine-tuning under label noise","consistency theory for regularized neural networks via sieve and minimal network methods"]},{"direction_index":2,"label":"D2","name":"Kernel and infinite-width limits","description":"Analyze neural networks through NTK, NNGP, random-feature, large-width, infinite-depth/width, lazy-training, and finite-width-correction limits.","inclusion":"Include atoms whose primary objective is deriving or using kernel-limit objects, infinite-width Gaussian-process correspondences, NTK spectra, lazy-regime training/generalization, random-feature asymptotics, or quantitative deviations from such limits.","exclusion":"If the work primarily studies feature learning beyond fixed kernels, ordinary optimization convergence, or functional-analysis foundations not centered on NTK/NNGP limits, assign to the relevant neighboring direction.","nearest_boundary":"D4","representative_topics":["generalization bounds for overparameterized 3-layer neural tangent kernel models with hidden-layer interactions","neural covariance sde for shaped infinite-depth-and-width feedforward network initialization","spectral bias bounds for deep networks via fisher information low effective rank in kernel regime"]},{"direction_index":3,"label":"D3","name":"Terminal representation geometry","description":"Study the limiting geometry of learned class representations and classifiers at late training, especially neural collapse and related terminal feature-collapse phenomena.","inclusion":"Include atoms whose primary objective is simplex/equiangular class geometry, class-mean collapse, classifier-feature duality, neural collapse under imbalance, label noise, normalization, robustness, transfer, or perturbation of terminal class geometry.","exclusion":"If the objective is general representation learning over training time, feature emergence, robustness certification, or broad generalization theory rather than terminal class geometry, assign to the neighboring representation or generalization direction.","nearest_boundary":"D4","representative_topics":["neural collapse as a terminal training state for generalization and transfer learning","theoretical analysis of neural collapse under label noise via memorization-dilation","invariant neural collapse geometry characterization under class imbalance via simplex-encoded label interpolation"]},{"direction_index":4,"label":"D4","name":"Training dynamics and optimizer mechanisms","description":"Characterize how neural-network training proceeds and what mechanisms govern convergence, representation formation, implicit regularization, loss geometry, sharpness, and phase behavior during optimization.","inclusion":"Include atoms whose primary objective is gradient-flow or SGD/GD dynamics, global/local convergence, loss landscapes, saddle or stationary-point structure, flatness/sharpness, edge-of-stability, implicit bias toward selected solutions, dropout or weight-decay regularization effects, feature-learning dynamics, lazy-to-rich transitions, spectral/simple-to-complex learning order, or mean-field training convergence.","exclusion":"If the main result is a kernel-limit asymptotic, a statistical generalization bound, terminal neural-collapse geometry, or a robustness property independent of training mechanism, assign to those directions instead.","nearest_boundary":"D2","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","stochastic gradient descent and weight decay bias toward low-rank weight matrix learning","theoretical analysis of feature learning emergence in neural networks via input structure exploitation"]},{"direction_index":5,"label":"D5","name":"Efficiency, sparsity, and learnability limits","description":"Establish theoretical limits and guarantees for efficient neural-network learning or representation, including sparse/pruned parameterizations and computational hardness of learning network classes.","inclusion":"Include atoms whose primary objective is sparse-network expressivity, pruning or lottery-ticket guarantees, low-rank or structured parameter efficiency, compact parameterizations, SQ or NP-hardness lower bounds, finite-sample identification, stationarity-testing hardness, or efficient algorithms for restricted neural classes.","exclusion":"If sparsity, low rank, or compactness is only the optimizer-selected outcome, assign to training dynamics; if parameter count is merely part of approximation-rate tradeoffs, assign to representation capacity.","nearest_boundary":"D4","representative_topics":["strong lottery ticket hypothesis generalization to arbitrary group-equivariant neural networks via pruning","theoretical analysis of random mask pruning via overparameterization and sparse network containment","strong lottery ticket hypothesis with epsilon-perturbation to relax over-parameterization requirements"]},{"direction_index":6,"label":"D6","name":"Probabilistic and physics-based network theory","description":"Explain neural networks using Bayesian inference, variational inference, ensemble theory, statistical mechanics, field theory, energy landscapes, entropy, and physical analogies.","inclusion":"Include atoms whose primary objective is posterior behavior, Bayesian neural-network asymptotics or sampling, variational Bayesian theory, statistical-mechanical learning, entropy or volume explanations, replica/field-theoretic finite-width analysis, Boltzmann/Hopfield energy models, or physics analogies for signal propagation and emergent network phenomena.","exclusion":"If probability appears only as a statistical learning bound, SGD noise model for flat minima, or mean-field convergence proof, assign to generalization or training dynamics instead.","nearest_boundary":"D1","representative_topics":["bayesian neural network posterior convergence rates in besov space with adaptive smoothness","repriorisation for accelerated mcmc sampling in wide bayesian neural networks","statistical mechanical theory of weight-distribution-constrained learning via optimal transport and information geometry"]},{"direction_index":7,"label":"D7","name":"Structured and nonstandard architectures","description":"Analyze neural-network theories whose primary objective is tied to architectural structure beyond standard feedforward networks, including graph, attention, sequence, recurrent, equivariant, invariant, reservoir, energy-based, biological, and specialized dynamical architectures.","inclusion":"Include atoms whose primary objective is Transformer or attention expressivity, GNN/WL/logical graph expressivity, recurrent or sequence computational power, memory and temporal dynamics, group-equivariant or invariant layer construction and universality, graphon or geometric graph theory, reservoir/spiking/energy-based/neural-ODE or biologically motivated architectures, and other specialized architectural computation theories.","exclusion":"If the same architecture is used only as an example for ordinary approximation, generalization, kernel limits, or optimizer convergence, assign to those broader objective directions instead.","nearest_boundary":"D0","representative_topics":["linear region bounds for graph convolutional network expressivity across depth","theoretical analysis of rank collapse and gradient propagation in transformer self-attention layers","complexity-theoretic characterization of log-precision transformers via parallelism tradeoff"]}],"split":"dev"}
{"episode_id":"F032_T2024-01","field_index":32,"field":"image restoration","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Illumination and artifact correction","description":"Restoring images whose main defect is incorrect illumination, exposure, dynamic range, shadows, reflections, overlays, compression/moire-like artifacts, or other localized photometric/superimposed degradations.","inclusion":"Assign atoms here when the primary objective is to correct lighting or photometric appearance, recover HDR/exposure detail, remove shadows or overlays, or eliminate named localized artifacts such as flare, reflection, highlight, watermark, JPEG, stripe, moire, document illumination, or under-display/camera artifacts.","exclusion":"If the dominant degradation is atmospheric or underwater scattering, assign to environmental visibility restoration; if it is generic sensor noise or modality-specific noise, assign to denoising and scientific-modality restoration; if the main goal is synthesizing missing content, assign to semantic completion and face restoration.","nearest_boundary":"Environmental visibility restoration","representative_topics":["intelligible unsupervised personalized low-light enhancement via interpretable attribution correlations","transformer-based low-light enhancement guided by structural prior estimation and adversarial learning","one-stage transformer for shadow removal via joint spatial attention and fourier residual blocks"]},{"direction_index":1,"label":"D1","name":"Resolution and fusion reconstruction","description":"Increasing spatial, spectral, angular, or perceptual resolution from one or more observations, including ordinary SISR, blind real-world SR, and auxiliary-view or cross-modal fusion-based reconstruction.","inclusion":"Assign atoms here when the primary objective is super-resolution or resolution-enhancing reconstruction, whether from a single low-resolution image, unknown real-world degradations, stereo/light-field/multi-view inputs, bursts, events, references, hyperspectral fusion, or cross-modal guidance.","exclusion":"If auxiliary observations are used mainly to remove weather, cloud, noise, or artifacts rather than enhance resolution, assign to the corresponding degradation program; if the contribution is a general-purpose restoration backbone across unrelated tasks, assign to unified adaptive restoration frameworks.","nearest_boundary":"Unified adaptive restoration frameworks","representative_topics":["patch-based super resolution via generalized gaussian mixture model and mmse reconstruction","dual-learning super-resolution balancing perceptual and l1 losses for real-world image restoration","generalizable latent degradation representations for blind super-resolution under novel degradations"]},{"direction_index":2,"label":"D2","name":"Denoising and scientific modalities","description":"Suppressing noise or acquisition-specific corruption while preserving detail, especially under self-supervised, real-noise, sensor-aware, medical, remote-sensing, microscopy, SAR, hyperspectral, polarized, Poisson, speckle, impulse, or other modality-shaped settings.","inclusion":"Assign atoms here when the primary objective is noise removal or restoration governed by sensor statistics, physical acquisition constraints, or scientific/medical modality requirements, including generic real-image denoising, self-supervised denoising, Poisson denoising, SAR despeckling, hyperspectral denoising, polarized denoising, impulse noise removal, microscopy, fundus, astronomical, terahertz, and related scientific image restoration.","exclusion":"If denoising is only a prior inside an explicit inverse-problem algorithm, assign to inverse-problem solvers; if the low-light/exposure correction objective dominates over noise modeling, assign to illumination and artifact correction; if the main goal is super-resolution, assign to resolution and fusion reconstruction.","nearest_boundary":"Inverse-problem solvers","representative_topics":["bayesian deep ensemble fusion for robust real-world image denoising under heterogeneous noise","test-time adaptation for real image denoising via meta-transfer learning and masked reconstruction","self-supervised sparse recurrent denoising for poisson noise in biomedical imaging"]},{"direction_index":3,"label":"D3","name":"Deblurring and geometric correction","description":"Recovering sharp, geometrically faithful images or videos from motion blur, defocus blur, spatially varying blur, video blur, turbulence-induced warping, refractive distortion, fisheye distortion, or related deformation.","inclusion":"Assign atoms here when the primary objective is blur removal, sharp-frame recovery, blur-map/deconvolution-centered restoration, video deblurring, event- or multi-frame-assisted deblurring, atmospheric turbulence mitigation, or correction of geometric distortion and refractive warping.","exclusion":"If scattering or visibility loss without dominant warping is the main degradation, assign to environmental visibility restoration; if explicit plug-and-play or variational solver design is the main contribution rather than deblurring itself, assign to inverse-problem solvers; if face-specific priors dominate, assign to semantic completion and face restoration.","nearest_boundary":"Environmental visibility restoration","representative_topics":["multi-scale frequency separation network for image deblurring via contrastive and cycle-consistency learning","non-blind spatially varying defocus blur removal using dual encoder-decoder networks","single-image defocus deblurring with spatially dynamic modulation guided by defocus map estimation"]},{"direction_index":4,"label":"D4","name":"Environmental visibility restoration","description":"Restoring scene visibility, color, and contrast under environmental media or weather, including atmospheric haze/fog/dust/sand, rain, snow, mixed adverse weather, and underwater absorption/scattering.","inclusion":"Assign atoms here when the primary objective is dehazing, defogging, dust or sandstorm removal, deraining, desnowing, joint weather removal, nighttime dehazing when scattering dominates, or underwater enhancement/restoration driven by water absorption, scattering, color cast, turbidity, marine particles, or underwater optics.","exclusion":"If the main issue is global underexposure rather than scattering or weather particles, assign to illumination and artifact correction; if the main issue is geometric turbulence, assign to deblurring and geometric correction; if the main objective is resolution enhancement of underwater or weather imagery, assign to this direction only when the medium/weather physics drives the research objective, otherwise to resolution and fusion reconstruction.","nearest_boundary":"Illumination and artifact correction","representative_topics":["haar wavelet reinforcement fusion for underwater image enhancement","self-supervised underwater image enhancement via degradation estimation and attention-guided high-intensity correction","transform-domain generative adversarial network for frequency-decoupled image dehazing"]},{"direction_index":5,"label":"D5","name":"Semantic completion and face restoration","description":"Restoring semantically structured content where high-level priors, identity, object structure, or plausible synthesis are central, including image inpainting, object removal, masked-region completion, and face-specific restoration.","inclusion":"Assign atoms here when the primary objective is filling missing or masked regions, removing objects by synthesizing plausible content, improving inpainted-region structure/texture consistency, or restoring faces using face-specific priors such as identity, parsing, codebooks, dictionaries, geometry, reference faces, or facial structure.","exclusion":"If a named overlay artifact is removed without substantial missing-region synthesis, assign to illumination and artifact correction; if a general SR/deblurring/denoising method is merely evaluated on faces without face-specific priors, assign to the corresponding task program; if GAN or diffusion is used chiefly as a general cross-task restoration framework, assign to unified adaptive restoration frameworks.","nearest_boundary":"Illumination and artifact correction","representative_topics":["wavelet prior attention and axial-transformer learning for artifact-free image inpainting","feature refinement for high-resolution neural image inpainting via multiscale consistency optimization","structure and texture joint recovery in deep image inpainting via hybrid gan and fourier convolutional design"]},{"direction_index":6,"label":"D6","name":"Unified adaptive restoration frameworks","description":"Developing broadly applicable restoration systems, priors, architectures, and training/deployment strategies intended to work across multiple degradations, tasks, or restoration settings rather than optimize one named degradation.","inclusion":"Assign atoms here when the primary objective is all-in-one or blind multi-degradation restoration, prompt/adaptive/unified restoration, cross-task backbone design, general restoration architecture or training strategy, model compression for restoration, degradation representation learning, or general generative-prior restoration using diffusion, score-based models, GANs, flows, or generative inversion across restoration tasks.","exclusion":"If a method-specific architecture or generative model is used mainly for one named task such as low-light, deblurring, inpainting, face restoration, or SR, assign to that task program; if explicit data-consistency optimization or plug-and-play solver structure is the main contribution, assign to inverse-problem solvers.","nearest_boundary":"Inverse-problem solvers","representative_topics":["multi-branch compound feature fusion for generalizable multi-degradation image restoration","hypernetwork-based adaptive image restoration with fixed-size single model","variational bayesian framework for blind image restoration with diverse unknown degradations"]},{"direction_index":7,"label":"D7","name":"Inverse-problem solvers","description":"Formulating and solving restoration as an explicit inverse problem with forward operators, data consistency, mathematical regularization, plug-and-play priors, proximal or unrolled optimization, posterior sampling, convergence analysis, or operator-faithful reconstruction.","inclusion":"Assign atoms here when the primary contribution is a restoration solver or formulation such as plug-and-play ADMM, RED, proximal algorithms, variational regularization, matrix completion, explicit deconvolution, compressed sensing, posterior sampling, algorithm unrolling, convergence/robustness analysis, or learned priors embedded in a data-consistent inverse-problem pipeline.","exclusion":"If the paper mainly proposes a task-specific network for deblurring, denoising, SR, or inpainting without emphasizing the inverse-problem solver, assign to that task program; if the work is a broad all-in-one architecture without explicit operator fidelity, assign to unified adaptive restoration frameworks.","nearest_boundary":"Unified adaptive restoration frameworks","representative_topics":["framelet local minimal prior for poissonian blurred image deconvolution with fractional operators","deep reinforcement learning denoisers for robust plug-and-play image restoration under model mismatch","variational joint image recovery using spatially-varying generalised gaussian priors"]}],"split":"dev"}
{"episode_id":"F032_T2024-07","field_index":32,"field":"image restoration","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Illumination and artifact correction","description":"Restoring images whose main defect is incorrect illumination, exposure, dynamic range, shadows, reflections, overlays, compression/moire-like artifacts, or other localized photometric/superimposed degradations.","inclusion":"Assign atoms here when the primary objective is to correct lighting or photometric appearance, recover HDR/exposure detail, remove shadows or overlays, or eliminate named localized artifacts such as flare, reflection, highlight, watermark, JPEG, stripe, moire, document illumination, or under-display/camera artifacts.","exclusion":"If the dominant degradation is atmospheric or underwater scattering, assign to environmental visibility restoration; if it is generic sensor noise or modality-specific noise, assign to denoising and scientific-modality restoration; if the main goal is synthesizing missing content, assign to semantic completion and face restoration.","nearest_boundary":"Environmental visibility restoration","representative_topics":["intelligible unsupervised personalized low-light enhancement via interpretable attribution correlations","transformer-based low-light enhancement guided by structural prior estimation and adversarial learning","one-stage transformer for shadow removal via joint spatial attention and fourier residual blocks"]},{"direction_index":1,"label":"D1","name":"Resolution and fusion reconstruction","description":"Increasing spatial, spectral, angular, or perceptual resolution from one or more observations, including ordinary SISR, blind real-world SR, and auxiliary-view or cross-modal fusion-based reconstruction.","inclusion":"Assign atoms here when the primary objective is super-resolution or resolution-enhancing reconstruction, whether from a single low-resolution image, unknown real-world degradations, stereo/light-field/multi-view inputs, bursts, events, references, hyperspectral fusion, or cross-modal guidance.","exclusion":"If auxiliary observations are used mainly to remove weather, cloud, noise, or artifacts rather than enhance resolution, assign to the corresponding degradation program; if the contribution is a general-purpose restoration backbone across unrelated tasks, assign to unified adaptive restoration frameworks.","nearest_boundary":"Unified adaptive restoration frameworks","representative_topics":["patch-based super resolution via generalized gaussian mixture model and mmse reconstruction","dual-learning super-resolution balancing perceptual and l1 losses for real-world image restoration","generalizable latent degradation representations for blind super-resolution under novel degradations"]},{"direction_index":2,"label":"D2","name":"Denoising and scientific modalities","description":"Suppressing noise or acquisition-specific corruption while preserving detail, especially under self-supervised, real-noise, sensor-aware, medical, remote-sensing, microscopy, SAR, hyperspectral, polarized, Poisson, speckle, impulse, or other modality-shaped settings.","inclusion":"Assign atoms here when the primary objective is noise removal or restoration governed by sensor statistics, physical acquisition constraints, or scientific/medical modality requirements, including generic real-image denoising, self-supervised denoising, Poisson denoising, SAR despeckling, hyperspectral denoising, polarized denoising, impulse noise removal, microscopy, fundus, astronomical, terahertz, and related scientific image restoration.","exclusion":"If denoising is only a prior inside an explicit inverse-problem algorithm, assign to inverse-problem solvers; if the low-light/exposure correction objective dominates over noise modeling, assign to illumination and artifact correction; if the main goal is super-resolution, assign to resolution and fusion reconstruction.","nearest_boundary":"Inverse-problem solvers","representative_topics":["bayesian deep ensemble fusion for robust real-world image denoising under heterogeneous noise","test-time adaptation for real image denoising via meta-transfer learning and masked reconstruction","self-supervised sparse recurrent denoising for poisson noise in biomedical imaging"]},{"direction_index":3,"label":"D3","name":"Deblurring and geometric correction","description":"Recovering sharp, geometrically faithful images or videos from motion blur, defocus blur, spatially varying blur, video blur, turbulence-induced warping, refractive distortion, fisheye distortion, or related deformation.","inclusion":"Assign atoms here when the primary objective is blur removal, sharp-frame recovery, blur-map/deconvolution-centered restoration, video deblurring, event- or multi-frame-assisted deblurring, atmospheric turbulence mitigation, or correction of geometric distortion and refractive warping.","exclusion":"If scattering or visibility loss without dominant warping is the main degradation, assign to environmental visibility restoration; if explicit plug-and-play or variational solver design is the main contribution rather than deblurring itself, assign to inverse-problem solvers; if face-specific priors dominate, assign to semantic completion and face restoration.","nearest_boundary":"Environmental visibility restoration","representative_topics":["multi-scale frequency separation network for image deblurring via contrastive and cycle-consistency learning","non-blind spatially varying defocus blur removal using dual encoder-decoder networks","single-image defocus deblurring with spatially dynamic modulation guided by defocus map estimation"]},{"direction_index":4,"label":"D4","name":"Environmental visibility restoration","description":"Restoring scene visibility, color, and contrast under environmental media or weather, including atmospheric haze/fog/dust/sand, rain, snow, mixed adverse weather, and underwater absorption/scattering.","inclusion":"Assign atoms here when the primary objective is dehazing, defogging, dust or sandstorm removal, deraining, desnowing, joint weather removal, nighttime dehazing when scattering dominates, or underwater enhancement/restoration driven by water absorption, scattering, color cast, turbidity, marine particles, or underwater optics.","exclusion":"If the main issue is global underexposure rather than scattering or weather particles, assign to illumination and artifact correction; if the main issue is geometric turbulence, assign to deblurring and geometric correction; if the main objective is resolution enhancement of underwater or weather imagery, assign to this direction only when the medium/weather physics drives the research objective, otherwise to resolution and fusion reconstruction.","nearest_boundary":"Illumination and artifact correction","representative_topics":["haar wavelet reinforcement fusion for underwater image enhancement","self-supervised underwater image enhancement via degradation estimation and attention-guided high-intensity correction","transform-domain generative adversarial network for frequency-decoupled image dehazing"]},{"direction_index":5,"label":"D5","name":"Semantic completion and face restoration","description":"Restoring semantically structured content where high-level priors, identity, object structure, or plausible synthesis are central, including image inpainting, object removal, masked-region completion, and face-specific restoration.","inclusion":"Assign atoms here when the primary objective is filling missing or masked regions, removing objects by synthesizing plausible content, improving inpainted-region structure/texture consistency, or restoring faces using face-specific priors such as identity, parsing, codebooks, dictionaries, geometry, reference faces, or facial structure.","exclusion":"If a named overlay artifact is removed without substantial missing-region synthesis, assign to illumination and artifact correction; if a general SR/deblurring/denoising method is merely evaluated on faces without face-specific priors, assign to the corresponding task program; if GAN or diffusion is used chiefly as a general cross-task restoration framework, assign to unified adaptive restoration frameworks.","nearest_boundary":"Illumination and artifact correction","representative_topics":["wavelet prior attention and axial-transformer learning for artifact-free image inpainting","feature refinement for high-resolution neural image inpainting via multiscale consistency optimization","structure and texture joint recovery in deep image inpainting via hybrid gan and fourier convolutional design"]},{"direction_index":6,"label":"D6","name":"Unified adaptive restoration frameworks","description":"Developing broadly applicable restoration systems, priors, architectures, and training/deployment strategies intended to work across multiple degradations, tasks, or restoration settings rather than optimize one named degradation.","inclusion":"Assign atoms here when the primary objective is all-in-one or blind multi-degradation restoration, prompt/adaptive/unified restoration, cross-task backbone design, general restoration architecture or training strategy, model compression for restoration, degradation representation learning, or general generative-prior restoration using diffusion, score-based models, GANs, flows, or generative inversion across restoration tasks.","exclusion":"If a method-specific architecture or generative model is used mainly for one named task such as low-light, deblurring, inpainting, face restoration, or SR, assign to that task program; if explicit data-consistency optimization or plug-and-play solver structure is the main contribution, assign to inverse-problem solvers.","nearest_boundary":"Inverse-problem solvers","representative_topics":["multi-branch compound feature fusion for generalizable multi-degradation image restoration","hypernetwork-based adaptive image restoration with fixed-size single model","variational bayesian framework for blind image restoration with diverse unknown degradations"]},{"direction_index":7,"label":"D7","name":"Inverse-problem solvers","description":"Formulating and solving restoration as an explicit inverse problem with forward operators, data consistency, mathematical regularization, plug-and-play priors, proximal or unrolled optimization, posterior sampling, convergence analysis, or operator-faithful reconstruction.","inclusion":"Assign atoms here when the primary contribution is a restoration solver or formulation such as plug-and-play ADMM, RED, proximal algorithms, variational regularization, matrix completion, explicit deconvolution, compressed sensing, posterior sampling, algorithm unrolling, convergence/robustness analysis, or learned priors embedded in a data-consistent inverse-problem pipeline.","exclusion":"If the paper mainly proposes a task-specific network for deblurring, denoising, SR, or inpainting without emphasizing the inverse-problem solver, assign to that task program; if the work is a broad all-in-one architecture without explicit operator fidelity, assign to unified adaptive restoration frameworks.","nearest_boundary":"Unified adaptive restoration frameworks","representative_topics":["framelet local minimal prior for poissonian blurred image deconvolution with fractional operators","deep reinforcement learning denoisers for robust plug-and-play image restoration under model mismatch","variational joint image recovery using spatially-varying generalised gaussian priors"]}],"split":"dev"}
{"episode_id":"F032_T2025-01","field_index":32,"field":"image restoration","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Illumination and artifact correction","description":"Restoring images whose main defect is incorrect illumination, exposure, dynamic range, shadows, reflections, overlays, compression/moire-like artifacts, or other localized photometric/superimposed degradations.","inclusion":"Assign atoms here when the primary objective is to correct lighting or photometric appearance, recover HDR/exposure detail, remove shadows or overlays, or eliminate named localized artifacts such as flare, reflection, highlight, watermark, JPEG, stripe, moire, document illumination, or under-display/camera artifacts.","exclusion":"If the dominant degradation is atmospheric or underwater scattering, assign to environmental visibility restoration; if it is generic sensor noise or modality-specific noise, assign to denoising and scientific-modality restoration; if the main goal is synthesizing missing content, assign to semantic completion and face restoration.","nearest_boundary":"Environmental visibility restoration","representative_topics":["intelligible unsupervised personalized low-light enhancement via interpretable attribution correlations","transformer-based low-light enhancement guided by structural prior estimation and adversarial learning","one-stage transformer for shadow removal via joint spatial attention and fourier residual blocks"]},{"direction_index":1,"label":"D1","name":"Resolution and fusion reconstruction","description":"Increasing spatial, spectral, angular, or perceptual resolution from one or more observations, including ordinary SISR, blind real-world SR, and auxiliary-view or cross-modal fusion-based reconstruction.","inclusion":"Assign atoms here when the primary objective is super-resolution or resolution-enhancing reconstruction, whether from a single low-resolution image, unknown real-world degradations, stereo/light-field/multi-view inputs, bursts, events, references, hyperspectral fusion, or cross-modal guidance.","exclusion":"If auxiliary observations are used mainly to remove weather, cloud, noise, or artifacts rather than enhance resolution, assign to the corresponding degradation program; if the contribution is a general-purpose restoration backbone across unrelated tasks, assign to unified adaptive restoration frameworks.","nearest_boundary":"Unified adaptive restoration frameworks","representative_topics":["patch-based super resolution via generalized gaussian mixture model and mmse reconstruction","dual-learning super-resolution balancing perceptual and l1 losses for real-world image restoration","generalizable latent degradation representations for blind super-resolution under novel degradations"]},{"direction_index":2,"label":"D2","name":"Denoising and scientific modalities","description":"Suppressing noise or acquisition-specific corruption while preserving detail, especially under self-supervised, real-noise, sensor-aware, medical, remote-sensing, microscopy, SAR, hyperspectral, polarized, Poisson, speckle, impulse, or other modality-shaped settings.","inclusion":"Assign atoms here when the primary objective is noise removal or restoration governed by sensor statistics, physical acquisition constraints, or scientific/medical modality requirements, including generic real-image denoising, self-supervised denoising, Poisson denoising, SAR despeckling, hyperspectral denoising, polarized denoising, impulse noise removal, microscopy, fundus, astronomical, terahertz, and related scientific image restoration.","exclusion":"If denoising is only a prior inside an explicit inverse-problem algorithm, assign to inverse-problem solvers; if the low-light/exposure correction objective dominates over noise modeling, assign to illumination and artifact correction; if the main goal is super-resolution, assign to resolution and fusion reconstruction.","nearest_boundary":"Inverse-problem solvers","representative_topics":["bayesian deep ensemble fusion for robust real-world image denoising under heterogeneous noise","test-time adaptation for real image denoising via meta-transfer learning and masked reconstruction","self-supervised sparse recurrent denoising for poisson noise in biomedical imaging"]},{"direction_index":3,"label":"D3","name":"Deblurring and geometric correction","description":"Recovering sharp, geometrically faithful images or videos from motion blur, defocus blur, spatially varying blur, video blur, turbulence-induced warping, refractive distortion, fisheye distortion, or related deformation.","inclusion":"Assign atoms here when the primary objective is blur removal, sharp-frame recovery, blur-map/deconvolution-centered restoration, video deblurring, event- or multi-frame-assisted deblurring, atmospheric turbulence mitigation, or correction of geometric distortion and refractive warping.","exclusion":"If scattering or visibility loss without dominant warping is the main degradation, assign to environmental visibility restoration; if explicit plug-and-play or variational solver design is the main contribution rather than deblurring itself, assign to inverse-problem solvers; if face-specific priors dominate, assign to semantic completion and face restoration.","nearest_boundary":"Environmental visibility restoration","representative_topics":["multi-scale frequency separation network for image deblurring via contrastive and cycle-consistency learning","non-blind spatially varying defocus blur removal using dual encoder-decoder networks","single-image defocus deblurring with spatially dynamic modulation guided by defocus map estimation"]},{"direction_index":4,"label":"D4","name":"Environmental visibility restoration","description":"Restoring scene visibility, color, and contrast under environmental media or weather, including atmospheric haze/fog/dust/sand, rain, snow, mixed adverse weather, and underwater absorption/scattering.","inclusion":"Assign atoms here when the primary objective is dehazing, defogging, dust or sandstorm removal, deraining, desnowing, joint weather removal, nighttime dehazing when scattering dominates, or underwater enhancement/restoration driven by water absorption, scattering, color cast, turbidity, marine particles, or underwater optics.","exclusion":"If the main issue is global underexposure rather than scattering or weather particles, assign to illumination and artifact correction; if the main issue is geometric turbulence, assign to deblurring and geometric correction; if the main objective is resolution enhancement of underwater or weather imagery, assign to this direction only when the medium/weather physics drives the research objective, otherwise to resolution and fusion reconstruction.","nearest_boundary":"Illumination and artifact correction","representative_topics":["haar wavelet reinforcement fusion for underwater image enhancement","self-supervised underwater image enhancement via degradation estimation and attention-guided high-intensity correction","transform-domain generative adversarial network for frequency-decoupled image dehazing"]},{"direction_index":5,"label":"D5","name":"Semantic completion and face restoration","description":"Restoring semantically structured content where high-level priors, identity, object structure, or plausible synthesis are central, including image inpainting, object removal, masked-region completion, and face-specific restoration.","inclusion":"Assign atoms here when the primary objective is filling missing or masked regions, removing objects by synthesizing plausible content, improving inpainted-region structure/texture consistency, or restoring faces using face-specific priors such as identity, parsing, codebooks, dictionaries, geometry, reference faces, or facial structure.","exclusion":"If a named overlay artifact is removed without substantial missing-region synthesis, assign to illumination and artifact correction; if a general SR/deblurring/denoising method is merely evaluated on faces without face-specific priors, assign to the corresponding task program; if GAN or diffusion is used chiefly as a general cross-task restoration framework, assign to unified adaptive restoration frameworks.","nearest_boundary":"Illumination and artifact correction","representative_topics":["wavelet prior attention and axial-transformer learning for artifact-free image inpainting","feature refinement for high-resolution neural image inpainting via multiscale consistency optimization","structure and texture joint recovery in deep image inpainting via hybrid gan and fourier convolutional design"]},{"direction_index":6,"label":"D6","name":"Unified adaptive restoration frameworks","description":"Developing broadly applicable restoration systems, priors, architectures, and training/deployment strategies intended to work across multiple degradations, tasks, or restoration settings rather than optimize one named degradation.","inclusion":"Assign atoms here when the primary objective is all-in-one or blind multi-degradation restoration, prompt/adaptive/unified restoration, cross-task backbone design, general restoration architecture or training strategy, model compression for restoration, degradation representation learning, or general generative-prior restoration using diffusion, score-based models, GANs, flows, or generative inversion across restoration tasks.","exclusion":"If a method-specific architecture or generative model is used mainly for one named task such as low-light, deblurring, inpainting, face restoration, or SR, assign to that task program; if explicit data-consistency optimization or plug-and-play solver structure is the main contribution, assign to inverse-problem solvers.","nearest_boundary":"Inverse-problem solvers","representative_topics":["multi-branch compound feature fusion for generalizable multi-degradation image restoration","hypernetwork-based adaptive image restoration with fixed-size single model","variational bayesian framework for blind image restoration with diverse unknown degradations"]},{"direction_index":7,"label":"D7","name":"Inverse-problem solvers","description":"Formulating and solving restoration as an explicit inverse problem with forward operators, data consistency, mathematical regularization, plug-and-play priors, proximal or unrolled optimization, posterior sampling, convergence analysis, or operator-faithful reconstruction.","inclusion":"Assign atoms here when the primary contribution is a restoration solver or formulation such as plug-and-play ADMM, RED, proximal algorithms, variational regularization, matrix completion, explicit deconvolution, compressed sensing, posterior sampling, algorithm unrolling, convergence/robustness analysis, or learned priors embedded in a data-consistent inverse-problem pipeline.","exclusion":"If the paper mainly proposes a task-specific network for deblurring, denoising, SR, or inpainting without emphasizing the inverse-problem solver, assign to that task program; if the work is a broad all-in-one architecture without explicit operator fidelity, assign to unified adaptive restoration frameworks.","nearest_boundary":"Unified adaptive restoration frameworks","representative_topics":["framelet local minimal prior for poissonian blurred image deconvolution with fractional operators","deep reinforcement learning denoisers for robust plug-and-play image restoration under model mismatch","variational joint image recovery using spatially-varying generalised gaussian priors"]}],"split":"dev"}
{"episode_id":"F032_T2025-07","field_index":32,"field":"image restoration","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Illumination and artifact correction","description":"Restoring images whose main defect is incorrect illumination, exposure, dynamic range, shadows, reflections, overlays, compression/moire-like artifacts, or other localized photometric/superimposed degradations.","inclusion":"Assign atoms here when the primary objective is to correct lighting or photometric appearance, recover HDR/exposure detail, remove shadows or overlays, or eliminate named localized artifacts such as flare, reflection, highlight, watermark, JPEG, stripe, moire, document illumination, or under-display/camera artifacts.","exclusion":"If the dominant degradation is atmospheric or underwater scattering, assign to environmental visibility restoration; if it is generic sensor noise or modality-specific noise, assign to denoising and scientific-modality restoration; if the main goal is synthesizing missing content, assign to semantic completion and face restoration.","nearest_boundary":"Environmental visibility restoration","representative_topics":["intelligible unsupervised personalized low-light enhancement via interpretable attribution correlations","transformer-based low-light enhancement guided by structural prior estimation and adversarial learning","one-stage transformer for shadow removal via joint spatial attention and fourier residual blocks"]},{"direction_index":1,"label":"D1","name":"Resolution and fusion reconstruction","description":"Increasing spatial, spectral, angular, or perceptual resolution from one or more observations, including ordinary SISR, blind real-world SR, and auxiliary-view or cross-modal fusion-based reconstruction.","inclusion":"Assign atoms here when the primary objective is super-resolution or resolution-enhancing reconstruction, whether from a single low-resolution image, unknown real-world degradations, stereo/light-field/multi-view inputs, bursts, events, references, hyperspectral fusion, or cross-modal guidance.","exclusion":"If auxiliary observations are used mainly to remove weather, cloud, noise, or artifacts rather than enhance resolution, assign to the corresponding degradation program; if the contribution is a general-purpose restoration backbone across unrelated tasks, assign to unified adaptive restoration frameworks.","nearest_boundary":"Unified adaptive restoration frameworks","representative_topics":["patch-based super resolution via generalized gaussian mixture model and mmse reconstruction","dual-learning super-resolution balancing perceptual and l1 losses for real-world image restoration","generalizable latent degradation representations for blind super-resolution under novel degradations"]},{"direction_index":2,"label":"D2","name":"Denoising and scientific modalities","description":"Suppressing noise or acquisition-specific corruption while preserving detail, especially under self-supervised, real-noise, sensor-aware, medical, remote-sensing, microscopy, SAR, hyperspectral, polarized, Poisson, speckle, impulse, or other modality-shaped settings.","inclusion":"Assign atoms here when the primary objective is noise removal or restoration governed by sensor statistics, physical acquisition constraints, or scientific/medical modality requirements, including generic real-image denoising, self-supervised denoising, Poisson denoising, SAR despeckling, hyperspectral denoising, polarized denoising, impulse noise removal, microscopy, fundus, astronomical, terahertz, and related scientific image restoration.","exclusion":"If denoising is only a prior inside an explicit inverse-problem algorithm, assign to inverse-problem solvers; if the low-light/exposure correction objective dominates over noise modeling, assign to illumination and artifact correction; if the main goal is super-resolution, assign to resolution and fusion reconstruction.","nearest_boundary":"Inverse-problem solvers","representative_topics":["bayesian deep ensemble fusion for robust real-world image denoising under heterogeneous noise","test-time adaptation for real image denoising via meta-transfer learning and masked reconstruction","self-supervised sparse recurrent denoising for poisson noise in biomedical imaging"]},{"direction_index":3,"label":"D3","name":"Deblurring and geometric correction","description":"Recovering sharp, geometrically faithful images or videos from motion blur, defocus blur, spatially varying blur, video blur, turbulence-induced warping, refractive distortion, fisheye distortion, or related deformation.","inclusion":"Assign atoms here when the primary objective is blur removal, sharp-frame recovery, blur-map/deconvolution-centered restoration, video deblurring, event- or multi-frame-assisted deblurring, atmospheric turbulence mitigation, or correction of geometric distortion and refractive warping.","exclusion":"If scattering or visibility loss without dominant warping is the main degradation, assign to environmental visibility restoration; if explicit plug-and-play or variational solver design is the main contribution rather than deblurring itself, assign to inverse-problem solvers; if face-specific priors dominate, assign to semantic completion and face restoration.","nearest_boundary":"Environmental visibility restoration","representative_topics":["multi-scale frequency separation network for image deblurring via contrastive and cycle-consistency learning","non-blind spatially varying defocus blur removal using dual encoder-decoder networks","single-image defocus deblurring with spatially dynamic modulation guided by defocus map estimation"]},{"direction_index":4,"label":"D4","name":"Environmental visibility restoration","description":"Restoring scene visibility, color, and contrast under environmental media or weather, including atmospheric haze/fog/dust/sand, rain, snow, mixed adverse weather, and underwater absorption/scattering.","inclusion":"Assign atoms here when the primary objective is dehazing, defogging, dust or sandstorm removal, deraining, desnowing, joint weather removal, nighttime dehazing when scattering dominates, or underwater enhancement/restoration driven by water absorption, scattering, color cast, turbidity, marine particles, or underwater optics.","exclusion":"If the main issue is global underexposure rather than scattering or weather particles, assign to illumination and artifact correction; if the main issue is geometric turbulence, assign to deblurring and geometric correction; if the main objective is resolution enhancement of underwater or weather imagery, assign to this direction only when the medium/weather physics drives the research objective, otherwise to resolution and fusion reconstruction.","nearest_boundary":"Illumination and artifact correction","representative_topics":["haar wavelet reinforcement fusion for underwater image enhancement","self-supervised underwater image enhancement via degradation estimation and attention-guided high-intensity correction","transform-domain generative adversarial network for frequency-decoupled image dehazing"]},{"direction_index":5,"label":"D5","name":"Semantic completion and face restoration","description":"Restoring semantically structured content where high-level priors, identity, object structure, or plausible synthesis are central, including image inpainting, object removal, masked-region completion, and face-specific restoration.","inclusion":"Assign atoms here when the primary objective is filling missing or masked regions, removing objects by synthesizing plausible content, improving inpainted-region structure/texture consistency, or restoring faces using face-specific priors such as identity, parsing, codebooks, dictionaries, geometry, reference faces, or facial structure.","exclusion":"If a named overlay artifact is removed without substantial missing-region synthesis, assign to illumination and artifact correction; if a general SR/deblurring/denoising method is merely evaluated on faces without face-specific priors, assign to the corresponding task program; if GAN or diffusion is used chiefly as a general cross-task restoration framework, assign to unified adaptive restoration frameworks.","nearest_boundary":"Illumination and artifact correction","representative_topics":["wavelet prior attention and axial-transformer learning for artifact-free image inpainting","feature refinement for high-resolution neural image inpainting via multiscale consistency optimization","structure and texture joint recovery in deep image inpainting via hybrid gan and fourier convolutional design"]},{"direction_index":6,"label":"D6","name":"Unified adaptive restoration frameworks","description":"Developing broadly applicable restoration systems, priors, architectures, and training/deployment strategies intended to work across multiple degradations, tasks, or restoration settings rather than optimize one named degradation.","inclusion":"Assign atoms here when the primary objective is all-in-one or blind multi-degradation restoration, prompt/adaptive/unified restoration, cross-task backbone design, general restoration architecture or training strategy, model compression for restoration, degradation representation learning, or general generative-prior restoration using diffusion, score-based models, GANs, flows, or generative inversion across restoration tasks.","exclusion":"If a method-specific architecture or generative model is used mainly for one named task such as low-light, deblurring, inpainting, face restoration, or SR, assign to that task program; if explicit data-consistency optimization or plug-and-play solver structure is the main contribution, assign to inverse-problem solvers.","nearest_boundary":"Inverse-problem solvers","representative_topics":["multi-branch compound feature fusion for generalizable multi-degradation image restoration","hypernetwork-based adaptive image restoration with fixed-size single model","variational bayesian framework for blind image restoration with diverse unknown degradations"]},{"direction_index":7,"label":"D7","name":"Inverse-problem solvers","description":"Formulating and solving restoration as an explicit inverse problem with forward operators, data consistency, mathematical regularization, plug-and-play priors, proximal or unrolled optimization, posterior sampling, convergence analysis, or operator-faithful reconstruction.","inclusion":"Assign atoms here when the primary contribution is a restoration solver or formulation such as plug-and-play ADMM, RED, proximal algorithms, variational regularization, matrix completion, explicit deconvolution, compressed sensing, posterior sampling, algorithm unrolling, convergence/robustness analysis, or learned priors embedded in a data-consistent inverse-problem pipeline.","exclusion":"If the paper mainly proposes a task-specific network for deblurring, denoising, SR, or inpainting without emphasizing the inverse-problem solver, assign to that task program; if the work is a broad all-in-one architecture without explicit operator fidelity, assign to unified adaptive restoration frameworks.","nearest_boundary":"Unified adaptive restoration frameworks","representative_topics":["framelet local minimal prior for poissonian blurred image deconvolution with fractional operators","deep reinforcement learning denoisers for robust plug-and-play image restoration under model mismatch","variational joint image recovery using spatially-varying generalised gaussian priors"]}],"split":"dev"}
{"episode_id":"F032_T2026-01","field_index":32,"field":"image restoration","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Illumination and artifact correction","description":"Restoring images whose main defect is incorrect illumination, exposure, dynamic range, shadows, reflections, overlays, compression/moire-like artifacts, or other localized photometric/superimposed degradations.","inclusion":"Assign atoms here when the primary objective is to correct lighting or photometric appearance, recover HDR/exposure detail, remove shadows or overlays, or eliminate named localized artifacts such as flare, reflection, highlight, watermark, JPEG, stripe, moire, document illumination, or under-display/camera artifacts.","exclusion":"If the dominant degradation is atmospheric or underwater scattering, assign to environmental visibility restoration; if it is generic sensor noise or modality-specific noise, assign to denoising and scientific-modality restoration; if the main goal is synthesizing missing content, assign to semantic completion and face restoration.","nearest_boundary":"Environmental visibility restoration","representative_topics":["intelligible unsupervised personalized low-light enhancement via interpretable attribution correlations","transformer-based low-light enhancement guided by structural prior estimation and adversarial learning","one-stage transformer for shadow removal via joint spatial attention and fourier residual blocks"]},{"direction_index":1,"label":"D1","name":"Resolution and fusion reconstruction","description":"Increasing spatial, spectral, angular, or perceptual resolution from one or more observations, including ordinary SISR, blind real-world SR, and auxiliary-view or cross-modal fusion-based reconstruction.","inclusion":"Assign atoms here when the primary objective is super-resolution or resolution-enhancing reconstruction, whether from a single low-resolution image, unknown real-world degradations, stereo/light-field/multi-view inputs, bursts, events, references, hyperspectral fusion, or cross-modal guidance.","exclusion":"If auxiliary observations are used mainly to remove weather, cloud, noise, or artifacts rather than enhance resolution, assign to the corresponding degradation program; if the contribution is a general-purpose restoration backbone across unrelated tasks, assign to unified adaptive restoration frameworks.","nearest_boundary":"Unified adaptive restoration frameworks","representative_topics":["patch-based super resolution via generalized gaussian mixture model and mmse reconstruction","dual-learning super-resolution balancing perceptual and l1 losses for real-world image restoration","generalizable latent degradation representations for blind super-resolution under novel degradations"]},{"direction_index":2,"label":"D2","name":"Denoising and scientific modalities","description":"Suppressing noise or acquisition-specific corruption while preserving detail, especially under self-supervised, real-noise, sensor-aware, medical, remote-sensing, microscopy, SAR, hyperspectral, polarized, Poisson, speckle, impulse, or other modality-shaped settings.","inclusion":"Assign atoms here when the primary objective is noise removal or restoration governed by sensor statistics, physical acquisition constraints, or scientific/medical modality requirements, including generic real-image denoising, self-supervised denoising, Poisson denoising, SAR despeckling, hyperspectral denoising, polarized denoising, impulse noise removal, microscopy, fundus, astronomical, terahertz, and related scientific image restoration.","exclusion":"If denoising is only a prior inside an explicit inverse-problem algorithm, assign to inverse-problem solvers; if the low-light/exposure correction objective dominates over noise modeling, assign to illumination and artifact correction; if the main goal is super-resolution, assign to resolution and fusion reconstruction.","nearest_boundary":"Inverse-problem solvers","representative_topics":["bayesian deep ensemble fusion for robust real-world image denoising under heterogeneous noise","test-time adaptation for real image denoising via meta-transfer learning and masked reconstruction","self-supervised sparse recurrent denoising for poisson noise in biomedical imaging"]},{"direction_index":3,"label":"D3","name":"Deblurring and geometric correction","description":"Recovering sharp, geometrically faithful images or videos from motion blur, defocus blur, spatially varying blur, video blur, turbulence-induced warping, refractive distortion, fisheye distortion, or related deformation.","inclusion":"Assign atoms here when the primary objective is blur removal, sharp-frame recovery, blur-map/deconvolution-centered restoration, video deblurring, event- or multi-frame-assisted deblurring, atmospheric turbulence mitigation, or correction of geometric distortion and refractive warping.","exclusion":"If scattering or visibility loss without dominant warping is the main degradation, assign to environmental visibility restoration; if explicit plug-and-play or variational solver design is the main contribution rather than deblurring itself, assign to inverse-problem solvers; if face-specific priors dominate, assign to semantic completion and face restoration.","nearest_boundary":"Environmental visibility restoration","representative_topics":["multi-scale frequency separation network for image deblurring via contrastive and cycle-consistency learning","non-blind spatially varying defocus blur removal using dual encoder-decoder networks","single-image defocus deblurring with spatially dynamic modulation guided by defocus map estimation"]},{"direction_index":4,"label":"D4","name":"Environmental visibility restoration","description":"Restoring scene visibility, color, and contrast under environmental media or weather, including atmospheric haze/fog/dust/sand, rain, snow, mixed adverse weather, and underwater absorption/scattering.","inclusion":"Assign atoms here when the primary objective is dehazing, defogging, dust or sandstorm removal, deraining, desnowing, joint weather removal, nighttime dehazing when scattering dominates, or underwater enhancement/restoration driven by water absorption, scattering, color cast, turbidity, marine particles, or underwater optics.","exclusion":"If the main issue is global underexposure rather than scattering or weather particles, assign to illumination and artifact correction; if the main issue is geometric turbulence, assign to deblurring and geometric correction; if the main objective is resolution enhancement of underwater or weather imagery, assign to this direction only when the medium/weather physics drives the research objective, otherwise to resolution and fusion reconstruction.","nearest_boundary":"Illumination and artifact correction","representative_topics":["haar wavelet reinforcement fusion for underwater image enhancement","self-supervised underwater image enhancement via degradation estimation and attention-guided high-intensity correction","transform-domain generative adversarial network for frequency-decoupled image dehazing"]},{"direction_index":5,"label":"D5","name":"Semantic completion and face restoration","description":"Restoring semantically structured content where high-level priors, identity, object structure, or plausible synthesis are central, including image inpainting, object removal, masked-region completion, and face-specific restoration.","inclusion":"Assign atoms here when the primary objective is filling missing or masked regions, removing objects by synthesizing plausible content, improving inpainted-region structure/texture consistency, or restoring faces using face-specific priors such as identity, parsing, codebooks, dictionaries, geometry, reference faces, or facial structure.","exclusion":"If a named overlay artifact is removed without substantial missing-region synthesis, assign to illumination and artifact correction; if a general SR/deblurring/denoising method is merely evaluated on faces without face-specific priors, assign to the corresponding task program; if GAN or diffusion is used chiefly as a general cross-task restoration framework, assign to unified adaptive restoration frameworks.","nearest_boundary":"Illumination and artifact correction","representative_topics":["wavelet prior attention and axial-transformer learning for artifact-free image inpainting","feature refinement for high-resolution neural image inpainting via multiscale consistency optimization","structure and texture joint recovery in deep image inpainting via hybrid gan and fourier convolutional design"]},{"direction_index":6,"label":"D6","name":"Unified adaptive restoration frameworks","description":"Developing broadly applicable restoration systems, priors, architectures, and training/deployment strategies intended to work across multiple degradations, tasks, or restoration settings rather than optimize one named degradation.","inclusion":"Assign atoms here when the primary objective is all-in-one or blind multi-degradation restoration, prompt/adaptive/unified restoration, cross-task backbone design, general restoration architecture or training strategy, model compression for restoration, degradation representation learning, or general generative-prior restoration using diffusion, score-based models, GANs, flows, or generative inversion across restoration tasks.","exclusion":"If a method-specific architecture or generative model is used mainly for one named task such as low-light, deblurring, inpainting, face restoration, or SR, assign to that task program; if explicit data-consistency optimization or plug-and-play solver structure is the main contribution, assign to inverse-problem solvers.","nearest_boundary":"Inverse-problem solvers","representative_topics":["multi-branch compound feature fusion for generalizable multi-degradation image restoration","hypernetwork-based adaptive image restoration with fixed-size single model","variational bayesian framework for blind image restoration with diverse unknown degradations"]},{"direction_index":7,"label":"D7","name":"Inverse-problem solvers","description":"Formulating and solving restoration as an explicit inverse problem with forward operators, data consistency, mathematical regularization, plug-and-play priors, proximal or unrolled optimization, posterior sampling, convergence analysis, or operator-faithful reconstruction.","inclusion":"Assign atoms here when the primary contribution is a restoration solver or formulation such as plug-and-play ADMM, RED, proximal algorithms, variational regularization, matrix completion, explicit deconvolution, compressed sensing, posterior sampling, algorithm unrolling, convergence/robustness analysis, or learned priors embedded in a data-consistent inverse-problem pipeline.","exclusion":"If the paper mainly proposes a task-specific network for deblurring, denoising, SR, or inpainting without emphasizing the inverse-problem solver, assign to that task program; if the work is a broad all-in-one architecture without explicit operator fidelity, assign to unified adaptive restoration frameworks.","nearest_boundary":"Unified adaptive restoration frameworks","representative_topics":["framelet local minimal prior for poissonian blurred image deconvolution with fractional operators","deep reinforcement learning denoisers for robust plug-and-play image restoration under model mismatch","variational joint image recovery using spatially-varying generalised gaussian priors"]}],"split":"dev"}
{"episode_id":"F037_T2024-01","field_index":37,"field":"anomaly detection","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"General normality modeling","description":"Develop general-purpose anomaly detectors by learning normality, density, boundaries, reconstruction, feature discrepancies, or anomaly-sensitive representations independent of a dominant application domain.","inclusion":"Assign atoms whose primary objective is a broadly reusable detection mechanism: reconstruction or generative scoring, diffusion/score modeling used as the detector, normalizing-flow or density scoring, one-class or isolation boundaries, self-supervised/contrastive representations, or teacher-student feature discrepancy.","exclusion":"If the same mechanisms are primarily used to solve a domain program such as medical imaging, industrial inspection, video events, logs, cyber systems, or temporal dependency modeling, assign to that domain/objective instead.","nearest_boundary":"Low-label and OOD generalization","representative_topics":["proximal error loss for spatially-aware anomaly detection and semantic feature learning","deep isolation forest with non-linear subspace partitioning for hard anomaly detection","locally adaptive autoencoder-based anomaly scoring via latent space reconstruction error analysis"]},{"direction_index":1,"label":"D1","name":"Low-label and OOD generalization","description":"Detect anomalies, outliers, or distribution shifts when anomaly examples, labels, target-category data, or task-specific training data are scarce or absent, including OOD detection around trained models and foundation-model transfer.","inclusion":"Assign atoms whose primary objective is few-shot, zero-shot, open-set, label-budgeted, prompt/foundation-model transfer, post-hoc OOD detection for trained predictors, or synthetic/outlier-exposure training for generalization under missing anomaly evidence.","exclusion":"If synthetic anomalies or few-shot settings are secondary to a specific domain such as industrial visual inspection or video surveillance, assign to that domain program unless scarce-label transfer is the central contribution.","nearest_boundary":"General normality modeling","representative_topics":["out-of-distribution detection via bidirectional gan latent space and minimum description length","unifying classifier inconsistency and autoencoder reconstruction error for out-of-distribution detection","registration-based few-shot anomaly detection with cross-category generalizable single model"]},{"direction_index":2,"label":"D2","name":"Video event anomaly detection","description":"Detect, localize, or recognize anomalous events, behaviors, trajectories, poses, or motion patterns in video streams.","inclusion":"Assign atoms whose primary objective is video anomaly localization or recognition, including weakly supervised surveillance, online video event scoring, trajectory prediction, crowd behavior analysis, appearance-motion modeling, and skeleton/pose-based human behavior anomaly detection.","exclusion":"Static image defect localization belongs to industrial inspection; medical image sequences belong to medical imaging when clinical pathology is primary; generic temporal numeric series belong to temporal and graph-structured dynamics.","nearest_boundary":"Industrial inspection and condition monitoring","representative_topics":["weakly supervised surveillance video anomaly detection using transformer and dilated convolution temporal features","multi-contextual transformer prediction for spatio-temporal video anomaly detection","trajectory prediction error-based unsupervised pedestrian anomaly detection in video"]},{"direction_index":3,"label":"D3","name":"Medical image anomaly localization","description":"Detect, segment, or localize abnormal anatomy, lesions, pathology, or clinically meaningful image abnormalities in medical images or medical image sequences.","inclusion":"Assign atoms whose primary objective is medical-imaging anomaly detection or segmentation in MRI, CT, X-ray, echocardiography, fundus, histopathology, blood-cell, or related clinical images, regardless of whether the mechanism is reconstruction, GAN, diffusion, transformer, or feature adaptation.","exclusion":"General-purpose generative or reconstruction detectors evaluated incidentally on medical data belong to general normality modeling; non-imaging physiological signals belong to temporal dynamics unless the clinical imaging objective dominates.","nearest_boundary":"General normality modeling","representative_topics":["diffusion model-based unsupervised brain anomaly detection and segmentation in medical imaging","progressive gan-based anomaly detection for medical images with limited labeled data","unsupervised 3d lung anomaly detection via virtual multi-view projection and 3d map restoration"]},{"direction_index":4,"label":"D4","name":"Industrial inspection and condition monitoring","description":"Detect product defects, surface flaws, manufacturing anomalies, anomalous sounds, vibration faults, and machine-condition abnormalities for industrial quality control and predictive maintenance.","inclusion":"Assign atoms whose primary objective is industrial visual inspection, surface/product defect localization, manufacturing quality monitoring, anomalous machine sound detection, vibration/acoustic fault detection, or sensor-based condition monitoring.","exclusion":"General image reconstruction or representation methods belong to general normality modeling when industrial benchmarks are only an evaluation venue; cyber-physical infrastructure security belongs to operational security and systems monitoring when attack or infrastructure safety is primary.","nearest_boundary":"Video event anomaly detection","representative_topics":["unified multi-class anomaly detection via query-aware reconstruction and masked attention","hierarchical conditional variational autoencoder for cross-domain acoustic anomaly detection in industrial monitoring","unsupervised anomalous sound detection with domain generalization under implicit domain shifts"]},{"direction_index":5,"label":"D5","name":"Temporal and graph-structured dynamics","description":"Detect anomalies by modeling temporal dependencies, cross-variable relations, streaming dynamics, or graph topology when the structure itself is the main scientific object.","inclusion":"Assign atoms whose primary objective is multivariate or univariate time-series anomaly detection, temporal dependency modeling, concept drift or online stream adaptation, continual anomaly detection, dynamic dependency graphs from sequences, or anomalous nodes, edges, subgraphs, and whole graphs in inherently graph-structured data.","exclusion":"Graph or temporal methods aimed mainly at logs, cyber networks, finance, industrial condition monitoring, or medical/video applications belong to those operational programs when the domain objective dominates.","nearest_boundary":"Operational security and systems monitoring","representative_topics":["adaptive model pooling for online deep anomaly detection under concept drift in evolving streams","calibrated nonparametric scan statistics for anomalous subgraph detection in large-scale graphs","intrinsic anomaly detection via adversarial and representation learning for multi-variate time series"]},{"direction_index":6,"label":"D6","name":"Operational security and systems monitoring","description":"Detect anomalous, malicious, faulty, fraudulent, or unsafe behavior in software systems, logs, networks, cyber-physical systems, financial operations, infrastructure, and distributed deployment settings.","inclusion":"Assign atoms whose primary objective is operational monitoring or protection: system logs and traces, cloud or HPC telemetry, network intrusion, NetFlow, botnets, CAN bus or IoT/CPS anomalies, critical infrastructure faults, financial or transaction fraud, federated/privacy-preserving deployment, edge/resource-constrained operation, or distributed anomaly detection.","exclusion":"Generic graph or time-series algorithms belong to temporal and graph-structured dynamics unless the operational security, log, financial, infrastructure, or deployment objective is primary; interpretability-only contributions belong to evaluation and trustworthy practice when explanation or accountability dominates.","nearest_boundary":"Temporal and graph-structured dynamics","representative_topics":["ensemble bagging of anomaly detection algorithms for early sewage pipe network fault detection","lightweight anomaly detection on microcontrollers using reservoir sampling and lof","log instruction-guided anomaly detection using source code-derived instructions for it operations"]},{"direction_index":7,"label":"D7","name":"Evaluation and trustworthy practice","description":"Build the methodological infrastructure and trust properties needed to evaluate, select, explain, audit, calibrate, or govern anomaly detectors.","inclusion":"Assign atoms whose primary objective is benchmarking, dataset or metric construction, surveys, reproducibility, hyperparameter/model selection, theoretical limits, interpretability, recourse, fairness, accountability, robustness to contamination or shift, active learning for label efficiency, or human-centered requirements for anomaly detection.","exclusion":"A new detector with routine experiments belongs to its substantive detector program; domain-specific explainability belongs here only when explanation, fairness, accountability, or evaluation methodology is the central scientific objective.","nearest_boundary":"General normality modeling","representative_topics":["scalable hyper-ensemble methods for robust deep outlier detection under hyperparameter sensitivity","benchmarking anomaly detection algorithms across supervision regimes and anomaly types","benchmarking unsupervised outlier node detection methods on static attributed graphs"]}],"split":"dev"}
{"episode_id":"F037_T2024-07","field_index":37,"field":"anomaly detection","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"General normality modeling","description":"Develop general-purpose anomaly detectors by learning normality, density, boundaries, reconstruction, feature discrepancies, or anomaly-sensitive representations independent of a dominant application domain.","inclusion":"Assign atoms whose primary objective is a broadly reusable detection mechanism: reconstruction or generative scoring, diffusion/score modeling used as the detector, normalizing-flow or density scoring, one-class or isolation boundaries, self-supervised/contrastive representations, or teacher-student feature discrepancy.","exclusion":"If the same mechanisms are primarily used to solve a domain program such as medical imaging, industrial inspection, video events, logs, cyber systems, or temporal dependency modeling, assign to that domain/objective instead.","nearest_boundary":"Low-label and OOD generalization","representative_topics":["proximal error loss for spatially-aware anomaly detection and semantic feature learning","deep isolation forest with non-linear subspace partitioning for hard anomaly detection","locally adaptive autoencoder-based anomaly scoring via latent space reconstruction error analysis"]},{"direction_index":1,"label":"D1","name":"Low-label and OOD generalization","description":"Detect anomalies, outliers, or distribution shifts when anomaly examples, labels, target-category data, or task-specific training data are scarce or absent, including OOD detection around trained models and foundation-model transfer.","inclusion":"Assign atoms whose primary objective is few-shot, zero-shot, open-set, label-budgeted, prompt/foundation-model transfer, post-hoc OOD detection for trained predictors, or synthetic/outlier-exposure training for generalization under missing anomaly evidence.","exclusion":"If synthetic anomalies or few-shot settings are secondary to a specific domain such as industrial visual inspection or video surveillance, assign to that domain program unless scarce-label transfer is the central contribution.","nearest_boundary":"General normality modeling","representative_topics":["out-of-distribution detection via bidirectional gan latent space and minimum description length","unifying classifier inconsistency and autoencoder reconstruction error for out-of-distribution detection","registration-based few-shot anomaly detection with cross-category generalizable single model"]},{"direction_index":2,"label":"D2","name":"Video event anomaly detection","description":"Detect, localize, or recognize anomalous events, behaviors, trajectories, poses, or motion patterns in video streams.","inclusion":"Assign atoms whose primary objective is video anomaly localization or recognition, including weakly supervised surveillance, online video event scoring, trajectory prediction, crowd behavior analysis, appearance-motion modeling, and skeleton/pose-based human behavior anomaly detection.","exclusion":"Static image defect localization belongs to industrial inspection; medical image sequences belong to medical imaging when clinical pathology is primary; generic temporal numeric series belong to temporal and graph-structured dynamics.","nearest_boundary":"Industrial inspection and condition monitoring","representative_topics":["weakly supervised surveillance video anomaly detection using transformer and dilated convolution temporal features","multi-contextual transformer prediction for spatio-temporal video anomaly detection","trajectory prediction error-based unsupervised pedestrian anomaly detection in video"]},{"direction_index":3,"label":"D3","name":"Medical image anomaly localization","description":"Detect, segment, or localize abnormal anatomy, lesions, pathology, or clinically meaningful image abnormalities in medical images or medical image sequences.","inclusion":"Assign atoms whose primary objective is medical-imaging anomaly detection or segmentation in MRI, CT, X-ray, echocardiography, fundus, histopathology, blood-cell, or related clinical images, regardless of whether the mechanism is reconstruction, GAN, diffusion, transformer, or feature adaptation.","exclusion":"General-purpose generative or reconstruction detectors evaluated incidentally on medical data belong to general normality modeling; non-imaging physiological signals belong to temporal dynamics unless the clinical imaging objective dominates.","nearest_boundary":"General normality modeling","representative_topics":["diffusion model-based unsupervised brain anomaly detection and segmentation in medical imaging","progressive gan-based anomaly detection for medical images with limited labeled data","unsupervised 3d lung anomaly detection via virtual multi-view projection and 3d map restoration"]},{"direction_index":4,"label":"D4","name":"Industrial inspection and condition monitoring","description":"Detect product defects, surface flaws, manufacturing anomalies, anomalous sounds, vibration faults, and machine-condition abnormalities for industrial quality control and predictive maintenance.","inclusion":"Assign atoms whose primary objective is industrial visual inspection, surface/product defect localization, manufacturing quality monitoring, anomalous machine sound detection, vibration/acoustic fault detection, or sensor-based condition monitoring.","exclusion":"General image reconstruction or representation methods belong to general normality modeling when industrial benchmarks are only an evaluation venue; cyber-physical infrastructure security belongs to operational security and systems monitoring when attack or infrastructure safety is primary.","nearest_boundary":"Video event anomaly detection","representative_topics":["unified multi-class anomaly detection via query-aware reconstruction and masked attention","hierarchical conditional variational autoencoder for cross-domain acoustic anomaly detection in industrial monitoring","unsupervised anomalous sound detection with domain generalization under implicit domain shifts"]},{"direction_index":5,"label":"D5","name":"Temporal and graph-structured dynamics","description":"Detect anomalies by modeling temporal dependencies, cross-variable relations, streaming dynamics, or graph topology when the structure itself is the main scientific object.","inclusion":"Assign atoms whose primary objective is multivariate or univariate time-series anomaly detection, temporal dependency modeling, concept drift or online stream adaptation, continual anomaly detection, dynamic dependency graphs from sequences, or anomalous nodes, edges, subgraphs, and whole graphs in inherently graph-structured data.","exclusion":"Graph or temporal methods aimed mainly at logs, cyber networks, finance, industrial condition monitoring, or medical/video applications belong to those operational programs when the domain objective dominates.","nearest_boundary":"Operational security and systems monitoring","representative_topics":["adaptive model pooling for online deep anomaly detection under concept drift in evolving streams","calibrated nonparametric scan statistics for anomalous subgraph detection in large-scale graphs","intrinsic anomaly detection via adversarial and representation learning for multi-variate time series"]},{"direction_index":6,"label":"D6","name":"Operational security and systems monitoring","description":"Detect anomalous, malicious, faulty, fraudulent, or unsafe behavior in software systems, logs, networks, cyber-physical systems, financial operations, infrastructure, and distributed deployment settings.","inclusion":"Assign atoms whose primary objective is operational monitoring or protection: system logs and traces, cloud or HPC telemetry, network intrusion, NetFlow, botnets, CAN bus or IoT/CPS anomalies, critical infrastructure faults, financial or transaction fraud, federated/privacy-preserving deployment, edge/resource-constrained operation, or distributed anomaly detection.","exclusion":"Generic graph or time-series algorithms belong to temporal and graph-structured dynamics unless the operational security, log, financial, infrastructure, or deployment objective is primary; interpretability-only contributions belong to evaluation and trustworthy practice when explanation or accountability dominates.","nearest_boundary":"Temporal and graph-structured dynamics","representative_topics":["ensemble bagging of anomaly detection algorithms for early sewage pipe network fault detection","lightweight anomaly detection on microcontrollers using reservoir sampling and lof","log instruction-guided anomaly detection using source code-derived instructions for it operations"]},{"direction_index":7,"label":"D7","name":"Evaluation and trustworthy practice","description":"Build the methodological infrastructure and trust properties needed to evaluate, select, explain, audit, calibrate, or govern anomaly detectors.","inclusion":"Assign atoms whose primary objective is benchmarking, dataset or metric construction, surveys, reproducibility, hyperparameter/model selection, theoretical limits, interpretability, recourse, fairness, accountability, robustness to contamination or shift, active learning for label efficiency, or human-centered requirements for anomaly detection.","exclusion":"A new detector with routine experiments belongs to its substantive detector program; domain-specific explainability belongs here only when explanation, fairness, accountability, or evaluation methodology is the central scientific objective.","nearest_boundary":"General normality modeling","representative_topics":["scalable hyper-ensemble methods for robust deep outlier detection under hyperparameter sensitivity","benchmarking anomaly detection algorithms across supervision regimes and anomaly types","benchmarking unsupervised outlier node detection methods on static attributed graphs"]}],"split":"dev"}
{"episode_id":"F037_T2025-01","field_index":37,"field":"anomaly detection","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"General normality modeling","description":"Develop general-purpose anomaly detectors by learning normality, density, boundaries, reconstruction, feature discrepancies, or anomaly-sensitive representations independent of a dominant application domain.","inclusion":"Assign atoms whose primary objective is a broadly reusable detection mechanism: reconstruction or generative scoring, diffusion/score modeling used as the detector, normalizing-flow or density scoring, one-class or isolation boundaries, self-supervised/contrastive representations, or teacher-student feature discrepancy.","exclusion":"If the same mechanisms are primarily used to solve a domain program such as medical imaging, industrial inspection, video events, logs, cyber systems, or temporal dependency modeling, assign to that domain/objective instead.","nearest_boundary":"Low-label and OOD generalization","representative_topics":["proximal error loss for spatially-aware anomaly detection and semantic feature learning","deep isolation forest with non-linear subspace partitioning for hard anomaly detection","locally adaptive autoencoder-based anomaly scoring via latent space reconstruction error analysis"]},{"direction_index":1,"label":"D1","name":"Low-label and OOD generalization","description":"Detect anomalies, outliers, or distribution shifts when anomaly examples, labels, target-category data, or task-specific training data are scarce or absent, including OOD detection around trained models and foundation-model transfer.","inclusion":"Assign atoms whose primary objective is few-shot, zero-shot, open-set, label-budgeted, prompt/foundation-model transfer, post-hoc OOD detection for trained predictors, or synthetic/outlier-exposure training for generalization under missing anomaly evidence.","exclusion":"If synthetic anomalies or few-shot settings are secondary to a specific domain such as industrial visual inspection or video surveillance, assign to that domain program unless scarce-label transfer is the central contribution.","nearest_boundary":"General normality modeling","representative_topics":["out-of-distribution detection via bidirectional gan latent space and minimum description length","unifying classifier inconsistency and autoencoder reconstruction error for out-of-distribution detection","registration-based few-shot anomaly detection with cross-category generalizable single model"]},{"direction_index":2,"label":"D2","name":"Video event anomaly detection","description":"Detect, localize, or recognize anomalous events, behaviors, trajectories, poses, or motion patterns in video streams.","inclusion":"Assign atoms whose primary objective is video anomaly localization or recognition, including weakly supervised surveillance, online video event scoring, trajectory prediction, crowd behavior analysis, appearance-motion modeling, and skeleton/pose-based human behavior anomaly detection.","exclusion":"Static image defect localization belongs to industrial inspection; medical image sequences belong to medical imaging when clinical pathology is primary; generic temporal numeric series belong to temporal and graph-structured dynamics.","nearest_boundary":"Industrial inspection and condition monitoring","representative_topics":["weakly supervised surveillance video anomaly detection using transformer and dilated convolution temporal features","multi-contextual transformer prediction for spatio-temporal video anomaly detection","trajectory prediction error-based unsupervised pedestrian anomaly detection in video"]},{"direction_index":3,"label":"D3","name":"Medical image anomaly localization","description":"Detect, segment, or localize abnormal anatomy, lesions, pathology, or clinically meaningful image abnormalities in medical images or medical image sequences.","inclusion":"Assign atoms whose primary objective is medical-imaging anomaly detection or segmentation in MRI, CT, X-ray, echocardiography, fundus, histopathology, blood-cell, or related clinical images, regardless of whether the mechanism is reconstruction, GAN, diffusion, transformer, or feature adaptation.","exclusion":"General-purpose generative or reconstruction detectors evaluated incidentally on medical data belong to general normality modeling; non-imaging physiological signals belong to temporal dynamics unless the clinical imaging objective dominates.","nearest_boundary":"General normality modeling","representative_topics":["diffusion model-based unsupervised brain anomaly detection and segmentation in medical imaging","progressive gan-based anomaly detection for medical images with limited labeled data","unsupervised 3d lung anomaly detection via virtual multi-view projection and 3d map restoration"]},{"direction_index":4,"label":"D4","name":"Industrial inspection and condition monitoring","description":"Detect product defects, surface flaws, manufacturing anomalies, anomalous sounds, vibration faults, and machine-condition abnormalities for industrial quality control and predictive maintenance.","inclusion":"Assign atoms whose primary objective is industrial visual inspection, surface/product defect localization, manufacturing quality monitoring, anomalous machine sound detection, vibration/acoustic fault detection, or sensor-based condition monitoring.","exclusion":"General image reconstruction or representation methods belong to general normality modeling when industrial benchmarks are only an evaluation venue; cyber-physical infrastructure security belongs to operational security and systems monitoring when attack or infrastructure safety is primary.","nearest_boundary":"Video event anomaly detection","representative_topics":["unified multi-class anomaly detection via query-aware reconstruction and masked attention","hierarchical conditional variational autoencoder for cross-domain acoustic anomaly detection in industrial monitoring","unsupervised anomalous sound detection with domain generalization under implicit domain shifts"]},{"direction_index":5,"label":"D5","name":"Temporal and graph-structured dynamics","description":"Detect anomalies by modeling temporal dependencies, cross-variable relations, streaming dynamics, or graph topology when the structure itself is the main scientific object.","inclusion":"Assign atoms whose primary objective is multivariate or univariate time-series anomaly detection, temporal dependency modeling, concept drift or online stream adaptation, continual anomaly detection, dynamic dependency graphs from sequences, or anomalous nodes, edges, subgraphs, and whole graphs in inherently graph-structured data.","exclusion":"Graph or temporal methods aimed mainly at logs, cyber networks, finance, industrial condition monitoring, or medical/video applications belong to those operational programs when the domain objective dominates.","nearest_boundary":"Operational security and systems monitoring","representative_topics":["adaptive model pooling for online deep anomaly detection under concept drift in evolving streams","calibrated nonparametric scan statistics for anomalous subgraph detection in large-scale graphs","intrinsic anomaly detection via adversarial and representation learning for multi-variate time series"]},{"direction_index":6,"label":"D6","name":"Operational security and systems monitoring","description":"Detect anomalous, malicious, faulty, fraudulent, or unsafe behavior in software systems, logs, networks, cyber-physical systems, financial operations, infrastructure, and distributed deployment settings.","inclusion":"Assign atoms whose primary objective is operational monitoring or protection: system logs and traces, cloud or HPC telemetry, network intrusion, NetFlow, botnets, CAN bus or IoT/CPS anomalies, critical infrastructure faults, financial or transaction fraud, federated/privacy-preserving deployment, edge/resource-constrained operation, or distributed anomaly detection.","exclusion":"Generic graph or time-series algorithms belong to temporal and graph-structured dynamics unless the operational security, log, financial, infrastructure, or deployment objective is primary; interpretability-only contributions belong to evaluation and trustworthy practice when explanation or accountability dominates.","nearest_boundary":"Temporal and graph-structured dynamics","representative_topics":["ensemble bagging of anomaly detection algorithms for early sewage pipe network fault detection","lightweight anomaly detection on microcontrollers using reservoir sampling and lof","log instruction-guided anomaly detection using source code-derived instructions for it operations"]},{"direction_index":7,"label":"D7","name":"Evaluation and trustworthy practice","description":"Build the methodological infrastructure and trust properties needed to evaluate, select, explain, audit, calibrate, or govern anomaly detectors.","inclusion":"Assign atoms whose primary objective is benchmarking, dataset or metric construction, surveys, reproducibility, hyperparameter/model selection, theoretical limits, interpretability, recourse, fairness, accountability, robustness to contamination or shift, active learning for label efficiency, or human-centered requirements for anomaly detection.","exclusion":"A new detector with routine experiments belongs to its substantive detector program; domain-specific explainability belongs here only when explanation, fairness, accountability, or evaluation methodology is the central scientific objective.","nearest_boundary":"General normality modeling","representative_topics":["scalable hyper-ensemble methods for robust deep outlier detection under hyperparameter sensitivity","benchmarking anomaly detection algorithms across supervision regimes and anomaly types","benchmarking unsupervised outlier node detection methods on static attributed graphs"]}],"split":"dev"}
{"episode_id":"F037_T2025-07","field_index":37,"field":"anomaly detection","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"General normality modeling","description":"Develop general-purpose anomaly detectors by learning normality, density, boundaries, reconstruction, feature discrepancies, or anomaly-sensitive representations independent of a dominant application domain.","inclusion":"Assign atoms whose primary objective is a broadly reusable detection mechanism: reconstruction or generative scoring, diffusion/score modeling used as the detector, normalizing-flow or density scoring, one-class or isolation boundaries, self-supervised/contrastive representations, or teacher-student feature discrepancy.","exclusion":"If the same mechanisms are primarily used to solve a domain program such as medical imaging, industrial inspection, video events, logs, cyber systems, or temporal dependency modeling, assign to that domain/objective instead.","nearest_boundary":"Low-label and OOD generalization","representative_topics":["proximal error loss for spatially-aware anomaly detection and semantic feature learning","deep isolation forest with non-linear subspace partitioning for hard anomaly detection","locally adaptive autoencoder-based anomaly scoring via latent space reconstruction error analysis"]},{"direction_index":1,"label":"D1","name":"Low-label and OOD generalization","description":"Detect anomalies, outliers, or distribution shifts when anomaly examples, labels, target-category data, or task-specific training data are scarce or absent, including OOD detection around trained models and foundation-model transfer.","inclusion":"Assign atoms whose primary objective is few-shot, zero-shot, open-set, label-budgeted, prompt/foundation-model transfer, post-hoc OOD detection for trained predictors, or synthetic/outlier-exposure training for generalization under missing anomaly evidence.","exclusion":"If synthetic anomalies or few-shot settings are secondary to a specific domain such as industrial visual inspection or video surveillance, assign to that domain program unless scarce-label transfer is the central contribution.","nearest_boundary":"General normality modeling","representative_topics":["out-of-distribution detection via bidirectional gan latent space and minimum description length","unifying classifier inconsistency and autoencoder reconstruction error for out-of-distribution detection","registration-based few-shot anomaly detection with cross-category generalizable single model"]},{"direction_index":2,"label":"D2","name":"Video event anomaly detection","description":"Detect, localize, or recognize anomalous events, behaviors, trajectories, poses, or motion patterns in video streams.","inclusion":"Assign atoms whose primary objective is video anomaly localization or recognition, including weakly supervised surveillance, online video event scoring, trajectory prediction, crowd behavior analysis, appearance-motion modeling, and skeleton/pose-based human behavior anomaly detection.","exclusion":"Static image defect localization belongs to industrial inspection; medical image sequences belong to medical imaging when clinical pathology is primary; generic temporal numeric series belong to temporal and graph-structured dynamics.","nearest_boundary":"Industrial inspection and condition monitoring","representative_topics":["weakly supervised surveillance video anomaly detection using transformer and dilated convolution temporal features","multi-contextual transformer prediction for spatio-temporal video anomaly detection","trajectory prediction error-based unsupervised pedestrian anomaly detection in video"]},{"direction_index":3,"label":"D3","name":"Medical image anomaly localization","description":"Detect, segment, or localize abnormal anatomy, lesions, pathology, or clinically meaningful image abnormalities in medical images or medical image sequences.","inclusion":"Assign atoms whose primary objective is medical-imaging anomaly detection or segmentation in MRI, CT, X-ray, echocardiography, fundus, histopathology, blood-cell, or related clinical images, regardless of whether the mechanism is reconstruction, GAN, diffusion, transformer, or feature adaptation.","exclusion":"General-purpose generative or reconstruction detectors evaluated incidentally on medical data belong to general normality modeling; non-imaging physiological signals belong to temporal dynamics unless the clinical imaging objective dominates.","nearest_boundary":"General normality modeling","representative_topics":["diffusion model-based unsupervised brain anomaly detection and segmentation in medical imaging","progressive gan-based anomaly detection for medical images with limited labeled data","unsupervised 3d lung anomaly detection via virtual multi-view projection and 3d map restoration"]},{"direction_index":4,"label":"D4","name":"Industrial inspection and condition monitoring","description":"Detect product defects, surface flaws, manufacturing anomalies, anomalous sounds, vibration faults, and machine-condition abnormalities for industrial quality control and predictive maintenance.","inclusion":"Assign atoms whose primary objective is industrial visual inspection, surface/product defect localization, manufacturing quality monitoring, anomalous machine sound detection, vibration/acoustic fault detection, or sensor-based condition monitoring.","exclusion":"General image reconstruction or representation methods belong to general normality modeling when industrial benchmarks are only an evaluation venue; cyber-physical infrastructure security belongs to operational security and systems monitoring when attack or infrastructure safety is primary.","nearest_boundary":"Video event anomaly detection","representative_topics":["unified multi-class anomaly detection via query-aware reconstruction and masked attention","hierarchical conditional variational autoencoder for cross-domain acoustic anomaly detection in industrial monitoring","unsupervised anomalous sound detection with domain generalization under implicit domain shifts"]},{"direction_index":5,"label":"D5","name":"Temporal and graph-structured dynamics","description":"Detect anomalies by modeling temporal dependencies, cross-variable relations, streaming dynamics, or graph topology when the structure itself is the main scientific object.","inclusion":"Assign atoms whose primary objective is multivariate or univariate time-series anomaly detection, temporal dependency modeling, concept drift or online stream adaptation, continual anomaly detection, dynamic dependency graphs from sequences, or anomalous nodes, edges, subgraphs, and whole graphs in inherently graph-structured data.","exclusion":"Graph or temporal methods aimed mainly at logs, cyber networks, finance, industrial condition monitoring, or medical/video applications belong to those operational programs when the domain objective dominates.","nearest_boundary":"Operational security and systems monitoring","representative_topics":["adaptive model pooling for online deep anomaly detection under concept drift in evolving streams","calibrated nonparametric scan statistics for anomalous subgraph detection in large-scale graphs","intrinsic anomaly detection via adversarial and representation learning for multi-variate time series"]},{"direction_index":6,"label":"D6","name":"Operational security and systems monitoring","description":"Detect anomalous, malicious, faulty, fraudulent, or unsafe behavior in software systems, logs, networks, cyber-physical systems, financial operations, infrastructure, and distributed deployment settings.","inclusion":"Assign atoms whose primary objective is operational monitoring or protection: system logs and traces, cloud or HPC telemetry, network intrusion, NetFlow, botnets, CAN bus or IoT/CPS anomalies, critical infrastructure faults, financial or transaction fraud, federated/privacy-preserving deployment, edge/resource-constrained operation, or distributed anomaly detection.","exclusion":"Generic graph or time-series algorithms belong to temporal and graph-structured dynamics unless the operational security, log, financial, infrastructure, or deployment objective is primary; interpretability-only contributions belong to evaluation and trustworthy practice when explanation or accountability dominates.","nearest_boundary":"Temporal and graph-structured dynamics","representative_topics":["ensemble bagging of anomaly detection algorithms for early sewage pipe network fault detection","lightweight anomaly detection on microcontrollers using reservoir sampling and lof","log instruction-guided anomaly detection using source code-derived instructions for it operations"]},{"direction_index":7,"label":"D7","name":"Evaluation and trustworthy practice","description":"Build the methodological infrastructure and trust properties needed to evaluate, select, explain, audit, calibrate, or govern anomaly detectors.","inclusion":"Assign atoms whose primary objective is benchmarking, dataset or metric construction, surveys, reproducibility, hyperparameter/model selection, theoretical limits, interpretability, recourse, fairness, accountability, robustness to contamination or shift, active learning for label efficiency, or human-centered requirements for anomaly detection.","exclusion":"A new detector with routine experiments belongs to its substantive detector program; domain-specific explainability belongs here only when explanation, fairness, accountability, or evaluation methodology is the central scientific objective.","nearest_boundary":"General normality modeling","representative_topics":["scalable hyper-ensemble methods for robust deep outlier detection under hyperparameter sensitivity","benchmarking anomaly detection algorithms across supervision regimes and anomaly types","benchmarking unsupervised outlier node detection methods on static attributed graphs"]}],"split":"dev"}
{"episode_id":"F037_T2026-01","field_index":37,"field":"anomaly detection","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"General normality modeling","description":"Develop general-purpose anomaly detectors by learning normality, density, boundaries, reconstruction, feature discrepancies, or anomaly-sensitive representations independent of a dominant application domain.","inclusion":"Assign atoms whose primary objective is a broadly reusable detection mechanism: reconstruction or generative scoring, diffusion/score modeling used as the detector, normalizing-flow or density scoring, one-class or isolation boundaries, self-supervised/contrastive representations, or teacher-student feature discrepancy.","exclusion":"If the same mechanisms are primarily used to solve a domain program such as medical imaging, industrial inspection, video events, logs, cyber systems, or temporal dependency modeling, assign to that domain/objective instead.","nearest_boundary":"Low-label and OOD generalization","representative_topics":["proximal error loss for spatially-aware anomaly detection and semantic feature learning","deep isolation forest with non-linear subspace partitioning for hard anomaly detection","locally adaptive autoencoder-based anomaly scoring via latent space reconstruction error analysis"]},{"direction_index":1,"label":"D1","name":"Low-label and OOD generalization","description":"Detect anomalies, outliers, or distribution shifts when anomaly examples, labels, target-category data, or task-specific training data are scarce or absent, including OOD detection around trained models and foundation-model transfer.","inclusion":"Assign atoms whose primary objective is few-shot, zero-shot, open-set, label-budgeted, prompt/foundation-model transfer, post-hoc OOD detection for trained predictors, or synthetic/outlier-exposure training for generalization under missing anomaly evidence.","exclusion":"If synthetic anomalies or few-shot settings are secondary to a specific domain such as industrial visual inspection or video surveillance, assign to that domain program unless scarce-label transfer is the central contribution.","nearest_boundary":"General normality modeling","representative_topics":["out-of-distribution detection via bidirectional gan latent space and minimum description length","unifying classifier inconsistency and autoencoder reconstruction error for out-of-distribution detection","registration-based few-shot anomaly detection with cross-category generalizable single model"]},{"direction_index":2,"label":"D2","name":"Video event anomaly detection","description":"Detect, localize, or recognize anomalous events, behaviors, trajectories, poses, or motion patterns in video streams.","inclusion":"Assign atoms whose primary objective is video anomaly localization or recognition, including weakly supervised surveillance, online video event scoring, trajectory prediction, crowd behavior analysis, appearance-motion modeling, and skeleton/pose-based human behavior anomaly detection.","exclusion":"Static image defect localization belongs to industrial inspection; medical image sequences belong to medical imaging when clinical pathology is primary; generic temporal numeric series belong to temporal and graph-structured dynamics.","nearest_boundary":"Industrial inspection and condition monitoring","representative_topics":["weakly supervised surveillance video anomaly detection using transformer and dilated convolution temporal features","multi-contextual transformer prediction for spatio-temporal video anomaly detection","trajectory prediction error-based unsupervised pedestrian anomaly detection in video"]},{"direction_index":3,"label":"D3","name":"Medical image anomaly localization","description":"Detect, segment, or localize abnormal anatomy, lesions, pathology, or clinically meaningful image abnormalities in medical images or medical image sequences.","inclusion":"Assign atoms whose primary objective is medical-imaging anomaly detection or segmentation in MRI, CT, X-ray, echocardiography, fundus, histopathology, blood-cell, or related clinical images, regardless of whether the mechanism is reconstruction, GAN, diffusion, transformer, or feature adaptation.","exclusion":"General-purpose generative or reconstruction detectors evaluated incidentally on medical data belong to general normality modeling; non-imaging physiological signals belong to temporal dynamics unless the clinical imaging objective dominates.","nearest_boundary":"General normality modeling","representative_topics":["diffusion model-based unsupervised brain anomaly detection and segmentation in medical imaging","progressive gan-based anomaly detection for medical images with limited labeled data","unsupervised 3d lung anomaly detection via virtual multi-view projection and 3d map restoration"]},{"direction_index":4,"label":"D4","name":"Industrial inspection and condition monitoring","description":"Detect product defects, surface flaws, manufacturing anomalies, anomalous sounds, vibration faults, and machine-condition abnormalities for industrial quality control and predictive maintenance.","inclusion":"Assign atoms whose primary objective is industrial visual inspection, surface/product defect localization, manufacturing quality monitoring, anomalous machine sound detection, vibration/acoustic fault detection, or sensor-based condition monitoring.","exclusion":"General image reconstruction or representation methods belong to general normality modeling when industrial benchmarks are only an evaluation venue; cyber-physical infrastructure security belongs to operational security and systems monitoring when attack or infrastructure safety is primary.","nearest_boundary":"Video event anomaly detection","representative_topics":["unified multi-class anomaly detection via query-aware reconstruction and masked attention","hierarchical conditional variational autoencoder for cross-domain acoustic anomaly detection in industrial monitoring","unsupervised anomalous sound detection with domain generalization under implicit domain shifts"]},{"direction_index":5,"label":"D5","name":"Temporal and graph-structured dynamics","description":"Detect anomalies by modeling temporal dependencies, cross-variable relations, streaming dynamics, or graph topology when the structure itself is the main scientific object.","inclusion":"Assign atoms whose primary objective is multivariate or univariate time-series anomaly detection, temporal dependency modeling, concept drift or online stream adaptation, continual anomaly detection, dynamic dependency graphs from sequences, or anomalous nodes, edges, subgraphs, and whole graphs in inherently graph-structured data.","exclusion":"Graph or temporal methods aimed mainly at logs, cyber networks, finance, industrial condition monitoring, or medical/video applications belong to those operational programs when the domain objective dominates.","nearest_boundary":"Operational security and systems monitoring","representative_topics":["adaptive model pooling for online deep anomaly detection under concept drift in evolving streams","calibrated nonparametric scan statistics for anomalous subgraph detection in large-scale graphs","intrinsic anomaly detection via adversarial and representation learning for multi-variate time series"]},{"direction_index":6,"label":"D6","name":"Operational security and systems monitoring","description":"Detect anomalous, malicious, faulty, fraudulent, or unsafe behavior in software systems, logs, networks, cyber-physical systems, financial operations, infrastructure, and distributed deployment settings.","inclusion":"Assign atoms whose primary objective is operational monitoring or protection: system logs and traces, cloud or HPC telemetry, network intrusion, NetFlow, botnets, CAN bus or IoT/CPS anomalies, critical infrastructure faults, financial or transaction fraud, federated/privacy-preserving deployment, edge/resource-constrained operation, or distributed anomaly detection.","exclusion":"Generic graph or time-series algorithms belong to temporal and graph-structured dynamics unless the operational security, log, financial, infrastructure, or deployment objective is primary; interpretability-only contributions belong to evaluation and trustworthy practice when explanation or accountability dominates.","nearest_boundary":"Temporal and graph-structured dynamics","representative_topics":["ensemble bagging of anomaly detection algorithms for early sewage pipe network fault detection","lightweight anomaly detection on microcontrollers using reservoir sampling and lof","log instruction-guided anomaly detection using source code-derived instructions for it operations"]},{"direction_index":7,"label":"D7","name":"Evaluation and trustworthy practice","description":"Build the methodological infrastructure and trust properties needed to evaluate, select, explain, audit, calibrate, or govern anomaly detectors.","inclusion":"Assign atoms whose primary objective is benchmarking, dataset or metric construction, surveys, reproducibility, hyperparameter/model selection, theoretical limits, interpretability, recourse, fairness, accountability, robustness to contamination or shift, active learning for label efficiency, or human-centered requirements for anomaly detection.","exclusion":"A new detector with routine experiments belongs to its substantive detector program; domain-specific explainability belongs here only when explanation, fairness, accountability, or evaluation methodology is the central scientific objective.","nearest_boundary":"General normality modeling","representative_topics":["scalable hyper-ensemble methods for robust deep outlier detection under hyperparameter sensitivity","benchmarking anomaly detection algorithms across supervision regimes and anomaly types","benchmarking unsupervised outlier node detection methods on static attributed graphs"]}],"split":"dev"}
{"episode_id":"F042_T2024-01","field_index":42,"field":"offline reinforcement learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Conservative policy improvement","description":"Model-free offline policy optimization that improves policies from fixed datasets while controlling ordinary behavior-policy shift, limited support, value overestimation, and out-of-distribution action selection, including domain adaptations whose central objective is offline policy improvement.","inclusion":"Assign atoms whose main objective is an offline policy optimizer based on pessimistic or conservative value estimation, behavior cloning or policy regularization, uncertainty penalties, support constraints, or practical deployment of such policy-improvement mechanisms in a concrete domain.","exclusion":"Use model-based offline planning when a learned dynamics or reward model is central; use assurance when the primary objective is safety, security, benchmarking, or evaluation rather than optimization.","nearest_boundary":"D1","representative_topics":["offline rl with explicit behavior cloning constraints on value-based policy learning","deep offline reinforcement learning with scalable uncertainty estimates via spi-bb policy optimization","mildly conservative q-learning for out-of-distribution generalization in offline reinforcement learning"]},{"direction_index":1,"label":"D1","name":"Model-based offline planning","description":"Offline RL that learns explicit environment models from static data and uses them for planning, imagination, counterfactual rollouts, or uncertainty-aware policy optimization.","inclusion":"Assign atoms whose core scientific objective is making learned transition, reward, causal, graph, or world models reliable for offline decision making through model uncertainty, pessimism, conservative rollouts, or model-based trajectory stitching.","exclusion":"Use conservative policy improvement when no explicit learned environment model is central; use data improvement when synthetic or transformed data is the main object rather than model-based planning.","nearest_boundary":"D0","representative_topics":["model generation with provable coverability for offline dynamics-aware policy learning","causal world models for offline reinforcement learning generalization","graph-structured world models for offline reinforcement learning in sparse-reward long-horizon environments"]},{"direction_index":2,"label":"D2","name":"Offline RL theory","description":"Theorem-driven characterization of offline RL feasibility, learnability, sample complexity, coverage requirements, structural assumptions, model selection, and minimax or instance-dependent guarantees.","inclusion":"Assign atoms whose primary objective is formal analysis of offline RL or OPE, including bounds, oracle inequalities, impossibility or identifiability conditions, distributionally robust formulations, and theory under linear, low-rank, Bellman-complete, or general function approximation structures.","exclusion":"Use the corresponding algorithmic direction when theory mainly supports a new optimizer, model-based planner, multi-agent method, or evaluation system.","nearest_boundary":"D0","representative_topics":["offline rl with trajectory-wise reward via pessimistic value iteration and reward decomposition","distributionally robust offline reinforcement learning with linear function approximation and non-asymptotic sample complexity","oracle inequalities for model selection in offline reinforcement learning with value function approximation"]},{"direction_index":3,"label":"D3","name":"Dataset utilization and generalization","description":"Research whose primary objective is to make fixed offline datasets more useful through augmentation, reweighting, relabeling, heterogeneous-data handling, representation of dataset structure, or broad algorithmic frameworks for partial coverage and generalization.","inclusion":"Assign atoms centered on transforming or exploiting offline data itself, including mixed-frequency logs, action-free trajectories, image or state augmentation, return rebalancing, counterfactual augmentation, federated or multi-source data, and general frameworks that organize offline data for improved learning without a full dynamics model as the central contribution.","exclusion":"Use model-based offline planning when an explicit learned dynamics model is the main decision-making engine; use structured policy learning when the main objective is hierarchical, sequence, diffusion, meta-task, or multi-agent policy structure.","nearest_boundary":"D7","representative_topics":["offline reinforcement learning with frequency-heterogeneous data via discretization-aware return scaling","state-conditioned image synthesis for data augmentation in offline image-based reinforcement learning","semi-supervised offline reinforcement learning via inverse dynamics proxy-labeling of action-free trajectories"]},{"direction_index":4,"label":"D4","name":"Offline-to-online transfer","description":"Using offline data to initialize, pretrain, constrain, or guide agents that subsequently learn or adapt through online interaction.","inclusion":"Assign atoms whose core objective is bridging offline training and online deployment or fine-tuning, including stable offline-to-online improvement, few-shot online adaptation, offline-pretrained exploration, hybrid offline-online sample complexity, and online learning from action-free offline pretraining.","exclusion":"Use purely offline optimization, modeling, data utilization, or structured policy learning when there is no subsequent online interaction or fine-tuning objective.","nearest_boundary":"D0","representative_topics":["offline rl pre-training for generalizable robot skill learning from few demonstrations","adaptive behavior cloning regularization for stable offline-to-online reinforcement learning fine-tuning","hybrid offline-online reinforcement learning with provably optimal sample complexity for linear mdps"]},{"direction_index":5,"label":"D5","name":"Evaluation and assurance","description":"Assessing, selecting, benchmarking, certifying, or hardening offline RL policies, algorithms, datasets, and systems before deployment.","inclusion":"Assign atoms whose primary objective is off-policy evaluation, confidence bounds, safe policy improvement, algorithm or hyperparameter selection, benchmark datasets, libraries, reproducibility protocols, robustness to perturbations, risk-aware deployment, dataset auditing, or security against poisoning and backdoors.","exclusion":"Use conservative policy improvement when uncertainty or pessimism is mainly used to train a better policy; use reward or causal inference when the main challenge is identifying the target signal or deconfounding observational data.","nearest_boundary":"D6","representative_topics":["hybrid value function estimation balancing offline data and learned model for off-policy evaluation and offline reinforcement learning","safe policy improvement via uncertainty-aware bootstrapping with provable safety guarantees","offline policy evaluation for deterministic ranking policies using position-biased inverse propensity score estimation"]},{"direction_index":6,"label":"D6","name":"Target-signal and causal inference","description":"Inferring, correcting, or identifying the reward, preference, imitation, value, or causal signal needed for offline policy evaluation or optimization when the logged data or reward labels are imperfect or confounded.","inclusion":"Assign atoms whose central objective is preference-based reward learning, offline imitation learning, imperfect reward correction, reward-free implicit value recovery, optimal-transport or discriminator-guided alignment, instrumental variables, deconfounding, propensity reasoning, sensitivity analysis, or policy learning under observed or unobserved confounding.","exclusion":"Use conservative policy improvement when demonstrations or rewards are ordinary behavior data with no separate inference problem; use evaluation and assurance when the main output is a policy value estimate or deployment certificate rather than inferred rewards or causal structure.","nearest_boundary":"D5","representative_topics":["discriminator-guided model-based offline imitation learning with suboptimal demonstration robustness","implicit multi-step contrastive value learning for offline rl without reward functions","offline preference-based reward learning via active learning queries and offline rl policy optimization"]},{"direction_index":7,"label":"D7","name":"Structured policy generalization","description":"Learning expressive or structured offline policies and planners that generalize across long horizons, modes, goals, tasks, agents, or interaction structures.","inclusion":"Assign atoms whose main objective is policy or planner structure beyond a standard single-task actor-critic, including transformer or return-conditioned sequence policies, diffusion or score-based policies and planners, hierarchical and goal-conditioned skills, latent plans, reusable primitives, offline meta-RL, task-context representation, and offline multi-agent coordination or equilibrium learning.","exclusion":"Use dataset utilization when the principal contribution is manipulating the dataset rather than the policy structure; use offline-to-online transfer when task adaptation primarily depends on later online interaction.","nearest_boundary":"D3","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","bootstrapped transformer for offline rl via self-generated trajectory data augmentation","diffusion model policy representation for offline reinforcement learning with behavior cloning and policy improvement"]}],"split":"dev"}
{"episode_id":"F042_T2024-07","field_index":42,"field":"offline reinforcement learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Conservative policy improvement","description":"Model-free offline policy optimization that improves policies from fixed datasets while controlling ordinary behavior-policy shift, limited support, value overestimation, and out-of-distribution action selection, including domain adaptations whose central objective is offline policy improvement.","inclusion":"Assign atoms whose main objective is an offline policy optimizer based on pessimistic or conservative value estimation, behavior cloning or policy regularization, uncertainty penalties, support constraints, or practical deployment of such policy-improvement mechanisms in a concrete domain.","exclusion":"Use model-based offline planning when a learned dynamics or reward model is central; use assurance when the primary objective is safety, security, benchmarking, or evaluation rather than optimization.","nearest_boundary":"D1","representative_topics":["offline rl with explicit behavior cloning constraints on value-based policy learning","deep offline reinforcement learning with scalable uncertainty estimates via spi-bb policy optimization","mildly conservative q-learning for out-of-distribution generalization in offline reinforcement learning"]},{"direction_index":1,"label":"D1","name":"Model-based offline planning","description":"Offline RL that learns explicit environment models from static data and uses them for planning, imagination, counterfactual rollouts, or uncertainty-aware policy optimization.","inclusion":"Assign atoms whose core scientific objective is making learned transition, reward, causal, graph, or world models reliable for offline decision making through model uncertainty, pessimism, conservative rollouts, or model-based trajectory stitching.","exclusion":"Use conservative policy improvement when no explicit learned environment model is central; use data improvement when synthetic or transformed data is the main object rather than model-based planning.","nearest_boundary":"D0","representative_topics":["model generation with provable coverability for offline dynamics-aware policy learning","causal world models for offline reinforcement learning generalization","graph-structured world models for offline reinforcement learning in sparse-reward long-horizon environments"]},{"direction_index":2,"label":"D2","name":"Offline RL theory","description":"Theorem-driven characterization of offline RL feasibility, learnability, sample complexity, coverage requirements, structural assumptions, model selection, and minimax or instance-dependent guarantees.","inclusion":"Assign atoms whose primary objective is formal analysis of offline RL or OPE, including bounds, oracle inequalities, impossibility or identifiability conditions, distributionally robust formulations, and theory under linear, low-rank, Bellman-complete, or general function approximation structures.","exclusion":"Use the corresponding algorithmic direction when theory mainly supports a new optimizer, model-based planner, multi-agent method, or evaluation system.","nearest_boundary":"D0","representative_topics":["offline rl with trajectory-wise reward via pessimistic value iteration and reward decomposition","distributionally robust offline reinforcement learning with linear function approximation and non-asymptotic sample complexity","oracle inequalities for model selection in offline reinforcement learning with value function approximation"]},{"direction_index":3,"label":"D3","name":"Dataset utilization and generalization","description":"Research whose primary objective is to make fixed offline datasets more useful through augmentation, reweighting, relabeling, heterogeneous-data handling, representation of dataset structure, or broad algorithmic frameworks for partial coverage and generalization.","inclusion":"Assign atoms centered on transforming or exploiting offline data itself, including mixed-frequency logs, action-free trajectories, image or state augmentation, return rebalancing, counterfactual augmentation, federated or multi-source data, and general frameworks that organize offline data for improved learning without a full dynamics model as the central contribution.","exclusion":"Use model-based offline planning when an explicit learned dynamics model is the main decision-making engine; use structured policy learning when the main objective is hierarchical, sequence, diffusion, meta-task, or multi-agent policy structure.","nearest_boundary":"D7","representative_topics":["offline reinforcement learning with frequency-heterogeneous data via discretization-aware return scaling","state-conditioned image synthesis for data augmentation in offline image-based reinforcement learning","semi-supervised offline reinforcement learning via inverse dynamics proxy-labeling of action-free trajectories"]},{"direction_index":4,"label":"D4","name":"Offline-to-online transfer","description":"Using offline data to initialize, pretrain, constrain, or guide agents that subsequently learn or adapt through online interaction.","inclusion":"Assign atoms whose core objective is bridging offline training and online deployment or fine-tuning, including stable offline-to-online improvement, few-shot online adaptation, offline-pretrained exploration, hybrid offline-online sample complexity, and online learning from action-free offline pretraining.","exclusion":"Use purely offline optimization, modeling, data utilization, or structured policy learning when there is no subsequent online interaction or fine-tuning objective.","nearest_boundary":"D0","representative_topics":["offline rl pre-training for generalizable robot skill learning from few demonstrations","adaptive behavior cloning regularization for stable offline-to-online reinforcement learning fine-tuning","hybrid offline-online reinforcement learning with provably optimal sample complexity for linear mdps"]},{"direction_index":5,"label":"D5","name":"Evaluation and assurance","description":"Assessing, selecting, benchmarking, certifying, or hardening offline RL policies, algorithms, datasets, and systems before deployment.","inclusion":"Assign atoms whose primary objective is off-policy evaluation, confidence bounds, safe policy improvement, algorithm or hyperparameter selection, benchmark datasets, libraries, reproducibility protocols, robustness to perturbations, risk-aware deployment, dataset auditing, or security against poisoning and backdoors.","exclusion":"Use conservative policy improvement when uncertainty or pessimism is mainly used to train a better policy; use reward or causal inference when the main challenge is identifying the target signal or deconfounding observational data.","nearest_boundary":"D6","representative_topics":["hybrid value function estimation balancing offline data and learned model for off-policy evaluation and offline reinforcement learning","safe policy improvement via uncertainty-aware bootstrapping with provable safety guarantees","offline policy evaluation for deterministic ranking policies using position-biased inverse propensity score estimation"]},{"direction_index":6,"label":"D6","name":"Target-signal and causal inference","description":"Inferring, correcting, or identifying the reward, preference, imitation, value, or causal signal needed for offline policy evaluation or optimization when the logged data or reward labels are imperfect or confounded.","inclusion":"Assign atoms whose central objective is preference-based reward learning, offline imitation learning, imperfect reward correction, reward-free implicit value recovery, optimal-transport or discriminator-guided alignment, instrumental variables, deconfounding, propensity reasoning, sensitivity analysis, or policy learning under observed or unobserved confounding.","exclusion":"Use conservative policy improvement when demonstrations or rewards are ordinary behavior data with no separate inference problem; use evaluation and assurance when the main output is a policy value estimate or deployment certificate rather than inferred rewards or causal structure.","nearest_boundary":"D5","representative_topics":["discriminator-guided model-based offline imitation learning with suboptimal demonstration robustness","implicit multi-step contrastive value learning for offline rl without reward functions","offline preference-based reward learning via active learning queries and offline rl policy optimization"]},{"direction_index":7,"label":"D7","name":"Structured policy generalization","description":"Learning expressive or structured offline policies and planners that generalize across long horizons, modes, goals, tasks, agents, or interaction structures.","inclusion":"Assign atoms whose main objective is policy or planner structure beyond a standard single-task actor-critic, including transformer or return-conditioned sequence policies, diffusion or score-based policies and planners, hierarchical and goal-conditioned skills, latent plans, reusable primitives, offline meta-RL, task-context representation, and offline multi-agent coordination or equilibrium learning.","exclusion":"Use dataset utilization when the principal contribution is manipulating the dataset rather than the policy structure; use offline-to-online transfer when task adaptation primarily depends on later online interaction.","nearest_boundary":"D3","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","bootstrapped transformer for offline rl via self-generated trajectory data augmentation","diffusion model policy representation for offline reinforcement learning with behavior cloning and policy improvement"]}],"split":"dev"}
{"episode_id":"F042_T2025-01","field_index":42,"field":"offline reinforcement learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Conservative policy improvement","description":"Model-free offline policy optimization that improves policies from fixed datasets while controlling ordinary behavior-policy shift, limited support, value overestimation, and out-of-distribution action selection, including domain adaptations whose central objective is offline policy improvement.","inclusion":"Assign atoms whose main objective is an offline policy optimizer based on pessimistic or conservative value estimation, behavior cloning or policy regularization, uncertainty penalties, support constraints, or practical deployment of such policy-improvement mechanisms in a concrete domain.","exclusion":"Use model-based offline planning when a learned dynamics or reward model is central; use assurance when the primary objective is safety, security, benchmarking, or evaluation rather than optimization.","nearest_boundary":"D1","representative_topics":["offline rl with explicit behavior cloning constraints on value-based policy learning","deep offline reinforcement learning with scalable uncertainty estimates via spi-bb policy optimization","mildly conservative q-learning for out-of-distribution generalization in offline reinforcement learning"]},{"direction_index":1,"label":"D1","name":"Model-based offline planning","description":"Offline RL that learns explicit environment models from static data and uses them for planning, imagination, counterfactual rollouts, or uncertainty-aware policy optimization.","inclusion":"Assign atoms whose core scientific objective is making learned transition, reward, causal, graph, or world models reliable for offline decision making through model uncertainty, pessimism, conservative rollouts, or model-based trajectory stitching.","exclusion":"Use conservative policy improvement when no explicit learned environment model is central; use data improvement when synthetic or transformed data is the main object rather than model-based planning.","nearest_boundary":"D0","representative_topics":["model generation with provable coverability for offline dynamics-aware policy learning","causal world models for offline reinforcement learning generalization","graph-structured world models for offline reinforcement learning in sparse-reward long-horizon environments"]},{"direction_index":2,"label":"D2","name":"Offline RL theory","description":"Theorem-driven characterization of offline RL feasibility, learnability, sample complexity, coverage requirements, structural assumptions, model selection, and minimax or instance-dependent guarantees.","inclusion":"Assign atoms whose primary objective is formal analysis of offline RL or OPE, including bounds, oracle inequalities, impossibility or identifiability conditions, distributionally robust formulations, and theory under linear, low-rank, Bellman-complete, or general function approximation structures.","exclusion":"Use the corresponding algorithmic direction when theory mainly supports a new optimizer, model-based planner, multi-agent method, or evaluation system.","nearest_boundary":"D0","representative_topics":["offline rl with trajectory-wise reward via pessimistic value iteration and reward decomposition","distributionally robust offline reinforcement learning with linear function approximation and non-asymptotic sample complexity","oracle inequalities for model selection in offline reinforcement learning with value function approximation"]},{"direction_index":3,"label":"D3","name":"Dataset utilization and generalization","description":"Research whose primary objective is to make fixed offline datasets more useful through augmentation, reweighting, relabeling, heterogeneous-data handling, representation of dataset structure, or broad algorithmic frameworks for partial coverage and generalization.","inclusion":"Assign atoms centered on transforming or exploiting offline data itself, including mixed-frequency logs, action-free trajectories, image or state augmentation, return rebalancing, counterfactual augmentation, federated or multi-source data, and general frameworks that organize offline data for improved learning without a full dynamics model as the central contribution.","exclusion":"Use model-based offline planning when an explicit learned dynamics model is the main decision-making engine; use structured policy learning when the main objective is hierarchical, sequence, diffusion, meta-task, or multi-agent policy structure.","nearest_boundary":"D7","representative_topics":["offline reinforcement learning with frequency-heterogeneous data via discretization-aware return scaling","state-conditioned image synthesis for data augmentation in offline image-based reinforcement learning","semi-supervised offline reinforcement learning via inverse dynamics proxy-labeling of action-free trajectories"]},{"direction_index":4,"label":"D4","name":"Offline-to-online transfer","description":"Using offline data to initialize, pretrain, constrain, or guide agents that subsequently learn or adapt through online interaction.","inclusion":"Assign atoms whose core objective is bridging offline training and online deployment or fine-tuning, including stable offline-to-online improvement, few-shot online adaptation, offline-pretrained exploration, hybrid offline-online sample complexity, and online learning from action-free offline pretraining.","exclusion":"Use purely offline optimization, modeling, data utilization, or structured policy learning when there is no subsequent online interaction or fine-tuning objective.","nearest_boundary":"D0","representative_topics":["offline rl pre-training for generalizable robot skill learning from few demonstrations","adaptive behavior cloning regularization for stable offline-to-online reinforcement learning fine-tuning","hybrid offline-online reinforcement learning with provably optimal sample complexity for linear mdps"]},{"direction_index":5,"label":"D5","name":"Evaluation and assurance","description":"Assessing, selecting, benchmarking, certifying, or hardening offline RL policies, algorithms, datasets, and systems before deployment.","inclusion":"Assign atoms whose primary objective is off-policy evaluation, confidence bounds, safe policy improvement, algorithm or hyperparameter selection, benchmark datasets, libraries, reproducibility protocols, robustness to perturbations, risk-aware deployment, dataset auditing, or security against poisoning and backdoors.","exclusion":"Use conservative policy improvement when uncertainty or pessimism is mainly used to train a better policy; use reward or causal inference when the main challenge is identifying the target signal or deconfounding observational data.","nearest_boundary":"D6","representative_topics":["hybrid value function estimation balancing offline data and learned model for off-policy evaluation and offline reinforcement learning","safe policy improvement via uncertainty-aware bootstrapping with provable safety guarantees","offline policy evaluation for deterministic ranking policies using position-biased inverse propensity score estimation"]},{"direction_index":6,"label":"D6","name":"Target-signal and causal inference","description":"Inferring, correcting, or identifying the reward, preference, imitation, value, or causal signal needed for offline policy evaluation or optimization when the logged data or reward labels are imperfect or confounded.","inclusion":"Assign atoms whose central objective is preference-based reward learning, offline imitation learning, imperfect reward correction, reward-free implicit value recovery, optimal-transport or discriminator-guided alignment, instrumental variables, deconfounding, propensity reasoning, sensitivity analysis, or policy learning under observed or unobserved confounding.","exclusion":"Use conservative policy improvement when demonstrations or rewards are ordinary behavior data with no separate inference problem; use evaluation and assurance when the main output is a policy value estimate or deployment certificate rather than inferred rewards or causal structure.","nearest_boundary":"D5","representative_topics":["discriminator-guided model-based offline imitation learning with suboptimal demonstration robustness","implicit multi-step contrastive value learning for offline rl without reward functions","offline preference-based reward learning via active learning queries and offline rl policy optimization"]},{"direction_index":7,"label":"D7","name":"Structured policy generalization","description":"Learning expressive or structured offline policies and planners that generalize across long horizons, modes, goals, tasks, agents, or interaction structures.","inclusion":"Assign atoms whose main objective is policy or planner structure beyond a standard single-task actor-critic, including transformer or return-conditioned sequence policies, diffusion or score-based policies and planners, hierarchical and goal-conditioned skills, latent plans, reusable primitives, offline meta-RL, task-context representation, and offline multi-agent coordination or equilibrium learning.","exclusion":"Use dataset utilization when the principal contribution is manipulating the dataset rather than the policy structure; use offline-to-online transfer when task adaptation primarily depends on later online interaction.","nearest_boundary":"D3","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","bootstrapped transformer for offline rl via self-generated trajectory data augmentation","diffusion model policy representation for offline reinforcement learning with behavior cloning and policy improvement"]}],"split":"dev"}
{"episode_id":"F042_T2025-07","field_index":42,"field":"offline reinforcement learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Conservative policy improvement","description":"Model-free offline policy optimization that improves policies from fixed datasets while controlling ordinary behavior-policy shift, limited support, value overestimation, and out-of-distribution action selection, including domain adaptations whose central objective is offline policy improvement.","inclusion":"Assign atoms whose main objective is an offline policy optimizer based on pessimistic or conservative value estimation, behavior cloning or policy regularization, uncertainty penalties, support constraints, or practical deployment of such policy-improvement mechanisms in a concrete domain.","exclusion":"Use model-based offline planning when a learned dynamics or reward model is central; use assurance when the primary objective is safety, security, benchmarking, or evaluation rather than optimization.","nearest_boundary":"D1","representative_topics":["offline rl with explicit behavior cloning constraints on value-based policy learning","deep offline reinforcement learning with scalable uncertainty estimates via spi-bb policy optimization","mildly conservative q-learning for out-of-distribution generalization in offline reinforcement learning"]},{"direction_index":1,"label":"D1","name":"Model-based offline planning","description":"Offline RL that learns explicit environment models from static data and uses them for planning, imagination, counterfactual rollouts, or uncertainty-aware policy optimization.","inclusion":"Assign atoms whose core scientific objective is making learned transition, reward, causal, graph, or world models reliable for offline decision making through model uncertainty, pessimism, conservative rollouts, or model-based trajectory stitching.","exclusion":"Use conservative policy improvement when no explicit learned environment model is central; use data improvement when synthetic or transformed data is the main object rather than model-based planning.","nearest_boundary":"D0","representative_topics":["model generation with provable coverability for offline dynamics-aware policy learning","causal world models for offline reinforcement learning generalization","graph-structured world models for offline reinforcement learning in sparse-reward long-horizon environments"]},{"direction_index":2,"label":"D2","name":"Offline RL theory","description":"Theorem-driven characterization of offline RL feasibility, learnability, sample complexity, coverage requirements, structural assumptions, model selection, and minimax or instance-dependent guarantees.","inclusion":"Assign atoms whose primary objective is formal analysis of offline RL or OPE, including bounds, oracle inequalities, impossibility or identifiability conditions, distributionally robust formulations, and theory under linear, low-rank, Bellman-complete, or general function approximation structures.","exclusion":"Use the corresponding algorithmic direction when theory mainly supports a new optimizer, model-based planner, multi-agent method, or evaluation system.","nearest_boundary":"D0","representative_topics":["offline rl with trajectory-wise reward via pessimistic value iteration and reward decomposition","distributionally robust offline reinforcement learning with linear function approximation and non-asymptotic sample complexity","oracle inequalities for model selection in offline reinforcement learning with value function approximation"]},{"direction_index":3,"label":"D3","name":"Dataset utilization and generalization","description":"Research whose primary objective is to make fixed offline datasets more useful through augmentation, reweighting, relabeling, heterogeneous-data handling, representation of dataset structure, or broad algorithmic frameworks for partial coverage and generalization.","inclusion":"Assign atoms centered on transforming or exploiting offline data itself, including mixed-frequency logs, action-free trajectories, image or state augmentation, return rebalancing, counterfactual augmentation, federated or multi-source data, and general frameworks that organize offline data for improved learning without a full dynamics model as the central contribution.","exclusion":"Use model-based offline planning when an explicit learned dynamics model is the main decision-making engine; use structured policy learning when the main objective is hierarchical, sequence, diffusion, meta-task, or multi-agent policy structure.","nearest_boundary":"D7","representative_topics":["offline reinforcement learning with frequency-heterogeneous data via discretization-aware return scaling","state-conditioned image synthesis for data augmentation in offline image-based reinforcement learning","semi-supervised offline reinforcement learning via inverse dynamics proxy-labeling of action-free trajectories"]},{"direction_index":4,"label":"D4","name":"Offline-to-online transfer","description":"Using offline data to initialize, pretrain, constrain, or guide agents that subsequently learn or adapt through online interaction.","inclusion":"Assign atoms whose core objective is bridging offline training and online deployment or fine-tuning, including stable offline-to-online improvement, few-shot online adaptation, offline-pretrained exploration, hybrid offline-online sample complexity, and online learning from action-free offline pretraining.","exclusion":"Use purely offline optimization, modeling, data utilization, or structured policy learning when there is no subsequent online interaction or fine-tuning objective.","nearest_boundary":"D0","representative_topics":["offline rl pre-training for generalizable robot skill learning from few demonstrations","adaptive behavior cloning regularization for stable offline-to-online reinforcement learning fine-tuning","hybrid offline-online reinforcement learning with provably optimal sample complexity for linear mdps"]},{"direction_index":5,"label":"D5","name":"Evaluation and assurance","description":"Assessing, selecting, benchmarking, certifying, or hardening offline RL policies, algorithms, datasets, and systems before deployment.","inclusion":"Assign atoms whose primary objective is off-policy evaluation, confidence bounds, safe policy improvement, algorithm or hyperparameter selection, benchmark datasets, libraries, reproducibility protocols, robustness to perturbations, risk-aware deployment, dataset auditing, or security against poisoning and backdoors.","exclusion":"Use conservative policy improvement when uncertainty or pessimism is mainly used to train a better policy; use reward or causal inference when the main challenge is identifying the target signal or deconfounding observational data.","nearest_boundary":"D6","representative_topics":["hybrid value function estimation balancing offline data and learned model for off-policy evaluation and offline reinforcement learning","safe policy improvement via uncertainty-aware bootstrapping with provable safety guarantees","offline policy evaluation for deterministic ranking policies using position-biased inverse propensity score estimation"]},{"direction_index":6,"label":"D6","name":"Target-signal and causal inference","description":"Inferring, correcting, or identifying the reward, preference, imitation, value, or causal signal needed for offline policy evaluation or optimization when the logged data or reward labels are imperfect or confounded.","inclusion":"Assign atoms whose central objective is preference-based reward learning, offline imitation learning, imperfect reward correction, reward-free implicit value recovery, optimal-transport or discriminator-guided alignment, instrumental variables, deconfounding, propensity reasoning, sensitivity analysis, or policy learning under observed or unobserved confounding.","exclusion":"Use conservative policy improvement when demonstrations or rewards are ordinary behavior data with no separate inference problem; use evaluation and assurance when the main output is a policy value estimate or deployment certificate rather than inferred rewards or causal structure.","nearest_boundary":"D5","representative_topics":["discriminator-guided model-based offline imitation learning with suboptimal demonstration robustness","implicit multi-step contrastive value learning for offline rl without reward functions","offline preference-based reward learning via active learning queries and offline rl policy optimization"]},{"direction_index":7,"label":"D7","name":"Structured policy generalization","description":"Learning expressive or structured offline policies and planners that generalize across long horizons, modes, goals, tasks, agents, or interaction structures.","inclusion":"Assign atoms whose main objective is policy or planner structure beyond a standard single-task actor-critic, including transformer or return-conditioned sequence policies, diffusion or score-based policies and planners, hierarchical and goal-conditioned skills, latent plans, reusable primitives, offline meta-RL, task-context representation, and offline multi-agent coordination or equilibrium learning.","exclusion":"Use dataset utilization when the principal contribution is manipulating the dataset rather than the policy structure; use offline-to-online transfer when task adaptation primarily depends on later online interaction.","nearest_boundary":"D3","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","bootstrapped transformer for offline rl via self-generated trajectory data augmentation","diffusion model policy representation for offline reinforcement learning with behavior cloning and policy improvement"]}],"split":"dev"}
{"episode_id":"F042_T2026-01","field_index":42,"field":"offline reinforcement learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Conservative policy improvement","description":"Model-free offline policy optimization that improves policies from fixed datasets while controlling ordinary behavior-policy shift, limited support, value overestimation, and out-of-distribution action selection, including domain adaptations whose central objective is offline policy improvement.","inclusion":"Assign atoms whose main objective is an offline policy optimizer based on pessimistic or conservative value estimation, behavior cloning or policy regularization, uncertainty penalties, support constraints, or practical deployment of such policy-improvement mechanisms in a concrete domain.","exclusion":"Use model-based offline planning when a learned dynamics or reward model is central; use assurance when the primary objective is safety, security, benchmarking, or evaluation rather than optimization.","nearest_boundary":"D1","representative_topics":["offline rl with explicit behavior cloning constraints on value-based policy learning","deep offline reinforcement learning with scalable uncertainty estimates via spi-bb policy optimization","mildly conservative q-learning for out-of-distribution generalization in offline reinforcement learning"]},{"direction_index":1,"label":"D1","name":"Model-based offline planning","description":"Offline RL that learns explicit environment models from static data and uses them for planning, imagination, counterfactual rollouts, or uncertainty-aware policy optimization.","inclusion":"Assign atoms whose core scientific objective is making learned transition, reward, causal, graph, or world models reliable for offline decision making through model uncertainty, pessimism, conservative rollouts, or model-based trajectory stitching.","exclusion":"Use conservative policy improvement when no explicit learned environment model is central; use data improvement when synthetic or transformed data is the main object rather than model-based planning.","nearest_boundary":"D0","representative_topics":["model generation with provable coverability for offline dynamics-aware policy learning","causal world models for offline reinforcement learning generalization","graph-structured world models for offline reinforcement learning in sparse-reward long-horizon environments"]},{"direction_index":2,"label":"D2","name":"Offline RL theory","description":"Theorem-driven characterization of offline RL feasibility, learnability, sample complexity, coverage requirements, structural assumptions, model selection, and minimax or instance-dependent guarantees.","inclusion":"Assign atoms whose primary objective is formal analysis of offline RL or OPE, including bounds, oracle inequalities, impossibility or identifiability conditions, distributionally robust formulations, and theory under linear, low-rank, Bellman-complete, or general function approximation structures.","exclusion":"Use the corresponding algorithmic direction when theory mainly supports a new optimizer, model-based planner, multi-agent method, or evaluation system.","nearest_boundary":"D0","representative_topics":["offline rl with trajectory-wise reward via pessimistic value iteration and reward decomposition","distributionally robust offline reinforcement learning with linear function approximation and non-asymptotic sample complexity","oracle inequalities for model selection in offline reinforcement learning with value function approximation"]},{"direction_index":3,"label":"D3","name":"Dataset utilization and generalization","description":"Research whose primary objective is to make fixed offline datasets more useful through augmentation, reweighting, relabeling, heterogeneous-data handling, representation of dataset structure, or broad algorithmic frameworks for partial coverage and generalization.","inclusion":"Assign atoms centered on transforming or exploiting offline data itself, including mixed-frequency logs, action-free trajectories, image or state augmentation, return rebalancing, counterfactual augmentation, federated or multi-source data, and general frameworks that organize offline data for improved learning without a full dynamics model as the central contribution.","exclusion":"Use model-based offline planning when an explicit learned dynamics model is the main decision-making engine; use structured policy learning when the main objective is hierarchical, sequence, diffusion, meta-task, or multi-agent policy structure.","nearest_boundary":"D7","representative_topics":["offline reinforcement learning with frequency-heterogeneous data via discretization-aware return scaling","state-conditioned image synthesis for data augmentation in offline image-based reinforcement learning","semi-supervised offline reinforcement learning via inverse dynamics proxy-labeling of action-free trajectories"]},{"direction_index":4,"label":"D4","name":"Offline-to-online transfer","description":"Using offline data to initialize, pretrain, constrain, or guide agents that subsequently learn or adapt through online interaction.","inclusion":"Assign atoms whose core objective is bridging offline training and online deployment or fine-tuning, including stable offline-to-online improvement, few-shot online adaptation, offline-pretrained exploration, hybrid offline-online sample complexity, and online learning from action-free offline pretraining.","exclusion":"Use purely offline optimization, modeling, data utilization, or structured policy learning when there is no subsequent online interaction or fine-tuning objective.","nearest_boundary":"D0","representative_topics":["offline rl pre-training for generalizable robot skill learning from few demonstrations","adaptive behavior cloning regularization for stable offline-to-online reinforcement learning fine-tuning","hybrid offline-online reinforcement learning with provably optimal sample complexity for linear mdps"]},{"direction_index":5,"label":"D5","name":"Evaluation and assurance","description":"Assessing, selecting, benchmarking, certifying, or hardening offline RL policies, algorithms, datasets, and systems before deployment.","inclusion":"Assign atoms whose primary objective is off-policy evaluation, confidence bounds, safe policy improvement, algorithm or hyperparameter selection, benchmark datasets, libraries, reproducibility protocols, robustness to perturbations, risk-aware deployment, dataset auditing, or security against poisoning and backdoors.","exclusion":"Use conservative policy improvement when uncertainty or pessimism is mainly used to train a better policy; use reward or causal inference when the main challenge is identifying the target signal or deconfounding observational data.","nearest_boundary":"D6","representative_topics":["hybrid value function estimation balancing offline data and learned model for off-policy evaluation and offline reinforcement learning","safe policy improvement via uncertainty-aware bootstrapping with provable safety guarantees","offline policy evaluation for deterministic ranking policies using position-biased inverse propensity score estimation"]},{"direction_index":6,"label":"D6","name":"Target-signal and causal inference","description":"Inferring, correcting, or identifying the reward, preference, imitation, value, or causal signal needed for offline policy evaluation or optimization when the logged data or reward labels are imperfect or confounded.","inclusion":"Assign atoms whose central objective is preference-based reward learning, offline imitation learning, imperfect reward correction, reward-free implicit value recovery, optimal-transport or discriminator-guided alignment, instrumental variables, deconfounding, propensity reasoning, sensitivity analysis, or policy learning under observed or unobserved confounding.","exclusion":"Use conservative policy improvement when demonstrations or rewards are ordinary behavior data with no separate inference problem; use evaluation and assurance when the main output is a policy value estimate or deployment certificate rather than inferred rewards or causal structure.","nearest_boundary":"D5","representative_topics":["discriminator-guided model-based offline imitation learning with suboptimal demonstration robustness","implicit multi-step contrastive value learning for offline rl without reward functions","offline preference-based reward learning via active learning queries and offline rl policy optimization"]},{"direction_index":7,"label":"D7","name":"Structured policy generalization","description":"Learning expressive or structured offline policies and planners that generalize across long horizons, modes, goals, tasks, agents, or interaction structures.","inclusion":"Assign atoms whose main objective is policy or planner structure beyond a standard single-task actor-critic, including transformer or return-conditioned sequence policies, diffusion or score-based policies and planners, hierarchical and goal-conditioned skills, latent plans, reusable primitives, offline meta-RL, task-context representation, and offline multi-agent coordination or equilibrium learning.","exclusion":"Use dataset utilization when the principal contribution is manipulating the dataset rather than the policy structure; use offline-to-online transfer when task adaptation primarily depends on later online interaction.","nearest_boundary":"D3","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","bootstrapped transformer for offline rl via self-generated trajectory data augmentation","diffusion model policy representation for offline reinforcement learning with behavior cloning and policy improvement"]}],"split":"dev"}
{"episode_id":"F045_T2024-01","field_index":45,"field":"3d object detection","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Multi-sensor inference fusion","description":"Improve 3D object detection by combining complementary sensor observations at inference time, especially LiDAR-camera, radar-camera, radar-only, and radar-inclusive multi-modal systems.","inclusion":"Assign atoms whose primary objective is a detector that uses multiple physical sensing modalities, or radar as a core modality, during inference for 3D detection, including feature-, BEV-, query-, proposal-, region-, or box-level fusion and radar-specific modeling.","exclusion":"Camera-only detectors using LiDAR only as training supervision belong to camera-only perception or supervision transfer; reduced-teacher-cost methods without multi-sensor inference belong to data supervision and transfer.","nearest_boundary":"Data supervision and transfer","representative_topics":["multi-scale lidar-camera fusion with depth-aware lifting and fine-grained voxel interaction for 3d object detection","lidar-camera fusion for two-stage 3d object detection using cross-modal attention in regions of interest","sparse center-based feature fusion for camera-lidar bev 3d object detection"]},{"direction_index":1,"label":"D1","name":"Camera-only 3D perception","description":"Estimate 3D object boxes from camera inputs without inference-time LiDAR or radar, covering monocular, stereo, surround-view, multi-camera, and camera-BEV formulations.","inclusion":"Assign atoms whose primary objective is 3D detection from images alone at inference time, including monocular/stereo depth and pose reasoning, pseudo-LiDAR from images, multi-view BEV construction, polar/frustum/query representations, and temporal camera-only BEV reasoning.","exclusion":"Methods using LiDAR, radar, or both as inference sensors belong to multi-sensor inference fusion; methods whose main novelty is privileged-modality teacher supervision rather than camera representation belong to data supervision and transfer.","nearest_boundary":"Multi-sensor inference fusion","representative_topics":["multi-level rgb-pseudo-lidar temporal fusion for per-object depth estimation in monocular 3d detection","ground plane prior for monocular 3d object depth estimation without auxiliary sensors","efficient pseudo-lidar 3d object detection via fast stereo matching and point cloud refinement"]},{"direction_index":2,"label":"D2","name":"LiDAR detector design and optimization","description":"Improve closed-set 3D detection primarily through LiDAR detector architectures, point/voxel/pillar/sparse representations, detection heads, proposal refinement, loss/assignment design, class-imbalance handling, and deployment-oriented efficiency.","inclusion":"Assign atoms whose dominant objective is the single-system detector design itself: LiDAR backbones, sparse or transformer encoders, geometric representations, proposal generation, target assignment, box optimization, cascaded refinement, loss functions, class/long-tail reweighting or sampling, runtime acceleration, edge deployment, and deadline-aware execution.","exclusion":"Temporal multi-frame reasoning belongs to temporal and joint perception; reduced labels, domain transfer, or teacher-student training belong to data supervision and transfer; reliability under corruptions or uncertainty belongs to robustness and safety.","nearest_boundary":"Temporal and joint perception","representative_topics":["sparse proposal-based end-to-end 3d object detection from point clouds using transformers","rotation-decoupled iou optimization for single-stage 3d object detection","full-degree-of-freedom lidar 3d object detection with ground-aware pose estimation on sloped terrain"]},{"direction_index":3,"label":"D3","name":"Data supervision and transfer","description":"Reduce annotation burden or improve transferability of 3D detectors through label-efficient learning, domain adaptation/generalization, and teacher-student knowledge transfer across modalities, frames, or representations.","inclusion":"Assign atoms whose primary objective is learning under limited, weak, unlabeled, shifted, or privileged-supervision regimes, including semi-supervised learning, self-supervised pretraining, pseudo-labeling, few-shot and active learning, cross-domain adaptation, synthetic-to-real transfer, multi-dataset generalization, and knowledge distillation.","exclusion":"If multiple sensors are jointly required at inference, assign to multi-sensor inference fusion; if the objective is measuring datasets or creating annotation tools rather than training a detector, assign to evaluation and annotation infrastructure.","nearest_boundary":"Multi-sensor inference fusion","representative_topics":["voxel masked autoencoding pre-training for sparse automotive lidar 3d object detection","unsupervised region-proposal contrastive pre-training for lidar 3d object detection","semi-supervised pseudo-labeling for image-based 3d object detection under monocular conditions"]},{"direction_index":4,"label":"D4","name":"Temporal and joint perception","description":"Use temporal sequences, tracking-coupled inference, multi-task perception, or cooperative multi-agent context as a primary program for 3D object detection.","inclusion":"Assign atoms whose main objective is exploiting time, motion, multi-frame aggregation, streaming context, spatio-temporal attention, joint detection-tracking, detection plus segmentation or panoptic perception, or collaborative/V2X multi-agent 3D detection.","exclusion":"Single-frame LiDAR architecture or efficiency work belongs to LiDAR detector design and optimization; camera-only static multi-view BEV construction belongs to camera-only 3D perception unless temporal or multi-task unification is the dominant contribution.","nearest_boundary":"LiDAR detector design and optimization","representative_topics":["spatiotemporal transformer and graph message passing for multi-frame lidar point cloud video object detection","temporal voxel feature aggregation for multi-frame 3d object detection using coarse-to-fine transformer fusion","efficient lidar 3d detection via spatial and temporal object relation modeling"]},{"direction_index":5,"label":"D5","name":"Robustness and safety assurance","description":"Assess or improve reliability of 3D detectors under uncertainty, label noise, occlusion, corruptions, adverse conditions, adversarial perturbations, and safety constraints.","inclusion":"Assign atoms whose primary objective is robustness benchmarking, uncertainty quantification, calibrated box prediction, label-uncertainty modeling, safety-constrained localization, corruption robustness, or attack/defense analysis for 3D detection.","exclusion":"Cross-domain source-target transfer belongs to data supervision and transfer; radar-centered weather robustness belongs to multi-sensor inference fusion when radar modeling is the main program.","nearest_boundary":"Data supervision and transfer","representative_topics":["efficient mimo uncertainty estimation for lidar-based 3d object detection under runtime constraints","generative label uncertainty estimation for robust 3d object detection via conditional variational autoencoders","safety-oriented 3d object detection with spatial coverage constraints for autonomous driving"]},{"direction_index":6,"label":"D6","name":"Open-vocabulary and language-guided detection","description":"Enable 3D detection or grounding for novel, unseen, language-specified, or open-set categories by aligning 3D representations with image, text, or foundation-model representations.","inclusion":"Assign atoms whose primary objective is open-vocabulary 3D detection, zero-shot 3D detection, language-guided 3D region grounding, CLIP-style or vision-language transfer for unseen classes, or foundation-model adaptation for category expansion.","exclusion":"Closed-set multi-modal fusion belongs to multi-sensor inference fusion; closed-set teacher-student transfer belongs to data supervision and transfer unless novel-category or language grounding is the central objective.","nearest_boundary":"Data supervision and transfer","representative_topics":["open-vocabulary 3d point cloud detection via cross-modal contrastive learning and image-level class supervision","open-vocabulary 3d point cloud detection via 2d-to-3d transfer and vision-language contrastive learning","multimodal language-guided 3d region grounding in lidar point clouds for autonomous driving"]},{"direction_index":7,"label":"D7","name":"Evaluation and annotation infrastructure","description":"Build measurement, benchmark, dataset, annotation, survey, and validation infrastructure for 3D object detection rather than proposing a detector-training method.","inclusion":"Assign atoms whose primary objective is benchmark construction, evaluation metric design, dataset curation, survey synthesis, annotation tooling, interactive annotation, labeling assistance, or offline validation infrastructure for 3D object detection.","exclusion":"Active learning or pseudo-labeling methods for training detectors belong to data supervision and transfer; robustness benchmarks paired primarily with detector reliability methods belong to robustness and safety assurance.","nearest_boundary":"Data supervision and transfer","representative_topics":["longitudinal error-tolerant evaluation metrics for camera-only 3d object detection","survey of 3d object detection methods for autonomous driving perception","full-degree-of-freedom 3d bounding box evaluation metrics for object detection"]}],"split":"dev"}
{"episode_id":"F045_T2024-07","field_index":45,"field":"3d object detection","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Multi-sensor inference fusion","description":"Improve 3D object detection by combining complementary sensor observations at inference time, especially LiDAR-camera, radar-camera, radar-only, and radar-inclusive multi-modal systems.","inclusion":"Assign atoms whose primary objective is a detector that uses multiple physical sensing modalities, or radar as a core modality, during inference for 3D detection, including feature-, BEV-, query-, proposal-, region-, or box-level fusion and radar-specific modeling.","exclusion":"Camera-only detectors using LiDAR only as training supervision belong to camera-only perception or supervision transfer; reduced-teacher-cost methods without multi-sensor inference belong to data supervision and transfer.","nearest_boundary":"Data supervision and transfer","representative_topics":["multi-scale lidar-camera fusion with depth-aware lifting and fine-grained voxel interaction for 3d object detection","lidar-camera fusion for two-stage 3d object detection using cross-modal attention in regions of interest","sparse center-based feature fusion for camera-lidar bev 3d object detection"]},{"direction_index":1,"label":"D1","name":"Camera-only 3D perception","description":"Estimate 3D object boxes from camera inputs without inference-time LiDAR or radar, covering monocular, stereo, surround-view, multi-camera, and camera-BEV formulations.","inclusion":"Assign atoms whose primary objective is 3D detection from images alone at inference time, including monocular/stereo depth and pose reasoning, pseudo-LiDAR from images, multi-view BEV construction, polar/frustum/query representations, and temporal camera-only BEV reasoning.","exclusion":"Methods using LiDAR, radar, or both as inference sensors belong to multi-sensor inference fusion; methods whose main novelty is privileged-modality teacher supervision rather than camera representation belong to data supervision and transfer.","nearest_boundary":"Multi-sensor inference fusion","representative_topics":["multi-level rgb-pseudo-lidar temporal fusion for per-object depth estimation in monocular 3d detection","ground plane prior for monocular 3d object depth estimation without auxiliary sensors","efficient pseudo-lidar 3d object detection via fast stereo matching and point cloud refinement"]},{"direction_index":2,"label":"D2","name":"LiDAR detector design and optimization","description":"Improve closed-set 3D detection primarily through LiDAR detector architectures, point/voxel/pillar/sparse representations, detection heads, proposal refinement, loss/assignment design, class-imbalance handling, and deployment-oriented efficiency.","inclusion":"Assign atoms whose dominant objective is the single-system detector design itself: LiDAR backbones, sparse or transformer encoders, geometric representations, proposal generation, target assignment, box optimization, cascaded refinement, loss functions, class/long-tail reweighting or sampling, runtime acceleration, edge deployment, and deadline-aware execution.","exclusion":"Temporal multi-frame reasoning belongs to temporal and joint perception; reduced labels, domain transfer, or teacher-student training belong to data supervision and transfer; reliability under corruptions or uncertainty belongs to robustness and safety.","nearest_boundary":"Temporal and joint perception","representative_topics":["sparse proposal-based end-to-end 3d object detection from point clouds using transformers","rotation-decoupled iou optimization for single-stage 3d object detection","full-degree-of-freedom lidar 3d object detection with ground-aware pose estimation on sloped terrain"]},{"direction_index":3,"label":"D3","name":"Data supervision and transfer","description":"Reduce annotation burden or improve transferability of 3D detectors through label-efficient learning, domain adaptation/generalization, and teacher-student knowledge transfer across modalities, frames, or representations.","inclusion":"Assign atoms whose primary objective is learning under limited, weak, unlabeled, shifted, or privileged-supervision regimes, including semi-supervised learning, self-supervised pretraining, pseudo-labeling, few-shot and active learning, cross-domain adaptation, synthetic-to-real transfer, multi-dataset generalization, and knowledge distillation.","exclusion":"If multiple sensors are jointly required at inference, assign to multi-sensor inference fusion; if the objective is measuring datasets or creating annotation tools rather than training a detector, assign to evaluation and annotation infrastructure.","nearest_boundary":"Multi-sensor inference fusion","representative_topics":["voxel masked autoencoding pre-training for sparse automotive lidar 3d object detection","unsupervised region-proposal contrastive pre-training for lidar 3d object detection","semi-supervised pseudo-labeling for image-based 3d object detection under monocular conditions"]},{"direction_index":4,"label":"D4","name":"Temporal and joint perception","description":"Use temporal sequences, tracking-coupled inference, multi-task perception, or cooperative multi-agent context as a primary program for 3D object detection.","inclusion":"Assign atoms whose main objective is exploiting time, motion, multi-frame aggregation, streaming context, spatio-temporal attention, joint detection-tracking, detection plus segmentation or panoptic perception, or collaborative/V2X multi-agent 3D detection.","exclusion":"Single-frame LiDAR architecture or efficiency work belongs to LiDAR detector design and optimization; camera-only static multi-view BEV construction belongs to camera-only 3D perception unless temporal or multi-task unification is the dominant contribution.","nearest_boundary":"LiDAR detector design and optimization","representative_topics":["spatiotemporal transformer and graph message passing for multi-frame lidar point cloud video object detection","temporal voxel feature aggregation for multi-frame 3d object detection using coarse-to-fine transformer fusion","efficient lidar 3d detection via spatial and temporal object relation modeling"]},{"direction_index":5,"label":"D5","name":"Robustness and safety assurance","description":"Assess or improve reliability of 3D detectors under uncertainty, label noise, occlusion, corruptions, adverse conditions, adversarial perturbations, and safety constraints.","inclusion":"Assign atoms whose primary objective is robustness benchmarking, uncertainty quantification, calibrated box prediction, label-uncertainty modeling, safety-constrained localization, corruption robustness, or attack/defense analysis for 3D detection.","exclusion":"Cross-domain source-target transfer belongs to data supervision and transfer; radar-centered weather robustness belongs to multi-sensor inference fusion when radar modeling is the main program.","nearest_boundary":"Data supervision and transfer","representative_topics":["efficient mimo uncertainty estimation for lidar-based 3d object detection under runtime constraints","generative label uncertainty estimation for robust 3d object detection via conditional variational autoencoders","safety-oriented 3d object detection with spatial coverage constraints for autonomous driving"]},{"direction_index":6,"label":"D6","name":"Open-vocabulary and language-guided detection","description":"Enable 3D detection or grounding for novel, unseen, language-specified, or open-set categories by aligning 3D representations with image, text, or foundation-model representations.","inclusion":"Assign atoms whose primary objective is open-vocabulary 3D detection, zero-shot 3D detection, language-guided 3D region grounding, CLIP-style or vision-language transfer for unseen classes, or foundation-model adaptation for category expansion.","exclusion":"Closed-set multi-modal fusion belongs to multi-sensor inference fusion; closed-set teacher-student transfer belongs to data supervision and transfer unless novel-category or language grounding is the central objective.","nearest_boundary":"Data supervision and transfer","representative_topics":["open-vocabulary 3d point cloud detection via cross-modal contrastive learning and image-level class supervision","open-vocabulary 3d point cloud detection via 2d-to-3d transfer and vision-language contrastive learning","multimodal language-guided 3d region grounding in lidar point clouds for autonomous driving"]},{"direction_index":7,"label":"D7","name":"Evaluation and annotation infrastructure","description":"Build measurement, benchmark, dataset, annotation, survey, and validation infrastructure for 3D object detection rather than proposing a detector-training method.","inclusion":"Assign atoms whose primary objective is benchmark construction, evaluation metric design, dataset curation, survey synthesis, annotation tooling, interactive annotation, labeling assistance, or offline validation infrastructure for 3D object detection.","exclusion":"Active learning or pseudo-labeling methods for training detectors belong to data supervision and transfer; robustness benchmarks paired primarily with detector reliability methods belong to robustness and safety assurance.","nearest_boundary":"Data supervision and transfer","representative_topics":["longitudinal error-tolerant evaluation metrics for camera-only 3d object detection","survey of 3d object detection methods for autonomous driving perception","full-degree-of-freedom 3d bounding box evaluation metrics for object detection"]}],"split":"dev"}
{"episode_id":"F045_T2025-01","field_index":45,"field":"3d object detection","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Multi-sensor inference fusion","description":"Improve 3D object detection by combining complementary sensor observations at inference time, especially LiDAR-camera, radar-camera, radar-only, and radar-inclusive multi-modal systems.","inclusion":"Assign atoms whose primary objective is a detector that uses multiple physical sensing modalities, or radar as a core modality, during inference for 3D detection, including feature-, BEV-, query-, proposal-, region-, or box-level fusion and radar-specific modeling.","exclusion":"Camera-only detectors using LiDAR only as training supervision belong to camera-only perception or supervision transfer; reduced-teacher-cost methods without multi-sensor inference belong to data supervision and transfer.","nearest_boundary":"Data supervision and transfer","representative_topics":["multi-scale lidar-camera fusion with depth-aware lifting and fine-grained voxel interaction for 3d object detection","lidar-camera fusion for two-stage 3d object detection using cross-modal attention in regions of interest","sparse center-based feature fusion for camera-lidar bev 3d object detection"]},{"direction_index":1,"label":"D1","name":"Camera-only 3D perception","description":"Estimate 3D object boxes from camera inputs without inference-time LiDAR or radar, covering monocular, stereo, surround-view, multi-camera, and camera-BEV formulations.","inclusion":"Assign atoms whose primary objective is 3D detection from images alone at inference time, including monocular/stereo depth and pose reasoning, pseudo-LiDAR from images, multi-view BEV construction, polar/frustum/query representations, and temporal camera-only BEV reasoning.","exclusion":"Methods using LiDAR, radar, or both as inference sensors belong to multi-sensor inference fusion; methods whose main novelty is privileged-modality teacher supervision rather than camera representation belong to data supervision and transfer.","nearest_boundary":"Multi-sensor inference fusion","representative_topics":["multi-level rgb-pseudo-lidar temporal fusion for per-object depth estimation in monocular 3d detection","ground plane prior for monocular 3d object depth estimation without auxiliary sensors","efficient pseudo-lidar 3d object detection via fast stereo matching and point cloud refinement"]},{"direction_index":2,"label":"D2","name":"LiDAR detector design and optimization","description":"Improve closed-set 3D detection primarily through LiDAR detector architectures, point/voxel/pillar/sparse representations, detection heads, proposal refinement, loss/assignment design, class-imbalance handling, and deployment-oriented efficiency.","inclusion":"Assign atoms whose dominant objective is the single-system detector design itself: LiDAR backbones, sparse or transformer encoders, geometric representations, proposal generation, target assignment, box optimization, cascaded refinement, loss functions, class/long-tail reweighting or sampling, runtime acceleration, edge deployment, and deadline-aware execution.","exclusion":"Temporal multi-frame reasoning belongs to temporal and joint perception; reduced labels, domain transfer, or teacher-student training belong to data supervision and transfer; reliability under corruptions or uncertainty belongs to robustness and safety.","nearest_boundary":"Temporal and joint perception","representative_topics":["sparse proposal-based end-to-end 3d object detection from point clouds using transformers","rotation-decoupled iou optimization for single-stage 3d object detection","full-degree-of-freedom lidar 3d object detection with ground-aware pose estimation on sloped terrain"]},{"direction_index":3,"label":"D3","name":"Data supervision and transfer","description":"Reduce annotation burden or improve transferability of 3D detectors through label-efficient learning, domain adaptation/generalization, and teacher-student knowledge transfer across modalities, frames, or representations.","inclusion":"Assign atoms whose primary objective is learning under limited, weak, unlabeled, shifted, or privileged-supervision regimes, including semi-supervised learning, self-supervised pretraining, pseudo-labeling, few-shot and active learning, cross-domain adaptation, synthetic-to-real transfer, multi-dataset generalization, and knowledge distillation.","exclusion":"If multiple sensors are jointly required at inference, assign to multi-sensor inference fusion; if the objective is measuring datasets or creating annotation tools rather than training a detector, assign to evaluation and annotation infrastructure.","nearest_boundary":"Multi-sensor inference fusion","representative_topics":["voxel masked autoencoding pre-training for sparse automotive lidar 3d object detection","unsupervised region-proposal contrastive pre-training for lidar 3d object detection","semi-supervised pseudo-labeling for image-based 3d object detection under monocular conditions"]},{"direction_index":4,"label":"D4","name":"Temporal and joint perception","description":"Use temporal sequences, tracking-coupled inference, multi-task perception, or cooperative multi-agent context as a primary program for 3D object detection.","inclusion":"Assign atoms whose main objective is exploiting time, motion, multi-frame aggregation, streaming context, spatio-temporal attention, joint detection-tracking, detection plus segmentation or panoptic perception, or collaborative/V2X multi-agent 3D detection.","exclusion":"Single-frame LiDAR architecture or efficiency work belongs to LiDAR detector design and optimization; camera-only static multi-view BEV construction belongs to camera-only 3D perception unless temporal or multi-task unification is the dominant contribution.","nearest_boundary":"LiDAR detector design and optimization","representative_topics":["spatiotemporal transformer and graph message passing for multi-frame lidar point cloud video object detection","temporal voxel feature aggregation for multi-frame 3d object detection using coarse-to-fine transformer fusion","efficient lidar 3d detection via spatial and temporal object relation modeling"]},{"direction_index":5,"label":"D5","name":"Robustness and safety assurance","description":"Assess or improve reliability of 3D detectors under uncertainty, label noise, occlusion, corruptions, adverse conditions, adversarial perturbations, and safety constraints.","inclusion":"Assign atoms whose primary objective is robustness benchmarking, uncertainty quantification, calibrated box prediction, label-uncertainty modeling, safety-constrained localization, corruption robustness, or attack/defense analysis for 3D detection.","exclusion":"Cross-domain source-target transfer belongs to data supervision and transfer; radar-centered weather robustness belongs to multi-sensor inference fusion when radar modeling is the main program.","nearest_boundary":"Data supervision and transfer","representative_topics":["efficient mimo uncertainty estimation for lidar-based 3d object detection under runtime constraints","generative label uncertainty estimation for robust 3d object detection via conditional variational autoencoders","safety-oriented 3d object detection with spatial coverage constraints for autonomous driving"]},{"direction_index":6,"label":"D6","name":"Open-vocabulary and language-guided detection","description":"Enable 3D detection or grounding for novel, unseen, language-specified, or open-set categories by aligning 3D representations with image, text, or foundation-model representations.","inclusion":"Assign atoms whose primary objective is open-vocabulary 3D detection, zero-shot 3D detection, language-guided 3D region grounding, CLIP-style or vision-language transfer for unseen classes, or foundation-model adaptation for category expansion.","exclusion":"Closed-set multi-modal fusion belongs to multi-sensor inference fusion; closed-set teacher-student transfer belongs to data supervision and transfer unless novel-category or language grounding is the central objective.","nearest_boundary":"Data supervision and transfer","representative_topics":["open-vocabulary 3d point cloud detection via cross-modal contrastive learning and image-level class supervision","open-vocabulary 3d point cloud detection via 2d-to-3d transfer and vision-language contrastive learning","multimodal language-guided 3d region grounding in lidar point clouds for autonomous driving"]},{"direction_index":7,"label":"D7","name":"Evaluation and annotation infrastructure","description":"Build measurement, benchmark, dataset, annotation, survey, and validation infrastructure for 3D object detection rather than proposing a detector-training method.","inclusion":"Assign atoms whose primary objective is benchmark construction, evaluation metric design, dataset curation, survey synthesis, annotation tooling, interactive annotation, labeling assistance, or offline validation infrastructure for 3D object detection.","exclusion":"Active learning or pseudo-labeling methods for training detectors belong to data supervision and transfer; robustness benchmarks paired primarily with detector reliability methods belong to robustness and safety assurance.","nearest_boundary":"Data supervision and transfer","representative_topics":["longitudinal error-tolerant evaluation metrics for camera-only 3d object detection","survey of 3d object detection methods for autonomous driving perception","full-degree-of-freedom 3d bounding box evaluation metrics for object detection"]}],"split":"dev"}
{"episode_id":"F045_T2025-07","field_index":45,"field":"3d object detection","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Multi-sensor inference fusion","description":"Improve 3D object detection by combining complementary sensor observations at inference time, especially LiDAR-camera, radar-camera, radar-only, and radar-inclusive multi-modal systems.","inclusion":"Assign atoms whose primary objective is a detector that uses multiple physical sensing modalities, or radar as a core modality, during inference for 3D detection, including feature-, BEV-, query-, proposal-, region-, or box-level fusion and radar-specific modeling.","exclusion":"Camera-only detectors using LiDAR only as training supervision belong to camera-only perception or supervision transfer; reduced-teacher-cost methods without multi-sensor inference belong to data supervision and transfer.","nearest_boundary":"Data supervision and transfer","representative_topics":["multi-scale lidar-camera fusion with depth-aware lifting and fine-grained voxel interaction for 3d object detection","lidar-camera fusion for two-stage 3d object detection using cross-modal attention in regions of interest","sparse center-based feature fusion for camera-lidar bev 3d object detection"]},{"direction_index":1,"label":"D1","name":"Camera-only 3D perception","description":"Estimate 3D object boxes from camera inputs without inference-time LiDAR or radar, covering monocular, stereo, surround-view, multi-camera, and camera-BEV formulations.","inclusion":"Assign atoms whose primary objective is 3D detection from images alone at inference time, including monocular/stereo depth and pose reasoning, pseudo-LiDAR from images, multi-view BEV construction, polar/frustum/query representations, and temporal camera-only BEV reasoning.","exclusion":"Methods using LiDAR, radar, or both as inference sensors belong to multi-sensor inference fusion; methods whose main novelty is privileged-modality teacher supervision rather than camera representation belong to data supervision and transfer.","nearest_boundary":"Multi-sensor inference fusion","representative_topics":["multi-level rgb-pseudo-lidar temporal fusion for per-object depth estimation in monocular 3d detection","ground plane prior for monocular 3d object depth estimation without auxiliary sensors","efficient pseudo-lidar 3d object detection via fast stereo matching and point cloud refinement"]},{"direction_index":2,"label":"D2","name":"LiDAR detector design and optimization","description":"Improve closed-set 3D detection primarily through LiDAR detector architectures, point/voxel/pillar/sparse representations, detection heads, proposal refinement, loss/assignment design, class-imbalance handling, and deployment-oriented efficiency.","inclusion":"Assign atoms whose dominant objective is the single-system detector design itself: LiDAR backbones, sparse or transformer encoders, geometric representations, proposal generation, target assignment, box optimization, cascaded refinement, loss functions, class/long-tail reweighting or sampling, runtime acceleration, edge deployment, and deadline-aware execution.","exclusion":"Temporal multi-frame reasoning belongs to temporal and joint perception; reduced labels, domain transfer, or teacher-student training belong to data supervision and transfer; reliability under corruptions or uncertainty belongs to robustness and safety.","nearest_boundary":"Temporal and joint perception","representative_topics":["sparse proposal-based end-to-end 3d object detection from point clouds using transformers","rotation-decoupled iou optimization for single-stage 3d object detection","full-degree-of-freedom lidar 3d object detection with ground-aware pose estimation on sloped terrain"]},{"direction_index":3,"label":"D3","name":"Data supervision and transfer","description":"Reduce annotation burden or improve transferability of 3D detectors through label-efficient learning, domain adaptation/generalization, and teacher-student knowledge transfer across modalities, frames, or representations.","inclusion":"Assign atoms whose primary objective is learning under limited, weak, unlabeled, shifted, or privileged-supervision regimes, including semi-supervised learning, self-supervised pretraining, pseudo-labeling, few-shot and active learning, cross-domain adaptation, synthetic-to-real transfer, multi-dataset generalization, and knowledge distillation.","exclusion":"If multiple sensors are jointly required at inference, assign to multi-sensor inference fusion; if the objective is measuring datasets or creating annotation tools rather than training a detector, assign to evaluation and annotation infrastructure.","nearest_boundary":"Multi-sensor inference fusion","representative_topics":["voxel masked autoencoding pre-training for sparse automotive lidar 3d object detection","unsupervised region-proposal contrastive pre-training for lidar 3d object detection","semi-supervised pseudo-labeling for image-based 3d object detection under monocular conditions"]},{"direction_index":4,"label":"D4","name":"Temporal and joint perception","description":"Use temporal sequences, tracking-coupled inference, multi-task perception, or cooperative multi-agent context as a primary program for 3D object detection.","inclusion":"Assign atoms whose main objective is exploiting time, motion, multi-frame aggregation, streaming context, spatio-temporal attention, joint detection-tracking, detection plus segmentation or panoptic perception, or collaborative/V2X multi-agent 3D detection.","exclusion":"Single-frame LiDAR architecture or efficiency work belongs to LiDAR detector design and optimization; camera-only static multi-view BEV construction belongs to camera-only 3D perception unless temporal or multi-task unification is the dominant contribution.","nearest_boundary":"LiDAR detector design and optimization","representative_topics":["spatiotemporal transformer and graph message passing for multi-frame lidar point cloud video object detection","temporal voxel feature aggregation for multi-frame 3d object detection using coarse-to-fine transformer fusion","efficient lidar 3d detection via spatial and temporal object relation modeling"]},{"direction_index":5,"label":"D5","name":"Robustness and safety assurance","description":"Assess or improve reliability of 3D detectors under uncertainty, label noise, occlusion, corruptions, adverse conditions, adversarial perturbations, and safety constraints.","inclusion":"Assign atoms whose primary objective is robustness benchmarking, uncertainty quantification, calibrated box prediction, label-uncertainty modeling, safety-constrained localization, corruption robustness, or attack/defense analysis for 3D detection.","exclusion":"Cross-domain source-target transfer belongs to data supervision and transfer; radar-centered weather robustness belongs to multi-sensor inference fusion when radar modeling is the main program.","nearest_boundary":"Data supervision and transfer","representative_topics":["efficient mimo uncertainty estimation for lidar-based 3d object detection under runtime constraints","generative label uncertainty estimation for robust 3d object detection via conditional variational autoencoders","safety-oriented 3d object detection with spatial coverage constraints for autonomous driving"]},{"direction_index":6,"label":"D6","name":"Open-vocabulary and language-guided detection","description":"Enable 3D detection or grounding for novel, unseen, language-specified, or open-set categories by aligning 3D representations with image, text, or foundation-model representations.","inclusion":"Assign atoms whose primary objective is open-vocabulary 3D detection, zero-shot 3D detection, language-guided 3D region grounding, CLIP-style or vision-language transfer for unseen classes, or foundation-model adaptation for category expansion.","exclusion":"Closed-set multi-modal fusion belongs to multi-sensor inference fusion; closed-set teacher-student transfer belongs to data supervision and transfer unless novel-category or language grounding is the central objective.","nearest_boundary":"Data supervision and transfer","representative_topics":["open-vocabulary 3d point cloud detection via cross-modal contrastive learning and image-level class supervision","open-vocabulary 3d point cloud detection via 2d-to-3d transfer and vision-language contrastive learning","multimodal language-guided 3d region grounding in lidar point clouds for autonomous driving"]},{"direction_index":7,"label":"D7","name":"Evaluation and annotation infrastructure","description":"Build measurement, benchmark, dataset, annotation, survey, and validation infrastructure for 3D object detection rather than proposing a detector-training method.","inclusion":"Assign atoms whose primary objective is benchmark construction, evaluation metric design, dataset curation, survey synthesis, annotation tooling, interactive annotation, labeling assistance, or offline validation infrastructure for 3D object detection.","exclusion":"Active learning or pseudo-labeling methods for training detectors belong to data supervision and transfer; robustness benchmarks paired primarily with detector reliability methods belong to robustness and safety assurance.","nearest_boundary":"Data supervision and transfer","representative_topics":["longitudinal error-tolerant evaluation metrics for camera-only 3d object detection","survey of 3d object detection methods for autonomous driving perception","full-degree-of-freedom 3d bounding box evaluation metrics for object detection"]}],"split":"dev"}
{"episode_id":"F045_T2026-01","field_index":45,"field":"3d object detection","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Multi-sensor inference fusion","description":"Improve 3D object detection by combining complementary sensor observations at inference time, especially LiDAR-camera, radar-camera, radar-only, and radar-inclusive multi-modal systems.","inclusion":"Assign atoms whose primary objective is a detector that uses multiple physical sensing modalities, or radar as a core modality, during inference for 3D detection, including feature-, BEV-, query-, proposal-, region-, or box-level fusion and radar-specific modeling.","exclusion":"Camera-only detectors using LiDAR only as training supervision belong to camera-only perception or supervision transfer; reduced-teacher-cost methods without multi-sensor inference belong to data supervision and transfer.","nearest_boundary":"Data supervision and transfer","representative_topics":["multi-scale lidar-camera fusion with depth-aware lifting and fine-grained voxel interaction for 3d object detection","lidar-camera fusion for two-stage 3d object detection using cross-modal attention in regions of interest","sparse center-based feature fusion for camera-lidar bev 3d object detection"]},{"direction_index":1,"label":"D1","name":"Camera-only 3D perception","description":"Estimate 3D object boxes from camera inputs without inference-time LiDAR or radar, covering monocular, stereo, surround-view, multi-camera, and camera-BEV formulations.","inclusion":"Assign atoms whose primary objective is 3D detection from images alone at inference time, including monocular/stereo depth and pose reasoning, pseudo-LiDAR from images, multi-view BEV construction, polar/frustum/query representations, and temporal camera-only BEV reasoning.","exclusion":"Methods using LiDAR, radar, or both as inference sensors belong to multi-sensor inference fusion; methods whose main novelty is privileged-modality teacher supervision rather than camera representation belong to data supervision and transfer.","nearest_boundary":"Multi-sensor inference fusion","representative_topics":["multi-level rgb-pseudo-lidar temporal fusion for per-object depth estimation in monocular 3d detection","ground plane prior for monocular 3d object depth estimation without auxiliary sensors","efficient pseudo-lidar 3d object detection via fast stereo matching and point cloud refinement"]},{"direction_index":2,"label":"D2","name":"LiDAR detector design and optimization","description":"Improve closed-set 3D detection primarily through LiDAR detector architectures, point/voxel/pillar/sparse representations, detection heads, proposal refinement, loss/assignment design, class-imbalance handling, and deployment-oriented efficiency.","inclusion":"Assign atoms whose dominant objective is the single-system detector design itself: LiDAR backbones, sparse or transformer encoders, geometric representations, proposal generation, target assignment, box optimization, cascaded refinement, loss functions, class/long-tail reweighting or sampling, runtime acceleration, edge deployment, and deadline-aware execution.","exclusion":"Temporal multi-frame reasoning belongs to temporal and joint perception; reduced labels, domain transfer, or teacher-student training belong to data supervision and transfer; reliability under corruptions or uncertainty belongs to robustness and safety.","nearest_boundary":"Temporal and joint perception","representative_topics":["sparse proposal-based end-to-end 3d object detection from point clouds using transformers","rotation-decoupled iou optimization for single-stage 3d object detection","full-degree-of-freedom lidar 3d object detection with ground-aware pose estimation on sloped terrain"]},{"direction_index":3,"label":"D3","name":"Data supervision and transfer","description":"Reduce annotation burden or improve transferability of 3D detectors through label-efficient learning, domain adaptation/generalization, and teacher-student knowledge transfer across modalities, frames, or representations.","inclusion":"Assign atoms whose primary objective is learning under limited, weak, unlabeled, shifted, or privileged-supervision regimes, including semi-supervised learning, self-supervised pretraining, pseudo-labeling, few-shot and active learning, cross-domain adaptation, synthetic-to-real transfer, multi-dataset generalization, and knowledge distillation.","exclusion":"If multiple sensors are jointly required at inference, assign to multi-sensor inference fusion; if the objective is measuring datasets or creating annotation tools rather than training a detector, assign to evaluation and annotation infrastructure.","nearest_boundary":"Multi-sensor inference fusion","representative_topics":["voxel masked autoencoding pre-training for sparse automotive lidar 3d object detection","unsupervised region-proposal contrastive pre-training for lidar 3d object detection","semi-supervised pseudo-labeling for image-based 3d object detection under monocular conditions"]},{"direction_index":4,"label":"D4","name":"Temporal and joint perception","description":"Use temporal sequences, tracking-coupled inference, multi-task perception, or cooperative multi-agent context as a primary program for 3D object detection.","inclusion":"Assign atoms whose main objective is exploiting time, motion, multi-frame aggregation, streaming context, spatio-temporal attention, joint detection-tracking, detection plus segmentation or panoptic perception, or collaborative/V2X multi-agent 3D detection.","exclusion":"Single-frame LiDAR architecture or efficiency work belongs to LiDAR detector design and optimization; camera-only static multi-view BEV construction belongs to camera-only 3D perception unless temporal or multi-task unification is the dominant contribution.","nearest_boundary":"LiDAR detector design and optimization","representative_topics":["spatiotemporal transformer and graph message passing for multi-frame lidar point cloud video object detection","temporal voxel feature aggregation for multi-frame 3d object detection using coarse-to-fine transformer fusion","efficient lidar 3d detection via spatial and temporal object relation modeling"]},{"direction_index":5,"label":"D5","name":"Robustness and safety assurance","description":"Assess or improve reliability of 3D detectors under uncertainty, label noise, occlusion, corruptions, adverse conditions, adversarial perturbations, and safety constraints.","inclusion":"Assign atoms whose primary objective is robustness benchmarking, uncertainty quantification, calibrated box prediction, label-uncertainty modeling, safety-constrained localization, corruption robustness, or attack/defense analysis for 3D detection.","exclusion":"Cross-domain source-target transfer belongs to data supervision and transfer; radar-centered weather robustness belongs to multi-sensor inference fusion when radar modeling is the main program.","nearest_boundary":"Data supervision and transfer","representative_topics":["efficient mimo uncertainty estimation for lidar-based 3d object detection under runtime constraints","generative label uncertainty estimation for robust 3d object detection via conditional variational autoencoders","safety-oriented 3d object detection with spatial coverage constraints for autonomous driving"]},{"direction_index":6,"label":"D6","name":"Open-vocabulary and language-guided detection","description":"Enable 3D detection or grounding for novel, unseen, language-specified, or open-set categories by aligning 3D representations with image, text, or foundation-model representations.","inclusion":"Assign atoms whose primary objective is open-vocabulary 3D detection, zero-shot 3D detection, language-guided 3D region grounding, CLIP-style or vision-language transfer for unseen classes, or foundation-model adaptation for category expansion.","exclusion":"Closed-set multi-modal fusion belongs to multi-sensor inference fusion; closed-set teacher-student transfer belongs to data supervision and transfer unless novel-category or language grounding is the central objective.","nearest_boundary":"Data supervision and transfer","representative_topics":["open-vocabulary 3d point cloud detection via cross-modal contrastive learning and image-level class supervision","open-vocabulary 3d point cloud detection via 2d-to-3d transfer and vision-language contrastive learning","multimodal language-guided 3d region grounding in lidar point clouds for autonomous driving"]},{"direction_index":7,"label":"D7","name":"Evaluation and annotation infrastructure","description":"Build measurement, benchmark, dataset, annotation, survey, and validation infrastructure for 3D object detection rather than proposing a detector-training method.","inclusion":"Assign atoms whose primary objective is benchmark construction, evaluation metric design, dataset curation, survey synthesis, annotation tooling, interactive annotation, labeling assistance, or offline validation infrastructure for 3D object detection.","exclusion":"Active learning or pseudo-labeling methods for training detectors belong to data supervision and transfer; robustness benchmarks paired primarily with detector reliability methods belong to robustness and safety assurance.","nearest_boundary":"Data supervision and transfer","representative_topics":["longitudinal error-tolerant evaluation metrics for camera-only 3d object detection","survey of 3d object detection methods for autonomous driving perception","full-degree-of-freedom 3d bounding box evaluation metrics for object detection"]}],"split":"dev"}
{"episode_id":"F047_T2024-01","field_index":47,"field":"robot manipulation","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Contact-Centric Dexterity and Grasping","description":"Acquiring reliable physical contact skills for grasping, picking, tactile refinement, constrained assembly, in-hand manipulation, multi-fingered hands, and bimanual coordination.","inclusion":"Include work whose main objective is to make contact with objects robustly and precisely: grasp-pose prediction, grasp success estimation, tactile or force-guided grasp refinement, dexterous in-hand control, bimanual manipulation, insertion, part mating, and contact-rich execution.","exclusion":"If the main objective is deformable-material modeling, assign to Deformable Object Manipulation; if the main objective is long-horizon sequencing or rearrangement rather than contact execution, assign to Skill Composition and Scene Rearrangement.","nearest_boundary":"Deformable Object Manipulation","representative_topics":["reinforcement learning benchmark for bimanual dexterous manipulation across skill levels and task settings","implicit value function representations for reactive dynamic object grasp planning via sampling-based mpc","simplified proportional control for adaptive grasping using high-sensitivity distributed tactile fingertips"]},{"direction_index":1,"label":"D1","name":"Language and Program Grounding","description":"Grounding natural language, multimodal prompts, foundation models, symbolic structures, or executable programs into manipulation goals, plans, rewards, waypoints, skills, or actions.","inclusion":"Include work whose main objective is enabling manipulation from linguistic, multimodal, neuro-symbolic, programmatic, or foundation-model reasoning signals, including language-conditioned policies, LLM/VLM-guided planning, program synthesis, and symbolic task-and-motion integration.","exclusion":"If the main objective is generic long-horizon skill chaining without language or program structure, assign to Skill Composition and Scene Rearrangement; if language is used only for online human correction in an interactive teaching loop, assign to Demonstration and Human Interaction.","nearest_boundary":"Skill Composition and Scene Rearrangement","representative_topics":["language-conditioned meta-learning for accelerated tool affordance-based manipulation policy adaptation","history-aware transformer policies for instruction-driven robotic manipulation with multi-view generalization","neurosymbolic modular architecture for language-guided visual reasoning in robot manipulation"]},{"direction_index":2,"label":"D2","name":"Demonstration and Human Interaction","description":"Acquiring, adapting, or executing manipulation skills through demonstrations, human videos, interactive teaching, shared autonomy, handovers, collaboration, or human-aware execution.","inclusion":"Include work whose main objective is learning from demonstrations or human involvement, including behavior cloning, human-to-robot video imitation, cross-embodiment transfer, active demonstration elicitation, online correction, human-in-the-loop deployment learning, collaborative assembly, and handover or co-manipulation.","exclusion":"If demonstrations are only a secondary data source for a dexterous, deformable, or language-program objective, assign to that objective; if the core contribution is a reusable dataset or benchmark rather than learning or interaction, assign to Representation, Data, and Evaluation Substrates.","nearest_boundary":"Language and Program Grounding","representative_topics":["one-shot robot manipulation learning from in-the-wild human demonstrations via intent priors","offline-to-online robot policy learning via advantage weighting and early termination","physics-informed differentiable rendering for robot manipulation skill learning from video"]},{"direction_index":3,"label":"D3","name":"Skill Composition and Scene Rearrangement","description":"Solving multi-step manipulation by composing skills, selecting subgoals, planning scene layouts, rearranging objects, generating target poses, or sequencing primitives over extended horizons.","inclusion":"Include work whose main objective is temporal or spatial organization: long-horizon skill decomposition, hierarchical planning, primitive chaining, object rearrangement, room or tabletop scene organization, assembly/rearrangement sequencing, stable reorientation, and generative goal or trajectory synthesis for multi-step manipulation.","exclusion":"If explicit language, symbolic program synthesis, or foundation-model reasoning is the main contribution, assign to Language and Program Grounding; if mobile base coordination is the main difficulty, assign to Embodied Mobility and Motion Control.","nearest_boundary":"Language and Program Grounding","representative_topics":["semantic-guided visual room rearrangement using voxel maps and search policies","dual-arm task planning for non-monotone tabletop object rearrangement with inter-arm handoffs","goal-free object rearrangement via learned target gradient fields for model-based and model-free planning"]},{"direction_index":4,"label":"D4","name":"Deformable Object Manipulation","description":"Perceiving, modeling, and controlling cloth, rope, cables, dough, food, bags, soft materials, and other objects whose changing shape is the central challenge.","inclusion":"Include work whose main objective is manipulation under deformable-object state uncertainty, occlusion, topology, dynamics, or material response, including folding, unfolding, untangling, shaping, routing, deformable state estimation, and deformable dynamics control.","exclusion":"If the target is rigid and the primary challenge is grasping or constrained contact, assign to Contact-Centric Dexterity and Grasping; if deformable manipulation is used mainly to study human collaboration, assign to Demonstration and Human Interaction.","nearest_boundary":"Contact-Centric Dexterity and Grasping","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","bilateral robot cable untangling using rgbd perception and composed manipulation primitives","learning from demonstration for deformable object manipulation balancing exploration and expert guidance"]},{"direction_index":5,"label":"D5","name":"Autonomous Policy Optimization and Dynamics","description":"Learning manipulation behavior through autonomous optimization, reward inference, self-improvement, model-based prediction, world models, dynamics learning, and sample-efficient control.","inclusion":"Include work whose main objective is improving policy acquisition or control through RL-style optimization, reward shaping or reward inference, offline-to-online self-improvement, model-based lookahead, structured world models, learned dynamics, physical priors, or predictive control.","exclusion":"If the primary objective is dexterous/contact hardware skill rather than the optimization or dynamics-learning program, assign to Contact-Centric Dexterity and Grasping; if the main objective is reusable perceptual representation or dataset construction, assign to Representation, Data, and Evaluation Substrates.","nearest_boundary":"Representation, Data, and Evaluation Substrates","representative_topics":["active inference for curiosity-driven exploration in sparse-reward robotic manipulation","decoupled masked world models for visual representation and dynamics learning in robot manipulation","model-based lookahead planning for sample-efficient long-horizon robot manipulation policies"]},{"direction_index":6,"label":"D6","name":"Representation, Data, and Evaluation Substrates","description":"Building reusable perceptual representations, geometric abstractions, multimodal pretraining schemes, generated data, datasets, simulations, and benchmark protocols that support manipulation generalization or evaluation.","inclusion":"Include work whose main objective is the reusable substrate rather than a particular manipulation task: 3D/geometric representation learning, visual or multimodal pretraining, affordance and keypoint representations, synthetic or augmented experience generation, benchmark datasets, simulators, and standardized evaluation resources.","exclusion":"If the learned representation is primarily a dynamics model for planning, assign to Autonomous Policy Optimization and Dynamics; if the generated output is a task plan or rearrangement goal rather than training/evaluation substrate, assign to Skill Composition and Scene Rearrangement.","nearest_boundary":"Autonomous Policy Optimization and Dynamics","representative_topics":["synthetic data generation for scene-aware ambidextrous bin picking via physics-based metaverse","end-to-end visual affordance learning via contact map prediction for generalizable robotic manipulation","se(3)-equivariant unsupervised keypoint learning for generalizable robot manipulation"]},{"direction_index":7,"label":"D7","name":"Embodied Mobility and Motion Control","description":"Coordinating robot embodiment, mobility, trajectory generation, and specialized platform constraints for manipulation-capable systems.","inclusion":"Include work whose main objective is feasible or optimized motion for manipulation embodiments, including mobile manipulation, whole-body or legged manipulation, aerial or space manipulation, base-arm coordination, collision-free manipulator motion planning, inverse kinematics or dynamics control, and specialized non-prehensile or tool-use primitives tied to platform constraints.","exclusion":"If the main objective is task-level rearrangement or skill sequencing rather than embodiment-level motion coordination, assign to Skill Composition and Scene Rearrangement; if the main objective is a grasp choice or contact policy, assign to Contact-Centric Dexterity and Grasping.","nearest_boundary":"Skill Composition and Scene Rearrangement","representative_topics":["decomposed reinforcement learning for mobile manipulation navigation in dynamic environments","reinforcement learning motion planning for free-float dual-arm manipulators tracking non-cooperative spinning objects","data-driven mobile manipulator systems for adaptive household tidying tasks"]}],"split":"dev"}
{"episode_id":"F047_T2024-07","field_index":47,"field":"robot manipulation","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Contact-Centric Dexterity and Grasping","description":"Acquiring reliable physical contact skills for grasping, picking, tactile refinement, constrained assembly, in-hand manipulation, multi-fingered hands, and bimanual coordination.","inclusion":"Include work whose main objective is to make contact with objects robustly and precisely: grasp-pose prediction, grasp success estimation, tactile or force-guided grasp refinement, dexterous in-hand control, bimanual manipulation, insertion, part mating, and contact-rich execution.","exclusion":"If the main objective is deformable-material modeling, assign to Deformable Object Manipulation; if the main objective is long-horizon sequencing or rearrangement rather than contact execution, assign to Skill Composition and Scene Rearrangement.","nearest_boundary":"Deformable Object Manipulation","representative_topics":["reinforcement learning benchmark for bimanual dexterous manipulation across skill levels and task settings","implicit value function representations for reactive dynamic object grasp planning via sampling-based mpc","simplified proportional control for adaptive grasping using high-sensitivity distributed tactile fingertips"]},{"direction_index":1,"label":"D1","name":"Language and Program Grounding","description":"Grounding natural language, multimodal prompts, foundation models, symbolic structures, or executable programs into manipulation goals, plans, rewards, waypoints, skills, or actions.","inclusion":"Include work whose main objective is enabling manipulation from linguistic, multimodal, neuro-symbolic, programmatic, or foundation-model reasoning signals, including language-conditioned policies, LLM/VLM-guided planning, program synthesis, and symbolic task-and-motion integration.","exclusion":"If the main objective is generic long-horizon skill chaining without language or program structure, assign to Skill Composition and Scene Rearrangement; if language is used only for online human correction in an interactive teaching loop, assign to Demonstration and Human Interaction.","nearest_boundary":"Skill Composition and Scene Rearrangement","representative_topics":["language-conditioned meta-learning for accelerated tool affordance-based manipulation policy adaptation","history-aware transformer policies for instruction-driven robotic manipulation with multi-view generalization","neurosymbolic modular architecture for language-guided visual reasoning in robot manipulation"]},{"direction_index":2,"label":"D2","name":"Demonstration and Human Interaction","description":"Acquiring, adapting, or executing manipulation skills through demonstrations, human videos, interactive teaching, shared autonomy, handovers, collaboration, or human-aware execution.","inclusion":"Include work whose main objective is learning from demonstrations or human involvement, including behavior cloning, human-to-robot video imitation, cross-embodiment transfer, active demonstration elicitation, online correction, human-in-the-loop deployment learning, collaborative assembly, and handover or co-manipulation.","exclusion":"If demonstrations are only a secondary data source for a dexterous, deformable, or language-program objective, assign to that objective; if the core contribution is a reusable dataset or benchmark rather than learning or interaction, assign to Representation, Data, and Evaluation Substrates.","nearest_boundary":"Language and Program Grounding","representative_topics":["one-shot robot manipulation learning from in-the-wild human demonstrations via intent priors","offline-to-online robot policy learning via advantage weighting and early termination","physics-informed differentiable rendering for robot manipulation skill learning from video"]},{"direction_index":3,"label":"D3","name":"Skill Composition and Scene Rearrangement","description":"Solving multi-step manipulation by composing skills, selecting subgoals, planning scene layouts, rearranging objects, generating target poses, or sequencing primitives over extended horizons.","inclusion":"Include work whose main objective is temporal or spatial organization: long-horizon skill decomposition, hierarchical planning, primitive chaining, object rearrangement, room or tabletop scene organization, assembly/rearrangement sequencing, stable reorientation, and generative goal or trajectory synthesis for multi-step manipulation.","exclusion":"If explicit language, symbolic program synthesis, or foundation-model reasoning is the main contribution, assign to Language and Program Grounding; if mobile base coordination is the main difficulty, assign to Embodied Mobility and Motion Control.","nearest_boundary":"Language and Program Grounding","representative_topics":["semantic-guided visual room rearrangement using voxel maps and search policies","dual-arm task planning for non-monotone tabletop object rearrangement with inter-arm handoffs","goal-free object rearrangement via learned target gradient fields for model-based and model-free planning"]},{"direction_index":4,"label":"D4","name":"Deformable Object Manipulation","description":"Perceiving, modeling, and controlling cloth, rope, cables, dough, food, bags, soft materials, and other objects whose changing shape is the central challenge.","inclusion":"Include work whose main objective is manipulation under deformable-object state uncertainty, occlusion, topology, dynamics, or material response, including folding, unfolding, untangling, shaping, routing, deformable state estimation, and deformable dynamics control.","exclusion":"If the target is rigid and the primary challenge is grasping or constrained contact, assign to Contact-Centric Dexterity and Grasping; if deformable manipulation is used mainly to study human collaboration, assign to Demonstration and Human Interaction.","nearest_boundary":"Contact-Centric Dexterity and Grasping","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","bilateral robot cable untangling using rgbd perception and composed manipulation primitives","learning from demonstration for deformable object manipulation balancing exploration and expert guidance"]},{"direction_index":5,"label":"D5","name":"Autonomous Policy Optimization and Dynamics","description":"Learning manipulation behavior through autonomous optimization, reward inference, self-improvement, model-based prediction, world models, dynamics learning, and sample-efficient control.","inclusion":"Include work whose main objective is improving policy acquisition or control through RL-style optimization, reward shaping or reward inference, offline-to-online self-improvement, model-based lookahead, structured world models, learned dynamics, physical priors, or predictive control.","exclusion":"If the primary objective is dexterous/contact hardware skill rather than the optimization or dynamics-learning program, assign to Contact-Centric Dexterity and Grasping; if the main objective is reusable perceptual representation or dataset construction, assign to Representation, Data, and Evaluation Substrates.","nearest_boundary":"Representation, Data, and Evaluation Substrates","representative_topics":["active inference for curiosity-driven exploration in sparse-reward robotic manipulation","decoupled masked world models for visual representation and dynamics learning in robot manipulation","model-based lookahead planning for sample-efficient long-horizon robot manipulation policies"]},{"direction_index":6,"label":"D6","name":"Representation, Data, and Evaluation Substrates","description":"Building reusable perceptual representations, geometric abstractions, multimodal pretraining schemes, generated data, datasets, simulations, and benchmark protocols that support manipulation generalization or evaluation.","inclusion":"Include work whose main objective is the reusable substrate rather than a particular manipulation task: 3D/geometric representation learning, visual or multimodal pretraining, affordance and keypoint representations, synthetic or augmented experience generation, benchmark datasets, simulators, and standardized evaluation resources.","exclusion":"If the learned representation is primarily a dynamics model for planning, assign to Autonomous Policy Optimization and Dynamics; if the generated output is a task plan or rearrangement goal rather than training/evaluation substrate, assign to Skill Composition and Scene Rearrangement.","nearest_boundary":"Autonomous Policy Optimization and Dynamics","representative_topics":["synthetic data generation for scene-aware ambidextrous bin picking via physics-based metaverse","end-to-end visual affordance learning via contact map prediction for generalizable robotic manipulation","se(3)-equivariant unsupervised keypoint learning for generalizable robot manipulation"]},{"direction_index":7,"label":"D7","name":"Embodied Mobility and Motion Control","description":"Coordinating robot embodiment, mobility, trajectory generation, and specialized platform constraints for manipulation-capable systems.","inclusion":"Include work whose main objective is feasible or optimized motion for manipulation embodiments, including mobile manipulation, whole-body or legged manipulation, aerial or space manipulation, base-arm coordination, collision-free manipulator motion planning, inverse kinematics or dynamics control, and specialized non-prehensile or tool-use primitives tied to platform constraints.","exclusion":"If the main objective is task-level rearrangement or skill sequencing rather than embodiment-level motion coordination, assign to Skill Composition and Scene Rearrangement; if the main objective is a grasp choice or contact policy, assign to Contact-Centric Dexterity and Grasping.","nearest_boundary":"Skill Composition and Scene Rearrangement","representative_topics":["decomposed reinforcement learning for mobile manipulation navigation in dynamic environments","reinforcement learning motion planning for free-float dual-arm manipulators tracking non-cooperative spinning objects","data-driven mobile manipulator systems for adaptive household tidying tasks"]}],"split":"dev"}
{"episode_id":"F047_T2025-01","field_index":47,"field":"robot manipulation","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Contact-Centric Dexterity and Grasping","description":"Acquiring reliable physical contact skills for grasping, picking, tactile refinement, constrained assembly, in-hand manipulation, multi-fingered hands, and bimanual coordination.","inclusion":"Include work whose main objective is to make contact with objects robustly and precisely: grasp-pose prediction, grasp success estimation, tactile or force-guided grasp refinement, dexterous in-hand control, bimanual manipulation, insertion, part mating, and contact-rich execution.","exclusion":"If the main objective is deformable-material modeling, assign to Deformable Object Manipulation; if the main objective is long-horizon sequencing or rearrangement rather than contact execution, assign to Skill Composition and Scene Rearrangement.","nearest_boundary":"Deformable Object Manipulation","representative_topics":["reinforcement learning benchmark for bimanual dexterous manipulation across skill levels and task settings","implicit value function representations for reactive dynamic object grasp planning via sampling-based mpc","simplified proportional control for adaptive grasping using high-sensitivity distributed tactile fingertips"]},{"direction_index":1,"label":"D1","name":"Language and Program Grounding","description":"Grounding natural language, multimodal prompts, foundation models, symbolic structures, or executable programs into manipulation goals, plans, rewards, waypoints, skills, or actions.","inclusion":"Include work whose main objective is enabling manipulation from linguistic, multimodal, neuro-symbolic, programmatic, or foundation-model reasoning signals, including language-conditioned policies, LLM/VLM-guided planning, program synthesis, and symbolic task-and-motion integration.","exclusion":"If the main objective is generic long-horizon skill chaining without language or program structure, assign to Skill Composition and Scene Rearrangement; if language is used only for online human correction in an interactive teaching loop, assign to Demonstration and Human Interaction.","nearest_boundary":"Skill Composition and Scene Rearrangement","representative_topics":["language-conditioned meta-learning for accelerated tool affordance-based manipulation policy adaptation","history-aware transformer policies for instruction-driven robotic manipulation with multi-view generalization","neurosymbolic modular architecture for language-guided visual reasoning in robot manipulation"]},{"direction_index":2,"label":"D2","name":"Demonstration and Human Interaction","description":"Acquiring, adapting, or executing manipulation skills through demonstrations, human videos, interactive teaching, shared autonomy, handovers, collaboration, or human-aware execution.","inclusion":"Include work whose main objective is learning from demonstrations or human involvement, including behavior cloning, human-to-robot video imitation, cross-embodiment transfer, active demonstration elicitation, online correction, human-in-the-loop deployment learning, collaborative assembly, and handover or co-manipulation.","exclusion":"If demonstrations are only a secondary data source for a dexterous, deformable, or language-program objective, assign to that objective; if the core contribution is a reusable dataset or benchmark rather than learning or interaction, assign to Representation, Data, and Evaluation Substrates.","nearest_boundary":"Language and Program Grounding","representative_topics":["one-shot robot manipulation learning from in-the-wild human demonstrations via intent priors","offline-to-online robot policy learning via advantage weighting and early termination","physics-informed differentiable rendering for robot manipulation skill learning from video"]},{"direction_index":3,"label":"D3","name":"Skill Composition and Scene Rearrangement","description":"Solving multi-step manipulation by composing skills, selecting subgoals, planning scene layouts, rearranging objects, generating target poses, or sequencing primitives over extended horizons.","inclusion":"Include work whose main objective is temporal or spatial organization: long-horizon skill decomposition, hierarchical planning, primitive chaining, object rearrangement, room or tabletop scene organization, assembly/rearrangement sequencing, stable reorientation, and generative goal or trajectory synthesis for multi-step manipulation.","exclusion":"If explicit language, symbolic program synthesis, or foundation-model reasoning is the main contribution, assign to Language and Program Grounding; if mobile base coordination is the main difficulty, assign to Embodied Mobility and Motion Control.","nearest_boundary":"Language and Program Grounding","representative_topics":["semantic-guided visual room rearrangement using voxel maps and search policies","dual-arm task planning for non-monotone tabletop object rearrangement with inter-arm handoffs","goal-free object rearrangement via learned target gradient fields for model-based and model-free planning"]},{"direction_index":4,"label":"D4","name":"Deformable Object Manipulation","description":"Perceiving, modeling, and controlling cloth, rope, cables, dough, food, bags, soft materials, and other objects whose changing shape is the central challenge.","inclusion":"Include work whose main objective is manipulation under deformable-object state uncertainty, occlusion, topology, dynamics, or material response, including folding, unfolding, untangling, shaping, routing, deformable state estimation, and deformable dynamics control.","exclusion":"If the target is rigid and the primary challenge is grasping or constrained contact, assign to Contact-Centric Dexterity and Grasping; if deformable manipulation is used mainly to study human collaboration, assign to Demonstration and Human Interaction.","nearest_boundary":"Contact-Centric Dexterity and Grasping","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","bilateral robot cable untangling using rgbd perception and composed manipulation primitives","learning from demonstration for deformable object manipulation balancing exploration and expert guidance"]},{"direction_index":5,"label":"D5","name":"Autonomous Policy Optimization and Dynamics","description":"Learning manipulation behavior through autonomous optimization, reward inference, self-improvement, model-based prediction, world models, dynamics learning, and sample-efficient control.","inclusion":"Include work whose main objective is improving policy acquisition or control through RL-style optimization, reward shaping or reward inference, offline-to-online self-improvement, model-based lookahead, structured world models, learned dynamics, physical priors, or predictive control.","exclusion":"If the primary objective is dexterous/contact hardware skill rather than the optimization or dynamics-learning program, assign to Contact-Centric Dexterity and Grasping; if the main objective is reusable perceptual representation or dataset construction, assign to Representation, Data, and Evaluation Substrates.","nearest_boundary":"Representation, Data, and Evaluation Substrates","representative_topics":["active inference for curiosity-driven exploration in sparse-reward robotic manipulation","decoupled masked world models for visual representation and dynamics learning in robot manipulation","model-based lookahead planning for sample-efficient long-horizon robot manipulation policies"]},{"direction_index":6,"label":"D6","name":"Representation, Data, and Evaluation Substrates","description":"Building reusable perceptual representations, geometric abstractions, multimodal pretraining schemes, generated data, datasets, simulations, and benchmark protocols that support manipulation generalization or evaluation.","inclusion":"Include work whose main objective is the reusable substrate rather than a particular manipulation task: 3D/geometric representation learning, visual or multimodal pretraining, affordance and keypoint representations, synthetic or augmented experience generation, benchmark datasets, simulators, and standardized evaluation resources.","exclusion":"If the learned representation is primarily a dynamics model for planning, assign to Autonomous Policy Optimization and Dynamics; if the generated output is a task plan or rearrangement goal rather than training/evaluation substrate, assign to Skill Composition and Scene Rearrangement.","nearest_boundary":"Autonomous Policy Optimization and Dynamics","representative_topics":["synthetic data generation for scene-aware ambidextrous bin picking via physics-based metaverse","end-to-end visual affordance learning via contact map prediction for generalizable robotic manipulation","se(3)-equivariant unsupervised keypoint learning for generalizable robot manipulation"]},{"direction_index":7,"label":"D7","name":"Embodied Mobility and Motion Control","description":"Coordinating robot embodiment, mobility, trajectory generation, and specialized platform constraints for manipulation-capable systems.","inclusion":"Include work whose main objective is feasible or optimized motion for manipulation embodiments, including mobile manipulation, whole-body or legged manipulation, aerial or space manipulation, base-arm coordination, collision-free manipulator motion planning, inverse kinematics or dynamics control, and specialized non-prehensile or tool-use primitives tied to platform constraints.","exclusion":"If the main objective is task-level rearrangement or skill sequencing rather than embodiment-level motion coordination, assign to Skill Composition and Scene Rearrangement; if the main objective is a grasp choice or contact policy, assign to Contact-Centric Dexterity and Grasping.","nearest_boundary":"Skill Composition and Scene Rearrangement","representative_topics":["decomposed reinforcement learning for mobile manipulation navigation in dynamic environments","reinforcement learning motion planning for free-float dual-arm manipulators tracking non-cooperative spinning objects","data-driven mobile manipulator systems for adaptive household tidying tasks"]}],"split":"dev"}
{"episode_id":"F047_T2025-07","field_index":47,"field":"robot manipulation","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Contact-Centric Dexterity and Grasping","description":"Acquiring reliable physical contact skills for grasping, picking, tactile refinement, constrained assembly, in-hand manipulation, multi-fingered hands, and bimanual coordination.","inclusion":"Include work whose main objective is to make contact with objects robustly and precisely: grasp-pose prediction, grasp success estimation, tactile or force-guided grasp refinement, dexterous in-hand control, bimanual manipulation, insertion, part mating, and contact-rich execution.","exclusion":"If the main objective is deformable-material modeling, assign to Deformable Object Manipulation; if the main objective is long-horizon sequencing or rearrangement rather than contact execution, assign to Skill Composition and Scene Rearrangement.","nearest_boundary":"Deformable Object Manipulation","representative_topics":["reinforcement learning benchmark for bimanual dexterous manipulation across skill levels and task settings","implicit value function representations for reactive dynamic object grasp planning via sampling-based mpc","simplified proportional control for adaptive grasping using high-sensitivity distributed tactile fingertips"]},{"direction_index":1,"label":"D1","name":"Language and Program Grounding","description":"Grounding natural language, multimodal prompts, foundation models, symbolic structures, or executable programs into manipulation goals, plans, rewards, waypoints, skills, or actions.","inclusion":"Include work whose main objective is enabling manipulation from linguistic, multimodal, neuro-symbolic, programmatic, or foundation-model reasoning signals, including language-conditioned policies, LLM/VLM-guided planning, program synthesis, and symbolic task-and-motion integration.","exclusion":"If the main objective is generic long-horizon skill chaining without language or program structure, assign to Skill Composition and Scene Rearrangement; if language is used only for online human correction in an interactive teaching loop, assign to Demonstration and Human Interaction.","nearest_boundary":"Skill Composition and Scene Rearrangement","representative_topics":["language-conditioned meta-learning for accelerated tool affordance-based manipulation policy adaptation","history-aware transformer policies for instruction-driven robotic manipulation with multi-view generalization","neurosymbolic modular architecture for language-guided visual reasoning in robot manipulation"]},{"direction_index":2,"label":"D2","name":"Demonstration and Human Interaction","description":"Acquiring, adapting, or executing manipulation skills through demonstrations, human videos, interactive teaching, shared autonomy, handovers, collaboration, or human-aware execution.","inclusion":"Include work whose main objective is learning from demonstrations or human involvement, including behavior cloning, human-to-robot video imitation, cross-embodiment transfer, active demonstration elicitation, online correction, human-in-the-loop deployment learning, collaborative assembly, and handover or co-manipulation.","exclusion":"If demonstrations are only a secondary data source for a dexterous, deformable, or language-program objective, assign to that objective; if the core contribution is a reusable dataset or benchmark rather than learning or interaction, assign to Representation, Data, and Evaluation Substrates.","nearest_boundary":"Language and Program Grounding","representative_topics":["one-shot robot manipulation learning from in-the-wild human demonstrations via intent priors","offline-to-online robot policy learning via advantage weighting and early termination","physics-informed differentiable rendering for robot manipulation skill learning from video"]},{"direction_index":3,"label":"D3","name":"Skill Composition and Scene Rearrangement","description":"Solving multi-step manipulation by composing skills, selecting subgoals, planning scene layouts, rearranging objects, generating target poses, or sequencing primitives over extended horizons.","inclusion":"Include work whose main objective is temporal or spatial organization: long-horizon skill decomposition, hierarchical planning, primitive chaining, object rearrangement, room or tabletop scene organization, assembly/rearrangement sequencing, stable reorientation, and generative goal or trajectory synthesis for multi-step manipulation.","exclusion":"If explicit language, symbolic program synthesis, or foundation-model reasoning is the main contribution, assign to Language and Program Grounding; if mobile base coordination is the main difficulty, assign to Embodied Mobility and Motion Control.","nearest_boundary":"Language and Program Grounding","representative_topics":["semantic-guided visual room rearrangement using voxel maps and search policies","dual-arm task planning for non-monotone tabletop object rearrangement with inter-arm handoffs","goal-free object rearrangement via learned target gradient fields for model-based and model-free planning"]},{"direction_index":4,"label":"D4","name":"Deformable Object Manipulation","description":"Perceiving, modeling, and controlling cloth, rope, cables, dough, food, bags, soft materials, and other objects whose changing shape is the central challenge.","inclusion":"Include work whose main objective is manipulation under deformable-object state uncertainty, occlusion, topology, dynamics, or material response, including folding, unfolding, untangling, shaping, routing, deformable state estimation, and deformable dynamics control.","exclusion":"If the target is rigid and the primary challenge is grasping or constrained contact, assign to Contact-Centric Dexterity and Grasping; if deformable manipulation is used mainly to study human collaboration, assign to Demonstration and Human Interaction.","nearest_boundary":"Contact-Centric Dexterity and Grasping","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","bilateral robot cable untangling using rgbd perception and composed manipulation primitives","learning from demonstration for deformable object manipulation balancing exploration and expert guidance"]},{"direction_index":5,"label":"D5","name":"Autonomous Policy Optimization and Dynamics","description":"Learning manipulation behavior through autonomous optimization, reward inference, self-improvement, model-based prediction, world models, dynamics learning, and sample-efficient control.","inclusion":"Include work whose main objective is improving policy acquisition or control through RL-style optimization, reward shaping or reward inference, offline-to-online self-improvement, model-based lookahead, structured world models, learned dynamics, physical priors, or predictive control.","exclusion":"If the primary objective is dexterous/contact hardware skill rather than the optimization or dynamics-learning program, assign to Contact-Centric Dexterity and Grasping; if the main objective is reusable perceptual representation or dataset construction, assign to Representation, Data, and Evaluation Substrates.","nearest_boundary":"Representation, Data, and Evaluation Substrates","representative_topics":["active inference for curiosity-driven exploration in sparse-reward robotic manipulation","decoupled masked world models for visual representation and dynamics learning in robot manipulation","model-based lookahead planning for sample-efficient long-horizon robot manipulation policies"]},{"direction_index":6,"label":"D6","name":"Representation, Data, and Evaluation Substrates","description":"Building reusable perceptual representations, geometric abstractions, multimodal pretraining schemes, generated data, datasets, simulations, and benchmark protocols that support manipulation generalization or evaluation.","inclusion":"Include work whose main objective is the reusable substrate rather than a particular manipulation task: 3D/geometric representation learning, visual or multimodal pretraining, affordance and keypoint representations, synthetic or augmented experience generation, benchmark datasets, simulators, and standardized evaluation resources.","exclusion":"If the learned representation is primarily a dynamics model for planning, assign to Autonomous Policy Optimization and Dynamics; if the generated output is a task plan or rearrangement goal rather than training/evaluation substrate, assign to Skill Composition and Scene Rearrangement.","nearest_boundary":"Autonomous Policy Optimization and Dynamics","representative_topics":["synthetic data generation for scene-aware ambidextrous bin picking via physics-based metaverse","end-to-end visual affordance learning via contact map prediction for generalizable robotic manipulation","se(3)-equivariant unsupervised keypoint learning for generalizable robot manipulation"]},{"direction_index":7,"label":"D7","name":"Embodied Mobility and Motion Control","description":"Coordinating robot embodiment, mobility, trajectory generation, and specialized platform constraints for manipulation-capable systems.","inclusion":"Include work whose main objective is feasible or optimized motion for manipulation embodiments, including mobile manipulation, whole-body or legged manipulation, aerial or space manipulation, base-arm coordination, collision-free manipulator motion planning, inverse kinematics or dynamics control, and specialized non-prehensile or tool-use primitives tied to platform constraints.","exclusion":"If the main objective is task-level rearrangement or skill sequencing rather than embodiment-level motion coordination, assign to Skill Composition and Scene Rearrangement; if the main objective is a grasp choice or contact policy, assign to Contact-Centric Dexterity and Grasping.","nearest_boundary":"Skill Composition and Scene Rearrangement","representative_topics":["decomposed reinforcement learning for mobile manipulation navigation in dynamic environments","reinforcement learning motion planning for free-float dual-arm manipulators tracking non-cooperative spinning objects","data-driven mobile manipulator systems for adaptive household tidying tasks"]}],"split":"dev"}
{"episode_id":"F047_T2026-01","field_index":47,"field":"robot manipulation","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Contact-Centric Dexterity and Grasping","description":"Acquiring reliable physical contact skills for grasping, picking, tactile refinement, constrained assembly, in-hand manipulation, multi-fingered hands, and bimanual coordination.","inclusion":"Include work whose main objective is to make contact with objects robustly and precisely: grasp-pose prediction, grasp success estimation, tactile or force-guided grasp refinement, dexterous in-hand control, bimanual manipulation, insertion, part mating, and contact-rich execution.","exclusion":"If the main objective is deformable-material modeling, assign to Deformable Object Manipulation; if the main objective is long-horizon sequencing or rearrangement rather than contact execution, assign to Skill Composition and Scene Rearrangement.","nearest_boundary":"Deformable Object Manipulation","representative_topics":["reinforcement learning benchmark for bimanual dexterous manipulation across skill levels and task settings","implicit value function representations for reactive dynamic object grasp planning via sampling-based mpc","simplified proportional control for adaptive grasping using high-sensitivity distributed tactile fingertips"]},{"direction_index":1,"label":"D1","name":"Language and Program Grounding","description":"Grounding natural language, multimodal prompts, foundation models, symbolic structures, or executable programs into manipulation goals, plans, rewards, waypoints, skills, or actions.","inclusion":"Include work whose main objective is enabling manipulation from linguistic, multimodal, neuro-symbolic, programmatic, or foundation-model reasoning signals, including language-conditioned policies, LLM/VLM-guided planning, program synthesis, and symbolic task-and-motion integration.","exclusion":"If the main objective is generic long-horizon skill chaining without language or program structure, assign to Skill Composition and Scene Rearrangement; if language is used only for online human correction in an interactive teaching loop, assign to Demonstration and Human Interaction.","nearest_boundary":"Skill Composition and Scene Rearrangement","representative_topics":["language-conditioned meta-learning for accelerated tool affordance-based manipulation policy adaptation","history-aware transformer policies for instruction-driven robotic manipulation with multi-view generalization","neurosymbolic modular architecture for language-guided visual reasoning in robot manipulation"]},{"direction_index":2,"label":"D2","name":"Demonstration and Human Interaction","description":"Acquiring, adapting, or executing manipulation skills through demonstrations, human videos, interactive teaching, shared autonomy, handovers, collaboration, or human-aware execution.","inclusion":"Include work whose main objective is learning from demonstrations or human involvement, including behavior cloning, human-to-robot video imitation, cross-embodiment transfer, active demonstration elicitation, online correction, human-in-the-loop deployment learning, collaborative assembly, and handover or co-manipulation.","exclusion":"If demonstrations are only a secondary data source for a dexterous, deformable, or language-program objective, assign to that objective; if the core contribution is a reusable dataset or benchmark rather than learning or interaction, assign to Representation, Data, and Evaluation Substrates.","nearest_boundary":"Language and Program Grounding","representative_topics":["one-shot robot manipulation learning from in-the-wild human demonstrations via intent priors","offline-to-online robot policy learning via advantage weighting and early termination","physics-informed differentiable rendering for robot manipulation skill learning from video"]},{"direction_index":3,"label":"D3","name":"Skill Composition and Scene Rearrangement","description":"Solving multi-step manipulation by composing skills, selecting subgoals, planning scene layouts, rearranging objects, generating target poses, or sequencing primitives over extended horizons.","inclusion":"Include work whose main objective is temporal or spatial organization: long-horizon skill decomposition, hierarchical planning, primitive chaining, object rearrangement, room or tabletop scene organization, assembly/rearrangement sequencing, stable reorientation, and generative goal or trajectory synthesis for multi-step manipulation.","exclusion":"If explicit language, symbolic program synthesis, or foundation-model reasoning is the main contribution, assign to Language and Program Grounding; if mobile base coordination is the main difficulty, assign to Embodied Mobility and Motion Control.","nearest_boundary":"Language and Program Grounding","representative_topics":["semantic-guided visual room rearrangement using voxel maps and search policies","dual-arm task planning for non-monotone tabletop object rearrangement with inter-arm handoffs","goal-free object rearrangement via learned target gradient fields for model-based and model-free planning"]},{"direction_index":4,"label":"D4","name":"Deformable Object Manipulation","description":"Perceiving, modeling, and controlling cloth, rope, cables, dough, food, bags, soft materials, and other objects whose changing shape is the central challenge.","inclusion":"Include work whose main objective is manipulation under deformable-object state uncertainty, occlusion, topology, dynamics, or material response, including folding, unfolding, untangling, shaping, routing, deformable state estimation, and deformable dynamics control.","exclusion":"If the target is rigid and the primary challenge is grasping or constrained contact, assign to Contact-Centric Dexterity and Grasping; if deformable manipulation is used mainly to study human collaboration, assign to Demonstration and Human Interaction.","nearest_boundary":"Contact-Centric Dexterity and Grasping","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","bilateral robot cable untangling using rgbd perception and composed manipulation primitives","learning from demonstration for deformable object manipulation balancing exploration and expert guidance"]},{"direction_index":5,"label":"D5","name":"Autonomous Policy Optimization and Dynamics","description":"Learning manipulation behavior through autonomous optimization, reward inference, self-improvement, model-based prediction, world models, dynamics learning, and sample-efficient control.","inclusion":"Include work whose main objective is improving policy acquisition or control through RL-style optimization, reward shaping or reward inference, offline-to-online self-improvement, model-based lookahead, structured world models, learned dynamics, physical priors, or predictive control.","exclusion":"If the primary objective is dexterous/contact hardware skill rather than the optimization or dynamics-learning program, assign to Contact-Centric Dexterity and Grasping; if the main objective is reusable perceptual representation or dataset construction, assign to Representation, Data, and Evaluation Substrates.","nearest_boundary":"Representation, Data, and Evaluation Substrates","representative_topics":["active inference for curiosity-driven exploration in sparse-reward robotic manipulation","decoupled masked world models for visual representation and dynamics learning in robot manipulation","model-based lookahead planning for sample-efficient long-horizon robot manipulation policies"]},{"direction_index":6,"label":"D6","name":"Representation, Data, and Evaluation Substrates","description":"Building reusable perceptual representations, geometric abstractions, multimodal pretraining schemes, generated data, datasets, simulations, and benchmark protocols that support manipulation generalization or evaluation.","inclusion":"Include work whose main objective is the reusable substrate rather than a particular manipulation task: 3D/geometric representation learning, visual or multimodal pretraining, affordance and keypoint representations, synthetic or augmented experience generation, benchmark datasets, simulators, and standardized evaluation resources.","exclusion":"If the learned representation is primarily a dynamics model for planning, assign to Autonomous Policy Optimization and Dynamics; if the generated output is a task plan or rearrangement goal rather than training/evaluation substrate, assign to Skill Composition and Scene Rearrangement.","nearest_boundary":"Autonomous Policy Optimization and Dynamics","representative_topics":["synthetic data generation for scene-aware ambidextrous bin picking via physics-based metaverse","end-to-end visual affordance learning via contact map prediction for generalizable robotic manipulation","se(3)-equivariant unsupervised keypoint learning for generalizable robot manipulation"]},{"direction_index":7,"label":"D7","name":"Embodied Mobility and Motion Control","description":"Coordinating robot embodiment, mobility, trajectory generation, and specialized platform constraints for manipulation-capable systems.","inclusion":"Include work whose main objective is feasible or optimized motion for manipulation embodiments, including mobile manipulation, whole-body or legged manipulation, aerial or space manipulation, base-arm coordination, collision-free manipulator motion planning, inverse kinematics or dynamics control, and specialized non-prehensile or tool-use primitives tied to platform constraints.","exclusion":"If the main objective is task-level rearrangement or skill sequencing rather than embodiment-level motion coordination, assign to Skill Composition and Scene Rearrangement; if the main objective is a grasp choice or contact policy, assign to Contact-Centric Dexterity and Grasping.","nearest_boundary":"Skill Composition and Scene Rearrangement","representative_topics":["decomposed reinforcement learning for mobile manipulation navigation in dynamic environments","reinforcement learning motion planning for free-float dual-arm manipulators tracking non-cooperative spinning objects","data-driven mobile manipulator systems for adaptive household tidying tasks"]}],"split":"dev"}
{"episode_id":"F048_T2024-01","field_index":48,"field":"differential privacy","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Private neural training and transfer","description":"Improving differentially private model training and adaptation of pretrained or public-data-supported models for accuracy, stability, scalability, and sample efficiency.","inclusion":"Include work whose main objective is to train or fine-tune predictive models under DP by modifying DP-SGD-style training, clipping, noise, compression, pruning, optimizers, hardware/software systems, public-data use, self-supervised or transfer pretraining, feature reuse, adapters, layer selection, or parameter-efficient fine-tuning.","exclusion":"If the primary output is a synthetic dataset rather than an adapted predictor, use private synthetic data generation; if the main result is an abstract optimization rate rather than a neural training or transfer method, use private optimization and sequential decision theory.","nearest_boundary":"Private synthetic data generation","representative_topics":["automatic gradient clipping for differentially private deep learning without hyperparameter tuning","low-rank sparse gradient updates for differentially private large-scale neural network training","hardware accelerator design for differentially private machine learning training workloads"]},{"direction_index":1,"label":"D1","name":"Private synthetic data generation","description":"Constructing differentially private synthetic data or private generative models that preserve distributional, statistical, or downstream task utility.","inclusion":"Include methods whose primary deliverable is a DP synthetic dataset, generator, data condensation mechanism, distribution summarizer, marginal or geometric synthesizer, modality-specific private generator, or post-processing method for improving synthetic-data utility.","exclusion":"If a pretrained generative model is fine-tuned mainly as a predictor or representation learner rather than to release synthetic data, use private neural training and transfer; if the result is a general estimator or inferential procedure, use private statistical estimation and query release.","nearest_boundary":"Private neural training and transfer","representative_topics":["dataset condensation as differentially private data generation for efficient privacy-preserving training","differentially private synthetic data generation for hierarchical query release with group-individual statistics","differentially private conditional gan for imbalanced tabular data with variable dependency capture"]},{"direction_index":2,"label":"D2","name":"Private federated learning","description":"Designing differentially private learning protocols over multiple clients, silos, devices, organizations, or parties.","inclusion":"Include work where the central objective is a federated or distributed learning protocol under DP, including private aggregation, local or user-level privacy, communication-efficient updates, quantization, secure aggregation, personalization, vertical or horizontal federated settings, and federated applications when the protocol is the main contribution.","exclusion":"If the distributed structure is mainly an abstract optimization topology without a federated learning protocol, use private optimization and sequential decision theory; if the performance metric is bandit regret or adaptive decision quality, use private optimization and sequential decision theory.","nearest_boundary":"Private optimization and sequential decision theory","representative_topics":["differential privacy and personalization in cross-silo federated learning with silo-specific sample-level constraints","local differential privacy federated learning via selective quantization and gradient subsampling","federated speech recognition via differential privacy for multi-organization asr"]},{"direction_index":3,"label":"D3","name":"Private optimization and sequential decision theory","description":"Developing and analyzing DP algorithms for abstract optimization objectives and adaptive decision processes, with rates, regret, convergence, or policy performance as the main scientific target.","inclusion":"Include private convex, nonconvex, minimax, saddle-point, distributed, online, heavy-tailed, user-level, and ERM/SCO optimization theory, as well as online learning, bandits, pricing, reinforcement learning, and sequential prediction where DP affects convergence, regret, sample complexity, or policy value.","exclusion":"If the contribution is primarily an empirical or systems improvement to neural DP-SGD training, use private neural training and transfer; if the optimized object is a specific statistical estimand rather than a general objective class or sequential decision problem, use private statistical estimation and query release.","nearest_boundary":"Private neural training and transfer","representative_topics":["differentially private stochastic minimax optimization with near-optimal privacy-loss trade-off via black-box algorithm framework","differentially private optimization with faster stationary point convergence rates in finite-sum and stochastic settings","dimension-independent generalization bounds for differentially private sgd via langevin convergence"]},{"direction_index":4,"label":"D4","name":"Private statistical estimation and query release","description":"Estimating statistical quantities, fitting structured statistical models, conducting inference, or releasing reusable analytic summaries under differential privacy.","inclusion":"Include private means, covariances, regression, clustering, matrix approximation, dependency measures, Bayesian or MCMC inference, label-private estimation, Shapley values, histograms, sketches, query answering, continual counting, and other statistic-release workloads when the main objective is accurate estimation or release of analytic summaries.","exclusion":"If the primary result is a broad learnability lower bound or optimality theorem independent of a concrete estimator or release workload, use privacy guarantees and foundations; if the output is a synthetic dataset intended to stand in for the original data, use private synthetic data generation.","nearest_boundary":"Privacy guarantees and foundations","representative_topics":["differentially private shapley value approximation via stratified sampling for data evaluation","user contribution bounding algorithms for differential privacy-preserving histogram estimation","differentially private multi-party linear regression with disjoint attribute ownership and eigenvalue-constrained mechanisms"]},{"direction_index":5,"label":"D5","name":"Privacy guarantees and foundations","description":"Characterizing, tightening, composing, amplifying, or proving fundamental limits of differential privacy guarantees and mechanisms.","inclusion":"Include privacy accountants, composition theorems, Rényi/Gaussian/zCDP analysis, privacy-loss distributions, amplification by subsampling, shuffling, iteration or decentralization, optimal mechanism analysis, learnability characterizations, sample-complexity separations, lower bounds, private sampling theory, and foundational mechanism results.","exclusion":"If privacy accounting only supports a specific learning, estimation, generation, or release algorithm, assign by that algorithm's primary objective; if the main goal is attack-driven leakage evaluation rather than formal guarantee derivation, use privacy risk, fairness, and deployment.","nearest_boundary":"Privacy risk, fairness, and deployment","representative_topics":["individual example privacy accounting in differentially private stochastic gradient descent","privacy amplification for distributed computation via shuffled check-in protocols and rényi differential privacy","edgeworth expansion for tight non-asymptotic differential privacy composition bounds"]},{"direction_index":6,"label":"D6","name":"Privacy risk, fairness, and deployment","description":"Assessing and improving the real-world behavior of DP systems with respect to leakage risk, attacks, robustness, fairness, calibration, utility, and practitioner usability.","inclusion":"Include empirical privacy auditing, membership inference, reconstruction or vote leakage, gradient leakage, attack robustness, privacy profiling, comparisons between DP and non-DP defenses, fairness and disparate impact of DP, calibration, selective prediction, hyperparameter or epsilon utility studies, practitioner guidance, and deployment-oriented evaluations.","exclusion":"If the main contribution is a formal privacy accountant or composition theorem, use privacy guarantees and foundations; if fairness or robustness is only a secondary metric for a new optimizer, generator, or federated protocol, assign to that primary algorithmic program.","nearest_boundary":"Privacy guarantees and foundations","representative_topics":["privacy guarantees and membership inference robustness of generative adversarial network outputs","epsilon-utility tradeoff analysis for practical differential privacy in private empirical risk minimization","accuracy-fairness tradeoffs in private learning under long-tailed data distributions"]},{"direction_index":7,"label":"D7","name":"Structured and domain-specific private mechanisms","description":"Designing DP mechanisms for data objects or domains whose structure defines the scientific problem, including graphs, language, relational structures, and specialized decision or interaction settings.","inclusion":"Include private graph learning, GNNs, graph embeddings, graph statistics, stochastic block models when graph structure is central, text rewriting, sanitization, de-identification, syntax-aware privatization, private text hashing or similarity, and specialized mechanisms such as voting, counterfactual explanations, contact tracing, secure multiplication, anomaly detection, swarm intelligence, or quantum ML when the domain structure determines the mechanism design.","exclusion":"If the graph is merely a communication network for federated or distributed training, use private federated learning; if a language model is fine-tuned mainly as a pretrained predictor rather than to transform or privatize text, use private neural training and transfer; if the specialized mechanism is primarily a general statistical estimator, use private statistical estimation and query release.","nearest_boundary":"Private statistical estimation and query release","representative_topics":["epsilon-differential privacy text sanitization with customized per-token token-level protection","differentially private personalized pagerank with bounded sensitivity for graph learning","open-source framework for reproducible and transparent differentially private text rewriting"]}],"split":"dev"}
{"episode_id":"F048_T2024-07","field_index":48,"field":"differential privacy","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Private neural training and transfer","description":"Improving differentially private model training and adaptation of pretrained or public-data-supported models for accuracy, stability, scalability, and sample efficiency.","inclusion":"Include work whose main objective is to train or fine-tune predictive models under DP by modifying DP-SGD-style training, clipping, noise, compression, pruning, optimizers, hardware/software systems, public-data use, self-supervised or transfer pretraining, feature reuse, adapters, layer selection, or parameter-efficient fine-tuning.","exclusion":"If the primary output is a synthetic dataset rather than an adapted predictor, use private synthetic data generation; if the main result is an abstract optimization rate rather than a neural training or transfer method, use private optimization and sequential decision theory.","nearest_boundary":"Private synthetic data generation","representative_topics":["automatic gradient clipping for differentially private deep learning without hyperparameter tuning","low-rank sparse gradient updates for differentially private large-scale neural network training","hardware accelerator design for differentially private machine learning training workloads"]},{"direction_index":1,"label":"D1","name":"Private synthetic data generation","description":"Constructing differentially private synthetic data or private generative models that preserve distributional, statistical, or downstream task utility.","inclusion":"Include methods whose primary deliverable is a DP synthetic dataset, generator, data condensation mechanism, distribution summarizer, marginal or geometric synthesizer, modality-specific private generator, or post-processing method for improving synthetic-data utility.","exclusion":"If a pretrained generative model is fine-tuned mainly as a predictor or representation learner rather than to release synthetic data, use private neural training and transfer; if the result is a general estimator or inferential procedure, use private statistical estimation and query release.","nearest_boundary":"Private neural training and transfer","representative_topics":["dataset condensation as differentially private data generation for efficient privacy-preserving training","differentially private synthetic data generation for hierarchical query release with group-individual statistics","differentially private conditional gan for imbalanced tabular data with variable dependency capture"]},{"direction_index":2,"label":"D2","name":"Private federated learning","description":"Designing differentially private learning protocols over multiple clients, silos, devices, organizations, or parties.","inclusion":"Include work where the central objective is a federated or distributed learning protocol under DP, including private aggregation, local or user-level privacy, communication-efficient updates, quantization, secure aggregation, personalization, vertical or horizontal federated settings, and federated applications when the protocol is the main contribution.","exclusion":"If the distributed structure is mainly an abstract optimization topology without a federated learning protocol, use private optimization and sequential decision theory; if the performance metric is bandit regret or adaptive decision quality, use private optimization and sequential decision theory.","nearest_boundary":"Private optimization and sequential decision theory","representative_topics":["differential privacy and personalization in cross-silo federated learning with silo-specific sample-level constraints","local differential privacy federated learning via selective quantization and gradient subsampling","federated speech recognition via differential privacy for multi-organization asr"]},{"direction_index":3,"label":"D3","name":"Private optimization and sequential decision theory","description":"Developing and analyzing DP algorithms for abstract optimization objectives and adaptive decision processes, with rates, regret, convergence, or policy performance as the main scientific target.","inclusion":"Include private convex, nonconvex, minimax, saddle-point, distributed, online, heavy-tailed, user-level, and ERM/SCO optimization theory, as well as online learning, bandits, pricing, reinforcement learning, and sequential prediction where DP affects convergence, regret, sample complexity, or policy value.","exclusion":"If the contribution is primarily an empirical or systems improvement to neural DP-SGD training, use private neural training and transfer; if the optimized object is a specific statistical estimand rather than a general objective class or sequential decision problem, use private statistical estimation and query release.","nearest_boundary":"Private neural training and transfer","representative_topics":["differentially private stochastic minimax optimization with near-optimal privacy-loss trade-off via black-box algorithm framework","differentially private optimization with faster stationary point convergence rates in finite-sum and stochastic settings","dimension-independent generalization bounds for differentially private sgd via langevin convergence"]},{"direction_index":4,"label":"D4","name":"Private statistical estimation and query release","description":"Estimating statistical quantities, fitting structured statistical models, conducting inference, or releasing reusable analytic summaries under differential privacy.","inclusion":"Include private means, covariances, regression, clustering, matrix approximation, dependency measures, Bayesian or MCMC inference, label-private estimation, Shapley values, histograms, sketches, query answering, continual counting, and other statistic-release workloads when the main objective is accurate estimation or release of analytic summaries.","exclusion":"If the primary result is a broad learnability lower bound or optimality theorem independent of a concrete estimator or release workload, use privacy guarantees and foundations; if the output is a synthetic dataset intended to stand in for the original data, use private synthetic data generation.","nearest_boundary":"Privacy guarantees and foundations","representative_topics":["differentially private shapley value approximation via stratified sampling for data evaluation","user contribution bounding algorithms for differential privacy-preserving histogram estimation","differentially private multi-party linear regression with disjoint attribute ownership and eigenvalue-constrained mechanisms"]},{"direction_index":5,"label":"D5","name":"Privacy guarantees and foundations","description":"Characterizing, tightening, composing, amplifying, or proving fundamental limits of differential privacy guarantees and mechanisms.","inclusion":"Include privacy accountants, composition theorems, Rényi/Gaussian/zCDP analysis, privacy-loss distributions, amplification by subsampling, shuffling, iteration or decentralization, optimal mechanism analysis, learnability characterizations, sample-complexity separations, lower bounds, private sampling theory, and foundational mechanism results.","exclusion":"If privacy accounting only supports a specific learning, estimation, generation, or release algorithm, assign by that algorithm's primary objective; if the main goal is attack-driven leakage evaluation rather than formal guarantee derivation, use privacy risk, fairness, and deployment.","nearest_boundary":"Privacy risk, fairness, and deployment","representative_topics":["individual example privacy accounting in differentially private stochastic gradient descent","privacy amplification for distributed computation via shuffled check-in protocols and rényi differential privacy","edgeworth expansion for tight non-asymptotic differential privacy composition bounds"]},{"direction_index":6,"label":"D6","name":"Privacy risk, fairness, and deployment","description":"Assessing and improving the real-world behavior of DP systems with respect to leakage risk, attacks, robustness, fairness, calibration, utility, and practitioner usability.","inclusion":"Include empirical privacy auditing, membership inference, reconstruction or vote leakage, gradient leakage, attack robustness, privacy profiling, comparisons between DP and non-DP defenses, fairness and disparate impact of DP, calibration, selective prediction, hyperparameter or epsilon utility studies, practitioner guidance, and deployment-oriented evaluations.","exclusion":"If the main contribution is a formal privacy accountant or composition theorem, use privacy guarantees and foundations; if fairness or robustness is only a secondary metric for a new optimizer, generator, or federated protocol, assign to that primary algorithmic program.","nearest_boundary":"Privacy guarantees and foundations","representative_topics":["privacy guarantees and membership inference robustness of generative adversarial network outputs","epsilon-utility tradeoff analysis for practical differential privacy in private empirical risk minimization","accuracy-fairness tradeoffs in private learning under long-tailed data distributions"]},{"direction_index":7,"label":"D7","name":"Structured and domain-specific private mechanisms","description":"Designing DP mechanisms for data objects or domains whose structure defines the scientific problem, including graphs, language, relational structures, and specialized decision or interaction settings.","inclusion":"Include private graph learning, GNNs, graph embeddings, graph statistics, stochastic block models when graph structure is central, text rewriting, sanitization, de-identification, syntax-aware privatization, private text hashing or similarity, and specialized mechanisms such as voting, counterfactual explanations, contact tracing, secure multiplication, anomaly detection, swarm intelligence, or quantum ML when the domain structure determines the mechanism design.","exclusion":"If the graph is merely a communication network for federated or distributed training, use private federated learning; if a language model is fine-tuned mainly as a pretrained predictor rather than to transform or privatize text, use private neural training and transfer; if the specialized mechanism is primarily a general statistical estimator, use private statistical estimation and query release.","nearest_boundary":"Private statistical estimation and query release","representative_topics":["epsilon-differential privacy text sanitization with customized per-token token-level protection","differentially private personalized pagerank with bounded sensitivity for graph learning","open-source framework for reproducible and transparent differentially private text rewriting"]}],"split":"dev"}
{"episode_id":"F048_T2025-01","field_index":48,"field":"differential privacy","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Private neural training and transfer","description":"Improving differentially private model training and adaptation of pretrained or public-data-supported models for accuracy, stability, scalability, and sample efficiency.","inclusion":"Include work whose main objective is to train or fine-tune predictive models under DP by modifying DP-SGD-style training, clipping, noise, compression, pruning, optimizers, hardware/software systems, public-data use, self-supervised or transfer pretraining, feature reuse, adapters, layer selection, or parameter-efficient fine-tuning.","exclusion":"If the primary output is a synthetic dataset rather than an adapted predictor, use private synthetic data generation; if the main result is an abstract optimization rate rather than a neural training or transfer method, use private optimization and sequential decision theory.","nearest_boundary":"Private synthetic data generation","representative_topics":["automatic gradient clipping for differentially private deep learning without hyperparameter tuning","low-rank sparse gradient updates for differentially private large-scale neural network training","hardware accelerator design for differentially private machine learning training workloads"]},{"direction_index":1,"label":"D1","name":"Private synthetic data generation","description":"Constructing differentially private synthetic data or private generative models that preserve distributional, statistical, or downstream task utility.","inclusion":"Include methods whose primary deliverable is a DP synthetic dataset, generator, data condensation mechanism, distribution summarizer, marginal or geometric synthesizer, modality-specific private generator, or post-processing method for improving synthetic-data utility.","exclusion":"If a pretrained generative model is fine-tuned mainly as a predictor or representation learner rather than to release synthetic data, use private neural training and transfer; if the result is a general estimator or inferential procedure, use private statistical estimation and query release.","nearest_boundary":"Private neural training and transfer","representative_topics":["dataset condensation as differentially private data generation for efficient privacy-preserving training","differentially private synthetic data generation for hierarchical query release with group-individual statistics","differentially private conditional gan for imbalanced tabular data with variable dependency capture"]},{"direction_index":2,"label":"D2","name":"Private federated learning","description":"Designing differentially private learning protocols over multiple clients, silos, devices, organizations, or parties.","inclusion":"Include work where the central objective is a federated or distributed learning protocol under DP, including private aggregation, local or user-level privacy, communication-efficient updates, quantization, secure aggregation, personalization, vertical or horizontal federated settings, and federated applications when the protocol is the main contribution.","exclusion":"If the distributed structure is mainly an abstract optimization topology without a federated learning protocol, use private optimization and sequential decision theory; if the performance metric is bandit regret or adaptive decision quality, use private optimization and sequential decision theory.","nearest_boundary":"Private optimization and sequential decision theory","representative_topics":["differential privacy and personalization in cross-silo federated learning with silo-specific sample-level constraints","local differential privacy federated learning via selective quantization and gradient subsampling","federated speech recognition via differential privacy for multi-organization asr"]},{"direction_index":3,"label":"D3","name":"Private optimization and sequential decision theory","description":"Developing and analyzing DP algorithms for abstract optimization objectives and adaptive decision processes, with rates, regret, convergence, or policy performance as the main scientific target.","inclusion":"Include private convex, nonconvex, minimax, saddle-point, distributed, online, heavy-tailed, user-level, and ERM/SCO optimization theory, as well as online learning, bandits, pricing, reinforcement learning, and sequential prediction where DP affects convergence, regret, sample complexity, or policy value.","exclusion":"If the contribution is primarily an empirical or systems improvement to neural DP-SGD training, use private neural training and transfer; if the optimized object is a specific statistical estimand rather than a general objective class or sequential decision problem, use private statistical estimation and query release.","nearest_boundary":"Private neural training and transfer","representative_topics":["differentially private stochastic minimax optimization with near-optimal privacy-loss trade-off via black-box algorithm framework","differentially private optimization with faster stationary point convergence rates in finite-sum and stochastic settings","dimension-independent generalization bounds for differentially private sgd via langevin convergence"]},{"direction_index":4,"label":"D4","name":"Private statistical estimation and query release","description":"Estimating statistical quantities, fitting structured statistical models, conducting inference, or releasing reusable analytic summaries under differential privacy.","inclusion":"Include private means, covariances, regression, clustering, matrix approximation, dependency measures, Bayesian or MCMC inference, label-private estimation, Shapley values, histograms, sketches, query answering, continual counting, and other statistic-release workloads when the main objective is accurate estimation or release of analytic summaries.","exclusion":"If the primary result is a broad learnability lower bound or optimality theorem independent of a concrete estimator or release workload, use privacy guarantees and foundations; if the output is a synthetic dataset intended to stand in for the original data, use private synthetic data generation.","nearest_boundary":"Privacy guarantees and foundations","representative_topics":["differentially private shapley value approximation via stratified sampling for data evaluation","user contribution bounding algorithms for differential privacy-preserving histogram estimation","differentially private multi-party linear regression with disjoint attribute ownership and eigenvalue-constrained mechanisms"]},{"direction_index":5,"label":"D5","name":"Privacy guarantees and foundations","description":"Characterizing, tightening, composing, amplifying, or proving fundamental limits of differential privacy guarantees and mechanisms.","inclusion":"Include privacy accountants, composition theorems, Rényi/Gaussian/zCDP analysis, privacy-loss distributions, amplification by subsampling, shuffling, iteration or decentralization, optimal mechanism analysis, learnability characterizations, sample-complexity separations, lower bounds, private sampling theory, and foundational mechanism results.","exclusion":"If privacy accounting only supports a specific learning, estimation, generation, or release algorithm, assign by that algorithm's primary objective; if the main goal is attack-driven leakage evaluation rather than formal guarantee derivation, use privacy risk, fairness, and deployment.","nearest_boundary":"Privacy risk, fairness, and deployment","representative_topics":["individual example privacy accounting in differentially private stochastic gradient descent","privacy amplification for distributed computation via shuffled check-in protocols and rényi differential privacy","edgeworth expansion for tight non-asymptotic differential privacy composition bounds"]},{"direction_index":6,"label":"D6","name":"Privacy risk, fairness, and deployment","description":"Assessing and improving the real-world behavior of DP systems with respect to leakage risk, attacks, robustness, fairness, calibration, utility, and practitioner usability.","inclusion":"Include empirical privacy auditing, membership inference, reconstruction or vote leakage, gradient leakage, attack robustness, privacy profiling, comparisons between DP and non-DP defenses, fairness and disparate impact of DP, calibration, selective prediction, hyperparameter or epsilon utility studies, practitioner guidance, and deployment-oriented evaluations.","exclusion":"If the main contribution is a formal privacy accountant or composition theorem, use privacy guarantees and foundations; if fairness or robustness is only a secondary metric for a new optimizer, generator, or federated protocol, assign to that primary algorithmic program.","nearest_boundary":"Privacy guarantees and foundations","representative_topics":["privacy guarantees and membership inference robustness of generative adversarial network outputs","epsilon-utility tradeoff analysis for practical differential privacy in private empirical risk minimization","accuracy-fairness tradeoffs in private learning under long-tailed data distributions"]},{"direction_index":7,"label":"D7","name":"Structured and domain-specific private mechanisms","description":"Designing DP mechanisms for data objects or domains whose structure defines the scientific problem, including graphs, language, relational structures, and specialized decision or interaction settings.","inclusion":"Include private graph learning, GNNs, graph embeddings, graph statistics, stochastic block models when graph structure is central, text rewriting, sanitization, de-identification, syntax-aware privatization, private text hashing or similarity, and specialized mechanisms such as voting, counterfactual explanations, contact tracing, secure multiplication, anomaly detection, swarm intelligence, or quantum ML when the domain structure determines the mechanism design.","exclusion":"If the graph is merely a communication network for federated or distributed training, use private federated learning; if a language model is fine-tuned mainly as a pretrained predictor rather than to transform or privatize text, use private neural training and transfer; if the specialized mechanism is primarily a general statistical estimator, use private statistical estimation and query release.","nearest_boundary":"Private statistical estimation and query release","representative_topics":["epsilon-differential privacy text sanitization with customized per-token token-level protection","differentially private personalized pagerank with bounded sensitivity for graph learning","open-source framework for reproducible and transparent differentially private text rewriting"]}],"split":"dev"}
{"episode_id":"F048_T2025-07","field_index":48,"field":"differential privacy","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Private neural training and transfer","description":"Improving differentially private model training and adaptation of pretrained or public-data-supported models for accuracy, stability, scalability, and sample efficiency.","inclusion":"Include work whose main objective is to train or fine-tune predictive models under DP by modifying DP-SGD-style training, clipping, noise, compression, pruning, optimizers, hardware/software systems, public-data use, self-supervised or transfer pretraining, feature reuse, adapters, layer selection, or parameter-efficient fine-tuning.","exclusion":"If the primary output is a synthetic dataset rather than an adapted predictor, use private synthetic data generation; if the main result is an abstract optimization rate rather than a neural training or transfer method, use private optimization and sequential decision theory.","nearest_boundary":"Private synthetic data generation","representative_topics":["automatic gradient clipping for differentially private deep learning without hyperparameter tuning","low-rank sparse gradient updates for differentially private large-scale neural network training","hardware accelerator design for differentially private machine learning training workloads"]},{"direction_index":1,"label":"D1","name":"Private synthetic data generation","description":"Constructing differentially private synthetic data or private generative models that preserve distributional, statistical, or downstream task utility.","inclusion":"Include methods whose primary deliverable is a DP synthetic dataset, generator, data condensation mechanism, distribution summarizer, marginal or geometric synthesizer, modality-specific private generator, or post-processing method for improving synthetic-data utility.","exclusion":"If a pretrained generative model is fine-tuned mainly as a predictor or representation learner rather than to release synthetic data, use private neural training and transfer; if the result is a general estimator or inferential procedure, use private statistical estimation and query release.","nearest_boundary":"Private neural training and transfer","representative_topics":["dataset condensation as differentially private data generation for efficient privacy-preserving training","differentially private synthetic data generation for hierarchical query release with group-individual statistics","differentially private conditional gan for imbalanced tabular data with variable dependency capture"]},{"direction_index":2,"label":"D2","name":"Private federated learning","description":"Designing differentially private learning protocols over multiple clients, silos, devices, organizations, or parties.","inclusion":"Include work where the central objective is a federated or distributed learning protocol under DP, including private aggregation, local or user-level privacy, communication-efficient updates, quantization, secure aggregation, personalization, vertical or horizontal federated settings, and federated applications when the protocol is the main contribution.","exclusion":"If the distributed structure is mainly an abstract optimization topology without a federated learning protocol, use private optimization and sequential decision theory; if the performance metric is bandit regret or adaptive decision quality, use private optimization and sequential decision theory.","nearest_boundary":"Private optimization and sequential decision theory","representative_topics":["differential privacy and personalization in cross-silo federated learning with silo-specific sample-level constraints","local differential privacy federated learning via selective quantization and gradient subsampling","federated speech recognition via differential privacy for multi-organization asr"]},{"direction_index":3,"label":"D3","name":"Private optimization and sequential decision theory","description":"Developing and analyzing DP algorithms for abstract optimization objectives and adaptive decision processes, with rates, regret, convergence, or policy performance as the main scientific target.","inclusion":"Include private convex, nonconvex, minimax, saddle-point, distributed, online, heavy-tailed, user-level, and ERM/SCO optimization theory, as well as online learning, bandits, pricing, reinforcement learning, and sequential prediction where DP affects convergence, regret, sample complexity, or policy value.","exclusion":"If the contribution is primarily an empirical or systems improvement to neural DP-SGD training, use private neural training and transfer; if the optimized object is a specific statistical estimand rather than a general objective class or sequential decision problem, use private statistical estimation and query release.","nearest_boundary":"Private neural training and transfer","representative_topics":["differentially private stochastic minimax optimization with near-optimal privacy-loss trade-off via black-box algorithm framework","differentially private optimization with faster stationary point convergence rates in finite-sum and stochastic settings","dimension-independent generalization bounds for differentially private sgd via langevin convergence"]},{"direction_index":4,"label":"D4","name":"Private statistical estimation and query release","description":"Estimating statistical quantities, fitting structured statistical models, conducting inference, or releasing reusable analytic summaries under differential privacy.","inclusion":"Include private means, covariances, regression, clustering, matrix approximation, dependency measures, Bayesian or MCMC inference, label-private estimation, Shapley values, histograms, sketches, query answering, continual counting, and other statistic-release workloads when the main objective is accurate estimation or release of analytic summaries.","exclusion":"If the primary result is a broad learnability lower bound or optimality theorem independent of a concrete estimator or release workload, use privacy guarantees and foundations; if the output is a synthetic dataset intended to stand in for the original data, use private synthetic data generation.","nearest_boundary":"Privacy guarantees and foundations","representative_topics":["differentially private shapley value approximation via stratified sampling for data evaluation","user contribution bounding algorithms for differential privacy-preserving histogram estimation","differentially private multi-party linear regression with disjoint attribute ownership and eigenvalue-constrained mechanisms"]},{"direction_index":5,"label":"D5","name":"Privacy guarantees and foundations","description":"Characterizing, tightening, composing, amplifying, or proving fundamental limits of differential privacy guarantees and mechanisms.","inclusion":"Include privacy accountants, composition theorems, Rényi/Gaussian/zCDP analysis, privacy-loss distributions, amplification by subsampling, shuffling, iteration or decentralization, optimal mechanism analysis, learnability characterizations, sample-complexity separations, lower bounds, private sampling theory, and foundational mechanism results.","exclusion":"If privacy accounting only supports a specific learning, estimation, generation, or release algorithm, assign by that algorithm's primary objective; if the main goal is attack-driven leakage evaluation rather than formal guarantee derivation, use privacy risk, fairness, and deployment.","nearest_boundary":"Privacy risk, fairness, and deployment","representative_topics":["individual example privacy accounting in differentially private stochastic gradient descent","privacy amplification for distributed computation via shuffled check-in protocols and rényi differential privacy","edgeworth expansion for tight non-asymptotic differential privacy composition bounds"]},{"direction_index":6,"label":"D6","name":"Privacy risk, fairness, and deployment","description":"Assessing and improving the real-world behavior of DP systems with respect to leakage risk, attacks, robustness, fairness, calibration, utility, and practitioner usability.","inclusion":"Include empirical privacy auditing, membership inference, reconstruction or vote leakage, gradient leakage, attack robustness, privacy profiling, comparisons between DP and non-DP defenses, fairness and disparate impact of DP, calibration, selective prediction, hyperparameter or epsilon utility studies, practitioner guidance, and deployment-oriented evaluations.","exclusion":"If the main contribution is a formal privacy accountant or composition theorem, use privacy guarantees and foundations; if fairness or robustness is only a secondary metric for a new optimizer, generator, or federated protocol, assign to that primary algorithmic program.","nearest_boundary":"Privacy guarantees and foundations","representative_topics":["privacy guarantees and membership inference robustness of generative adversarial network outputs","epsilon-utility tradeoff analysis for practical differential privacy in private empirical risk minimization","accuracy-fairness tradeoffs in private learning under long-tailed data distributions"]},{"direction_index":7,"label":"D7","name":"Structured and domain-specific private mechanisms","description":"Designing DP mechanisms for data objects or domains whose structure defines the scientific problem, including graphs, language, relational structures, and specialized decision or interaction settings.","inclusion":"Include private graph learning, GNNs, graph embeddings, graph statistics, stochastic block models when graph structure is central, text rewriting, sanitization, de-identification, syntax-aware privatization, private text hashing or similarity, and specialized mechanisms such as voting, counterfactual explanations, contact tracing, secure multiplication, anomaly detection, swarm intelligence, or quantum ML when the domain structure determines the mechanism design.","exclusion":"If the graph is merely a communication network for federated or distributed training, use private federated learning; if a language model is fine-tuned mainly as a pretrained predictor rather than to transform or privatize text, use private neural training and transfer; if the specialized mechanism is primarily a general statistical estimator, use private statistical estimation and query release.","nearest_boundary":"Private statistical estimation and query release","representative_topics":["epsilon-differential privacy text sanitization with customized per-token token-level protection","differentially private personalized pagerank with bounded sensitivity for graph learning","open-source framework for reproducible and transparent differentially private text rewriting"]}],"split":"dev"}
{"episode_id":"F048_T2026-01","field_index":48,"field":"differential privacy","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Private neural training and transfer","description":"Improving differentially private model training and adaptation of pretrained or public-data-supported models for accuracy, stability, scalability, and sample efficiency.","inclusion":"Include work whose main objective is to train or fine-tune predictive models under DP by modifying DP-SGD-style training, clipping, noise, compression, pruning, optimizers, hardware/software systems, public-data use, self-supervised or transfer pretraining, feature reuse, adapters, layer selection, or parameter-efficient fine-tuning.","exclusion":"If the primary output is a synthetic dataset rather than an adapted predictor, use private synthetic data generation; if the main result is an abstract optimization rate rather than a neural training or transfer method, use private optimization and sequential decision theory.","nearest_boundary":"Private synthetic data generation","representative_topics":["automatic gradient clipping for differentially private deep learning without hyperparameter tuning","low-rank sparse gradient updates for differentially private large-scale neural network training","hardware accelerator design for differentially private machine learning training workloads"]},{"direction_index":1,"label":"D1","name":"Private synthetic data generation","description":"Constructing differentially private synthetic data or private generative models that preserve distributional, statistical, or downstream task utility.","inclusion":"Include methods whose primary deliverable is a DP synthetic dataset, generator, data condensation mechanism, distribution summarizer, marginal or geometric synthesizer, modality-specific private generator, or post-processing method for improving synthetic-data utility.","exclusion":"If a pretrained generative model is fine-tuned mainly as a predictor or representation learner rather than to release synthetic data, use private neural training and transfer; if the result is a general estimator or inferential procedure, use private statistical estimation and query release.","nearest_boundary":"Private neural training and transfer","representative_topics":["dataset condensation as differentially private data generation for efficient privacy-preserving training","differentially private synthetic data generation for hierarchical query release with group-individual statistics","differentially private conditional gan for imbalanced tabular data with variable dependency capture"]},{"direction_index":2,"label":"D2","name":"Private federated learning","description":"Designing differentially private learning protocols over multiple clients, silos, devices, organizations, or parties.","inclusion":"Include work where the central objective is a federated or distributed learning protocol under DP, including private aggregation, local or user-level privacy, communication-efficient updates, quantization, secure aggregation, personalization, vertical or horizontal federated settings, and federated applications when the protocol is the main contribution.","exclusion":"If the distributed structure is mainly an abstract optimization topology without a federated learning protocol, use private optimization and sequential decision theory; if the performance metric is bandit regret or adaptive decision quality, use private optimization and sequential decision theory.","nearest_boundary":"Private optimization and sequential decision theory","representative_topics":["differential privacy and personalization in cross-silo federated learning with silo-specific sample-level constraints","local differential privacy federated learning via selective quantization and gradient subsampling","federated speech recognition via differential privacy for multi-organization asr"]},{"direction_index":3,"label":"D3","name":"Private optimization and sequential decision theory","description":"Developing and analyzing DP algorithms for abstract optimization objectives and adaptive decision processes, with rates, regret, convergence, or policy performance as the main scientific target.","inclusion":"Include private convex, nonconvex, minimax, saddle-point, distributed, online, heavy-tailed, user-level, and ERM/SCO optimization theory, as well as online learning, bandits, pricing, reinforcement learning, and sequential prediction where DP affects convergence, regret, sample complexity, or policy value.","exclusion":"If the contribution is primarily an empirical or systems improvement to neural DP-SGD training, use private neural training and transfer; if the optimized object is a specific statistical estimand rather than a general objective class or sequential decision problem, use private statistical estimation and query release.","nearest_boundary":"Private neural training and transfer","representative_topics":["differentially private stochastic minimax optimization with near-optimal privacy-loss trade-off via black-box algorithm framework","differentially private optimization with faster stationary point convergence rates in finite-sum and stochastic settings","dimension-independent generalization bounds for differentially private sgd via langevin convergence"]},{"direction_index":4,"label":"D4","name":"Private statistical estimation and query release","description":"Estimating statistical quantities, fitting structured statistical models, conducting inference, or releasing reusable analytic summaries under differential privacy.","inclusion":"Include private means, covariances, regression, clustering, matrix approximation, dependency measures, Bayesian or MCMC inference, label-private estimation, Shapley values, histograms, sketches, query answering, continual counting, and other statistic-release workloads when the main objective is accurate estimation or release of analytic summaries.","exclusion":"If the primary result is a broad learnability lower bound or optimality theorem independent of a concrete estimator or release workload, use privacy guarantees and foundations; if the output is a synthetic dataset intended to stand in for the original data, use private synthetic data generation.","nearest_boundary":"Privacy guarantees and foundations","representative_topics":["differentially private shapley value approximation via stratified sampling for data evaluation","user contribution bounding algorithms for differential privacy-preserving histogram estimation","differentially private multi-party linear regression with disjoint attribute ownership and eigenvalue-constrained mechanisms"]},{"direction_index":5,"label":"D5","name":"Privacy guarantees and foundations","description":"Characterizing, tightening, composing, amplifying, or proving fundamental limits of differential privacy guarantees and mechanisms.","inclusion":"Include privacy accountants, composition theorems, Rényi/Gaussian/zCDP analysis, privacy-loss distributions, amplification by subsampling, shuffling, iteration or decentralization, optimal mechanism analysis, learnability characterizations, sample-complexity separations, lower bounds, private sampling theory, and foundational mechanism results.","exclusion":"If privacy accounting only supports a specific learning, estimation, generation, or release algorithm, assign by that algorithm's primary objective; if the main goal is attack-driven leakage evaluation rather than formal guarantee derivation, use privacy risk, fairness, and deployment.","nearest_boundary":"Privacy risk, fairness, and deployment","representative_topics":["individual example privacy accounting in differentially private stochastic gradient descent","privacy amplification for distributed computation via shuffled check-in protocols and rényi differential privacy","edgeworth expansion for tight non-asymptotic differential privacy composition bounds"]},{"direction_index":6,"label":"D6","name":"Privacy risk, fairness, and deployment","description":"Assessing and improving the real-world behavior of DP systems with respect to leakage risk, attacks, robustness, fairness, calibration, utility, and practitioner usability.","inclusion":"Include empirical privacy auditing, membership inference, reconstruction or vote leakage, gradient leakage, attack robustness, privacy profiling, comparisons between DP and non-DP defenses, fairness and disparate impact of DP, calibration, selective prediction, hyperparameter or epsilon utility studies, practitioner guidance, and deployment-oriented evaluations.","exclusion":"If the main contribution is a formal privacy accountant or composition theorem, use privacy guarantees and foundations; if fairness or robustness is only a secondary metric for a new optimizer, generator, or federated protocol, assign to that primary algorithmic program.","nearest_boundary":"Privacy guarantees and foundations","representative_topics":["privacy guarantees and membership inference robustness of generative adversarial network outputs","epsilon-utility tradeoff analysis for practical differential privacy in private empirical risk minimization","accuracy-fairness tradeoffs in private learning under long-tailed data distributions"]},{"direction_index":7,"label":"D7","name":"Structured and domain-specific private mechanisms","description":"Designing DP mechanisms for data objects or domains whose structure defines the scientific problem, including graphs, language, relational structures, and specialized decision or interaction settings.","inclusion":"Include private graph learning, GNNs, graph embeddings, graph statistics, stochastic block models when graph structure is central, text rewriting, sanitization, de-identification, syntax-aware privatization, private text hashing or similarity, and specialized mechanisms such as voting, counterfactual explanations, contact tracing, secure multiplication, anomaly detection, swarm intelligence, or quantum ML when the domain structure determines the mechanism design.","exclusion":"If the graph is merely a communication network for federated or distributed training, use private federated learning; if a language model is fine-tuned mainly as a pretrained predictor rather than to transform or privatize text, use private neural training and transfer; if the specialized mechanism is primarily a general statistical estimator, use private statistical estimation and query release.","nearest_boundary":"Private statistical estimation and query release","representative_topics":["epsilon-differential privacy text sanitization with customized per-token token-level protection","differentially private personalized pagerank with bounded sensitivity for graph learning","open-source framework for reproducible and transparent differentially private text rewriting"]}],"split":"dev"}
{"episode_id":"F051_T2024-01","field_index":51,"field":"3d scene understanding","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"3D Object Detection and Scalable Perception","description":"Estimating 3D object boxes and object locations from camera, LiDAR, radar, or fused sensor inputs, including efficient sparse backbones, BEV systems, distillation, and unified multi-task architectures when detection-centered perception is a main objective.","inclusion":"Assign atoms whose primary program is 3D object detection or scalable/unified perception architectures principally evaluated through detection or closely coupled detection-segmentation perception outputs.","exclusion":"Dense voxel occupancy, free-space completion, or future occupancy grids belong to D2; pure semantic or instance mask prediction without boxes belongs to D1.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["temporal modeling for unified multi-task 3d perception from multi-camera images","ground plane prior for monocular 3d object depth estimation without auxiliary sensors","polar coordinate transformer for multi-camera 3d object detection in bird's-eye-view"]},{"direction_index":1,"label":"D1","name":"3D Segmentation and Semantic Scene Fields","description":"Assigning semantic, instance, panoptic, part, object-centric, or neural-field-based semantic labels to 3D scene elements, including supervised, weakly supervised, few-shot, domain-adaptive, neural-rendering, and object-decomposition approaches.","inclusion":"Assign atoms whose main deliverable is per-point, per-voxel, per-mesh, per-region, object-instance, panoptic, part-level, or neural-field semantic decomposition of 3D scenes.","exclusion":"Bounding-box-centered object localization belongs to D0; open-vocabulary or language/foundation-model semantic transfer as the primary scientific objective belongs to D4.","nearest_boundary":"D4 Language, Open-Vocabulary, and Structured Reasoning","representative_topics":["weakly supervised 3d instance segmentation via bounding box parameter voting and clustering","hierarchical part segmentation of objects in real-world noisy indoor 3d scans","multi-property neural radiance field synthesis for unified 3d scene understanding"]},{"direction_index":2,"label":"D2","name":"Dense Occupancy and Scene Completion","description":"Predicting dense 3D occupancy, free space, semantic occupancy, completed volumetric scenes, or occupancy-flow grids from partial observations in indoor, outdoor, driving, or cooperative settings.","inclusion":"Assign atoms whose primary output is a dense voxel, grid, or volumetric scene state representing occupied/free/semantic space, including static completion and future occupancy forecasting.","exclusion":"Point-wise motion, scene flow, moving-object segmentation, or temporal 4D analysis without occupancy as the main representation belongs to D3; surface/layout reconstruction without dense occupancy belongs to D5.","nearest_boundary":"D3 Dynamic 4D Scene Understanding","representative_topics":["hierarchical spatial-temporal networks for joint occupancy and flow prediction in 3d scenes","multi-modal hierarchical transformer for joint occupancy and flow field prediction in autonomous driving","allo-centric occupancy grid prediction for long-term urban traffic scene forecasting"]},{"direction_index":3,"label":"D3","name":"Dynamic 4D Scene Understanding","description":"Understanding temporal change, motion, correspondence, moving objects, temporal segmentation, scene flow, tracking, change detection, and evolving 4D scene representations.","inclusion":"Assign atoms whose primary objective is motion or temporal scene analysis, including dense scene flow, moving-object segmentation, 3D change detection, 4D point-cloud segmentation, temporal compression, dynamic object-level mapping, or sequential correspondence.","exclusion":"Future scene state represented mainly as dense occupancy grids belongs to D2; static segmentation or static reconstruction belongs to D1 or D5 respectively.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["unsupervised 3d scene flow estimation from monocular video using depth and pose priors","receding horizon moving object segmentation in 3d lidar using sparse 4d temporal-spatial convolutions","temporal variation-aware 4d point cloud segmentation for autonomous driving"]},{"direction_index":4,"label":"D4","name":"Language, Open-Vocabulary, and Structured Reasoning","description":"Connecting 3D scenes to language, open vocabularies, foundation models, large language models, visual grounding, VQA, dense captioning, cross-modal pretraining, and explicit relational scene graphs or structured semantic reasoning.","inclusion":"Assign atoms whose primary program is language-conditioned localization or answering, open-vocabulary/zero-shot 3D perception, 3D-language representation learning, LLM integration, dense captioning as a semantic-language task, or explicit object-relation graph prediction.","exclusion":"Closed-vocabulary segmentation architectures without language/foundation transfer belong to D1; dataset construction whose main contribution is a resource rather than a language/reasoning model belongs to D8.","nearest_boundary":"D1 3D Segmentation and Semantic Scene Fields","representative_topics":["phrase-aware 3d visual grounding with fine-grained contextual phrase-object alignment","zero-shot open-world 3d scene understanding via semantic abstraction of 2d vision-language models","word-level text decoupling and dense alignment for 3d point cloud visual grounding"]},{"direction_index":5,"label":"D5","name":"Geometric Reconstruction and Layout Recovery","description":"Recovering 3D scene geometry, surfaces, planes, room layouts, object shapes, articulated structures, camera/scan alignment, and structural decompositions from images, RGB-D streams, videos, scans, or partial observations.","inclusion":"Assign atoms whose primary objective is faithful geometric reconstruction, layout estimation, plane or structural recovery, shape completion from observations, active view selection for reconstruction, or articulation model discovery.","exclusion":"Dense semantic occupancy prediction belongs to D2; generative synthesis or controllable editing of novel scene content belongs to D6.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["real-time 3d plane reconstruction from posed monocular video via incremental fragment fusion","mcts-based proposal selection and refinement for indoor scene layout reconstruction","panoramic 3d room layout estimation via learnable manhattan hough transform on cubemap tiles"]},{"direction_index":6,"label":"D6","name":"Generative Scene Synthesis and Editing","description":"Generating, completing, inpainting, optimizing, planning, or controllably editing 3D scene content, layouts, object arrangements, human-scene interactions, or neural scene representations using generative priors.","inclusion":"Assign atoms whose primary objective is creating or modifying 3D scene content rather than recognizing, localizing, segmenting, or faithfully reconstructing an observed scene.","exclusion":"Observed-scene reconstruction or layout recovery without generative modification belongs to D5; discriminative semantic occupancy completion belongs to D2.","nearest_boundary":"D5 Geometric Reconstruction and Layout Recovery","representative_topics":["2d-supervised 3d scene prior learning via autoregressive latent decoding of layout and geometry","human-motion-aware autoregressive 3d indoor scene generation with furniture layout inference","free-form 3d scene inpainting with dual-stream gan for large diverse missing regions"]},{"direction_index":7,"label":"D7","name":"Reusable Representations and Data Resources","description":"Building enabling foundations for 3D scene understanding through self-supervised or unsupervised representation learning, masked or contrastive pretraining, pseudo-labeling, object-centric feature learning, datasets, benchmarks, annotation pipelines, and resource-level evaluation protocols.","inclusion":"Assign atoms whose main contribution is a reusable 3D representation objective or a data/benchmark/annotation resource intended to support multiple downstream 3D understanding tasks.","exclusion":"Task-specific segmentation, detection, reconstruction, grounding, or occupancy methods using datasets or pretraining only as tools belong to their respective task directions.","nearest_boundary":"D1 3D Segmentation and Semantic Scene Fields","representative_topics":["invariance analysis and contrastive pre-training for self-supervised 3d vision representations","cross-modal pretraining of image transformers for 3d point cloud classification and segmentation","large-scale 3d indoor object annotation dataset with texture-guided boundary semantics for navigation"]}],"split":"dev"}
{"episode_id":"F051_T2024-07","field_index":51,"field":"3d scene understanding","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"3D Object Detection and Scalable Perception","description":"Estimating 3D object boxes and object locations from camera, LiDAR, radar, or fused sensor inputs, including efficient sparse backbones, BEV systems, distillation, and unified multi-task architectures when detection-centered perception is a main objective.","inclusion":"Assign atoms whose primary program is 3D object detection or scalable/unified perception architectures principally evaluated through detection or closely coupled detection-segmentation perception outputs.","exclusion":"Dense voxel occupancy, free-space completion, or future occupancy grids belong to D2; pure semantic or instance mask prediction without boxes belongs to D1.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["temporal modeling for unified multi-task 3d perception from multi-camera images","ground plane prior for monocular 3d object depth estimation without auxiliary sensors","polar coordinate transformer for multi-camera 3d object detection in bird's-eye-view"]},{"direction_index":1,"label":"D1","name":"3D Segmentation and Semantic Scene Fields","description":"Assigning semantic, instance, panoptic, part, object-centric, or neural-field-based semantic labels to 3D scene elements, including supervised, weakly supervised, few-shot, domain-adaptive, neural-rendering, and object-decomposition approaches.","inclusion":"Assign atoms whose main deliverable is per-point, per-voxel, per-mesh, per-region, object-instance, panoptic, part-level, or neural-field semantic decomposition of 3D scenes.","exclusion":"Bounding-box-centered object localization belongs to D0; open-vocabulary or language/foundation-model semantic transfer as the primary scientific objective belongs to D4.","nearest_boundary":"D4 Language, Open-Vocabulary, and Structured Reasoning","representative_topics":["weakly supervised 3d instance segmentation via bounding box parameter voting and clustering","hierarchical part segmentation of objects in real-world noisy indoor 3d scans","multi-property neural radiance field synthesis for unified 3d scene understanding"]},{"direction_index":2,"label":"D2","name":"Dense Occupancy and Scene Completion","description":"Predicting dense 3D occupancy, free space, semantic occupancy, completed volumetric scenes, or occupancy-flow grids from partial observations in indoor, outdoor, driving, or cooperative settings.","inclusion":"Assign atoms whose primary output is a dense voxel, grid, or volumetric scene state representing occupied/free/semantic space, including static completion and future occupancy forecasting.","exclusion":"Point-wise motion, scene flow, moving-object segmentation, or temporal 4D analysis without occupancy as the main representation belongs to D3; surface/layout reconstruction without dense occupancy belongs to D5.","nearest_boundary":"D3 Dynamic 4D Scene Understanding","representative_topics":["hierarchical spatial-temporal networks for joint occupancy and flow prediction in 3d scenes","multi-modal hierarchical transformer for joint occupancy and flow field prediction in autonomous driving","allo-centric occupancy grid prediction for long-term urban traffic scene forecasting"]},{"direction_index":3,"label":"D3","name":"Dynamic 4D Scene Understanding","description":"Understanding temporal change, motion, correspondence, moving objects, temporal segmentation, scene flow, tracking, change detection, and evolving 4D scene representations.","inclusion":"Assign atoms whose primary objective is motion or temporal scene analysis, including dense scene flow, moving-object segmentation, 3D change detection, 4D point-cloud segmentation, temporal compression, dynamic object-level mapping, or sequential correspondence.","exclusion":"Future scene state represented mainly as dense occupancy grids belongs to D2; static segmentation or static reconstruction belongs to D1 or D5 respectively.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["unsupervised 3d scene flow estimation from monocular video using depth and pose priors","receding horizon moving object segmentation in 3d lidar using sparse 4d temporal-spatial convolutions","temporal variation-aware 4d point cloud segmentation for autonomous driving"]},{"direction_index":4,"label":"D4","name":"Language, Open-Vocabulary, and Structured Reasoning","description":"Connecting 3D scenes to language, open vocabularies, foundation models, large language models, visual grounding, VQA, dense captioning, cross-modal pretraining, and explicit relational scene graphs or structured semantic reasoning.","inclusion":"Assign atoms whose primary program is language-conditioned localization or answering, open-vocabulary/zero-shot 3D perception, 3D-language representation learning, LLM integration, dense captioning as a semantic-language task, or explicit object-relation graph prediction.","exclusion":"Closed-vocabulary segmentation architectures without language/foundation transfer belong to D1; dataset construction whose main contribution is a resource rather than a language/reasoning model belongs to D8.","nearest_boundary":"D1 3D Segmentation and Semantic Scene Fields","representative_topics":["phrase-aware 3d visual grounding with fine-grained contextual phrase-object alignment","zero-shot open-world 3d scene understanding via semantic abstraction of 2d vision-language models","word-level text decoupling and dense alignment for 3d point cloud visual grounding"]},{"direction_index":5,"label":"D5","name":"Geometric Reconstruction and Layout Recovery","description":"Recovering 3D scene geometry, surfaces, planes, room layouts, object shapes, articulated structures, camera/scan alignment, and structural decompositions from images, RGB-D streams, videos, scans, or partial observations.","inclusion":"Assign atoms whose primary objective is faithful geometric reconstruction, layout estimation, plane or structural recovery, shape completion from observations, active view selection for reconstruction, or articulation model discovery.","exclusion":"Dense semantic occupancy prediction belongs to D2; generative synthesis or controllable editing of novel scene content belongs to D6.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["real-time 3d plane reconstruction from posed monocular video via incremental fragment fusion","mcts-based proposal selection and refinement for indoor scene layout reconstruction","panoramic 3d room layout estimation via learnable manhattan hough transform on cubemap tiles"]},{"direction_index":6,"label":"D6","name":"Generative Scene Synthesis and Editing","description":"Generating, completing, inpainting, optimizing, planning, or controllably editing 3D scene content, layouts, object arrangements, human-scene interactions, or neural scene representations using generative priors.","inclusion":"Assign atoms whose primary objective is creating or modifying 3D scene content rather than recognizing, localizing, segmenting, or faithfully reconstructing an observed scene.","exclusion":"Observed-scene reconstruction or layout recovery without generative modification belongs to D5; discriminative semantic occupancy completion belongs to D2.","nearest_boundary":"D5 Geometric Reconstruction and Layout Recovery","representative_topics":["2d-supervised 3d scene prior learning via autoregressive latent decoding of layout and geometry","human-motion-aware autoregressive 3d indoor scene generation with furniture layout inference","free-form 3d scene inpainting with dual-stream gan for large diverse missing regions"]},{"direction_index":7,"label":"D7","name":"Reusable Representations and Data Resources","description":"Building enabling foundations for 3D scene understanding through self-supervised or unsupervised representation learning, masked or contrastive pretraining, pseudo-labeling, object-centric feature learning, datasets, benchmarks, annotation pipelines, and resource-level evaluation protocols.","inclusion":"Assign atoms whose main contribution is a reusable 3D representation objective or a data/benchmark/annotation resource intended to support multiple downstream 3D understanding tasks.","exclusion":"Task-specific segmentation, detection, reconstruction, grounding, or occupancy methods using datasets or pretraining only as tools belong to their respective task directions.","nearest_boundary":"D1 3D Segmentation and Semantic Scene Fields","representative_topics":["invariance analysis and contrastive pre-training for self-supervised 3d vision representations","cross-modal pretraining of image transformers for 3d point cloud classification and segmentation","large-scale 3d indoor object annotation dataset with texture-guided boundary semantics for navigation"]}],"split":"dev"}
{"episode_id":"F051_T2025-01","field_index":51,"field":"3d scene understanding","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"3D Object Detection and Scalable Perception","description":"Estimating 3D object boxes and object locations from camera, LiDAR, radar, or fused sensor inputs, including efficient sparse backbones, BEV systems, distillation, and unified multi-task architectures when detection-centered perception is a main objective.","inclusion":"Assign atoms whose primary program is 3D object detection or scalable/unified perception architectures principally evaluated through detection or closely coupled detection-segmentation perception outputs.","exclusion":"Dense voxel occupancy, free-space completion, or future occupancy grids belong to D2; pure semantic or instance mask prediction without boxes belongs to D1.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["temporal modeling for unified multi-task 3d perception from multi-camera images","ground plane prior for monocular 3d object depth estimation without auxiliary sensors","polar coordinate transformer for multi-camera 3d object detection in bird's-eye-view"]},{"direction_index":1,"label":"D1","name":"3D Segmentation and Semantic Scene Fields","description":"Assigning semantic, instance, panoptic, part, object-centric, or neural-field-based semantic labels to 3D scene elements, including supervised, weakly supervised, few-shot, domain-adaptive, neural-rendering, and object-decomposition approaches.","inclusion":"Assign atoms whose main deliverable is per-point, per-voxel, per-mesh, per-region, object-instance, panoptic, part-level, or neural-field semantic decomposition of 3D scenes.","exclusion":"Bounding-box-centered object localization belongs to D0; open-vocabulary or language/foundation-model semantic transfer as the primary scientific objective belongs to D4.","nearest_boundary":"D4 Language, Open-Vocabulary, and Structured Reasoning","representative_topics":["weakly supervised 3d instance segmentation via bounding box parameter voting and clustering","hierarchical part segmentation of objects in real-world noisy indoor 3d scans","multi-property neural radiance field synthesis for unified 3d scene understanding"]},{"direction_index":2,"label":"D2","name":"Dense Occupancy and Scene Completion","description":"Predicting dense 3D occupancy, free space, semantic occupancy, completed volumetric scenes, or occupancy-flow grids from partial observations in indoor, outdoor, driving, or cooperative settings.","inclusion":"Assign atoms whose primary output is a dense voxel, grid, or volumetric scene state representing occupied/free/semantic space, including static completion and future occupancy forecasting.","exclusion":"Point-wise motion, scene flow, moving-object segmentation, or temporal 4D analysis without occupancy as the main representation belongs to D3; surface/layout reconstruction without dense occupancy belongs to D5.","nearest_boundary":"D3 Dynamic 4D Scene Understanding","representative_topics":["hierarchical spatial-temporal networks for joint occupancy and flow prediction in 3d scenes","multi-modal hierarchical transformer for joint occupancy and flow field prediction in autonomous driving","allo-centric occupancy grid prediction for long-term urban traffic scene forecasting"]},{"direction_index":3,"label":"D3","name":"Dynamic 4D Scene Understanding","description":"Understanding temporal change, motion, correspondence, moving objects, temporal segmentation, scene flow, tracking, change detection, and evolving 4D scene representations.","inclusion":"Assign atoms whose primary objective is motion or temporal scene analysis, including dense scene flow, moving-object segmentation, 3D change detection, 4D point-cloud segmentation, temporal compression, dynamic object-level mapping, or sequential correspondence.","exclusion":"Future scene state represented mainly as dense occupancy grids belongs to D2; static segmentation or static reconstruction belongs to D1 or D5 respectively.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["unsupervised 3d scene flow estimation from monocular video using depth and pose priors","receding horizon moving object segmentation in 3d lidar using sparse 4d temporal-spatial convolutions","temporal variation-aware 4d point cloud segmentation for autonomous driving"]},{"direction_index":4,"label":"D4","name":"Language, Open-Vocabulary, and Structured Reasoning","description":"Connecting 3D scenes to language, open vocabularies, foundation models, large language models, visual grounding, VQA, dense captioning, cross-modal pretraining, and explicit relational scene graphs or structured semantic reasoning.","inclusion":"Assign atoms whose primary program is language-conditioned localization or answering, open-vocabulary/zero-shot 3D perception, 3D-language representation learning, LLM integration, dense captioning as a semantic-language task, or explicit object-relation graph prediction.","exclusion":"Closed-vocabulary segmentation architectures without language/foundation transfer belong to D1; dataset construction whose main contribution is a resource rather than a language/reasoning model belongs to D8.","nearest_boundary":"D1 3D Segmentation and Semantic Scene Fields","representative_topics":["phrase-aware 3d visual grounding with fine-grained contextual phrase-object alignment","zero-shot open-world 3d scene understanding via semantic abstraction of 2d vision-language models","word-level text decoupling and dense alignment for 3d point cloud visual grounding"]},{"direction_index":5,"label":"D5","name":"Geometric Reconstruction and Layout Recovery","description":"Recovering 3D scene geometry, surfaces, planes, room layouts, object shapes, articulated structures, camera/scan alignment, and structural decompositions from images, RGB-D streams, videos, scans, or partial observations.","inclusion":"Assign atoms whose primary objective is faithful geometric reconstruction, layout estimation, plane or structural recovery, shape completion from observations, active view selection for reconstruction, or articulation model discovery.","exclusion":"Dense semantic occupancy prediction belongs to D2; generative synthesis or controllable editing of novel scene content belongs to D6.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["real-time 3d plane reconstruction from posed monocular video via incremental fragment fusion","mcts-based proposal selection and refinement for indoor scene layout reconstruction","panoramic 3d room layout estimation via learnable manhattan hough transform on cubemap tiles"]},{"direction_index":6,"label":"D6","name":"Generative Scene Synthesis and Editing","description":"Generating, completing, inpainting, optimizing, planning, or controllably editing 3D scene content, layouts, object arrangements, human-scene interactions, or neural scene representations using generative priors.","inclusion":"Assign atoms whose primary objective is creating or modifying 3D scene content rather than recognizing, localizing, segmenting, or faithfully reconstructing an observed scene.","exclusion":"Observed-scene reconstruction or layout recovery without generative modification belongs to D5; discriminative semantic occupancy completion belongs to D2.","nearest_boundary":"D5 Geometric Reconstruction and Layout Recovery","representative_topics":["2d-supervised 3d scene prior learning via autoregressive latent decoding of layout and geometry","human-motion-aware autoregressive 3d indoor scene generation with furniture layout inference","free-form 3d scene inpainting with dual-stream gan for large diverse missing regions"]},{"direction_index":7,"label":"D7","name":"Reusable Representations and Data Resources","description":"Building enabling foundations for 3D scene understanding through self-supervised or unsupervised representation learning, masked or contrastive pretraining, pseudo-labeling, object-centric feature learning, datasets, benchmarks, annotation pipelines, and resource-level evaluation protocols.","inclusion":"Assign atoms whose main contribution is a reusable 3D representation objective or a data/benchmark/annotation resource intended to support multiple downstream 3D understanding tasks.","exclusion":"Task-specific segmentation, detection, reconstruction, grounding, or occupancy methods using datasets or pretraining only as tools belong to their respective task directions.","nearest_boundary":"D1 3D Segmentation and Semantic Scene Fields","representative_topics":["invariance analysis and contrastive pre-training for self-supervised 3d vision representations","cross-modal pretraining of image transformers for 3d point cloud classification and segmentation","large-scale 3d indoor object annotation dataset with texture-guided boundary semantics for navigation"]}],"split":"dev"}
{"episode_id":"F051_T2025-07","field_index":51,"field":"3d scene understanding","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"3D Object Detection and Scalable Perception","description":"Estimating 3D object boxes and object locations from camera, LiDAR, radar, or fused sensor inputs, including efficient sparse backbones, BEV systems, distillation, and unified multi-task architectures when detection-centered perception is a main objective.","inclusion":"Assign atoms whose primary program is 3D object detection or scalable/unified perception architectures principally evaluated through detection or closely coupled detection-segmentation perception outputs.","exclusion":"Dense voxel occupancy, free-space completion, or future occupancy grids belong to D2; pure semantic or instance mask prediction without boxes belongs to D1.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["temporal modeling for unified multi-task 3d perception from multi-camera images","ground plane prior for monocular 3d object depth estimation without auxiliary sensors","polar coordinate transformer for multi-camera 3d object detection in bird's-eye-view"]},{"direction_index":1,"label":"D1","name":"3D Segmentation and Semantic Scene Fields","description":"Assigning semantic, instance, panoptic, part, object-centric, or neural-field-based semantic labels to 3D scene elements, including supervised, weakly supervised, few-shot, domain-adaptive, neural-rendering, and object-decomposition approaches.","inclusion":"Assign atoms whose main deliverable is per-point, per-voxel, per-mesh, per-region, object-instance, panoptic, part-level, or neural-field semantic decomposition of 3D scenes.","exclusion":"Bounding-box-centered object localization belongs to D0; open-vocabulary or language/foundation-model semantic transfer as the primary scientific objective belongs to D4.","nearest_boundary":"D4 Language, Open-Vocabulary, and Structured Reasoning","representative_topics":["weakly supervised 3d instance segmentation via bounding box parameter voting and clustering","hierarchical part segmentation of objects in real-world noisy indoor 3d scans","multi-property neural radiance field synthesis for unified 3d scene understanding"]},{"direction_index":2,"label":"D2","name":"Dense Occupancy and Scene Completion","description":"Predicting dense 3D occupancy, free space, semantic occupancy, completed volumetric scenes, or occupancy-flow grids from partial observations in indoor, outdoor, driving, or cooperative settings.","inclusion":"Assign atoms whose primary output is a dense voxel, grid, or volumetric scene state representing occupied/free/semantic space, including static completion and future occupancy forecasting.","exclusion":"Point-wise motion, scene flow, moving-object segmentation, or temporal 4D analysis without occupancy as the main representation belongs to D3; surface/layout reconstruction without dense occupancy belongs to D5.","nearest_boundary":"D3 Dynamic 4D Scene Understanding","representative_topics":["hierarchical spatial-temporal networks for joint occupancy and flow prediction in 3d scenes","multi-modal hierarchical transformer for joint occupancy and flow field prediction in autonomous driving","allo-centric occupancy grid prediction for long-term urban traffic scene forecasting"]},{"direction_index":3,"label":"D3","name":"Dynamic 4D Scene Understanding","description":"Understanding temporal change, motion, correspondence, moving objects, temporal segmentation, scene flow, tracking, change detection, and evolving 4D scene representations.","inclusion":"Assign atoms whose primary objective is motion or temporal scene analysis, including dense scene flow, moving-object segmentation, 3D change detection, 4D point-cloud segmentation, temporal compression, dynamic object-level mapping, or sequential correspondence.","exclusion":"Future scene state represented mainly as dense occupancy grids belongs to D2; static segmentation or static reconstruction belongs to D1 or D5 respectively.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["unsupervised 3d scene flow estimation from monocular video using depth and pose priors","receding horizon moving object segmentation in 3d lidar using sparse 4d temporal-spatial convolutions","temporal variation-aware 4d point cloud segmentation for autonomous driving"]},{"direction_index":4,"label":"D4","name":"Language, Open-Vocabulary, and Structured Reasoning","description":"Connecting 3D scenes to language, open vocabularies, foundation models, large language models, visual grounding, VQA, dense captioning, cross-modal pretraining, and explicit relational scene graphs or structured semantic reasoning.","inclusion":"Assign atoms whose primary program is language-conditioned localization or answering, open-vocabulary/zero-shot 3D perception, 3D-language representation learning, LLM integration, dense captioning as a semantic-language task, or explicit object-relation graph prediction.","exclusion":"Closed-vocabulary segmentation architectures without language/foundation transfer belong to D1; dataset construction whose main contribution is a resource rather than a language/reasoning model belongs to D8.","nearest_boundary":"D1 3D Segmentation and Semantic Scene Fields","representative_topics":["phrase-aware 3d visual grounding with fine-grained contextual phrase-object alignment","zero-shot open-world 3d scene understanding via semantic abstraction of 2d vision-language models","word-level text decoupling and dense alignment for 3d point cloud visual grounding"]},{"direction_index":5,"label":"D5","name":"Geometric Reconstruction and Layout Recovery","description":"Recovering 3D scene geometry, surfaces, planes, room layouts, object shapes, articulated structures, camera/scan alignment, and structural decompositions from images, RGB-D streams, videos, scans, or partial observations.","inclusion":"Assign atoms whose primary objective is faithful geometric reconstruction, layout estimation, plane or structural recovery, shape completion from observations, active view selection for reconstruction, or articulation model discovery.","exclusion":"Dense semantic occupancy prediction belongs to D2; generative synthesis or controllable editing of novel scene content belongs to D6.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["real-time 3d plane reconstruction from posed monocular video via incremental fragment fusion","mcts-based proposal selection and refinement for indoor scene layout reconstruction","panoramic 3d room layout estimation via learnable manhattan hough transform on cubemap tiles"]},{"direction_index":6,"label":"D6","name":"Generative Scene Synthesis and Editing","description":"Generating, completing, inpainting, optimizing, planning, or controllably editing 3D scene content, layouts, object arrangements, human-scene interactions, or neural scene representations using generative priors.","inclusion":"Assign atoms whose primary objective is creating or modifying 3D scene content rather than recognizing, localizing, segmenting, or faithfully reconstructing an observed scene.","exclusion":"Observed-scene reconstruction or layout recovery without generative modification belongs to D5; discriminative semantic occupancy completion belongs to D2.","nearest_boundary":"D5 Geometric Reconstruction and Layout Recovery","representative_topics":["2d-supervised 3d scene prior learning via autoregressive latent decoding of layout and geometry","human-motion-aware autoregressive 3d indoor scene generation with furniture layout inference","free-form 3d scene inpainting with dual-stream gan for large diverse missing regions"]},{"direction_index":7,"label":"D7","name":"Reusable Representations and Data Resources","description":"Building enabling foundations for 3D scene understanding through self-supervised or unsupervised representation learning, masked or contrastive pretraining, pseudo-labeling, object-centric feature learning, datasets, benchmarks, annotation pipelines, and resource-level evaluation protocols.","inclusion":"Assign atoms whose main contribution is a reusable 3D representation objective or a data/benchmark/annotation resource intended to support multiple downstream 3D understanding tasks.","exclusion":"Task-specific segmentation, detection, reconstruction, grounding, or occupancy methods using datasets or pretraining only as tools belong to their respective task directions.","nearest_boundary":"D1 3D Segmentation and Semantic Scene Fields","representative_topics":["invariance analysis and contrastive pre-training for self-supervised 3d vision representations","cross-modal pretraining of image transformers for 3d point cloud classification and segmentation","large-scale 3d indoor object annotation dataset with texture-guided boundary semantics for navigation"]}],"split":"dev"}
{"episode_id":"F051_T2026-01","field_index":51,"field":"3d scene understanding","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"3D Object Detection and Scalable Perception","description":"Estimating 3D object boxes and object locations from camera, LiDAR, radar, or fused sensor inputs, including efficient sparse backbones, BEV systems, distillation, and unified multi-task architectures when detection-centered perception is a main objective.","inclusion":"Assign atoms whose primary program is 3D object detection or scalable/unified perception architectures principally evaluated through detection or closely coupled detection-segmentation perception outputs.","exclusion":"Dense voxel occupancy, free-space completion, or future occupancy grids belong to D2; pure semantic or instance mask prediction without boxes belongs to D1.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["temporal modeling for unified multi-task 3d perception from multi-camera images","ground plane prior for monocular 3d object depth estimation without auxiliary sensors","polar coordinate transformer for multi-camera 3d object detection in bird's-eye-view"]},{"direction_index":1,"label":"D1","name":"3D Segmentation and Semantic Scene Fields","description":"Assigning semantic, instance, panoptic, part, object-centric, or neural-field-based semantic labels to 3D scene elements, including supervised, weakly supervised, few-shot, domain-adaptive, neural-rendering, and object-decomposition approaches.","inclusion":"Assign atoms whose main deliverable is per-point, per-voxel, per-mesh, per-region, object-instance, panoptic, part-level, or neural-field semantic decomposition of 3D scenes.","exclusion":"Bounding-box-centered object localization belongs to D0; open-vocabulary or language/foundation-model semantic transfer as the primary scientific objective belongs to D4.","nearest_boundary":"D4 Language, Open-Vocabulary, and Structured Reasoning","representative_topics":["weakly supervised 3d instance segmentation via bounding box parameter voting and clustering","hierarchical part segmentation of objects in real-world noisy indoor 3d scans","multi-property neural radiance field synthesis for unified 3d scene understanding"]},{"direction_index":2,"label":"D2","name":"Dense Occupancy and Scene Completion","description":"Predicting dense 3D occupancy, free space, semantic occupancy, completed volumetric scenes, or occupancy-flow grids from partial observations in indoor, outdoor, driving, or cooperative settings.","inclusion":"Assign atoms whose primary output is a dense voxel, grid, or volumetric scene state representing occupied/free/semantic space, including static completion and future occupancy forecasting.","exclusion":"Point-wise motion, scene flow, moving-object segmentation, or temporal 4D analysis without occupancy as the main representation belongs to D3; surface/layout reconstruction without dense occupancy belongs to D5.","nearest_boundary":"D3 Dynamic 4D Scene Understanding","representative_topics":["hierarchical spatial-temporal networks for joint occupancy and flow prediction in 3d scenes","multi-modal hierarchical transformer for joint occupancy and flow field prediction in autonomous driving","allo-centric occupancy grid prediction for long-term urban traffic scene forecasting"]},{"direction_index":3,"label":"D3","name":"Dynamic 4D Scene Understanding","description":"Understanding temporal change, motion, correspondence, moving objects, temporal segmentation, scene flow, tracking, change detection, and evolving 4D scene representations.","inclusion":"Assign atoms whose primary objective is motion or temporal scene analysis, including dense scene flow, moving-object segmentation, 3D change detection, 4D point-cloud segmentation, temporal compression, dynamic object-level mapping, or sequential correspondence.","exclusion":"Future scene state represented mainly as dense occupancy grids belongs to D2; static segmentation or static reconstruction belongs to D1 or D5 respectively.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["unsupervised 3d scene flow estimation from monocular video using depth and pose priors","receding horizon moving object segmentation in 3d lidar using sparse 4d temporal-spatial convolutions","temporal variation-aware 4d point cloud segmentation for autonomous driving"]},{"direction_index":4,"label":"D4","name":"Language, Open-Vocabulary, and Structured Reasoning","description":"Connecting 3D scenes to language, open vocabularies, foundation models, large language models, visual grounding, VQA, dense captioning, cross-modal pretraining, and explicit relational scene graphs or structured semantic reasoning.","inclusion":"Assign atoms whose primary program is language-conditioned localization or answering, open-vocabulary/zero-shot 3D perception, 3D-language representation learning, LLM integration, dense captioning as a semantic-language task, or explicit object-relation graph prediction.","exclusion":"Closed-vocabulary segmentation architectures without language/foundation transfer belong to D1; dataset construction whose main contribution is a resource rather than a language/reasoning model belongs to D8.","nearest_boundary":"D1 3D Segmentation and Semantic Scene Fields","representative_topics":["phrase-aware 3d visual grounding with fine-grained contextual phrase-object alignment","zero-shot open-world 3d scene understanding via semantic abstraction of 2d vision-language models","word-level text decoupling and dense alignment for 3d point cloud visual grounding"]},{"direction_index":5,"label":"D5","name":"Geometric Reconstruction and Layout Recovery","description":"Recovering 3D scene geometry, surfaces, planes, room layouts, object shapes, articulated structures, camera/scan alignment, and structural decompositions from images, RGB-D streams, videos, scans, or partial observations.","inclusion":"Assign atoms whose primary objective is faithful geometric reconstruction, layout estimation, plane or structural recovery, shape completion from observations, active view selection for reconstruction, or articulation model discovery.","exclusion":"Dense semantic occupancy prediction belongs to D2; generative synthesis or controllable editing of novel scene content belongs to D6.","nearest_boundary":"D2 Dense Occupancy and Scene Completion","representative_topics":["real-time 3d plane reconstruction from posed monocular video via incremental fragment fusion","mcts-based proposal selection and refinement for indoor scene layout reconstruction","panoramic 3d room layout estimation via learnable manhattan hough transform on cubemap tiles"]},{"direction_index":6,"label":"D6","name":"Generative Scene Synthesis and Editing","description":"Generating, completing, inpainting, optimizing, planning, or controllably editing 3D scene content, layouts, object arrangements, human-scene interactions, or neural scene representations using generative priors.","inclusion":"Assign atoms whose primary objective is creating or modifying 3D scene content rather than recognizing, localizing, segmenting, or faithfully reconstructing an observed scene.","exclusion":"Observed-scene reconstruction or layout recovery without generative modification belongs to D5; discriminative semantic occupancy completion belongs to D2.","nearest_boundary":"D5 Geometric Reconstruction and Layout Recovery","representative_topics":["2d-supervised 3d scene prior learning via autoregressive latent decoding of layout and geometry","human-motion-aware autoregressive 3d indoor scene generation with furniture layout inference","free-form 3d scene inpainting with dual-stream gan for large diverse missing regions"]},{"direction_index":7,"label":"D7","name":"Reusable Representations and Data Resources","description":"Building enabling foundations for 3D scene understanding through self-supervised or unsupervised representation learning, masked or contrastive pretraining, pseudo-labeling, object-centric feature learning, datasets, benchmarks, annotation pipelines, and resource-level evaluation protocols.","inclusion":"Assign atoms whose main contribution is a reusable 3D representation objective or a data/benchmark/annotation resource intended to support multiple downstream 3D understanding tasks.","exclusion":"Task-specific segmentation, detection, reconstruction, grounding, or occupancy methods using datasets or pretraining only as tools belong to their respective task directions.","nearest_boundary":"D1 3D Segmentation and Semantic Scene Fields","representative_topics":["invariance analysis and contrastive pre-training for self-supervised 3d vision representations","cross-modal pretraining of image transformers for 3d point cloud classification and segmentation","large-scale 3d indoor object annotation dataset with texture-guided boundary semantics for navigation"]}],"split":"dev"}
{"episode_id":"F058_T2024-01","field_index":58,"field":"privacy-preserving machine learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Confidential computation and verification","description":"Design systems and protocols that execute or verify ML computations while keeping inputs, model states, client updates, or computation traces confidential and/or integrity-protected.","inclusion":"Include work whose primary objective is secure private inference, secure or robust aggregation, TEE/confidential-computing ML, zero-knowledge verification, information-flow control, homomorphic/secret-shared computation, or protocol co-design for confidential ML execution.","exclusion":"If the main objective is improving DP learning utility rather than secure execution, assign to D2; if the main objective is federated optimization or split-learning privacy rather than cryptographic/hardware verification, assign to D1.","nearest_boundary":"Collaborative private learning","representative_topics":["homomorphic encryption approximation for privacy-preserving transformer inference on ciphertext","privacy-preserving federated embedding aggregation with joint entity and client protection","tee-gpu cooperative execution for privacy and integrity preserving dnn training via matrix masking"]},{"direction_index":1,"label":"D1","name":"Collaborative private learning","description":"Enable privacy-preserving learning across distributed parties, clients, vertical feature holders, split architectures, and domain deployments while managing communication, personalization, heterogeneity, and leakage from shared training signals.","inclusion":"Include work whose primary objective is privacy in FL, split learning, vertical FL, distributed collaborative training, secure or private update exchange when tied to FL protocols, domain-adapted FL systems, communication-efficient private FL, gradient leakage defenses in FL, or privacy-utility evaluation of collaborative learning.","exclusion":"If cryptographic execution or formal verification is the central contribution, assign to D0; if the contribution is a general DP optimizer independent of collaborative architecture, assign to D2; if graph structure is the primary privacy object, assign to D6.","nearest_boundary":"Confidential computation and verification","representative_topics":["federated learning defense via synthetic data generation to obscure client data from gradient leakage attacks","group privacy via metric obfuscation for personalized federated learning with heterogeneous clients","communication-compressed private federated learning via unified local gradient estimator framework"]},{"direction_index":2,"label":"D2","name":"Private learning algorithms and theory","description":"Develop differentially private algorithms, utility-improving training procedures, statistical estimators, optimization theory, and sequential decision methods with formal privacy-performance guarantees.","inclusion":"Include work whose primary objective is DP-SGD or private stochastic optimization, clipping/noise/variance methods, private statistical estimation, private regression or distribution learning, DP learning theory, private bandits or reinforcement learning, and general private optimization rates or lower bounds.","exclusion":"If the main contribution is privacy-loss accounting or auditing rather than a learning algorithm, assign to D3; if the algorithm is primarily an FL protocol or deployment, assign to D1; if the output is a synthetic releasable dataset, assign to D4.","nearest_boundary":"Privacy guarantees, accounting, and auditing","representative_topics":["differentially private offline reinforcement learning with instance-dependent learning bounds","differentially private optimization with faster stationary point convergence rates in finite-sum and stochastic settings","dimension-independent generalization bounds for differentially private sgd via langevin convergence"]},{"direction_index":3,"label":"D3","name":"Privacy guarantees, accounting, and auditing","description":"Measure, compose, estimate, audit, and validate privacy guarantees and privacy loss in ML systems, especially differential privacy mechanisms.","inclusion":"Include work whose primary objective is DP accounting, composition, privacy amplification, individual privacy accounting, empirical privacy estimation, privacy auditing frameworks, reproducibility of privacy claims, or systematic verification of achieved privacy levels.","exclusion":"If membership inference is the main attack or defense contribution, assign to D5; if accounting is only a tool supporting a new optimizer or estimator, assign to D2.","nearest_boundary":"Leakage attacks, defenses, and deletion","representative_topics":["individual example privacy accounting in differentially private stochastic gradient descent","privacy amplification for distributed computation via shuffled check-in protocols and rényi differential privacy","edgeworth expansion for tight non-asymptotic differential privacy composition bounds"]},{"direction_index":4,"label":"D4","name":"Private synthetic data release","description":"Generate, publish, evaluate, or govern synthetic datasets and data releases that protect individuals while preserving statistical or downstream-task utility.","inclusion":"Include work whose primary objective is DP synthetic data, private data publishing, private generative modeling for releasable datasets, dataset condensation for privacy, synthetic tabular/text/image/time-series release, and evaluation or risk analysis of private synthetic data.","exclusion":"If synthetic data is only an auxiliary defense for FL leakage, assign to D1; if the work studies membership attacks on generators rather than release construction, assign to D5; if it is perceptual de-identification without formal release-oriented synthesis, assign to D6.","nearest_boundary":"Leakage attacks, defenses, and deletion","representative_topics":["dataset condensation as differentially private data generation for efficient privacy-preserving training","differentially private synthetic data generation for hierarchical query release with group-individual statistics","synthetic data-driven face recognition under strict privacy constraints via generative adversarial networks"]},{"direction_index":5,"label":"D5","name":"Leakage attacks, defenses, and deletion","description":"Discover, analyze, benchmark, mitigate, or reverse privacy leakage from trained models and training pipelines, including attacks on membership, reconstruction, attributes, properties, functionality, and residual training-data influence.","inclusion":"Include work whose primary objective is membership inference, model inversion, gradient or data reconstruction, attribute/property/distribution inference, model extraction or functionality stealing, memorization leakage outside a modality-specific program, defenses against these attacks, and machine unlearning or data deletion.","exclusion":"If the main objective is estimating a formal DP budget or auditing a claimed guarantee rather than creating an attack/defense, assign to D3; if language-model leakage is addressed as part of a broader LLM/text privacy program, assign to D6; if the central objective is fairness-privacy trade-off rather than leakage exploitation, assign to D7.","nearest_boundary":"Privacy guarantees, accounting, and auditing","representative_topics":["privacy guarantees and membership inference robustness of generative adversarial network outputs","data stealing attacks on medical image models exported from remote data lakes","bilateral dependency optimization for model-inversion attack defense in classifiers"]},{"direction_index":6,"label":"D6","name":"Modality-tailored privacy mechanisms","description":"Protect privacy in ML systems where the data modality, structure, or domain semantics shape the mechanism, including language/text, graphs, biomedical data, perceptual signals, biometrics, and other domain-specific non-FL settings.","inclusion":"Include work whose primary objective is privacy for LLMs or text, graph data or GNNs, medical/clinical data, visual/audio/biometric signals, perceptual obfuscation, de-identification, domain-specific DP mechanisms, privacy-preserving explanations, or content-aware privacy transformations.","exclusion":"If the mechanism is a general DP optimizer or estimator without modality-specific design, assign to D2; if the work is primarily FL deployment across clients, assign to D1; if the primary contribution is an attack taxonomy independent of modality, assign to D5.","nearest_boundary":"Private learning algorithms and theory","representative_topics":["differentially private sparse compression of fine-tuned language models","privacy-preserving human action recognition via camera lens-based video degradation","training data extraction and privacy leakage in text classification models"]},{"direction_index":7,"label":"D7","name":"Privacy, fairness, consent, and utility","description":"Understand and optimize interactions among privacy mechanisms, group fairness, consent, trustworthiness, and utility across ML systems.","inclusion":"Include work whose primary objective is analyzing or mitigating disparate impact of privacy mechanisms, jointly private and fair training, privacy-fairness-utility bounds, consent-aware learning, or relationships between fairness constraints and privacy leakage.","exclusion":"If fairness is absent and the objective is only DP utility optimization, assign to D2; if the work mainly attacks sensitive attributes rather than optimizing fairness/privacy trade-offs, assign to D5; if biomedical or other domain context dominates without fairness as the primary objective, assign to D6.","nearest_boundary":"Private learning algorithms and theory","representative_topics":["accuracy-fairness tradeoffs in private learning under long-tailed data distributions","reducing disparate impact in differentially private training via gradient misalignment correction","privacy-preserving trustworthy machine learning balancing fairness, robustness, and explainability"]}],"split":"dev"}
{"episode_id":"F058_T2024-07","field_index":58,"field":"privacy-preserving machine learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Confidential computation and verification","description":"Design systems and protocols that execute or verify ML computations while keeping inputs, model states, client updates, or computation traces confidential and/or integrity-protected.","inclusion":"Include work whose primary objective is secure private inference, secure or robust aggregation, TEE/confidential-computing ML, zero-knowledge verification, information-flow control, homomorphic/secret-shared computation, or protocol co-design for confidential ML execution.","exclusion":"If the main objective is improving DP learning utility rather than secure execution, assign to D2; if the main objective is federated optimization or split-learning privacy rather than cryptographic/hardware verification, assign to D1.","nearest_boundary":"Collaborative private learning","representative_topics":["homomorphic encryption approximation for privacy-preserving transformer inference on ciphertext","privacy-preserving federated embedding aggregation with joint entity and client protection","tee-gpu cooperative execution for privacy and integrity preserving dnn training via matrix masking"]},{"direction_index":1,"label":"D1","name":"Collaborative private learning","description":"Enable privacy-preserving learning across distributed parties, clients, vertical feature holders, split architectures, and domain deployments while managing communication, personalization, heterogeneity, and leakage from shared training signals.","inclusion":"Include work whose primary objective is privacy in FL, split learning, vertical FL, distributed collaborative training, secure or private update exchange when tied to FL protocols, domain-adapted FL systems, communication-efficient private FL, gradient leakage defenses in FL, or privacy-utility evaluation of collaborative learning.","exclusion":"If cryptographic execution or formal verification is the central contribution, assign to D0; if the contribution is a general DP optimizer independent of collaborative architecture, assign to D2; if graph structure is the primary privacy object, assign to D6.","nearest_boundary":"Confidential computation and verification","representative_topics":["federated learning defense via synthetic data generation to obscure client data from gradient leakage attacks","group privacy via metric obfuscation for personalized federated learning with heterogeneous clients","communication-compressed private federated learning via unified local gradient estimator framework"]},{"direction_index":2,"label":"D2","name":"Private learning algorithms and theory","description":"Develop differentially private algorithms, utility-improving training procedures, statistical estimators, optimization theory, and sequential decision methods with formal privacy-performance guarantees.","inclusion":"Include work whose primary objective is DP-SGD or private stochastic optimization, clipping/noise/variance methods, private statistical estimation, private regression or distribution learning, DP learning theory, private bandits or reinforcement learning, and general private optimization rates or lower bounds.","exclusion":"If the main contribution is privacy-loss accounting or auditing rather than a learning algorithm, assign to D3; if the algorithm is primarily an FL protocol or deployment, assign to D1; if the output is a synthetic releasable dataset, assign to D4.","nearest_boundary":"Privacy guarantees, accounting, and auditing","representative_topics":["differentially private offline reinforcement learning with instance-dependent learning bounds","differentially private optimization with faster stationary point convergence rates in finite-sum and stochastic settings","dimension-independent generalization bounds for differentially private sgd via langevin convergence"]},{"direction_index":3,"label":"D3","name":"Privacy guarantees, accounting, and auditing","description":"Measure, compose, estimate, audit, and validate privacy guarantees and privacy loss in ML systems, especially differential privacy mechanisms.","inclusion":"Include work whose primary objective is DP accounting, composition, privacy amplification, individual privacy accounting, empirical privacy estimation, privacy auditing frameworks, reproducibility of privacy claims, or systematic verification of achieved privacy levels.","exclusion":"If membership inference is the main attack or defense contribution, assign to D5; if accounting is only a tool supporting a new optimizer or estimator, assign to D2.","nearest_boundary":"Leakage attacks, defenses, and deletion","representative_topics":["individual example privacy accounting in differentially private stochastic gradient descent","privacy amplification for distributed computation via shuffled check-in protocols and rényi differential privacy","edgeworth expansion for tight non-asymptotic differential privacy composition bounds"]},{"direction_index":4,"label":"D4","name":"Private synthetic data release","description":"Generate, publish, evaluate, or govern synthetic datasets and data releases that protect individuals while preserving statistical or downstream-task utility.","inclusion":"Include work whose primary objective is DP synthetic data, private data publishing, private generative modeling for releasable datasets, dataset condensation for privacy, synthetic tabular/text/image/time-series release, and evaluation or risk analysis of private synthetic data.","exclusion":"If synthetic data is only an auxiliary defense for FL leakage, assign to D1; if the work studies membership attacks on generators rather than release construction, assign to D5; if it is perceptual de-identification without formal release-oriented synthesis, assign to D6.","nearest_boundary":"Leakage attacks, defenses, and deletion","representative_topics":["dataset condensation as differentially private data generation for efficient privacy-preserving training","differentially private synthetic data generation for hierarchical query release with group-individual statistics","synthetic data-driven face recognition under strict privacy constraints via generative adversarial networks"]},{"direction_index":5,"label":"D5","name":"Leakage attacks, defenses, and deletion","description":"Discover, analyze, benchmark, mitigate, or reverse privacy leakage from trained models and training pipelines, including attacks on membership, reconstruction, attributes, properties, functionality, and residual training-data influence.","inclusion":"Include work whose primary objective is membership inference, model inversion, gradient or data reconstruction, attribute/property/distribution inference, model extraction or functionality stealing, memorization leakage outside a modality-specific program, defenses against these attacks, and machine unlearning or data deletion.","exclusion":"If the main objective is estimating a formal DP budget or auditing a claimed guarantee rather than creating an attack/defense, assign to D3; if language-model leakage is addressed as part of a broader LLM/text privacy program, assign to D6; if the central objective is fairness-privacy trade-off rather than leakage exploitation, assign to D7.","nearest_boundary":"Privacy guarantees, accounting, and auditing","representative_topics":["privacy guarantees and membership inference robustness of generative adversarial network outputs","data stealing attacks on medical image models exported from remote data lakes","bilateral dependency optimization for model-inversion attack defense in classifiers"]},{"direction_index":6,"label":"D6","name":"Modality-tailored privacy mechanisms","description":"Protect privacy in ML systems where the data modality, structure, or domain semantics shape the mechanism, including language/text, graphs, biomedical data, perceptual signals, biometrics, and other domain-specific non-FL settings.","inclusion":"Include work whose primary objective is privacy for LLMs or text, graph data or GNNs, medical/clinical data, visual/audio/biometric signals, perceptual obfuscation, de-identification, domain-specific DP mechanisms, privacy-preserving explanations, or content-aware privacy transformations.","exclusion":"If the mechanism is a general DP optimizer or estimator without modality-specific design, assign to D2; if the work is primarily FL deployment across clients, assign to D1; if the primary contribution is an attack taxonomy independent of modality, assign to D5.","nearest_boundary":"Private learning algorithms and theory","representative_topics":["differentially private sparse compression of fine-tuned language models","privacy-preserving human action recognition via camera lens-based video degradation","training data extraction and privacy leakage in text classification models"]},{"direction_index":7,"label":"D7","name":"Privacy, fairness, consent, and utility","description":"Understand and optimize interactions among privacy mechanisms, group fairness, consent, trustworthiness, and utility across ML systems.","inclusion":"Include work whose primary objective is analyzing or mitigating disparate impact of privacy mechanisms, jointly private and fair training, privacy-fairness-utility bounds, consent-aware learning, or relationships between fairness constraints and privacy leakage.","exclusion":"If fairness is absent and the objective is only DP utility optimization, assign to D2; if the work mainly attacks sensitive attributes rather than optimizing fairness/privacy trade-offs, assign to D5; if biomedical or other domain context dominates without fairness as the primary objective, assign to D6.","nearest_boundary":"Private learning algorithms and theory","representative_topics":["accuracy-fairness tradeoffs in private learning under long-tailed data distributions","reducing disparate impact in differentially private training via gradient misalignment correction","privacy-preserving trustworthy machine learning balancing fairness, robustness, and explainability"]}],"split":"dev"}
{"episode_id":"F058_T2025-01","field_index":58,"field":"privacy-preserving machine learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Confidential computation and verification","description":"Design systems and protocols that execute or verify ML computations while keeping inputs, model states, client updates, or computation traces confidential and/or integrity-protected.","inclusion":"Include work whose primary objective is secure private inference, secure or robust aggregation, TEE/confidential-computing ML, zero-knowledge verification, information-flow control, homomorphic/secret-shared computation, or protocol co-design for confidential ML execution.","exclusion":"If the main objective is improving DP learning utility rather than secure execution, assign to D2; if the main objective is federated optimization or split-learning privacy rather than cryptographic/hardware verification, assign to D1.","nearest_boundary":"Collaborative private learning","representative_topics":["homomorphic encryption approximation for privacy-preserving transformer inference on ciphertext","privacy-preserving federated embedding aggregation with joint entity and client protection","tee-gpu cooperative execution for privacy and integrity preserving dnn training via matrix masking"]},{"direction_index":1,"label":"D1","name":"Collaborative private learning","description":"Enable privacy-preserving learning across distributed parties, clients, vertical feature holders, split architectures, and domain deployments while managing communication, personalization, heterogeneity, and leakage from shared training signals.","inclusion":"Include work whose primary objective is privacy in FL, split learning, vertical FL, distributed collaborative training, secure or private update exchange when tied to FL protocols, domain-adapted FL systems, communication-efficient private FL, gradient leakage defenses in FL, or privacy-utility evaluation of collaborative learning.","exclusion":"If cryptographic execution or formal verification is the central contribution, assign to D0; if the contribution is a general DP optimizer independent of collaborative architecture, assign to D2; if graph structure is the primary privacy object, assign to D6.","nearest_boundary":"Confidential computation and verification","representative_topics":["federated learning defense via synthetic data generation to obscure client data from gradient leakage attacks","group privacy via metric obfuscation for personalized federated learning with heterogeneous clients","communication-compressed private federated learning via unified local gradient estimator framework"]},{"direction_index":2,"label":"D2","name":"Private learning algorithms and theory","description":"Develop differentially private algorithms, utility-improving training procedures, statistical estimators, optimization theory, and sequential decision methods with formal privacy-performance guarantees.","inclusion":"Include work whose primary objective is DP-SGD or private stochastic optimization, clipping/noise/variance methods, private statistical estimation, private regression or distribution learning, DP learning theory, private bandits or reinforcement learning, and general private optimization rates or lower bounds.","exclusion":"If the main contribution is privacy-loss accounting or auditing rather than a learning algorithm, assign to D3; if the algorithm is primarily an FL protocol or deployment, assign to D1; if the output is a synthetic releasable dataset, assign to D4.","nearest_boundary":"Privacy guarantees, accounting, and auditing","representative_topics":["differentially private offline reinforcement learning with instance-dependent learning bounds","differentially private optimization with faster stationary point convergence rates in finite-sum and stochastic settings","dimension-independent generalization bounds for differentially private sgd via langevin convergence"]},{"direction_index":3,"label":"D3","name":"Privacy guarantees, accounting, and auditing","description":"Measure, compose, estimate, audit, and validate privacy guarantees and privacy loss in ML systems, especially differential privacy mechanisms.","inclusion":"Include work whose primary objective is DP accounting, composition, privacy amplification, individual privacy accounting, empirical privacy estimation, privacy auditing frameworks, reproducibility of privacy claims, or systematic verification of achieved privacy levels.","exclusion":"If membership inference is the main attack or defense contribution, assign to D5; if accounting is only a tool supporting a new optimizer or estimator, assign to D2.","nearest_boundary":"Leakage attacks, defenses, and deletion","representative_topics":["individual example privacy accounting in differentially private stochastic gradient descent","privacy amplification for distributed computation via shuffled check-in protocols and rényi differential privacy","edgeworth expansion for tight non-asymptotic differential privacy composition bounds"]},{"direction_index":4,"label":"D4","name":"Private synthetic data release","description":"Generate, publish, evaluate, or govern synthetic datasets and data releases that protect individuals while preserving statistical or downstream-task utility.","inclusion":"Include work whose primary objective is DP synthetic data, private data publishing, private generative modeling for releasable datasets, dataset condensation for privacy, synthetic tabular/text/image/time-series release, and evaluation or risk analysis of private synthetic data.","exclusion":"If synthetic data is only an auxiliary defense for FL leakage, assign to D1; if the work studies membership attacks on generators rather than release construction, assign to D5; if it is perceptual de-identification without formal release-oriented synthesis, assign to D6.","nearest_boundary":"Leakage attacks, defenses, and deletion","representative_topics":["dataset condensation as differentially private data generation for efficient privacy-preserving training","differentially private synthetic data generation for hierarchical query release with group-individual statistics","synthetic data-driven face recognition under strict privacy constraints via generative adversarial networks"]},{"direction_index":5,"label":"D5","name":"Leakage attacks, defenses, and deletion","description":"Discover, analyze, benchmark, mitigate, or reverse privacy leakage from trained models and training pipelines, including attacks on membership, reconstruction, attributes, properties, functionality, and residual training-data influence.","inclusion":"Include work whose primary objective is membership inference, model inversion, gradient or data reconstruction, attribute/property/distribution inference, model extraction or functionality stealing, memorization leakage outside a modality-specific program, defenses against these attacks, and machine unlearning or data deletion.","exclusion":"If the main objective is estimating a formal DP budget or auditing a claimed guarantee rather than creating an attack/defense, assign to D3; if language-model leakage is addressed as part of a broader LLM/text privacy program, assign to D6; if the central objective is fairness-privacy trade-off rather than leakage exploitation, assign to D7.","nearest_boundary":"Privacy guarantees, accounting, and auditing","representative_topics":["privacy guarantees and membership inference robustness of generative adversarial network outputs","data stealing attacks on medical image models exported from remote data lakes","bilateral dependency optimization for model-inversion attack defense in classifiers"]},{"direction_index":6,"label":"D6","name":"Modality-tailored privacy mechanisms","description":"Protect privacy in ML systems where the data modality, structure, or domain semantics shape the mechanism, including language/text, graphs, biomedical data, perceptual signals, biometrics, and other domain-specific non-FL settings.","inclusion":"Include work whose primary objective is privacy for LLMs or text, graph data or GNNs, medical/clinical data, visual/audio/biometric signals, perceptual obfuscation, de-identification, domain-specific DP mechanisms, privacy-preserving explanations, or content-aware privacy transformations.","exclusion":"If the mechanism is a general DP optimizer or estimator without modality-specific design, assign to D2; if the work is primarily FL deployment across clients, assign to D1; if the primary contribution is an attack taxonomy independent of modality, assign to D5.","nearest_boundary":"Private learning algorithms and theory","representative_topics":["differentially private sparse compression of fine-tuned language models","privacy-preserving human action recognition via camera lens-based video degradation","training data extraction and privacy leakage in text classification models"]},{"direction_index":7,"label":"D7","name":"Privacy, fairness, consent, and utility","description":"Understand and optimize interactions among privacy mechanisms, group fairness, consent, trustworthiness, and utility across ML systems.","inclusion":"Include work whose primary objective is analyzing or mitigating disparate impact of privacy mechanisms, jointly private and fair training, privacy-fairness-utility bounds, consent-aware learning, or relationships between fairness constraints and privacy leakage.","exclusion":"If fairness is absent and the objective is only DP utility optimization, assign to D2; if the work mainly attacks sensitive attributes rather than optimizing fairness/privacy trade-offs, assign to D5; if biomedical or other domain context dominates without fairness as the primary objective, assign to D6.","nearest_boundary":"Private learning algorithms and theory","representative_topics":["accuracy-fairness tradeoffs in private learning under long-tailed data distributions","reducing disparate impact in differentially private training via gradient misalignment correction","privacy-preserving trustworthy machine learning balancing fairness, robustness, and explainability"]}],"split":"dev"}
{"episode_id":"F058_T2025-07","field_index":58,"field":"privacy-preserving machine learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Confidential computation and verification","description":"Design systems and protocols that execute or verify ML computations while keeping inputs, model states, client updates, or computation traces confidential and/or integrity-protected.","inclusion":"Include work whose primary objective is secure private inference, secure or robust aggregation, TEE/confidential-computing ML, zero-knowledge verification, information-flow control, homomorphic/secret-shared computation, or protocol co-design for confidential ML execution.","exclusion":"If the main objective is improving DP learning utility rather than secure execution, assign to D2; if the main objective is federated optimization or split-learning privacy rather than cryptographic/hardware verification, assign to D1.","nearest_boundary":"Collaborative private learning","representative_topics":["homomorphic encryption approximation for privacy-preserving transformer inference on ciphertext","privacy-preserving federated embedding aggregation with joint entity and client protection","tee-gpu cooperative execution for privacy and integrity preserving dnn training via matrix masking"]},{"direction_index":1,"label":"D1","name":"Collaborative private learning","description":"Enable privacy-preserving learning across distributed parties, clients, vertical feature holders, split architectures, and domain deployments while managing communication, personalization, heterogeneity, and leakage from shared training signals.","inclusion":"Include work whose primary objective is privacy in FL, split learning, vertical FL, distributed collaborative training, secure or private update exchange when tied to FL protocols, domain-adapted FL systems, communication-efficient private FL, gradient leakage defenses in FL, or privacy-utility evaluation of collaborative learning.","exclusion":"If cryptographic execution or formal verification is the central contribution, assign to D0; if the contribution is a general DP optimizer independent of collaborative architecture, assign to D2; if graph structure is the primary privacy object, assign to D6.","nearest_boundary":"Confidential computation and verification","representative_topics":["federated learning defense via synthetic data generation to obscure client data from gradient leakage attacks","group privacy via metric obfuscation for personalized federated learning with heterogeneous clients","communication-compressed private federated learning via unified local gradient estimator framework"]},{"direction_index":2,"label":"D2","name":"Private learning algorithms and theory","description":"Develop differentially private algorithms, utility-improving training procedures, statistical estimators, optimization theory, and sequential decision methods with formal privacy-performance guarantees.","inclusion":"Include work whose primary objective is DP-SGD or private stochastic optimization, clipping/noise/variance methods, private statistical estimation, private regression or distribution learning, DP learning theory, private bandits or reinforcement learning, and general private optimization rates or lower bounds.","exclusion":"If the main contribution is privacy-loss accounting or auditing rather than a learning algorithm, assign to D3; if the algorithm is primarily an FL protocol or deployment, assign to D1; if the output is a synthetic releasable dataset, assign to D4.","nearest_boundary":"Privacy guarantees, accounting, and auditing","representative_topics":["differentially private offline reinforcement learning with instance-dependent learning bounds","differentially private optimization with faster stationary point convergence rates in finite-sum and stochastic settings","dimension-independent generalization bounds for differentially private sgd via langevin convergence"]},{"direction_index":3,"label":"D3","name":"Privacy guarantees, accounting, and auditing","description":"Measure, compose, estimate, audit, and validate privacy guarantees and privacy loss in ML systems, especially differential privacy mechanisms.","inclusion":"Include work whose primary objective is DP accounting, composition, privacy amplification, individual privacy accounting, empirical privacy estimation, privacy auditing frameworks, reproducibility of privacy claims, or systematic verification of achieved privacy levels.","exclusion":"If membership inference is the main attack or defense contribution, assign to D5; if accounting is only a tool supporting a new optimizer or estimator, assign to D2.","nearest_boundary":"Leakage attacks, defenses, and deletion","representative_topics":["individual example privacy accounting in differentially private stochastic gradient descent","privacy amplification for distributed computation via shuffled check-in protocols and rényi differential privacy","edgeworth expansion for tight non-asymptotic differential privacy composition bounds"]},{"direction_index":4,"label":"D4","name":"Private synthetic data release","description":"Generate, publish, evaluate, or govern synthetic datasets and data releases that protect individuals while preserving statistical or downstream-task utility.","inclusion":"Include work whose primary objective is DP synthetic data, private data publishing, private generative modeling for releasable datasets, dataset condensation for privacy, synthetic tabular/text/image/time-series release, and evaluation or risk analysis of private synthetic data.","exclusion":"If synthetic data is only an auxiliary defense for FL leakage, assign to D1; if the work studies membership attacks on generators rather than release construction, assign to D5; if it is perceptual de-identification without formal release-oriented synthesis, assign to D6.","nearest_boundary":"Leakage attacks, defenses, and deletion","representative_topics":["dataset condensation as differentially private data generation for efficient privacy-preserving training","differentially private synthetic data generation for hierarchical query release with group-individual statistics","synthetic data-driven face recognition under strict privacy constraints via generative adversarial networks"]},{"direction_index":5,"label":"D5","name":"Leakage attacks, defenses, and deletion","description":"Discover, analyze, benchmark, mitigate, or reverse privacy leakage from trained models and training pipelines, including attacks on membership, reconstruction, attributes, properties, functionality, and residual training-data influence.","inclusion":"Include work whose primary objective is membership inference, model inversion, gradient or data reconstruction, attribute/property/distribution inference, model extraction or functionality stealing, memorization leakage outside a modality-specific program, defenses against these attacks, and machine unlearning or data deletion.","exclusion":"If the main objective is estimating a formal DP budget or auditing a claimed guarantee rather than creating an attack/defense, assign to D3; if language-model leakage is addressed as part of a broader LLM/text privacy program, assign to D6; if the central objective is fairness-privacy trade-off rather than leakage exploitation, assign to D7.","nearest_boundary":"Privacy guarantees, accounting, and auditing","representative_topics":["privacy guarantees and membership inference robustness of generative adversarial network outputs","data stealing attacks on medical image models exported from remote data lakes","bilateral dependency optimization for model-inversion attack defense in classifiers"]},{"direction_index":6,"label":"D6","name":"Modality-tailored privacy mechanisms","description":"Protect privacy in ML systems where the data modality, structure, or domain semantics shape the mechanism, including language/text, graphs, biomedical data, perceptual signals, biometrics, and other domain-specific non-FL settings.","inclusion":"Include work whose primary objective is privacy for LLMs or text, graph data or GNNs, medical/clinical data, visual/audio/biometric signals, perceptual obfuscation, de-identification, domain-specific DP mechanisms, privacy-preserving explanations, or content-aware privacy transformations.","exclusion":"If the mechanism is a general DP optimizer or estimator without modality-specific design, assign to D2; if the work is primarily FL deployment across clients, assign to D1; if the primary contribution is an attack taxonomy independent of modality, assign to D5.","nearest_boundary":"Private learning algorithms and theory","representative_topics":["differentially private sparse compression of fine-tuned language models","privacy-preserving human action recognition via camera lens-based video degradation","training data extraction and privacy leakage in text classification models"]},{"direction_index":7,"label":"D7","name":"Privacy, fairness, consent, and utility","description":"Understand and optimize interactions among privacy mechanisms, group fairness, consent, trustworthiness, and utility across ML systems.","inclusion":"Include work whose primary objective is analyzing or mitigating disparate impact of privacy mechanisms, jointly private and fair training, privacy-fairness-utility bounds, consent-aware learning, or relationships between fairness constraints and privacy leakage.","exclusion":"If fairness is absent and the objective is only DP utility optimization, assign to D2; if the work mainly attacks sensitive attributes rather than optimizing fairness/privacy trade-offs, assign to D5; if biomedical or other domain context dominates without fairness as the primary objective, assign to D6.","nearest_boundary":"Private learning algorithms and theory","representative_topics":["accuracy-fairness tradeoffs in private learning under long-tailed data distributions","reducing disparate impact in differentially private training via gradient misalignment correction","privacy-preserving trustworthy machine learning balancing fairness, robustness, and explainability"]}],"split":"dev"}
{"episode_id":"F058_T2026-01","field_index":58,"field":"privacy-preserving machine learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Confidential computation and verification","description":"Design systems and protocols that execute or verify ML computations while keeping inputs, model states, client updates, or computation traces confidential and/or integrity-protected.","inclusion":"Include work whose primary objective is secure private inference, secure or robust aggregation, TEE/confidential-computing ML, zero-knowledge verification, information-flow control, homomorphic/secret-shared computation, or protocol co-design for confidential ML execution.","exclusion":"If the main objective is improving DP learning utility rather than secure execution, assign to D2; if the main objective is federated optimization or split-learning privacy rather than cryptographic/hardware verification, assign to D1.","nearest_boundary":"Collaborative private learning","representative_topics":["homomorphic encryption approximation for privacy-preserving transformer inference on ciphertext","privacy-preserving federated embedding aggregation with joint entity and client protection","tee-gpu cooperative execution for privacy and integrity preserving dnn training via matrix masking"]},{"direction_index":1,"label":"D1","name":"Collaborative private learning","description":"Enable privacy-preserving learning across distributed parties, clients, vertical feature holders, split architectures, and domain deployments while managing communication, personalization, heterogeneity, and leakage from shared training signals.","inclusion":"Include work whose primary objective is privacy in FL, split learning, vertical FL, distributed collaborative training, secure or private update exchange when tied to FL protocols, domain-adapted FL systems, communication-efficient private FL, gradient leakage defenses in FL, or privacy-utility evaluation of collaborative learning.","exclusion":"If cryptographic execution or formal verification is the central contribution, assign to D0; if the contribution is a general DP optimizer independent of collaborative architecture, assign to D2; if graph structure is the primary privacy object, assign to D6.","nearest_boundary":"Confidential computation and verification","representative_topics":["federated learning defense via synthetic data generation to obscure client data from gradient leakage attacks","group privacy via metric obfuscation for personalized federated learning with heterogeneous clients","communication-compressed private federated learning via unified local gradient estimator framework"]},{"direction_index":2,"label":"D2","name":"Private learning algorithms and theory","description":"Develop differentially private algorithms, utility-improving training procedures, statistical estimators, optimization theory, and sequential decision methods with formal privacy-performance guarantees.","inclusion":"Include work whose primary objective is DP-SGD or private stochastic optimization, clipping/noise/variance methods, private statistical estimation, private regression or distribution learning, DP learning theory, private bandits or reinforcement learning, and general private optimization rates or lower bounds.","exclusion":"If the main contribution is privacy-loss accounting or auditing rather than a learning algorithm, assign to D3; if the algorithm is primarily an FL protocol or deployment, assign to D1; if the output is a synthetic releasable dataset, assign to D4.","nearest_boundary":"Privacy guarantees, accounting, and auditing","representative_topics":["differentially private offline reinforcement learning with instance-dependent learning bounds","differentially private optimization with faster stationary point convergence rates in finite-sum and stochastic settings","dimension-independent generalization bounds for differentially private sgd via langevin convergence"]},{"direction_index":3,"label":"D3","name":"Privacy guarantees, accounting, and auditing","description":"Measure, compose, estimate, audit, and validate privacy guarantees and privacy loss in ML systems, especially differential privacy mechanisms.","inclusion":"Include work whose primary objective is DP accounting, composition, privacy amplification, individual privacy accounting, empirical privacy estimation, privacy auditing frameworks, reproducibility of privacy claims, or systematic verification of achieved privacy levels.","exclusion":"If membership inference is the main attack or defense contribution, assign to D5; if accounting is only a tool supporting a new optimizer or estimator, assign to D2.","nearest_boundary":"Leakage attacks, defenses, and deletion","representative_topics":["individual example privacy accounting in differentially private stochastic gradient descent","privacy amplification for distributed computation via shuffled check-in protocols and rényi differential privacy","edgeworth expansion for tight non-asymptotic differential privacy composition bounds"]},{"direction_index":4,"label":"D4","name":"Private synthetic data release","description":"Generate, publish, evaluate, or govern synthetic datasets and data releases that protect individuals while preserving statistical or downstream-task utility.","inclusion":"Include work whose primary objective is DP synthetic data, private data publishing, private generative modeling for releasable datasets, dataset condensation for privacy, synthetic tabular/text/image/time-series release, and evaluation or risk analysis of private synthetic data.","exclusion":"If synthetic data is only an auxiliary defense for FL leakage, assign to D1; if the work studies membership attacks on generators rather than release construction, assign to D5; if it is perceptual de-identification without formal release-oriented synthesis, assign to D6.","nearest_boundary":"Leakage attacks, defenses, and deletion","representative_topics":["dataset condensation as differentially private data generation for efficient privacy-preserving training","differentially private synthetic data generation for hierarchical query release with group-individual statistics","synthetic data-driven face recognition under strict privacy constraints via generative adversarial networks"]},{"direction_index":5,"label":"D5","name":"Leakage attacks, defenses, and deletion","description":"Discover, analyze, benchmark, mitigate, or reverse privacy leakage from trained models and training pipelines, including attacks on membership, reconstruction, attributes, properties, functionality, and residual training-data influence.","inclusion":"Include work whose primary objective is membership inference, model inversion, gradient or data reconstruction, attribute/property/distribution inference, model extraction or functionality stealing, memorization leakage outside a modality-specific program, defenses against these attacks, and machine unlearning or data deletion.","exclusion":"If the main objective is estimating a formal DP budget or auditing a claimed guarantee rather than creating an attack/defense, assign to D3; if language-model leakage is addressed as part of a broader LLM/text privacy program, assign to D6; if the central objective is fairness-privacy trade-off rather than leakage exploitation, assign to D7.","nearest_boundary":"Privacy guarantees, accounting, and auditing","representative_topics":["privacy guarantees and membership inference robustness of generative adversarial network outputs","data stealing attacks on medical image models exported from remote data lakes","bilateral dependency optimization for model-inversion attack defense in classifiers"]},{"direction_index":6,"label":"D6","name":"Modality-tailored privacy mechanisms","description":"Protect privacy in ML systems where the data modality, structure, or domain semantics shape the mechanism, including language/text, graphs, biomedical data, perceptual signals, biometrics, and other domain-specific non-FL settings.","inclusion":"Include work whose primary objective is privacy for LLMs or text, graph data or GNNs, medical/clinical data, visual/audio/biometric signals, perceptual obfuscation, de-identification, domain-specific DP mechanisms, privacy-preserving explanations, or content-aware privacy transformations.","exclusion":"If the mechanism is a general DP optimizer or estimator without modality-specific design, assign to D2; if the work is primarily FL deployment across clients, assign to D1; if the primary contribution is an attack taxonomy independent of modality, assign to D5.","nearest_boundary":"Private learning algorithms and theory","representative_topics":["differentially private sparse compression of fine-tuned language models","privacy-preserving human action recognition via camera lens-based video degradation","training data extraction and privacy leakage in text classification models"]},{"direction_index":7,"label":"D7","name":"Privacy, fairness, consent, and utility","description":"Understand and optimize interactions among privacy mechanisms, group fairness, consent, trustworthiness, and utility across ML systems.","inclusion":"Include work whose primary objective is analyzing or mitigating disparate impact of privacy mechanisms, jointly private and fair training, privacy-fairness-utility bounds, consent-aware learning, or relationships between fairness constraints and privacy leakage.","exclusion":"If fairness is absent and the objective is only DP utility optimization, assign to D2; if the work mainly attacks sensitive attributes rather than optimizing fairness/privacy trade-offs, assign to D5; if biomedical or other domain context dominates without fairness as the primary objective, assign to D6.","nearest_boundary":"Private learning algorithms and theory","representative_topics":["accuracy-fairness tradeoffs in private learning under long-tailed data distributions","reducing disparate impact in differentially private training via gradient misalignment correction","privacy-preserving trustworthy machine learning balancing fairness, robustness, and explainability"]}],"split":"dev"}
{"episode_id":"F061_T2024-01","field_index":61,"field":"knowledge distillation","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"General transfer principles and signals","description":"Research whose primary objective is to define, analyze, select, or align the knowledge signal transferred from teacher to student in broadly applicable distillation settings, including objective theory, logits, temperature behavior, intermediate representations, relations, affinities, gradients, adaptive weighting, multi-teacher aggregation, and online collaborative mechanisms.","inclusion":"Include work when the central contribution is a general-purpose distillation objective, explanation, optimization mechanism, representation or relation-alignment signal, or source-combination strategy that is not primarily driven by a deployment constraint, data-access constraint, reliability target, or domain-specific structure.","exclusion":"If the main objective is making a small or mismatched student deployable under resource constraints, assign to D1; if the main objective is missing or compact data, assign to D2; if the main objective is robustness, calibration, or fairness, assign to D4.","nearest_boundary":"Capacity- and resource-constrained compression","representative_topics":["theoretical and empirical analysis of self-distillation via loss landscape geometry and generalization","information-theoretic analysis of knowledge points in distillation for classification","revising knowledge distillation loss via cross-entropy decomposition and soft target learning"]},{"direction_index":1,"label":"D1","name":"Capacity- and resource-constrained compression","description":"Research whose primary objective is to make distillation work for substantially smaller, faster, cheaper, or architecturally mismatched students through capacity-gap bridging, student-oriented teachers, architecture search, hardware-aware orchestration, quantization, edge deployment, cost allocation, or large-scale distillation systems.","inclusion":"Include work when the dominant problem is teacher-student capacity or architecture mismatch, student architecture selection, deployment cost, training/inference efficiency, budget-aware compact-model training, or systems support for large-scale distillation.","exclusion":"If the main novelty is a general feature/relation/objective rather than the capacity or resource constraint, assign to D0; if efficiency is specific to generative sampling, assign to D7; if efficiency is inseparable from a language, graph, or dense-prediction task program, assign to the relevant task-adaptation direction.","nearest_boundary":"General transfer principles and signals","representative_topics":["dynamic prior knowledge distillation for large teacher-to-small student capacity gaps","architecture-aware knowledge distillation via bayesian optimization for student model search","sparse teacher pruning via prediction uncertainty enlargement for knowledge distillation"]},{"direction_index":2,"label":"D2","name":"Data-constrained and compact-data distillation","description":"Research whose primary objective is to reduce dependence on original full training data, either by distilling without source data, exploiting unlabeled or weakly labeled data through pseudo-target refinement, or synthesizing/selecting compact datasets and coresets that replace large training sets.","inclusion":"Include work when lack of original data, limited labeled data, pseudo-label quality, substitute-query generation, synthetic sample generation, unlabeled-data exploitation, or compact dataset construction is the defining research problem.","exclusion":"If the supervision comes mainly from the model itself without a data-access or pseudo-label objective, assign to D3; if a generator or diffusion model is itself the object being compressed, assign to D7; if data are ordinary full training data and the novelty is the transfer signal, assign to D0.","nearest_boundary":"Self- and teacher-free distillation","representative_topics":["dynamic data-free knowledge distillation with adaptive easy-to-hard pseudo-label learning","adversarial data-free knowledge distillation with momentum generator for distribution shift","data-free knowledge distillation for regression via adversarial synthetic data generation"]},{"direction_index":3,"label":"D3","name":"Self- and teacher-free distillation","description":"Research whose primary objective is to improve models using their own predictions, earlier or deeper states, peer branches, augmented views, dropout variants, mutual online signals, or teacher-free internal supervision rather than a separate stronger fixed teacher.","inclusion":"Include work when no external pretrained teacher is central and the distillation signal is produced by the same model, a self-ensemble, online peers, shallow/deep exits, augmentation-induced views, or teacher-free mutual learning.","exclusion":"If the central challenge is pseudo-label selection from unlabeled data or missing source data, assign to D2; if multiple distinct teachers or ensembles are coordinated as external sources, assign to D0.","nearest_boundary":"Data-constrained and compact-data distillation","representative_topics":["teacher-free online knowledge distillation via mutual contrastive learning across layers","efficient self-distillation with zipf-distribution label smoothing for fine-tuning","parameter-free self-knowledge distillation via dropout-based posterior regularization"]},{"direction_index":4,"label":"D4","name":"Reliable and equitable distillation","description":"Research whose primary objective is to ensure distilled students are robust, calibrated, fair, reliable, or distribution-shift resistant, including adversarial robustness, OOD generalization, label-noise handling, worst-class performance, confidence gating, non-IID correction, spurious-correlation reduction, and subgroup reliability.","inclusion":"Include work when the explicit target of the distillation method or evaluation is robustness, calibration, fairness, worst-case behavior, long-tail reliability, domain-shift resistance, or trustworthy behavior beyond average predictive accuracy.","exclusion":"If robustness or calibration is only used to explain a general objective, assign to D0; if distribution shift arises mainly because original data are unavailable for data-free querying, assign to D2; if the primary issue is capacity mismatch rather than reliability, assign to D1.","nearest_boundary":"General transfer principles and signals","representative_topics":["shape-aware inductive bias distillation for generalization and adversarial robustness","robust knowledge distillation for worst-class performance under long-tailed distributions","generalizable neural vehicle routing via adaptive multi-distribution knowledge distillation"]},{"direction_index":5,"label":"D5","name":"Language, speech, and reasoning distillation","description":"Research whose primary objective is to distill linguistic, multilingual, speech/audio, sequence-generation, LLM, instruction-following, rationale, commonsense, chain-of-thought, or explicit symbolic knowledge capabilities into smaller or more efficient sequence and communication models.","inclusion":"Include work when the defining objective is compressing or transferring capabilities of language, multilingual, translation, ASR, speech, audio, LLM, or reasoning-oriented models, including cross-lingual retention, multilingual transfer, speech-model compression, rationale distillation, and explicit-knowledge transfer.","exclusion":"If the core obstacle is cross-modal transfer outside language/speech capability preservation, assign to D6; if the main goal is generic architecture or resource search without a sequence-model capability target, assign to D1; if the teacher/student is primarily a generative image model, assign to D7.","nearest_boundary":"Structured task and cross-domain adaptation","representative_topics":["retrieval-augmented query intent understanding via multi-stage llm distillation","knowledge distillation architectures combining multi-teacher training and lexical knowledge extraction for efficient language models","diverse small language model distillation from large-scale diverse instruction tuning"]},{"direction_index":6,"label":"D6","name":"Structured task and cross-domain adaptation","description":"Research whose primary objective is to adapt distillation to task structures or source-target domains whose geometry, outputs, modalities, or application constraints dominate the transfer problem, including dense visual prediction, graph-structured learning, cross-modal or privileged-source transfer, and domain-specialized non-language tasks.","inclusion":"Include work when the main challenge is transferring spatial, boundary, geometric, graph-structural, modality-mismatched, privileged-source, cross-task, or domain-constrained knowledge rather than proposing a domain-agnostic distillation mechanism.","exclusion":"If the task structure is ordinary language, multilingual, speech, or reasoning capability transfer, assign to D5; if the task is generative model sampling or diffusion/GAN compression, assign to D7; if the same alignment method is presented as generic across domains, assign to D0.","nearest_boundary":"Language, speech, and reasoning distillation","representative_topics":["point-to-voxel knowledge distillation for efficient lidar semantic segmentation with structural affinity","normalized feature distillation for semantic segmentation without manual knowledge design","heterogeneous knowledge distillation for object detection via assistant detection heads"]},{"direction_index":7,"label":"D7","name":"Generative-model distillation","description":"Research whose primary objective is to distill, compress, accelerate, or transfer generative models, especially diffusion, denoising, score-based, GAN, StyleGAN, or implicit generator teachers, into faster, fewer-step, smaller, or otherwise more efficient generative students.","inclusion":"Include work when the teacher or student being distilled is itself a generative model and the main goal is generative sampling efficiency, score or denoising transfer, diffusion step reduction, distribution matching, GAN compression, or universal transfer from pretrained generative teachers.","exclusion":"If a generator is used only to synthesize inputs for a non-generative student under missing-data constraints, assign to D2; if diffusion or denoising is only an auxiliary feature-processing mechanism for ordinary discriminative transfer, assign to D0 or D6 according to the target task.","nearest_boundary":"Data-constrained and compact-data distillation","representative_topics":["stylegan distillation with latent-direction loss for semantic consistency across teacher-student models","information-theoretic gan compression via variational energy-based mutual information distillation","survey of knowledge distillation methods for diffusion models into neural vector fields and implicit generators"]}],"split":"dev"}
{"episode_id":"F061_T2024-07","field_index":61,"field":"knowledge distillation","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"General transfer principles and signals","description":"Research whose primary objective is to define, analyze, select, or align the knowledge signal transferred from teacher to student in broadly applicable distillation settings, including objective theory, logits, temperature behavior, intermediate representations, relations, affinities, gradients, adaptive weighting, multi-teacher aggregation, and online collaborative mechanisms.","inclusion":"Include work when the central contribution is a general-purpose distillation objective, explanation, optimization mechanism, representation or relation-alignment signal, or source-combination strategy that is not primarily driven by a deployment constraint, data-access constraint, reliability target, or domain-specific structure.","exclusion":"If the main objective is making a small or mismatched student deployable under resource constraints, assign to D1; if the main objective is missing or compact data, assign to D2; if the main objective is robustness, calibration, or fairness, assign to D4.","nearest_boundary":"Capacity- and resource-constrained compression","representative_topics":["theoretical and empirical analysis of self-distillation via loss landscape geometry and generalization","information-theoretic analysis of knowledge points in distillation for classification","revising knowledge distillation loss via cross-entropy decomposition and soft target learning"]},{"direction_index":1,"label":"D1","name":"Capacity- and resource-constrained compression","description":"Research whose primary objective is to make distillation work for substantially smaller, faster, cheaper, or architecturally mismatched students through capacity-gap bridging, student-oriented teachers, architecture search, hardware-aware orchestration, quantization, edge deployment, cost allocation, or large-scale distillation systems.","inclusion":"Include work when the dominant problem is teacher-student capacity or architecture mismatch, student architecture selection, deployment cost, training/inference efficiency, budget-aware compact-model training, or systems support for large-scale distillation.","exclusion":"If the main novelty is a general feature/relation/objective rather than the capacity or resource constraint, assign to D0; if efficiency is specific to generative sampling, assign to D7; if efficiency is inseparable from a language, graph, or dense-prediction task program, assign to the relevant task-adaptation direction.","nearest_boundary":"General transfer principles and signals","representative_topics":["dynamic prior knowledge distillation for large teacher-to-small student capacity gaps","architecture-aware knowledge distillation via bayesian optimization for student model search","sparse teacher pruning via prediction uncertainty enlargement for knowledge distillation"]},{"direction_index":2,"label":"D2","name":"Data-constrained and compact-data distillation","description":"Research whose primary objective is to reduce dependence on original full training data, either by distilling without source data, exploiting unlabeled or weakly labeled data through pseudo-target refinement, or synthesizing/selecting compact datasets and coresets that replace large training sets.","inclusion":"Include work when lack of original data, limited labeled data, pseudo-label quality, substitute-query generation, synthetic sample generation, unlabeled-data exploitation, or compact dataset construction is the defining research problem.","exclusion":"If the supervision comes mainly from the model itself without a data-access or pseudo-label objective, assign to D3; if a generator or diffusion model is itself the object being compressed, assign to D7; if data are ordinary full training data and the novelty is the transfer signal, assign to D0.","nearest_boundary":"Self- and teacher-free distillation","representative_topics":["dynamic data-free knowledge distillation with adaptive easy-to-hard pseudo-label learning","adversarial data-free knowledge distillation with momentum generator for distribution shift","data-free knowledge distillation for regression via adversarial synthetic data generation"]},{"direction_index":3,"label":"D3","name":"Self- and teacher-free distillation","description":"Research whose primary objective is to improve models using their own predictions, earlier or deeper states, peer branches, augmented views, dropout variants, mutual online signals, or teacher-free internal supervision rather than a separate stronger fixed teacher.","inclusion":"Include work when no external pretrained teacher is central and the distillation signal is produced by the same model, a self-ensemble, online peers, shallow/deep exits, augmentation-induced views, or teacher-free mutual learning.","exclusion":"If the central challenge is pseudo-label selection from unlabeled data or missing source data, assign to D2; if multiple distinct teachers or ensembles are coordinated as external sources, assign to D0.","nearest_boundary":"Data-constrained and compact-data distillation","representative_topics":["teacher-free online knowledge distillation via mutual contrastive learning across layers","efficient self-distillation with zipf-distribution label smoothing for fine-tuning","parameter-free self-knowledge distillation via dropout-based posterior regularization"]},{"direction_index":4,"label":"D4","name":"Reliable and equitable distillation","description":"Research whose primary objective is to ensure distilled students are robust, calibrated, fair, reliable, or distribution-shift resistant, including adversarial robustness, OOD generalization, label-noise handling, worst-class performance, confidence gating, non-IID correction, spurious-correlation reduction, and subgroup reliability.","inclusion":"Include work when the explicit target of the distillation method or evaluation is robustness, calibration, fairness, worst-case behavior, long-tail reliability, domain-shift resistance, or trustworthy behavior beyond average predictive accuracy.","exclusion":"If robustness or calibration is only used to explain a general objective, assign to D0; if distribution shift arises mainly because original data are unavailable for data-free querying, assign to D2; if the primary issue is capacity mismatch rather than reliability, assign to D1.","nearest_boundary":"General transfer principles and signals","representative_topics":["shape-aware inductive bias distillation for generalization and adversarial robustness","robust knowledge distillation for worst-class performance under long-tailed distributions","generalizable neural vehicle routing via adaptive multi-distribution knowledge distillation"]},{"direction_index":5,"label":"D5","name":"Language, speech, and reasoning distillation","description":"Research whose primary objective is to distill linguistic, multilingual, speech/audio, sequence-generation, LLM, instruction-following, rationale, commonsense, chain-of-thought, or explicit symbolic knowledge capabilities into smaller or more efficient sequence and communication models.","inclusion":"Include work when the defining objective is compressing or transferring capabilities of language, multilingual, translation, ASR, speech, audio, LLM, or reasoning-oriented models, including cross-lingual retention, multilingual transfer, speech-model compression, rationale distillation, and explicit-knowledge transfer.","exclusion":"If the core obstacle is cross-modal transfer outside language/speech capability preservation, assign to D6; if the main goal is generic architecture or resource search without a sequence-model capability target, assign to D1; if the teacher/student is primarily a generative image model, assign to D7.","nearest_boundary":"Structured task and cross-domain adaptation","representative_topics":["retrieval-augmented query intent understanding via multi-stage llm distillation","knowledge distillation architectures combining multi-teacher training and lexical knowledge extraction for efficient language models","diverse small language model distillation from large-scale diverse instruction tuning"]},{"direction_index":6,"label":"D6","name":"Structured task and cross-domain adaptation","description":"Research whose primary objective is to adapt distillation to task structures or source-target domains whose geometry, outputs, modalities, or application constraints dominate the transfer problem, including dense visual prediction, graph-structured learning, cross-modal or privileged-source transfer, and domain-specialized non-language tasks.","inclusion":"Include work when the main challenge is transferring spatial, boundary, geometric, graph-structural, modality-mismatched, privileged-source, cross-task, or domain-constrained knowledge rather than proposing a domain-agnostic distillation mechanism.","exclusion":"If the task structure is ordinary language, multilingual, speech, or reasoning capability transfer, assign to D5; if the task is generative model sampling or diffusion/GAN compression, assign to D7; if the same alignment method is presented as generic across domains, assign to D0.","nearest_boundary":"Language, speech, and reasoning distillation","representative_topics":["point-to-voxel knowledge distillation for efficient lidar semantic segmentation with structural affinity","normalized feature distillation for semantic segmentation without manual knowledge design","heterogeneous knowledge distillation for object detection via assistant detection heads"]},{"direction_index":7,"label":"D7","name":"Generative-model distillation","description":"Research whose primary objective is to distill, compress, accelerate, or transfer generative models, especially diffusion, denoising, score-based, GAN, StyleGAN, or implicit generator teachers, into faster, fewer-step, smaller, or otherwise more efficient generative students.","inclusion":"Include work when the teacher or student being distilled is itself a generative model and the main goal is generative sampling efficiency, score or denoising transfer, diffusion step reduction, distribution matching, GAN compression, or universal transfer from pretrained generative teachers.","exclusion":"If a generator is used only to synthesize inputs for a non-generative student under missing-data constraints, assign to D2; if diffusion or denoising is only an auxiliary feature-processing mechanism for ordinary discriminative transfer, assign to D0 or D6 according to the target task.","nearest_boundary":"Data-constrained and compact-data distillation","representative_topics":["stylegan distillation with latent-direction loss for semantic consistency across teacher-student models","information-theoretic gan compression via variational energy-based mutual information distillation","survey of knowledge distillation methods for diffusion models into neural vector fields and implicit generators"]}],"split":"dev"}
{"episode_id":"F061_T2025-01","field_index":61,"field":"knowledge distillation","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"General transfer principles and signals","description":"Research whose primary objective is to define, analyze, select, or align the knowledge signal transferred from teacher to student in broadly applicable distillation settings, including objective theory, logits, temperature behavior, intermediate representations, relations, affinities, gradients, adaptive weighting, multi-teacher aggregation, and online collaborative mechanisms.","inclusion":"Include work when the central contribution is a general-purpose distillation objective, explanation, optimization mechanism, representation or relation-alignment signal, or source-combination strategy that is not primarily driven by a deployment constraint, data-access constraint, reliability target, or domain-specific structure.","exclusion":"If the main objective is making a small or mismatched student deployable under resource constraints, assign to D1; if the main objective is missing or compact data, assign to D2; if the main objective is robustness, calibration, or fairness, assign to D4.","nearest_boundary":"Capacity- and resource-constrained compression","representative_topics":["theoretical and empirical analysis of self-distillation via loss landscape geometry and generalization","information-theoretic analysis of knowledge points in distillation for classification","revising knowledge distillation loss via cross-entropy decomposition and soft target learning"]},{"direction_index":1,"label":"D1","name":"Capacity- and resource-constrained compression","description":"Research whose primary objective is to make distillation work for substantially smaller, faster, cheaper, or architecturally mismatched students through capacity-gap bridging, student-oriented teachers, architecture search, hardware-aware orchestration, quantization, edge deployment, cost allocation, or large-scale distillation systems.","inclusion":"Include work when the dominant problem is teacher-student capacity or architecture mismatch, student architecture selection, deployment cost, training/inference efficiency, budget-aware compact-model training, or systems support for large-scale distillation.","exclusion":"If the main novelty is a general feature/relation/objective rather than the capacity or resource constraint, assign to D0; if efficiency is specific to generative sampling, assign to D7; if efficiency is inseparable from a language, graph, or dense-prediction task program, assign to the relevant task-adaptation direction.","nearest_boundary":"General transfer principles and signals","representative_topics":["dynamic prior knowledge distillation for large teacher-to-small student capacity gaps","architecture-aware knowledge distillation via bayesian optimization for student model search","sparse teacher pruning via prediction uncertainty enlargement for knowledge distillation"]},{"direction_index":2,"label":"D2","name":"Data-constrained and compact-data distillation","description":"Research whose primary objective is to reduce dependence on original full training data, either by distilling without source data, exploiting unlabeled or weakly labeled data through pseudo-target refinement, or synthesizing/selecting compact datasets and coresets that replace large training sets.","inclusion":"Include work when lack of original data, limited labeled data, pseudo-label quality, substitute-query generation, synthetic sample generation, unlabeled-data exploitation, or compact dataset construction is the defining research problem.","exclusion":"If the supervision comes mainly from the model itself without a data-access or pseudo-label objective, assign to D3; if a generator or diffusion model is itself the object being compressed, assign to D7; if data are ordinary full training data and the novelty is the transfer signal, assign to D0.","nearest_boundary":"Self- and teacher-free distillation","representative_topics":["dynamic data-free knowledge distillation with adaptive easy-to-hard pseudo-label learning","adversarial data-free knowledge distillation with momentum generator for distribution shift","data-free knowledge distillation for regression via adversarial synthetic data generation"]},{"direction_index":3,"label":"D3","name":"Self- and teacher-free distillation","description":"Research whose primary objective is to improve models using their own predictions, earlier or deeper states, peer branches, augmented views, dropout variants, mutual online signals, or teacher-free internal supervision rather than a separate stronger fixed teacher.","inclusion":"Include work when no external pretrained teacher is central and the distillation signal is produced by the same model, a self-ensemble, online peers, shallow/deep exits, augmentation-induced views, or teacher-free mutual learning.","exclusion":"If the central challenge is pseudo-label selection from unlabeled data or missing source data, assign to D2; if multiple distinct teachers or ensembles are coordinated as external sources, assign to D0.","nearest_boundary":"Data-constrained and compact-data distillation","representative_topics":["teacher-free online knowledge distillation via mutual contrastive learning across layers","efficient self-distillation with zipf-distribution label smoothing for fine-tuning","parameter-free self-knowledge distillation via dropout-based posterior regularization"]},{"direction_index":4,"label":"D4","name":"Reliable and equitable distillation","description":"Research whose primary objective is to ensure distilled students are robust, calibrated, fair, reliable, or distribution-shift resistant, including adversarial robustness, OOD generalization, label-noise handling, worst-class performance, confidence gating, non-IID correction, spurious-correlation reduction, and subgroup reliability.","inclusion":"Include work when the explicit target of the distillation method or evaluation is robustness, calibration, fairness, worst-case behavior, long-tail reliability, domain-shift resistance, or trustworthy behavior beyond average predictive accuracy.","exclusion":"If robustness or calibration is only used to explain a general objective, assign to D0; if distribution shift arises mainly because original data are unavailable for data-free querying, assign to D2; if the primary issue is capacity mismatch rather than reliability, assign to D1.","nearest_boundary":"General transfer principles and signals","representative_topics":["shape-aware inductive bias distillation for generalization and adversarial robustness","robust knowledge distillation for worst-class performance under long-tailed distributions","generalizable neural vehicle routing via adaptive multi-distribution knowledge distillation"]},{"direction_index":5,"label":"D5","name":"Language, speech, and reasoning distillation","description":"Research whose primary objective is to distill linguistic, multilingual, speech/audio, sequence-generation, LLM, instruction-following, rationale, commonsense, chain-of-thought, or explicit symbolic knowledge capabilities into smaller or more efficient sequence and communication models.","inclusion":"Include work when the defining objective is compressing or transferring capabilities of language, multilingual, translation, ASR, speech, audio, LLM, or reasoning-oriented models, including cross-lingual retention, multilingual transfer, speech-model compression, rationale distillation, and explicit-knowledge transfer.","exclusion":"If the core obstacle is cross-modal transfer outside language/speech capability preservation, assign to D6; if the main goal is generic architecture or resource search without a sequence-model capability target, assign to D1; if the teacher/student is primarily a generative image model, assign to D7.","nearest_boundary":"Structured task and cross-domain adaptation","representative_topics":["retrieval-augmented query intent understanding via multi-stage llm distillation","knowledge distillation architectures combining multi-teacher training and lexical knowledge extraction for efficient language models","diverse small language model distillation from large-scale diverse instruction tuning"]},{"direction_index":6,"label":"D6","name":"Structured task and cross-domain adaptation","description":"Research whose primary objective is to adapt distillation to task structures or source-target domains whose geometry, outputs, modalities, or application constraints dominate the transfer problem, including dense visual prediction, graph-structured learning, cross-modal or privileged-source transfer, and domain-specialized non-language tasks.","inclusion":"Include work when the main challenge is transferring spatial, boundary, geometric, graph-structural, modality-mismatched, privileged-source, cross-task, or domain-constrained knowledge rather than proposing a domain-agnostic distillation mechanism.","exclusion":"If the task structure is ordinary language, multilingual, speech, or reasoning capability transfer, assign to D5; if the task is generative model sampling or diffusion/GAN compression, assign to D7; if the same alignment method is presented as generic across domains, assign to D0.","nearest_boundary":"Language, speech, and reasoning distillation","representative_topics":["point-to-voxel knowledge distillation for efficient lidar semantic segmentation with structural affinity","normalized feature distillation for semantic segmentation without manual knowledge design","heterogeneous knowledge distillation for object detection via assistant detection heads"]},{"direction_index":7,"label":"D7","name":"Generative-model distillation","description":"Research whose primary objective is to distill, compress, accelerate, or transfer generative models, especially diffusion, denoising, score-based, GAN, StyleGAN, or implicit generator teachers, into faster, fewer-step, smaller, or otherwise more efficient generative students.","inclusion":"Include work when the teacher or student being distilled is itself a generative model and the main goal is generative sampling efficiency, score or denoising transfer, diffusion step reduction, distribution matching, GAN compression, or universal transfer from pretrained generative teachers.","exclusion":"If a generator is used only to synthesize inputs for a non-generative student under missing-data constraints, assign to D2; if diffusion or denoising is only an auxiliary feature-processing mechanism for ordinary discriminative transfer, assign to D0 or D6 according to the target task.","nearest_boundary":"Data-constrained and compact-data distillation","representative_topics":["stylegan distillation with latent-direction loss for semantic consistency across teacher-student models","information-theoretic gan compression via variational energy-based mutual information distillation","survey of knowledge distillation methods for diffusion models into neural vector fields and implicit generators"]}],"split":"dev"}
{"episode_id":"F061_T2025-07","field_index":61,"field":"knowledge distillation","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"General transfer principles and signals","description":"Research whose primary objective is to define, analyze, select, or align the knowledge signal transferred from teacher to student in broadly applicable distillation settings, including objective theory, logits, temperature behavior, intermediate representations, relations, affinities, gradients, adaptive weighting, multi-teacher aggregation, and online collaborative mechanisms.","inclusion":"Include work when the central contribution is a general-purpose distillation objective, explanation, optimization mechanism, representation or relation-alignment signal, or source-combination strategy that is not primarily driven by a deployment constraint, data-access constraint, reliability target, or domain-specific structure.","exclusion":"If the main objective is making a small or mismatched student deployable under resource constraints, assign to D1; if the main objective is missing or compact data, assign to D2; if the main objective is robustness, calibration, or fairness, assign to D4.","nearest_boundary":"Capacity- and resource-constrained compression","representative_topics":["theoretical and empirical analysis of self-distillation via loss landscape geometry and generalization","information-theoretic analysis of knowledge points in distillation for classification","revising knowledge distillation loss via cross-entropy decomposition and soft target learning"]},{"direction_index":1,"label":"D1","name":"Capacity- and resource-constrained compression","description":"Research whose primary objective is to make distillation work for substantially smaller, faster, cheaper, or architecturally mismatched students through capacity-gap bridging, student-oriented teachers, architecture search, hardware-aware orchestration, quantization, edge deployment, cost allocation, or large-scale distillation systems.","inclusion":"Include work when the dominant problem is teacher-student capacity or architecture mismatch, student architecture selection, deployment cost, training/inference efficiency, budget-aware compact-model training, or systems support for large-scale distillation.","exclusion":"If the main novelty is a general feature/relation/objective rather than the capacity or resource constraint, assign to D0; if efficiency is specific to generative sampling, assign to D7; if efficiency is inseparable from a language, graph, or dense-prediction task program, assign to the relevant task-adaptation direction.","nearest_boundary":"General transfer principles and signals","representative_topics":["dynamic prior knowledge distillation for large teacher-to-small student capacity gaps","architecture-aware knowledge distillation via bayesian optimization for student model search","sparse teacher pruning via prediction uncertainty enlargement for knowledge distillation"]},{"direction_index":2,"label":"D2","name":"Data-constrained and compact-data distillation","description":"Research whose primary objective is to reduce dependence on original full training data, either by distilling without source data, exploiting unlabeled or weakly labeled data through pseudo-target refinement, or synthesizing/selecting compact datasets and coresets that replace large training sets.","inclusion":"Include work when lack of original data, limited labeled data, pseudo-label quality, substitute-query generation, synthetic sample generation, unlabeled-data exploitation, or compact dataset construction is the defining research problem.","exclusion":"If the supervision comes mainly from the model itself without a data-access or pseudo-label objective, assign to D3; if a generator or diffusion model is itself the object being compressed, assign to D7; if data are ordinary full training data and the novelty is the transfer signal, assign to D0.","nearest_boundary":"Self- and teacher-free distillation","representative_topics":["dynamic data-free knowledge distillation with adaptive easy-to-hard pseudo-label learning","adversarial data-free knowledge distillation with momentum generator for distribution shift","data-free knowledge distillation for regression via adversarial synthetic data generation"]},{"direction_index":3,"label":"D3","name":"Self- and teacher-free distillation","description":"Research whose primary objective is to improve models using their own predictions, earlier or deeper states, peer branches, augmented views, dropout variants, mutual online signals, or teacher-free internal supervision rather than a separate stronger fixed teacher.","inclusion":"Include work when no external pretrained teacher is central and the distillation signal is produced by the same model, a self-ensemble, online peers, shallow/deep exits, augmentation-induced views, or teacher-free mutual learning.","exclusion":"If the central challenge is pseudo-label selection from unlabeled data or missing source data, assign to D2; if multiple distinct teachers or ensembles are coordinated as external sources, assign to D0.","nearest_boundary":"Data-constrained and compact-data distillation","representative_topics":["teacher-free online knowledge distillation via mutual contrastive learning across layers","efficient self-distillation with zipf-distribution label smoothing for fine-tuning","parameter-free self-knowledge distillation via dropout-based posterior regularization"]},{"direction_index":4,"label":"D4","name":"Reliable and equitable distillation","description":"Research whose primary objective is to ensure distilled students are robust, calibrated, fair, reliable, or distribution-shift resistant, including adversarial robustness, OOD generalization, label-noise handling, worst-class performance, confidence gating, non-IID correction, spurious-correlation reduction, and subgroup reliability.","inclusion":"Include work when the explicit target of the distillation method or evaluation is robustness, calibration, fairness, worst-case behavior, long-tail reliability, domain-shift resistance, or trustworthy behavior beyond average predictive accuracy.","exclusion":"If robustness or calibration is only used to explain a general objective, assign to D0; if distribution shift arises mainly because original data are unavailable for data-free querying, assign to D2; if the primary issue is capacity mismatch rather than reliability, assign to D1.","nearest_boundary":"General transfer principles and signals","representative_topics":["shape-aware inductive bias distillation for generalization and adversarial robustness","robust knowledge distillation for worst-class performance under long-tailed distributions","generalizable neural vehicle routing via adaptive multi-distribution knowledge distillation"]},{"direction_index":5,"label":"D5","name":"Language, speech, and reasoning distillation","description":"Research whose primary objective is to distill linguistic, multilingual, speech/audio, sequence-generation, LLM, instruction-following, rationale, commonsense, chain-of-thought, or explicit symbolic knowledge capabilities into smaller or more efficient sequence and communication models.","inclusion":"Include work when the defining objective is compressing or transferring capabilities of language, multilingual, translation, ASR, speech, audio, LLM, or reasoning-oriented models, including cross-lingual retention, multilingual transfer, speech-model compression, rationale distillation, and explicit-knowledge transfer.","exclusion":"If the core obstacle is cross-modal transfer outside language/speech capability preservation, assign to D6; if the main goal is generic architecture or resource search without a sequence-model capability target, assign to D1; if the teacher/student is primarily a generative image model, assign to D7.","nearest_boundary":"Structured task and cross-domain adaptation","representative_topics":["retrieval-augmented query intent understanding via multi-stage llm distillation","knowledge distillation architectures combining multi-teacher training and lexical knowledge extraction for efficient language models","diverse small language model distillation from large-scale diverse instruction tuning"]},{"direction_index":6,"label":"D6","name":"Structured task and cross-domain adaptation","description":"Research whose primary objective is to adapt distillation to task structures or source-target domains whose geometry, outputs, modalities, or application constraints dominate the transfer problem, including dense visual prediction, graph-structured learning, cross-modal or privileged-source transfer, and domain-specialized non-language tasks.","inclusion":"Include work when the main challenge is transferring spatial, boundary, geometric, graph-structural, modality-mismatched, privileged-source, cross-task, or domain-constrained knowledge rather than proposing a domain-agnostic distillation mechanism.","exclusion":"If the task structure is ordinary language, multilingual, speech, or reasoning capability transfer, assign to D5; if the task is generative model sampling or diffusion/GAN compression, assign to D7; if the same alignment method is presented as generic across domains, assign to D0.","nearest_boundary":"Language, speech, and reasoning distillation","representative_topics":["point-to-voxel knowledge distillation for efficient lidar semantic segmentation with structural affinity","normalized feature distillation for semantic segmentation without manual knowledge design","heterogeneous knowledge distillation for object detection via assistant detection heads"]},{"direction_index":7,"label":"D7","name":"Generative-model distillation","description":"Research whose primary objective is to distill, compress, accelerate, or transfer generative models, especially diffusion, denoising, score-based, GAN, StyleGAN, or implicit generator teachers, into faster, fewer-step, smaller, or otherwise more efficient generative students.","inclusion":"Include work when the teacher or student being distilled is itself a generative model and the main goal is generative sampling efficiency, score or denoising transfer, diffusion step reduction, distribution matching, GAN compression, or universal transfer from pretrained generative teachers.","exclusion":"If a generator is used only to synthesize inputs for a non-generative student under missing-data constraints, assign to D2; if diffusion or denoising is only an auxiliary feature-processing mechanism for ordinary discriminative transfer, assign to D0 or D6 according to the target task.","nearest_boundary":"Data-constrained and compact-data distillation","representative_topics":["stylegan distillation with latent-direction loss for semantic consistency across teacher-student models","information-theoretic gan compression via variational energy-based mutual information distillation","survey of knowledge distillation methods for diffusion models into neural vector fields and implicit generators"]}],"split":"dev"}
{"episode_id":"F061_T2026-01","field_index":61,"field":"knowledge distillation","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"General transfer principles and signals","description":"Research whose primary objective is to define, analyze, select, or align the knowledge signal transferred from teacher to student in broadly applicable distillation settings, including objective theory, logits, temperature behavior, intermediate representations, relations, affinities, gradients, adaptive weighting, multi-teacher aggregation, and online collaborative mechanisms.","inclusion":"Include work when the central contribution is a general-purpose distillation objective, explanation, optimization mechanism, representation or relation-alignment signal, or source-combination strategy that is not primarily driven by a deployment constraint, data-access constraint, reliability target, or domain-specific structure.","exclusion":"If the main objective is making a small or mismatched student deployable under resource constraints, assign to D1; if the main objective is missing or compact data, assign to D2; if the main objective is robustness, calibration, or fairness, assign to D4.","nearest_boundary":"Capacity- and resource-constrained compression","representative_topics":["theoretical and empirical analysis of self-distillation via loss landscape geometry and generalization","information-theoretic analysis of knowledge points in distillation for classification","revising knowledge distillation loss via cross-entropy decomposition and soft target learning"]},{"direction_index":1,"label":"D1","name":"Capacity- and resource-constrained compression","description":"Research whose primary objective is to make distillation work for substantially smaller, faster, cheaper, or architecturally mismatched students through capacity-gap bridging, student-oriented teachers, architecture search, hardware-aware orchestration, quantization, edge deployment, cost allocation, or large-scale distillation systems.","inclusion":"Include work when the dominant problem is teacher-student capacity or architecture mismatch, student architecture selection, deployment cost, training/inference efficiency, budget-aware compact-model training, or systems support for large-scale distillation.","exclusion":"If the main novelty is a general feature/relation/objective rather than the capacity or resource constraint, assign to D0; if efficiency is specific to generative sampling, assign to D7; if efficiency is inseparable from a language, graph, or dense-prediction task program, assign to the relevant task-adaptation direction.","nearest_boundary":"General transfer principles and signals","representative_topics":["dynamic prior knowledge distillation for large teacher-to-small student capacity gaps","architecture-aware knowledge distillation via bayesian optimization for student model search","sparse teacher pruning via prediction uncertainty enlargement for knowledge distillation"]},{"direction_index":2,"label":"D2","name":"Data-constrained and compact-data distillation","description":"Research whose primary objective is to reduce dependence on original full training data, either by distilling without source data, exploiting unlabeled or weakly labeled data through pseudo-target refinement, or synthesizing/selecting compact datasets and coresets that replace large training sets.","inclusion":"Include work when lack of original data, limited labeled data, pseudo-label quality, substitute-query generation, synthetic sample generation, unlabeled-data exploitation, or compact dataset construction is the defining research problem.","exclusion":"If the supervision comes mainly from the model itself without a data-access or pseudo-label objective, assign to D3; if a generator or diffusion model is itself the object being compressed, assign to D7; if data are ordinary full training data and the novelty is the transfer signal, assign to D0.","nearest_boundary":"Self- and teacher-free distillation","representative_topics":["dynamic data-free knowledge distillation with adaptive easy-to-hard pseudo-label learning","adversarial data-free knowledge distillation with momentum generator for distribution shift","data-free knowledge distillation for regression via adversarial synthetic data generation"]},{"direction_index":3,"label":"D3","name":"Self- and teacher-free distillation","description":"Research whose primary objective is to improve models using their own predictions, earlier or deeper states, peer branches, augmented views, dropout variants, mutual online signals, or teacher-free internal supervision rather than a separate stronger fixed teacher.","inclusion":"Include work when no external pretrained teacher is central and the distillation signal is produced by the same model, a self-ensemble, online peers, shallow/deep exits, augmentation-induced views, or teacher-free mutual learning.","exclusion":"If the central challenge is pseudo-label selection from unlabeled data or missing source data, assign to D2; if multiple distinct teachers or ensembles are coordinated as external sources, assign to D0.","nearest_boundary":"Data-constrained and compact-data distillation","representative_topics":["teacher-free online knowledge distillation via mutual contrastive learning across layers","efficient self-distillation with zipf-distribution label smoothing for fine-tuning","parameter-free self-knowledge distillation via dropout-based posterior regularization"]},{"direction_index":4,"label":"D4","name":"Reliable and equitable distillation","description":"Research whose primary objective is to ensure distilled students are robust, calibrated, fair, reliable, or distribution-shift resistant, including adversarial robustness, OOD generalization, label-noise handling, worst-class performance, confidence gating, non-IID correction, spurious-correlation reduction, and subgroup reliability.","inclusion":"Include work when the explicit target of the distillation method or evaluation is robustness, calibration, fairness, worst-case behavior, long-tail reliability, domain-shift resistance, or trustworthy behavior beyond average predictive accuracy.","exclusion":"If robustness or calibration is only used to explain a general objective, assign to D0; if distribution shift arises mainly because original data are unavailable for data-free querying, assign to D2; if the primary issue is capacity mismatch rather than reliability, assign to D1.","nearest_boundary":"General transfer principles and signals","representative_topics":["shape-aware inductive bias distillation for generalization and adversarial robustness","robust knowledge distillation for worst-class performance under long-tailed distributions","generalizable neural vehicle routing via adaptive multi-distribution knowledge distillation"]},{"direction_index":5,"label":"D5","name":"Language, speech, and reasoning distillation","description":"Research whose primary objective is to distill linguistic, multilingual, speech/audio, sequence-generation, LLM, instruction-following, rationale, commonsense, chain-of-thought, or explicit symbolic knowledge capabilities into smaller or more efficient sequence and communication models.","inclusion":"Include work when the defining objective is compressing or transferring capabilities of language, multilingual, translation, ASR, speech, audio, LLM, or reasoning-oriented models, including cross-lingual retention, multilingual transfer, speech-model compression, rationale distillation, and explicit-knowledge transfer.","exclusion":"If the core obstacle is cross-modal transfer outside language/speech capability preservation, assign to D6; if the main goal is generic architecture or resource search without a sequence-model capability target, assign to D1; if the teacher/student is primarily a generative image model, assign to D7.","nearest_boundary":"Structured task and cross-domain adaptation","representative_topics":["retrieval-augmented query intent understanding via multi-stage llm distillation","knowledge distillation architectures combining multi-teacher training and lexical knowledge extraction for efficient language models","diverse small language model distillation from large-scale diverse instruction tuning"]},{"direction_index":6,"label":"D6","name":"Structured task and cross-domain adaptation","description":"Research whose primary objective is to adapt distillation to task structures or source-target domains whose geometry, outputs, modalities, or application constraints dominate the transfer problem, including dense visual prediction, graph-structured learning, cross-modal or privileged-source transfer, and domain-specialized non-language tasks.","inclusion":"Include work when the main challenge is transferring spatial, boundary, geometric, graph-structural, modality-mismatched, privileged-source, cross-task, or domain-constrained knowledge rather than proposing a domain-agnostic distillation mechanism.","exclusion":"If the task structure is ordinary language, multilingual, speech, or reasoning capability transfer, assign to D5; if the task is generative model sampling or diffusion/GAN compression, assign to D7; if the same alignment method is presented as generic across domains, assign to D0.","nearest_boundary":"Language, speech, and reasoning distillation","representative_topics":["point-to-voxel knowledge distillation for efficient lidar semantic segmentation with structural affinity","normalized feature distillation for semantic segmentation without manual knowledge design","heterogeneous knowledge distillation for object detection via assistant detection heads"]},{"direction_index":7,"label":"D7","name":"Generative-model distillation","description":"Research whose primary objective is to distill, compress, accelerate, or transfer generative models, especially diffusion, denoising, score-based, GAN, StyleGAN, or implicit generator teachers, into faster, fewer-step, smaller, or otherwise more efficient generative students.","inclusion":"Include work when the teacher or student being distilled is itself a generative model and the main goal is generative sampling efficiency, score or denoising transfer, diffusion step reduction, distribution matching, GAN compression, or universal transfer from pretrained generative teachers.","exclusion":"If a generator is used only to synthesize inputs for a non-generative student under missing-data constraints, assign to D2; if diffusion or denoising is only an auxiliary feature-processing mechanism for ordinary discriminative transfer, assign to D0 or D6 according to the target task.","nearest_boundary":"Data-constrained and compact-data distillation","representative_topics":["stylegan distillation with latent-direction loss for semantic consistency across teacher-student models","information-theoretic gan compression via variational energy-based mutual information distillation","survey of knowledge distillation methods for diffusion models into neural vector fields and implicit generators"]}],"split":"dev"}
{"episode_id":"F063_T2024-01","field_index":63,"field":"graph representation learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Self-supervised pre-training and transfer","description":"Learning transferable graph representations with label-free or weakly supervised objectives, including contrastive learning, masked or generative pre-training, clustering-driven pretexts, prompt tuning, and language/text alignment when the objective is transfer across tasks.","inclusion":"Assign atoms whose primary objective is to design graph pre-training objectives, augmentations, positive/negative sampling, reconstruction or masking losses, clustering pseudo-labels, graph prompts, adapters, or language-graph alignment for downstream transfer.","exclusion":"If the dominant objective is temporal evolution, scalability, robustness, heterophily, or architectural expressivity rather than the pre-training or transfer program, assign to that corresponding direction.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["graph contrastive learning with interpolation augmentation and correlation reduction for semi-supervised node classification","rationale-aware graph contrastive learning via invariant instance-discrimination discovery","similarity-aware positive instance sampling for graph contrastive pre-training with hierarchical encoding"]},{"direction_index":1,"label":"D1","name":"Temporal and dynamic graph learning","description":"Learning representations from graphs whose edges, nodes, attributes, labels, or interaction events evolve over time, including snapshot, continuous-time, streaming, and event-based settings.","inclusion":"Assign atoms whose primary objective is modeling temporal dependency, structural evolution, event sequences, temporal memory, time-aware updates, or scalable representation maintenance for dynamic graphs.","exclusion":"Static self-supervised, heterogeneous, geometric, or scalable methods belong elsewhere unless time-varying graph evolution is the main research objective.","nearest_boundary":"Scalable graph learning systems","representative_topics":["incremental graph representation updates for scalable dynamic graph neural networks","temporal graph neural networks using depth-first search for path-oriented temporal aggregation","asynchronous dynamic graph embedding via time-aware transformer on temporal edge sequences"]},{"direction_index":2,"label":"D2","name":"Core GNN architecture and long-range representation","description":"Designing graph neural architectures and aggregation mechanisms to improve expressivity, graph/subgraph readout, long-range propagation, transformer-style attention, pooling, and algorithmic or structural inductive bias.","inclusion":"Assign atoms whose primary objective is architectural representation design: mitigating over-smoothing or over-squashing, improving WL/subgraph/path expressivity, designing graph transformers, structural or positional encodings, graph pooling, adaptive readout, hierarchical coarsening, or graph/subgraph aggregation.","exclusion":"If the main objective is heterophily, typed/higher-order graph structure, non-Euclidean geometry, system throughput, or self-supervised pre-training rather than architecture, assign to that more specific direction.","nearest_boundary":"Specialized graph structure and geometric learning","representative_topics":["scalable graph attention multi-layer perceptron for over-smoothing in sparse large-scale graphs","mitigating feature overcorrelation and oversmoothing in deep graph neural networks","theoretical analysis of subgraph gnn expressive power via symmetry group reduction and unified message passing"]},{"direction_index":3,"label":"D3","name":"Specialized graph structure and geometric learning","description":"Learning representations for graphs whose scientific challenge is non-standard relational structure, heterophily, higher-order or typed relations, multimodal node information, or non-Euclidean/geometric embedding spaces.","inclusion":"Assign atoms whose primary objective is handling heterophilic neighborhoods, heterogeneous or multiplex relation types, hyperedges, signed/directed/property graphs, multimodal feature fusion on graphs, hyperbolic/Riemannian/quasimetric/sheaf geometry, or curvature-aware representation spaces.","exclusion":"If relation or geometry is only incidental to a general architecture, temporal, SSL, or systems contribution, assign to that primary objective instead.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["decoupled self-supervised learning for node representation on non-homophilous graphs via latent semantic modeling","inductive graph embedding with joint node and edge property aggregation via biased neighbor sampling","random walk aggregation for graph neural networks on heterophily graphs"]},{"direction_index":4,"label":"D4","name":"Scalable graph learning systems","description":"Scaling graph representation learning through efficient algorithms, sampling, summarization, optimization, distributed execution, hardware-aware storage/loading, caching, memory reduction, and training or inference throughput improvements.","inclusion":"Assign atoms whose primary objective is computational scalability, systems design, mini-batch or sampling efficiency, graph summarization for embedding scalability, distributed or out-of-core GNN training, hardware utilization, or training-time prediction.","exclusion":"Temporal update efficiency belongs to temporal graph learning when graph evolution is central; transformer-specific attention sparsification belongs to core architecture when the architecture is the main contribution.","nearest_boundary":"Temporal and dynamic graph learning","representative_topics":["graph metric-based prediction of gnn training time for acceleration and design selection","theoretical analysis of graph summarization for scalable node embedding learning","scalable graph convolutional networks via sampling-free minibatch gradient methods"]},{"direction_index":5,"label":"D5","name":"Topology optimization and reliable representations","description":"Learning, refining, generating, or defending graph representations and graph structures when topology or data are noisy, missing, adversarial, uncertain, imbalanced, private, shifting, or sequentially changing.","inclusion":"Assign atoms whose primary objective is graph structure learning, adjacency refinement, latent topology inference, graph similarity structure inference, robust hypergraph or graph topology optimization, adversarial robustness, privacy, fairness, imbalance, uncertainty, causal robustness, continual learning, or graph generation/completion as a topology modeling objective.","exclusion":"Contrastive or SSL methods belong to self-supervised pre-training if robustness is not the primary objective; heterophily-specific topology changes belong to specialized graph structure learning.","nearest_boundary":"Self-supervised pre-training and transfer","representative_topics":["dummy node augmentation for expressive graph kernels and neural network structure learning","unsupervised graph structure refinement for robust graph neural network defense","transferable backdoor attacks on gnns via surrogate-guided perturbation-based dataset poisoning"]},{"direction_index":6,"label":"D6","name":"Spectral, walk, and structural embeddings","description":"Producing graph, node, or subgraph embeddings from non-message-passing structural objectives such as spectra, walks, transition matrices, diffusion signatures, matrix factorization, latent distances, compact graph codes, or structural descriptors.","inclusion":"Assign atoms whose primary objective is deriving embeddings or structural encodings from random walks, spectral decompositions, transition probabilities, diffusion, factorization, graph summarization, Prufer or other combinatorial encodings, or variational/autoencoding embeddings when treated as general unsupervised embedding rather than SSL transfer.","exclusion":"Spectral filters for heterophily, positional encodings for GNN expressivity, or reconstruction objectives for pre-training transfer belong to their respective primary programs.","nearest_boundary":"Self-supervised pre-training and transfer","representative_topics":["theoretical analysis of graph summarization for scalable node embedding learning","scalable diffusion-based structural node embeddings capturing directional information on directed graphs","graph node embedding via anonymous walk-based transition probability matrix for topology-aware classification"]},{"direction_index":7,"label":"D7","name":"Evaluation, theory, and domain utility","description":"Understanding, benchmarking, explaining, and validating graph representation learning, including theoretical guarantees, benchmark taxonomies, interpretability, graph-information utility, surveys, and domain-grounded evaluation where the main objective is assessing or justifying representation usefulness.","inclusion":"Assign atoms whose primary objective is generalization theory, benchmark design, evaluation methodology, interpretability, architectural taxonomies, measuring when graph structure helps, field surveys, or domain-specific evaluation/adaptation framed around representational utility rather than a new general mechanism.","exclusion":"Mechanism-specific theory belongs to that mechanism's direction, such as contrastive loss theory, oversmoothing analysis, heterophily filtering theory, or expressivity analysis tied directly to new architecture design.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["survey of attention-based graph neural network architectures for graph-structured data","theoretical generalization bounds for gnns interpolating bandlimited functions via sampling-theorem connections","taxonomy of graph benchmarks via perturbation-based sensitivity profiling for gnn evaluation"]}],"split":"dev"}
{"episode_id":"F063_T2024-07","field_index":63,"field":"graph representation learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Self-supervised pre-training and transfer","description":"Learning transferable graph representations with label-free or weakly supervised objectives, including contrastive learning, masked or generative pre-training, clustering-driven pretexts, prompt tuning, and language/text alignment when the objective is transfer across tasks.","inclusion":"Assign atoms whose primary objective is to design graph pre-training objectives, augmentations, positive/negative sampling, reconstruction or masking losses, clustering pseudo-labels, graph prompts, adapters, or language-graph alignment for downstream transfer.","exclusion":"If the dominant objective is temporal evolution, scalability, robustness, heterophily, or architectural expressivity rather than the pre-training or transfer program, assign to that corresponding direction.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["graph contrastive learning with interpolation augmentation and correlation reduction for semi-supervised node classification","rationale-aware graph contrastive learning via invariant instance-discrimination discovery","similarity-aware positive instance sampling for graph contrastive pre-training with hierarchical encoding"]},{"direction_index":1,"label":"D1","name":"Temporal and dynamic graph learning","description":"Learning representations from graphs whose edges, nodes, attributes, labels, or interaction events evolve over time, including snapshot, continuous-time, streaming, and event-based settings.","inclusion":"Assign atoms whose primary objective is modeling temporal dependency, structural evolution, event sequences, temporal memory, time-aware updates, or scalable representation maintenance for dynamic graphs.","exclusion":"Static self-supervised, heterogeneous, geometric, or scalable methods belong elsewhere unless time-varying graph evolution is the main research objective.","nearest_boundary":"Scalable graph learning systems","representative_topics":["incremental graph representation updates for scalable dynamic graph neural networks","temporal graph neural networks using depth-first search for path-oriented temporal aggregation","asynchronous dynamic graph embedding via time-aware transformer on temporal edge sequences"]},{"direction_index":2,"label":"D2","name":"Core GNN architecture and long-range representation","description":"Designing graph neural architectures and aggregation mechanisms to improve expressivity, graph/subgraph readout, long-range propagation, transformer-style attention, pooling, and algorithmic or structural inductive bias.","inclusion":"Assign atoms whose primary objective is architectural representation design: mitigating over-smoothing or over-squashing, improving WL/subgraph/path expressivity, designing graph transformers, structural or positional encodings, graph pooling, adaptive readout, hierarchical coarsening, or graph/subgraph aggregation.","exclusion":"If the main objective is heterophily, typed/higher-order graph structure, non-Euclidean geometry, system throughput, or self-supervised pre-training rather than architecture, assign to that more specific direction.","nearest_boundary":"Specialized graph structure and geometric learning","representative_topics":["scalable graph attention multi-layer perceptron for over-smoothing in sparse large-scale graphs","mitigating feature overcorrelation and oversmoothing in deep graph neural networks","theoretical analysis of subgraph gnn expressive power via symmetry group reduction and unified message passing"]},{"direction_index":3,"label":"D3","name":"Specialized graph structure and geometric learning","description":"Learning representations for graphs whose scientific challenge is non-standard relational structure, heterophily, higher-order or typed relations, multimodal node information, or non-Euclidean/geometric embedding spaces.","inclusion":"Assign atoms whose primary objective is handling heterophilic neighborhoods, heterogeneous or multiplex relation types, hyperedges, signed/directed/property graphs, multimodal feature fusion on graphs, hyperbolic/Riemannian/quasimetric/sheaf geometry, or curvature-aware representation spaces.","exclusion":"If relation or geometry is only incidental to a general architecture, temporal, SSL, or systems contribution, assign to that primary objective instead.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["decoupled self-supervised learning for node representation on non-homophilous graphs via latent semantic modeling","inductive graph embedding with joint node and edge property aggregation via biased neighbor sampling","random walk aggregation for graph neural networks on heterophily graphs"]},{"direction_index":4,"label":"D4","name":"Scalable graph learning systems","description":"Scaling graph representation learning through efficient algorithms, sampling, summarization, optimization, distributed execution, hardware-aware storage/loading, caching, memory reduction, and training or inference throughput improvements.","inclusion":"Assign atoms whose primary objective is computational scalability, systems design, mini-batch or sampling efficiency, graph summarization for embedding scalability, distributed or out-of-core GNN training, hardware utilization, or training-time prediction.","exclusion":"Temporal update efficiency belongs to temporal graph learning when graph evolution is central; transformer-specific attention sparsification belongs to core architecture when the architecture is the main contribution.","nearest_boundary":"Temporal and dynamic graph learning","representative_topics":["graph metric-based prediction of gnn training time for acceleration and design selection","theoretical analysis of graph summarization for scalable node embedding learning","scalable graph convolutional networks via sampling-free minibatch gradient methods"]},{"direction_index":5,"label":"D5","name":"Topology optimization and reliable representations","description":"Learning, refining, generating, or defending graph representations and graph structures when topology or data are noisy, missing, adversarial, uncertain, imbalanced, private, shifting, or sequentially changing.","inclusion":"Assign atoms whose primary objective is graph structure learning, adjacency refinement, latent topology inference, graph similarity structure inference, robust hypergraph or graph topology optimization, adversarial robustness, privacy, fairness, imbalance, uncertainty, causal robustness, continual learning, or graph generation/completion as a topology modeling objective.","exclusion":"Contrastive or SSL methods belong to self-supervised pre-training if robustness is not the primary objective; heterophily-specific topology changes belong to specialized graph structure learning.","nearest_boundary":"Self-supervised pre-training and transfer","representative_topics":["dummy node augmentation for expressive graph kernels and neural network structure learning","unsupervised graph structure refinement for robust graph neural network defense","transferable backdoor attacks on gnns via surrogate-guided perturbation-based dataset poisoning"]},{"direction_index":6,"label":"D6","name":"Spectral, walk, and structural embeddings","description":"Producing graph, node, or subgraph embeddings from non-message-passing structural objectives such as spectra, walks, transition matrices, diffusion signatures, matrix factorization, latent distances, compact graph codes, or structural descriptors.","inclusion":"Assign atoms whose primary objective is deriving embeddings or structural encodings from random walks, spectral decompositions, transition probabilities, diffusion, factorization, graph summarization, Prufer or other combinatorial encodings, or variational/autoencoding embeddings when treated as general unsupervised embedding rather than SSL transfer.","exclusion":"Spectral filters for heterophily, positional encodings for GNN expressivity, or reconstruction objectives for pre-training transfer belong to their respective primary programs.","nearest_boundary":"Self-supervised pre-training and transfer","representative_topics":["theoretical analysis of graph summarization for scalable node embedding learning","scalable diffusion-based structural node embeddings capturing directional information on directed graphs","graph node embedding via anonymous walk-based transition probability matrix for topology-aware classification"]},{"direction_index":7,"label":"D7","name":"Evaluation, theory, and domain utility","description":"Understanding, benchmarking, explaining, and validating graph representation learning, including theoretical guarantees, benchmark taxonomies, interpretability, graph-information utility, surveys, and domain-grounded evaluation where the main objective is assessing or justifying representation usefulness.","inclusion":"Assign atoms whose primary objective is generalization theory, benchmark design, evaluation methodology, interpretability, architectural taxonomies, measuring when graph structure helps, field surveys, or domain-specific evaluation/adaptation framed around representational utility rather than a new general mechanism.","exclusion":"Mechanism-specific theory belongs to that mechanism's direction, such as contrastive loss theory, oversmoothing analysis, heterophily filtering theory, or expressivity analysis tied directly to new architecture design.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["survey of attention-based graph neural network architectures for graph-structured data","theoretical generalization bounds for gnns interpolating bandlimited functions via sampling-theorem connections","taxonomy of graph benchmarks via perturbation-based sensitivity profiling for gnn evaluation"]}],"split":"dev"}
{"episode_id":"F063_T2025-01","field_index":63,"field":"graph representation learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Self-supervised pre-training and transfer","description":"Learning transferable graph representations with label-free or weakly supervised objectives, including contrastive learning, masked or generative pre-training, clustering-driven pretexts, prompt tuning, and language/text alignment when the objective is transfer across tasks.","inclusion":"Assign atoms whose primary objective is to design graph pre-training objectives, augmentations, positive/negative sampling, reconstruction or masking losses, clustering pseudo-labels, graph prompts, adapters, or language-graph alignment for downstream transfer.","exclusion":"If the dominant objective is temporal evolution, scalability, robustness, heterophily, or architectural expressivity rather than the pre-training or transfer program, assign to that corresponding direction.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["graph contrastive learning with interpolation augmentation and correlation reduction for semi-supervised node classification","rationale-aware graph contrastive learning via invariant instance-discrimination discovery","similarity-aware positive instance sampling for graph contrastive pre-training with hierarchical encoding"]},{"direction_index":1,"label":"D1","name":"Temporal and dynamic graph learning","description":"Learning representations from graphs whose edges, nodes, attributes, labels, or interaction events evolve over time, including snapshot, continuous-time, streaming, and event-based settings.","inclusion":"Assign atoms whose primary objective is modeling temporal dependency, structural evolution, event sequences, temporal memory, time-aware updates, or scalable representation maintenance for dynamic graphs.","exclusion":"Static self-supervised, heterogeneous, geometric, or scalable methods belong elsewhere unless time-varying graph evolution is the main research objective.","nearest_boundary":"Scalable graph learning systems","representative_topics":["incremental graph representation updates for scalable dynamic graph neural networks","temporal graph neural networks using depth-first search for path-oriented temporal aggregation","asynchronous dynamic graph embedding via time-aware transformer on temporal edge sequences"]},{"direction_index":2,"label":"D2","name":"Core GNN architecture and long-range representation","description":"Designing graph neural architectures and aggregation mechanisms to improve expressivity, graph/subgraph readout, long-range propagation, transformer-style attention, pooling, and algorithmic or structural inductive bias.","inclusion":"Assign atoms whose primary objective is architectural representation design: mitigating over-smoothing or over-squashing, improving WL/subgraph/path expressivity, designing graph transformers, structural or positional encodings, graph pooling, adaptive readout, hierarchical coarsening, or graph/subgraph aggregation.","exclusion":"If the main objective is heterophily, typed/higher-order graph structure, non-Euclidean geometry, system throughput, or self-supervised pre-training rather than architecture, assign to that more specific direction.","nearest_boundary":"Specialized graph structure and geometric learning","representative_topics":["scalable graph attention multi-layer perceptron for over-smoothing in sparse large-scale graphs","mitigating feature overcorrelation and oversmoothing in deep graph neural networks","theoretical analysis of subgraph gnn expressive power via symmetry group reduction and unified message passing"]},{"direction_index":3,"label":"D3","name":"Specialized graph structure and geometric learning","description":"Learning representations for graphs whose scientific challenge is non-standard relational structure, heterophily, higher-order or typed relations, multimodal node information, or non-Euclidean/geometric embedding spaces.","inclusion":"Assign atoms whose primary objective is handling heterophilic neighborhoods, heterogeneous or multiplex relation types, hyperedges, signed/directed/property graphs, multimodal feature fusion on graphs, hyperbolic/Riemannian/quasimetric/sheaf geometry, or curvature-aware representation spaces.","exclusion":"If relation or geometry is only incidental to a general architecture, temporal, SSL, or systems contribution, assign to that primary objective instead.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["decoupled self-supervised learning for node representation on non-homophilous graphs via latent semantic modeling","inductive graph embedding with joint node and edge property aggregation via biased neighbor sampling","random walk aggregation for graph neural networks on heterophily graphs"]},{"direction_index":4,"label":"D4","name":"Scalable graph learning systems","description":"Scaling graph representation learning through efficient algorithms, sampling, summarization, optimization, distributed execution, hardware-aware storage/loading, caching, memory reduction, and training or inference throughput improvements.","inclusion":"Assign atoms whose primary objective is computational scalability, systems design, mini-batch or sampling efficiency, graph summarization for embedding scalability, distributed or out-of-core GNN training, hardware utilization, or training-time prediction.","exclusion":"Temporal update efficiency belongs to temporal graph learning when graph evolution is central; transformer-specific attention sparsification belongs to core architecture when the architecture is the main contribution.","nearest_boundary":"Temporal and dynamic graph learning","representative_topics":["graph metric-based prediction of gnn training time for acceleration and design selection","theoretical analysis of graph summarization for scalable node embedding learning","scalable graph convolutional networks via sampling-free minibatch gradient methods"]},{"direction_index":5,"label":"D5","name":"Topology optimization and reliable representations","description":"Learning, refining, generating, or defending graph representations and graph structures when topology or data are noisy, missing, adversarial, uncertain, imbalanced, private, shifting, or sequentially changing.","inclusion":"Assign atoms whose primary objective is graph structure learning, adjacency refinement, latent topology inference, graph similarity structure inference, robust hypergraph or graph topology optimization, adversarial robustness, privacy, fairness, imbalance, uncertainty, causal robustness, continual learning, or graph generation/completion as a topology modeling objective.","exclusion":"Contrastive or SSL methods belong to self-supervised pre-training if robustness is not the primary objective; heterophily-specific topology changes belong to specialized graph structure learning.","nearest_boundary":"Self-supervised pre-training and transfer","representative_topics":["dummy node augmentation for expressive graph kernels and neural network structure learning","unsupervised graph structure refinement for robust graph neural network defense","transferable backdoor attacks on gnns via surrogate-guided perturbation-based dataset poisoning"]},{"direction_index":6,"label":"D6","name":"Spectral, walk, and structural embeddings","description":"Producing graph, node, or subgraph embeddings from non-message-passing structural objectives such as spectra, walks, transition matrices, diffusion signatures, matrix factorization, latent distances, compact graph codes, or structural descriptors.","inclusion":"Assign atoms whose primary objective is deriving embeddings or structural encodings from random walks, spectral decompositions, transition probabilities, diffusion, factorization, graph summarization, Prufer or other combinatorial encodings, or variational/autoencoding embeddings when treated as general unsupervised embedding rather than SSL transfer.","exclusion":"Spectral filters for heterophily, positional encodings for GNN expressivity, or reconstruction objectives for pre-training transfer belong to their respective primary programs.","nearest_boundary":"Self-supervised pre-training and transfer","representative_topics":["theoretical analysis of graph summarization for scalable node embedding learning","scalable diffusion-based structural node embeddings capturing directional information on directed graphs","graph node embedding via anonymous walk-based transition probability matrix for topology-aware classification"]},{"direction_index":7,"label":"D7","name":"Evaluation, theory, and domain utility","description":"Understanding, benchmarking, explaining, and validating graph representation learning, including theoretical guarantees, benchmark taxonomies, interpretability, graph-information utility, surveys, and domain-grounded evaluation where the main objective is assessing or justifying representation usefulness.","inclusion":"Assign atoms whose primary objective is generalization theory, benchmark design, evaluation methodology, interpretability, architectural taxonomies, measuring when graph structure helps, field surveys, or domain-specific evaluation/adaptation framed around representational utility rather than a new general mechanism.","exclusion":"Mechanism-specific theory belongs to that mechanism's direction, such as contrastive loss theory, oversmoothing analysis, heterophily filtering theory, or expressivity analysis tied directly to new architecture design.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["survey of attention-based graph neural network architectures for graph-structured data","theoretical generalization bounds for gnns interpolating bandlimited functions via sampling-theorem connections","taxonomy of graph benchmarks via perturbation-based sensitivity profiling for gnn evaluation"]}],"split":"dev"}
{"episode_id":"F063_T2025-07","field_index":63,"field":"graph representation learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Self-supervised pre-training and transfer","description":"Learning transferable graph representations with label-free or weakly supervised objectives, including contrastive learning, masked or generative pre-training, clustering-driven pretexts, prompt tuning, and language/text alignment when the objective is transfer across tasks.","inclusion":"Assign atoms whose primary objective is to design graph pre-training objectives, augmentations, positive/negative sampling, reconstruction or masking losses, clustering pseudo-labels, graph prompts, adapters, or language-graph alignment for downstream transfer.","exclusion":"If the dominant objective is temporal evolution, scalability, robustness, heterophily, or architectural expressivity rather than the pre-training or transfer program, assign to that corresponding direction.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["graph contrastive learning with interpolation augmentation and correlation reduction for semi-supervised node classification","rationale-aware graph contrastive learning via invariant instance-discrimination discovery","similarity-aware positive instance sampling for graph contrastive pre-training with hierarchical encoding"]},{"direction_index":1,"label":"D1","name":"Temporal and dynamic graph learning","description":"Learning representations from graphs whose edges, nodes, attributes, labels, or interaction events evolve over time, including snapshot, continuous-time, streaming, and event-based settings.","inclusion":"Assign atoms whose primary objective is modeling temporal dependency, structural evolution, event sequences, temporal memory, time-aware updates, or scalable representation maintenance for dynamic graphs.","exclusion":"Static self-supervised, heterogeneous, geometric, or scalable methods belong elsewhere unless time-varying graph evolution is the main research objective.","nearest_boundary":"Scalable graph learning systems","representative_topics":["incremental graph representation updates for scalable dynamic graph neural networks","temporal graph neural networks using depth-first search for path-oriented temporal aggregation","asynchronous dynamic graph embedding via time-aware transformer on temporal edge sequences"]},{"direction_index":2,"label":"D2","name":"Core GNN architecture and long-range representation","description":"Designing graph neural architectures and aggregation mechanisms to improve expressivity, graph/subgraph readout, long-range propagation, transformer-style attention, pooling, and algorithmic or structural inductive bias.","inclusion":"Assign atoms whose primary objective is architectural representation design: mitigating over-smoothing or over-squashing, improving WL/subgraph/path expressivity, designing graph transformers, structural or positional encodings, graph pooling, adaptive readout, hierarchical coarsening, or graph/subgraph aggregation.","exclusion":"If the main objective is heterophily, typed/higher-order graph structure, non-Euclidean geometry, system throughput, or self-supervised pre-training rather than architecture, assign to that more specific direction.","nearest_boundary":"Specialized graph structure and geometric learning","representative_topics":["scalable graph attention multi-layer perceptron for over-smoothing in sparse large-scale graphs","mitigating feature overcorrelation and oversmoothing in deep graph neural networks","theoretical analysis of subgraph gnn expressive power via symmetry group reduction and unified message passing"]},{"direction_index":3,"label":"D3","name":"Specialized graph structure and geometric learning","description":"Learning representations for graphs whose scientific challenge is non-standard relational structure, heterophily, higher-order or typed relations, multimodal node information, or non-Euclidean/geometric embedding spaces.","inclusion":"Assign atoms whose primary objective is handling heterophilic neighborhoods, heterogeneous or multiplex relation types, hyperedges, signed/directed/property graphs, multimodal feature fusion on graphs, hyperbolic/Riemannian/quasimetric/sheaf geometry, or curvature-aware representation spaces.","exclusion":"If relation or geometry is only incidental to a general architecture, temporal, SSL, or systems contribution, assign to that primary objective instead.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["decoupled self-supervised learning for node representation on non-homophilous graphs via latent semantic modeling","inductive graph embedding with joint node and edge property aggregation via biased neighbor sampling","random walk aggregation for graph neural networks on heterophily graphs"]},{"direction_index":4,"label":"D4","name":"Scalable graph learning systems","description":"Scaling graph representation learning through efficient algorithms, sampling, summarization, optimization, distributed execution, hardware-aware storage/loading, caching, memory reduction, and training or inference throughput improvements.","inclusion":"Assign atoms whose primary objective is computational scalability, systems design, mini-batch or sampling efficiency, graph summarization for embedding scalability, distributed or out-of-core GNN training, hardware utilization, or training-time prediction.","exclusion":"Temporal update efficiency belongs to temporal graph learning when graph evolution is central; transformer-specific attention sparsification belongs to core architecture when the architecture is the main contribution.","nearest_boundary":"Temporal and dynamic graph learning","representative_topics":["graph metric-based prediction of gnn training time for acceleration and design selection","theoretical analysis of graph summarization for scalable node embedding learning","scalable graph convolutional networks via sampling-free minibatch gradient methods"]},{"direction_index":5,"label":"D5","name":"Topology optimization and reliable representations","description":"Learning, refining, generating, or defending graph representations and graph structures when topology or data are noisy, missing, adversarial, uncertain, imbalanced, private, shifting, or sequentially changing.","inclusion":"Assign atoms whose primary objective is graph structure learning, adjacency refinement, latent topology inference, graph similarity structure inference, robust hypergraph or graph topology optimization, adversarial robustness, privacy, fairness, imbalance, uncertainty, causal robustness, continual learning, or graph generation/completion as a topology modeling objective.","exclusion":"Contrastive or SSL methods belong to self-supervised pre-training if robustness is not the primary objective; heterophily-specific topology changes belong to specialized graph structure learning.","nearest_boundary":"Self-supervised pre-training and transfer","representative_topics":["dummy node augmentation for expressive graph kernels and neural network structure learning","unsupervised graph structure refinement for robust graph neural network defense","transferable backdoor attacks on gnns via surrogate-guided perturbation-based dataset poisoning"]},{"direction_index":6,"label":"D6","name":"Spectral, walk, and structural embeddings","description":"Producing graph, node, or subgraph embeddings from non-message-passing structural objectives such as spectra, walks, transition matrices, diffusion signatures, matrix factorization, latent distances, compact graph codes, or structural descriptors.","inclusion":"Assign atoms whose primary objective is deriving embeddings or structural encodings from random walks, spectral decompositions, transition probabilities, diffusion, factorization, graph summarization, Prufer or other combinatorial encodings, or variational/autoencoding embeddings when treated as general unsupervised embedding rather than SSL transfer.","exclusion":"Spectral filters for heterophily, positional encodings for GNN expressivity, or reconstruction objectives for pre-training transfer belong to their respective primary programs.","nearest_boundary":"Self-supervised pre-training and transfer","representative_topics":["theoretical analysis of graph summarization for scalable node embedding learning","scalable diffusion-based structural node embeddings capturing directional information on directed graphs","graph node embedding via anonymous walk-based transition probability matrix for topology-aware classification"]},{"direction_index":7,"label":"D7","name":"Evaluation, theory, and domain utility","description":"Understanding, benchmarking, explaining, and validating graph representation learning, including theoretical guarantees, benchmark taxonomies, interpretability, graph-information utility, surveys, and domain-grounded evaluation where the main objective is assessing or justifying representation usefulness.","inclusion":"Assign atoms whose primary objective is generalization theory, benchmark design, evaluation methodology, interpretability, architectural taxonomies, measuring when graph structure helps, field surveys, or domain-specific evaluation/adaptation framed around representational utility rather than a new general mechanism.","exclusion":"Mechanism-specific theory belongs to that mechanism's direction, such as contrastive loss theory, oversmoothing analysis, heterophily filtering theory, or expressivity analysis tied directly to new architecture design.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["survey of attention-based graph neural network architectures for graph-structured data","theoretical generalization bounds for gnns interpolating bandlimited functions via sampling-theorem connections","taxonomy of graph benchmarks via perturbation-based sensitivity profiling for gnn evaluation"]}],"split":"dev"}
{"episode_id":"F063_T2026-01","field_index":63,"field":"graph representation learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Self-supervised pre-training and transfer","description":"Learning transferable graph representations with label-free or weakly supervised objectives, including contrastive learning, masked or generative pre-training, clustering-driven pretexts, prompt tuning, and language/text alignment when the objective is transfer across tasks.","inclusion":"Assign atoms whose primary objective is to design graph pre-training objectives, augmentations, positive/negative sampling, reconstruction or masking losses, clustering pseudo-labels, graph prompts, adapters, or language-graph alignment for downstream transfer.","exclusion":"If the dominant objective is temporal evolution, scalability, robustness, heterophily, or architectural expressivity rather than the pre-training or transfer program, assign to that corresponding direction.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["graph contrastive learning with interpolation augmentation and correlation reduction for semi-supervised node classification","rationale-aware graph contrastive learning via invariant instance-discrimination discovery","similarity-aware positive instance sampling for graph contrastive pre-training with hierarchical encoding"]},{"direction_index":1,"label":"D1","name":"Temporal and dynamic graph learning","description":"Learning representations from graphs whose edges, nodes, attributes, labels, or interaction events evolve over time, including snapshot, continuous-time, streaming, and event-based settings.","inclusion":"Assign atoms whose primary objective is modeling temporal dependency, structural evolution, event sequences, temporal memory, time-aware updates, or scalable representation maintenance for dynamic graphs.","exclusion":"Static self-supervised, heterogeneous, geometric, or scalable methods belong elsewhere unless time-varying graph evolution is the main research objective.","nearest_boundary":"Scalable graph learning systems","representative_topics":["incremental graph representation updates for scalable dynamic graph neural networks","temporal graph neural networks using depth-first search for path-oriented temporal aggregation","asynchronous dynamic graph embedding via time-aware transformer on temporal edge sequences"]},{"direction_index":2,"label":"D2","name":"Core GNN architecture and long-range representation","description":"Designing graph neural architectures and aggregation mechanisms to improve expressivity, graph/subgraph readout, long-range propagation, transformer-style attention, pooling, and algorithmic or structural inductive bias.","inclusion":"Assign atoms whose primary objective is architectural representation design: mitigating over-smoothing or over-squashing, improving WL/subgraph/path expressivity, designing graph transformers, structural or positional encodings, graph pooling, adaptive readout, hierarchical coarsening, or graph/subgraph aggregation.","exclusion":"If the main objective is heterophily, typed/higher-order graph structure, non-Euclidean geometry, system throughput, or self-supervised pre-training rather than architecture, assign to that more specific direction.","nearest_boundary":"Specialized graph structure and geometric learning","representative_topics":["scalable graph attention multi-layer perceptron for over-smoothing in sparse large-scale graphs","mitigating feature overcorrelation and oversmoothing in deep graph neural networks","theoretical analysis of subgraph gnn expressive power via symmetry group reduction and unified message passing"]},{"direction_index":3,"label":"D3","name":"Specialized graph structure and geometric learning","description":"Learning representations for graphs whose scientific challenge is non-standard relational structure, heterophily, higher-order or typed relations, multimodal node information, or non-Euclidean/geometric embedding spaces.","inclusion":"Assign atoms whose primary objective is handling heterophilic neighborhoods, heterogeneous or multiplex relation types, hyperedges, signed/directed/property graphs, multimodal feature fusion on graphs, hyperbolic/Riemannian/quasimetric/sheaf geometry, or curvature-aware representation spaces.","exclusion":"If relation or geometry is only incidental to a general architecture, temporal, SSL, or systems contribution, assign to that primary objective instead.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["decoupled self-supervised learning for node representation on non-homophilous graphs via latent semantic modeling","inductive graph embedding with joint node and edge property aggregation via biased neighbor sampling","random walk aggregation for graph neural networks on heterophily graphs"]},{"direction_index":4,"label":"D4","name":"Scalable graph learning systems","description":"Scaling graph representation learning through efficient algorithms, sampling, summarization, optimization, distributed execution, hardware-aware storage/loading, caching, memory reduction, and training or inference throughput improvements.","inclusion":"Assign atoms whose primary objective is computational scalability, systems design, mini-batch or sampling efficiency, graph summarization for embedding scalability, distributed or out-of-core GNN training, hardware utilization, or training-time prediction.","exclusion":"Temporal update efficiency belongs to temporal graph learning when graph evolution is central; transformer-specific attention sparsification belongs to core architecture when the architecture is the main contribution.","nearest_boundary":"Temporal and dynamic graph learning","representative_topics":["graph metric-based prediction of gnn training time for acceleration and design selection","theoretical analysis of graph summarization for scalable node embedding learning","scalable graph convolutional networks via sampling-free minibatch gradient methods"]},{"direction_index":5,"label":"D5","name":"Topology optimization and reliable representations","description":"Learning, refining, generating, or defending graph representations and graph structures when topology or data are noisy, missing, adversarial, uncertain, imbalanced, private, shifting, or sequentially changing.","inclusion":"Assign atoms whose primary objective is graph structure learning, adjacency refinement, latent topology inference, graph similarity structure inference, robust hypergraph or graph topology optimization, adversarial robustness, privacy, fairness, imbalance, uncertainty, causal robustness, continual learning, or graph generation/completion as a topology modeling objective.","exclusion":"Contrastive or SSL methods belong to self-supervised pre-training if robustness is not the primary objective; heterophily-specific topology changes belong to specialized graph structure learning.","nearest_boundary":"Self-supervised pre-training and transfer","representative_topics":["dummy node augmentation for expressive graph kernels and neural network structure learning","unsupervised graph structure refinement for robust graph neural network defense","transferable backdoor attacks on gnns via surrogate-guided perturbation-based dataset poisoning"]},{"direction_index":6,"label":"D6","name":"Spectral, walk, and structural embeddings","description":"Producing graph, node, or subgraph embeddings from non-message-passing structural objectives such as spectra, walks, transition matrices, diffusion signatures, matrix factorization, latent distances, compact graph codes, or structural descriptors.","inclusion":"Assign atoms whose primary objective is deriving embeddings or structural encodings from random walks, spectral decompositions, transition probabilities, diffusion, factorization, graph summarization, Prufer or other combinatorial encodings, or variational/autoencoding embeddings when treated as general unsupervised embedding rather than SSL transfer.","exclusion":"Spectral filters for heterophily, positional encodings for GNN expressivity, or reconstruction objectives for pre-training transfer belong to their respective primary programs.","nearest_boundary":"Self-supervised pre-training and transfer","representative_topics":["theoretical analysis of graph summarization for scalable node embedding learning","scalable diffusion-based structural node embeddings capturing directional information on directed graphs","graph node embedding via anonymous walk-based transition probability matrix for topology-aware classification"]},{"direction_index":7,"label":"D7","name":"Evaluation, theory, and domain utility","description":"Understanding, benchmarking, explaining, and validating graph representation learning, including theoretical guarantees, benchmark taxonomies, interpretability, graph-information utility, surveys, and domain-grounded evaluation where the main objective is assessing or justifying representation usefulness.","inclusion":"Assign atoms whose primary objective is generalization theory, benchmark design, evaluation methodology, interpretability, architectural taxonomies, measuring when graph structure helps, field surveys, or domain-specific evaluation/adaptation framed around representational utility rather than a new general mechanism.","exclusion":"Mechanism-specific theory belongs to that mechanism's direction, such as contrastive loss theory, oversmoothing analysis, heterophily filtering theory, or expressivity analysis tied directly to new architecture design.","nearest_boundary":"Core GNN architecture and long-range representation","representative_topics":["survey of attention-based graph neural network architectures for graph-structured data","theoretical generalization bounds for gnns interpolating bandlimited functions via sampling-theorem connections","taxonomy of graph benchmarks via perturbation-based sensitivity profiling for gnn evaluation"]}],"split":"dev"}
{"episode_id":"F065_T2024-01","field_index":65,"field":"image super-resolution","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Robust real-world and weak-supervision SR","description":"Super-resolution whose primary objective is to work under unknown real-world degradations, domain shift, missing paired supervision, or to evaluate robustness under such mismatches.","inclusion":"Include atoms focused on blind degradation modeling or estimation, real-world degradation robustness, unpaired/self-supervised/source-free/domain-adaptive SR, test-time adaptation driven by mismatch, and systematic robustness or benchmark evaluation of real-world SR.","exclusion":"If the main mechanism is a generative prior for perceptual synthesis, assign to generative perceptual SR; if the defining constraint is a medical, scientific, sensor, or semantic domain, assign to domain-constrained SR; if the main goal is architecture quality under standard paired SISR, assign to representation architectures.","nearest_boundary":"Generative perceptual SR","representative_topics":["real-world single image super-resolution under rainy weather degradation","blind super-resolution via degradation-guided meta-restoration with joint consistency loss","blind super-resolution via meta-learning and implicit degradation representation extraction without groundtruth supervision"]},{"direction_index":1,"label":"D1","name":"Efficient deployable SR","description":"Super-resolution whose primary objective is reducing computation, memory, latency, energy, bit-width, or model size while preserving reconstruction quality.","inclusion":"Include lightweight networks, pruning, quantization, distillation, sparse or dynamic computation, mobile/NPU/edge deployment, and real-time SR when resource reduction is the dominant research program.","exclusion":"If efficiency is only incidental and the main contribution is feature modeling or reconstruction quality, assign to representation architectures; if temporal modeling dominates in video despite efficient components, assign to temporal sequence SR.","nearest_boundary":"Representation and scaling architectures","representative_topics":["structured sparsity pruning for efficient video super-resolution on resource-constrained devices","lightweight transformer-cnn hybrid for contextual-aware single image super-resolution","lightweight super-resolution via dual regression constraints and channel-level model compression"]},{"direction_index":2,"label":"D2","name":"Representation and scaling architectures","description":"Super-resolution whose primary objective is improving single-image reconstruction through architectures, feature representations, priors, optimization, or continuous/arbitrary-scale image representations.","inclusion":"Include general SISR architectures, attention or transformer/CNN hybrids, dictionary or sparse priors, self-similarity, frequency or wavelet modeling, loss/training designs, implicit neural representations, invertible rescaling, and arbitrary-scale or coordinate-continuous SR when the central goal is better image representation or scale generalization.","exclusion":"If reduced resource use is the explicit objective, assign to efficient deployable SR; if unknown real-world degradation or supervision scarcity defines the problem, assign to robust real-world and weak-supervision SR; if GAN/diffusion perceptual synthesis dominates, assign to generative perceptual SR.","nearest_boundary":"Efficient deployable SR","representative_topics":["patch-based super resolution via generalized gaussian mixture model and mmse reconstruction","hierarchical self-similarity learning for aliasing suppression in single image super-resolution","explicit dictionary learning for out-of-domain robustness in single-image super-resolution"]},{"direction_index":3,"label":"D3","name":"Generative perceptual SR","description":"Super-resolution whose primary objective is synthesizing perceptually plausible high-frequency detail or managing the fidelity-perception tradeoff through GAN, diffusion, flow, or other generative priors.","inclusion":"Include pretrained GAN prior inversion or latent banks, diffusion/score-based SR, conditional generative sampling, adversarially trained perceptual SR, GAN artifact handling, and real-world SR when the dominant contribution is generative synthesis rather than explicit degradation estimation.","exclusion":"If explicit degradation estimation or adaptation is the core objective, assign to robust real-world and weak-supervision SR; if medical or other domain constraints dominate over the generative mechanism, assign to the corresponding domain program.","nearest_boundary":"Robust real-world and weak-supervision SR","representative_topics":["pre-trained gan latent bank for perceptual image super-resolution without runtime inversion","gan-prior large-scale super-resolution via latent space graph-based multi-relation disentanglement and code generation","latent diffusion models for single image super-resolution with multimodal adversarial denoising"]},{"direction_index":4,"label":"D4","name":"Temporal sequence SR","description":"Super-resolution whose primary objective is recovering high-resolution images or videos from temporally related frames by modeling motion, alignment, propagation, fusion, and temporal consistency.","inclusion":"Include video SR, burst SR, recurrent or transformer sequence fusion, multi-frame alignment, temporal consistency learning, motion compensation, raw or exposure-bracketed burst reconstruction, and compressed video SR when temporal modeling is central.","exclusion":"If auxiliary views are non-temporal references or stereo/light-field geometry, assign to auxiliary-view and pipeline-conditioned SR; if resource reduction is the dominant goal, assign to efficient deployable SR.","nearest_boundary":"Auxiliary-view and pipeline-conditioned SR","representative_topics":["parametric 3d spatiotemporal filters for joint video super-resolution, denoising, and illumination enhancement","patch alignment for efficient multi-frame attention in video super-resolution transformers","super-resolution and dynamic range reconstruction from raw multi-exposure bursts using physics-based inverse optimization"]},{"direction_index":5,"label":"D5","name":"Biomedical SR","description":"Super-resolution whose primary objective is reconstructing medical or biomedical imagery while respecting modality-specific acquisition, anatomy, volumetric structure, contrast, biological texture, or clinical utility.","inclusion":"Include MRI, CT, ultrasound, OCT/OCTA, photoacoustic, microscopy, histopathology, medical volumes, multi-contrast medical fusion, unpaired medical SR, and medical-domain GAN/diffusion/transformer SR when biomedical constraints define the problem.","exclusion":"If the target is non-biomedical scientific, remote-sensing, thermal, depth, or semantic content, assign to domain-constrained non-medical SR; if a general method is merely evaluated on medical data without biomedical constraints, assign to representation architectures.","nearest_boundary":"Domain-constrained non-medical SR","representative_topics":["transformer-based volumetric super-resolution for anisotropic ct with real-paired datasets","generative adversarial network super-resolution for 2d brain mri enhancement","hybrid cnn-vision transformer super-resolution of mri with structure and texture constraints"]},{"direction_index":6,"label":"D6","name":"Domain-constrained non-medical SR","description":"Super-resolution whose primary objective is satisfying the physics, sensor geometry, spectral structure, semantic structure, or validation criteria of a non-medical specialized domain.","inclusion":"Include remote sensing, satellite, hyperspectral/multispectral, SAR, thermal, depth/time-of-flight, scientific simulations, dynamical systems, calorimetry, face, text, screen content, and other structured semantic or sensor domains when domain constraints are primary.","exclusion":"If the modality is medical or biomedical, assign to biomedical SR; if an external reference image or graphics/codec pipeline is the defining source of structure, assign to auxiliary-view and pipeline-conditioned SR; if domain shift or missing pairs is the main objective independent of domain constraints, assign to robust real-world and weak-supervision SR.","nearest_boundary":"Biomedical SR","representative_topics":["thermal image super-resolution via bilateral network with channel splitting and transformer fusion","stochastic attribute modeling for uncertainty-aware face super-resolution","real-world benchmark for multi-image super-resolution of satellite remote sensing imagery"]},{"direction_index":7,"label":"D7","name":"Auxiliary-view and pipeline-conditioned SR","description":"Super-resolution whose primary objective is exploiting external views, reference images, geometry, or production-pipeline structure such as codecs, tone mapping, rasterization, rendering buffers, or graphics metadata.","inclusion":"Include reference-based, multi-reference, stereo, light-field, multi-view, deformable correspondence or texture-transfer SR, codec-assisted or compressed-content SR, joint SR with inverse tone mapping, rasterized/game/rendered-content upscaling, and g-buffer or motion-vector guided rendering SR when these auxiliary sources or pipeline structures define the problem.","exclusion":"If temporally adjacent frames are fused primarily for video consistency, assign to temporal sequence SR; if compression is treated only as an unknown degradation without exploiting codec or pipeline structure, assign to robust real-world and weak-supervision SR; if the auxiliary source is a domain sensor whose physics dominates, assign to domain-constrained non-medical SR.","nearest_boundary":"Temporal sequence SR","representative_topics":["joint super-resolution and inverse tone-mapping via hierarchical feature decomposition aggregation","geometry-aware reference synthesis for multi-view image super-resolution with large-angle transformations","reference-based super-resolution via deformable attention and multi-scale texture feature transfer"]}],"split":"dev"}
{"episode_id":"F065_T2024-07","field_index":65,"field":"image super-resolution","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Robust real-world and weak-supervision SR","description":"Super-resolution whose primary objective is to work under unknown real-world degradations, domain shift, missing paired supervision, or to evaluate robustness under such mismatches.","inclusion":"Include atoms focused on blind degradation modeling or estimation, real-world degradation robustness, unpaired/self-supervised/source-free/domain-adaptive SR, test-time adaptation driven by mismatch, and systematic robustness or benchmark evaluation of real-world SR.","exclusion":"If the main mechanism is a generative prior for perceptual synthesis, assign to generative perceptual SR; if the defining constraint is a medical, scientific, sensor, or semantic domain, assign to domain-constrained SR; if the main goal is architecture quality under standard paired SISR, assign to representation architectures.","nearest_boundary":"Generative perceptual SR","representative_topics":["real-world single image super-resolution under rainy weather degradation","blind super-resolution via degradation-guided meta-restoration with joint consistency loss","blind super-resolution via meta-learning and implicit degradation representation extraction without groundtruth supervision"]},{"direction_index":1,"label":"D1","name":"Efficient deployable SR","description":"Super-resolution whose primary objective is reducing computation, memory, latency, energy, bit-width, or model size while preserving reconstruction quality.","inclusion":"Include lightweight networks, pruning, quantization, distillation, sparse or dynamic computation, mobile/NPU/edge deployment, and real-time SR when resource reduction is the dominant research program.","exclusion":"If efficiency is only incidental and the main contribution is feature modeling or reconstruction quality, assign to representation architectures; if temporal modeling dominates in video despite efficient components, assign to temporal sequence SR.","nearest_boundary":"Representation and scaling architectures","representative_topics":["structured sparsity pruning for efficient video super-resolution on resource-constrained devices","lightweight transformer-cnn hybrid for contextual-aware single image super-resolution","lightweight super-resolution via dual regression constraints and channel-level model compression"]},{"direction_index":2,"label":"D2","name":"Representation and scaling architectures","description":"Super-resolution whose primary objective is improving single-image reconstruction through architectures, feature representations, priors, optimization, or continuous/arbitrary-scale image representations.","inclusion":"Include general SISR architectures, attention or transformer/CNN hybrids, dictionary or sparse priors, self-similarity, frequency or wavelet modeling, loss/training designs, implicit neural representations, invertible rescaling, and arbitrary-scale or coordinate-continuous SR when the central goal is better image representation or scale generalization.","exclusion":"If reduced resource use is the explicit objective, assign to efficient deployable SR; if unknown real-world degradation or supervision scarcity defines the problem, assign to robust real-world and weak-supervision SR; if GAN/diffusion perceptual synthesis dominates, assign to generative perceptual SR.","nearest_boundary":"Efficient deployable SR","representative_topics":["patch-based super resolution via generalized gaussian mixture model and mmse reconstruction","hierarchical self-similarity learning for aliasing suppression in single image super-resolution","explicit dictionary learning for out-of-domain robustness in single-image super-resolution"]},{"direction_index":3,"label":"D3","name":"Generative perceptual SR","description":"Super-resolution whose primary objective is synthesizing perceptually plausible high-frequency detail or managing the fidelity-perception tradeoff through GAN, diffusion, flow, or other generative priors.","inclusion":"Include pretrained GAN prior inversion or latent banks, diffusion/score-based SR, conditional generative sampling, adversarially trained perceptual SR, GAN artifact handling, and real-world SR when the dominant contribution is generative synthesis rather than explicit degradation estimation.","exclusion":"If explicit degradation estimation or adaptation is the core objective, assign to robust real-world and weak-supervision SR; if medical or other domain constraints dominate over the generative mechanism, assign to the corresponding domain program.","nearest_boundary":"Robust real-world and weak-supervision SR","representative_topics":["pre-trained gan latent bank for perceptual image super-resolution without runtime inversion","gan-prior large-scale super-resolution via latent space graph-based multi-relation disentanglement and code generation","latent diffusion models for single image super-resolution with multimodal adversarial denoising"]},{"direction_index":4,"label":"D4","name":"Temporal sequence SR","description":"Super-resolution whose primary objective is recovering high-resolution images or videos from temporally related frames by modeling motion, alignment, propagation, fusion, and temporal consistency.","inclusion":"Include video SR, burst SR, recurrent or transformer sequence fusion, multi-frame alignment, temporal consistency learning, motion compensation, raw or exposure-bracketed burst reconstruction, and compressed video SR when temporal modeling is central.","exclusion":"If auxiliary views are non-temporal references or stereo/light-field geometry, assign to auxiliary-view and pipeline-conditioned SR; if resource reduction is the dominant goal, assign to efficient deployable SR.","nearest_boundary":"Auxiliary-view and pipeline-conditioned SR","representative_topics":["parametric 3d spatiotemporal filters for joint video super-resolution, denoising, and illumination enhancement","patch alignment for efficient multi-frame attention in video super-resolution transformers","super-resolution and dynamic range reconstruction from raw multi-exposure bursts using physics-based inverse optimization"]},{"direction_index":5,"label":"D5","name":"Biomedical SR","description":"Super-resolution whose primary objective is reconstructing medical or biomedical imagery while respecting modality-specific acquisition, anatomy, volumetric structure, contrast, biological texture, or clinical utility.","inclusion":"Include MRI, CT, ultrasound, OCT/OCTA, photoacoustic, microscopy, histopathology, medical volumes, multi-contrast medical fusion, unpaired medical SR, and medical-domain GAN/diffusion/transformer SR when biomedical constraints define the problem.","exclusion":"If the target is non-biomedical scientific, remote-sensing, thermal, depth, or semantic content, assign to domain-constrained non-medical SR; if a general method is merely evaluated on medical data without biomedical constraints, assign to representation architectures.","nearest_boundary":"Domain-constrained non-medical SR","representative_topics":["transformer-based volumetric super-resolution for anisotropic ct with real-paired datasets","generative adversarial network super-resolution for 2d brain mri enhancement","hybrid cnn-vision transformer super-resolution of mri with structure and texture constraints"]},{"direction_index":6,"label":"D6","name":"Domain-constrained non-medical SR","description":"Super-resolution whose primary objective is satisfying the physics, sensor geometry, spectral structure, semantic structure, or validation criteria of a non-medical specialized domain.","inclusion":"Include remote sensing, satellite, hyperspectral/multispectral, SAR, thermal, depth/time-of-flight, scientific simulations, dynamical systems, calorimetry, face, text, screen content, and other structured semantic or sensor domains when domain constraints are primary.","exclusion":"If the modality is medical or biomedical, assign to biomedical SR; if an external reference image or graphics/codec pipeline is the defining source of structure, assign to auxiliary-view and pipeline-conditioned SR; if domain shift or missing pairs is the main objective independent of domain constraints, assign to robust real-world and weak-supervision SR.","nearest_boundary":"Biomedical SR","representative_topics":["thermal image super-resolution via bilateral network with channel splitting and transformer fusion","stochastic attribute modeling for uncertainty-aware face super-resolution","real-world benchmark for multi-image super-resolution of satellite remote sensing imagery"]},{"direction_index":7,"label":"D7","name":"Auxiliary-view and pipeline-conditioned SR","description":"Super-resolution whose primary objective is exploiting external views, reference images, geometry, or production-pipeline structure such as codecs, tone mapping, rasterization, rendering buffers, or graphics metadata.","inclusion":"Include reference-based, multi-reference, stereo, light-field, multi-view, deformable correspondence or texture-transfer SR, codec-assisted or compressed-content SR, joint SR with inverse tone mapping, rasterized/game/rendered-content upscaling, and g-buffer or motion-vector guided rendering SR when these auxiliary sources or pipeline structures define the problem.","exclusion":"If temporally adjacent frames are fused primarily for video consistency, assign to temporal sequence SR; if compression is treated only as an unknown degradation without exploiting codec or pipeline structure, assign to robust real-world and weak-supervision SR; if the auxiliary source is a domain sensor whose physics dominates, assign to domain-constrained non-medical SR.","nearest_boundary":"Temporal sequence SR","representative_topics":["joint super-resolution and inverse tone-mapping via hierarchical feature decomposition aggregation","geometry-aware reference synthesis for multi-view image super-resolution with large-angle transformations","reference-based super-resolution via deformable attention and multi-scale texture feature transfer"]}],"split":"dev"}
{"episode_id":"F065_T2025-01","field_index":65,"field":"image super-resolution","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Robust real-world and weak-supervision SR","description":"Super-resolution whose primary objective is to work under unknown real-world degradations, domain shift, missing paired supervision, or to evaluate robustness under such mismatches.","inclusion":"Include atoms focused on blind degradation modeling or estimation, real-world degradation robustness, unpaired/self-supervised/source-free/domain-adaptive SR, test-time adaptation driven by mismatch, and systematic robustness or benchmark evaluation of real-world SR.","exclusion":"If the main mechanism is a generative prior for perceptual synthesis, assign to generative perceptual SR; if the defining constraint is a medical, scientific, sensor, or semantic domain, assign to domain-constrained SR; if the main goal is architecture quality under standard paired SISR, assign to representation architectures.","nearest_boundary":"Generative perceptual SR","representative_topics":["real-world single image super-resolution under rainy weather degradation","blind super-resolution via degradation-guided meta-restoration with joint consistency loss","blind super-resolution via meta-learning and implicit degradation representation extraction without groundtruth supervision"]},{"direction_index":1,"label":"D1","name":"Efficient deployable SR","description":"Super-resolution whose primary objective is reducing computation, memory, latency, energy, bit-width, or model size while preserving reconstruction quality.","inclusion":"Include lightweight networks, pruning, quantization, distillation, sparse or dynamic computation, mobile/NPU/edge deployment, and real-time SR when resource reduction is the dominant research program.","exclusion":"If efficiency is only incidental and the main contribution is feature modeling or reconstruction quality, assign to representation architectures; if temporal modeling dominates in video despite efficient components, assign to temporal sequence SR.","nearest_boundary":"Representation and scaling architectures","representative_topics":["structured sparsity pruning for efficient video super-resolution on resource-constrained devices","lightweight transformer-cnn hybrid for contextual-aware single image super-resolution","lightweight super-resolution via dual regression constraints and channel-level model compression"]},{"direction_index":2,"label":"D2","name":"Representation and scaling architectures","description":"Super-resolution whose primary objective is improving single-image reconstruction through architectures, feature representations, priors, optimization, or continuous/arbitrary-scale image representations.","inclusion":"Include general SISR architectures, attention or transformer/CNN hybrids, dictionary or sparse priors, self-similarity, frequency or wavelet modeling, loss/training designs, implicit neural representations, invertible rescaling, and arbitrary-scale or coordinate-continuous SR when the central goal is better image representation or scale generalization.","exclusion":"If reduced resource use is the explicit objective, assign to efficient deployable SR; if unknown real-world degradation or supervision scarcity defines the problem, assign to robust real-world and weak-supervision SR; if GAN/diffusion perceptual synthesis dominates, assign to generative perceptual SR.","nearest_boundary":"Efficient deployable SR","representative_topics":["patch-based super resolution via generalized gaussian mixture model and mmse reconstruction","hierarchical self-similarity learning for aliasing suppression in single image super-resolution","explicit dictionary learning for out-of-domain robustness in single-image super-resolution"]},{"direction_index":3,"label":"D3","name":"Generative perceptual SR","description":"Super-resolution whose primary objective is synthesizing perceptually plausible high-frequency detail or managing the fidelity-perception tradeoff through GAN, diffusion, flow, or other generative priors.","inclusion":"Include pretrained GAN prior inversion or latent banks, diffusion/score-based SR, conditional generative sampling, adversarially trained perceptual SR, GAN artifact handling, and real-world SR when the dominant contribution is generative synthesis rather than explicit degradation estimation.","exclusion":"If explicit degradation estimation or adaptation is the core objective, assign to robust real-world and weak-supervision SR; if medical or other domain constraints dominate over the generative mechanism, assign to the corresponding domain program.","nearest_boundary":"Robust real-world and weak-supervision SR","representative_topics":["pre-trained gan latent bank for perceptual image super-resolution without runtime inversion","gan-prior large-scale super-resolution via latent space graph-based multi-relation disentanglement and code generation","latent diffusion models for single image super-resolution with multimodal adversarial denoising"]},{"direction_index":4,"label":"D4","name":"Temporal sequence SR","description":"Super-resolution whose primary objective is recovering high-resolution images or videos from temporally related frames by modeling motion, alignment, propagation, fusion, and temporal consistency.","inclusion":"Include video SR, burst SR, recurrent or transformer sequence fusion, multi-frame alignment, temporal consistency learning, motion compensation, raw or exposure-bracketed burst reconstruction, and compressed video SR when temporal modeling is central.","exclusion":"If auxiliary views are non-temporal references or stereo/light-field geometry, assign to auxiliary-view and pipeline-conditioned SR; if resource reduction is the dominant goal, assign to efficient deployable SR.","nearest_boundary":"Auxiliary-view and pipeline-conditioned SR","representative_topics":["parametric 3d spatiotemporal filters for joint video super-resolution, denoising, and illumination enhancement","patch alignment for efficient multi-frame attention in video super-resolution transformers","super-resolution and dynamic range reconstruction from raw multi-exposure bursts using physics-based inverse optimization"]},{"direction_index":5,"label":"D5","name":"Biomedical SR","description":"Super-resolution whose primary objective is reconstructing medical or biomedical imagery while respecting modality-specific acquisition, anatomy, volumetric structure, contrast, biological texture, or clinical utility.","inclusion":"Include MRI, CT, ultrasound, OCT/OCTA, photoacoustic, microscopy, histopathology, medical volumes, multi-contrast medical fusion, unpaired medical SR, and medical-domain GAN/diffusion/transformer SR when biomedical constraints define the problem.","exclusion":"If the target is non-biomedical scientific, remote-sensing, thermal, depth, or semantic content, assign to domain-constrained non-medical SR; if a general method is merely evaluated on medical data without biomedical constraints, assign to representation architectures.","nearest_boundary":"Domain-constrained non-medical SR","representative_topics":["transformer-based volumetric super-resolution for anisotropic ct with real-paired datasets","generative adversarial network super-resolution for 2d brain mri enhancement","hybrid cnn-vision transformer super-resolution of mri with structure and texture constraints"]},{"direction_index":6,"label":"D6","name":"Domain-constrained non-medical SR","description":"Super-resolution whose primary objective is satisfying the physics, sensor geometry, spectral structure, semantic structure, or validation criteria of a non-medical specialized domain.","inclusion":"Include remote sensing, satellite, hyperspectral/multispectral, SAR, thermal, depth/time-of-flight, scientific simulations, dynamical systems, calorimetry, face, text, screen content, and other structured semantic or sensor domains when domain constraints are primary.","exclusion":"If the modality is medical or biomedical, assign to biomedical SR; if an external reference image or graphics/codec pipeline is the defining source of structure, assign to auxiliary-view and pipeline-conditioned SR; if domain shift or missing pairs is the main objective independent of domain constraints, assign to robust real-world and weak-supervision SR.","nearest_boundary":"Biomedical SR","representative_topics":["thermal image super-resolution via bilateral network with channel splitting and transformer fusion","stochastic attribute modeling for uncertainty-aware face super-resolution","real-world benchmark for multi-image super-resolution of satellite remote sensing imagery"]},{"direction_index":7,"label":"D7","name":"Auxiliary-view and pipeline-conditioned SR","description":"Super-resolution whose primary objective is exploiting external views, reference images, geometry, or production-pipeline structure such as codecs, tone mapping, rasterization, rendering buffers, or graphics metadata.","inclusion":"Include reference-based, multi-reference, stereo, light-field, multi-view, deformable correspondence or texture-transfer SR, codec-assisted or compressed-content SR, joint SR with inverse tone mapping, rasterized/game/rendered-content upscaling, and g-buffer or motion-vector guided rendering SR when these auxiliary sources or pipeline structures define the problem.","exclusion":"If temporally adjacent frames are fused primarily for video consistency, assign to temporal sequence SR; if compression is treated only as an unknown degradation without exploiting codec or pipeline structure, assign to robust real-world and weak-supervision SR; if the auxiliary source is a domain sensor whose physics dominates, assign to domain-constrained non-medical SR.","nearest_boundary":"Temporal sequence SR","representative_topics":["joint super-resolution and inverse tone-mapping via hierarchical feature decomposition aggregation","geometry-aware reference synthesis for multi-view image super-resolution with large-angle transformations","reference-based super-resolution via deformable attention and multi-scale texture feature transfer"]}],"split":"dev"}
{"episode_id":"F065_T2025-07","field_index":65,"field":"image super-resolution","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Robust real-world and weak-supervision SR","description":"Super-resolution whose primary objective is to work under unknown real-world degradations, domain shift, missing paired supervision, or to evaluate robustness under such mismatches.","inclusion":"Include atoms focused on blind degradation modeling or estimation, real-world degradation robustness, unpaired/self-supervised/source-free/domain-adaptive SR, test-time adaptation driven by mismatch, and systematic robustness or benchmark evaluation of real-world SR.","exclusion":"If the main mechanism is a generative prior for perceptual synthesis, assign to generative perceptual SR; if the defining constraint is a medical, scientific, sensor, or semantic domain, assign to domain-constrained SR; if the main goal is architecture quality under standard paired SISR, assign to representation architectures.","nearest_boundary":"Generative perceptual SR","representative_topics":["real-world single image super-resolution under rainy weather degradation","blind super-resolution via degradation-guided meta-restoration with joint consistency loss","blind super-resolution via meta-learning and implicit degradation representation extraction without groundtruth supervision"]},{"direction_index":1,"label":"D1","name":"Efficient deployable SR","description":"Super-resolution whose primary objective is reducing computation, memory, latency, energy, bit-width, or model size while preserving reconstruction quality.","inclusion":"Include lightweight networks, pruning, quantization, distillation, sparse or dynamic computation, mobile/NPU/edge deployment, and real-time SR when resource reduction is the dominant research program.","exclusion":"If efficiency is only incidental and the main contribution is feature modeling or reconstruction quality, assign to representation architectures; if temporal modeling dominates in video despite efficient components, assign to temporal sequence SR.","nearest_boundary":"Representation and scaling architectures","representative_topics":["structured sparsity pruning for efficient video super-resolution on resource-constrained devices","lightweight transformer-cnn hybrid for contextual-aware single image super-resolution","lightweight super-resolution via dual regression constraints and channel-level model compression"]},{"direction_index":2,"label":"D2","name":"Representation and scaling architectures","description":"Super-resolution whose primary objective is improving single-image reconstruction through architectures, feature representations, priors, optimization, or continuous/arbitrary-scale image representations.","inclusion":"Include general SISR architectures, attention or transformer/CNN hybrids, dictionary or sparse priors, self-similarity, frequency or wavelet modeling, loss/training designs, implicit neural representations, invertible rescaling, and arbitrary-scale or coordinate-continuous SR when the central goal is better image representation or scale generalization.","exclusion":"If reduced resource use is the explicit objective, assign to efficient deployable SR; if unknown real-world degradation or supervision scarcity defines the problem, assign to robust real-world and weak-supervision SR; if GAN/diffusion perceptual synthesis dominates, assign to generative perceptual SR.","nearest_boundary":"Efficient deployable SR","representative_topics":["patch-based super resolution via generalized gaussian mixture model and mmse reconstruction","hierarchical self-similarity learning for aliasing suppression in single image super-resolution","explicit dictionary learning for out-of-domain robustness in single-image super-resolution"]},{"direction_index":3,"label":"D3","name":"Generative perceptual SR","description":"Super-resolution whose primary objective is synthesizing perceptually plausible high-frequency detail or managing the fidelity-perception tradeoff through GAN, diffusion, flow, or other generative priors.","inclusion":"Include pretrained GAN prior inversion or latent banks, diffusion/score-based SR, conditional generative sampling, adversarially trained perceptual SR, GAN artifact handling, and real-world SR when the dominant contribution is generative synthesis rather than explicit degradation estimation.","exclusion":"If explicit degradation estimation or adaptation is the core objective, assign to robust real-world and weak-supervision SR; if medical or other domain constraints dominate over the generative mechanism, assign to the corresponding domain program.","nearest_boundary":"Robust real-world and weak-supervision SR","representative_topics":["pre-trained gan latent bank for perceptual image super-resolution without runtime inversion","gan-prior large-scale super-resolution via latent space graph-based multi-relation disentanglement and code generation","latent diffusion models for single image super-resolution with multimodal adversarial denoising"]},{"direction_index":4,"label":"D4","name":"Temporal sequence SR","description":"Super-resolution whose primary objective is recovering high-resolution images or videos from temporally related frames by modeling motion, alignment, propagation, fusion, and temporal consistency.","inclusion":"Include video SR, burst SR, recurrent or transformer sequence fusion, multi-frame alignment, temporal consistency learning, motion compensation, raw or exposure-bracketed burst reconstruction, and compressed video SR when temporal modeling is central.","exclusion":"If auxiliary views are non-temporal references or stereo/light-field geometry, assign to auxiliary-view and pipeline-conditioned SR; if resource reduction is the dominant goal, assign to efficient deployable SR.","nearest_boundary":"Auxiliary-view and pipeline-conditioned SR","representative_topics":["parametric 3d spatiotemporal filters for joint video super-resolution, denoising, and illumination enhancement","patch alignment for efficient multi-frame attention in video super-resolution transformers","super-resolution and dynamic range reconstruction from raw multi-exposure bursts using physics-based inverse optimization"]},{"direction_index":5,"label":"D5","name":"Biomedical SR","description":"Super-resolution whose primary objective is reconstructing medical or biomedical imagery while respecting modality-specific acquisition, anatomy, volumetric structure, contrast, biological texture, or clinical utility.","inclusion":"Include MRI, CT, ultrasound, OCT/OCTA, photoacoustic, microscopy, histopathology, medical volumes, multi-contrast medical fusion, unpaired medical SR, and medical-domain GAN/diffusion/transformer SR when biomedical constraints define the problem.","exclusion":"If the target is non-biomedical scientific, remote-sensing, thermal, depth, or semantic content, assign to domain-constrained non-medical SR; if a general method is merely evaluated on medical data without biomedical constraints, assign to representation architectures.","nearest_boundary":"Domain-constrained non-medical SR","representative_topics":["transformer-based volumetric super-resolution for anisotropic ct with real-paired datasets","generative adversarial network super-resolution for 2d brain mri enhancement","hybrid cnn-vision transformer super-resolution of mri with structure and texture constraints"]},{"direction_index":6,"label":"D6","name":"Domain-constrained non-medical SR","description":"Super-resolution whose primary objective is satisfying the physics, sensor geometry, spectral structure, semantic structure, or validation criteria of a non-medical specialized domain.","inclusion":"Include remote sensing, satellite, hyperspectral/multispectral, SAR, thermal, depth/time-of-flight, scientific simulations, dynamical systems, calorimetry, face, text, screen content, and other structured semantic or sensor domains when domain constraints are primary.","exclusion":"If the modality is medical or biomedical, assign to biomedical SR; if an external reference image or graphics/codec pipeline is the defining source of structure, assign to auxiliary-view and pipeline-conditioned SR; if domain shift or missing pairs is the main objective independent of domain constraints, assign to robust real-world and weak-supervision SR.","nearest_boundary":"Biomedical SR","representative_topics":["thermal image super-resolution via bilateral network with channel splitting and transformer fusion","stochastic attribute modeling for uncertainty-aware face super-resolution","real-world benchmark for multi-image super-resolution of satellite remote sensing imagery"]},{"direction_index":7,"label":"D7","name":"Auxiliary-view and pipeline-conditioned SR","description":"Super-resolution whose primary objective is exploiting external views, reference images, geometry, or production-pipeline structure such as codecs, tone mapping, rasterization, rendering buffers, or graphics metadata.","inclusion":"Include reference-based, multi-reference, stereo, light-field, multi-view, deformable correspondence or texture-transfer SR, codec-assisted or compressed-content SR, joint SR with inverse tone mapping, rasterized/game/rendered-content upscaling, and g-buffer or motion-vector guided rendering SR when these auxiliary sources or pipeline structures define the problem.","exclusion":"If temporally adjacent frames are fused primarily for video consistency, assign to temporal sequence SR; if compression is treated only as an unknown degradation without exploiting codec or pipeline structure, assign to robust real-world and weak-supervision SR; if the auxiliary source is a domain sensor whose physics dominates, assign to domain-constrained non-medical SR.","nearest_boundary":"Temporal sequence SR","representative_topics":["joint super-resolution and inverse tone-mapping via hierarchical feature decomposition aggregation","geometry-aware reference synthesis for multi-view image super-resolution with large-angle transformations","reference-based super-resolution via deformable attention and multi-scale texture feature transfer"]}],"split":"dev"}
{"episode_id":"F065_T2026-01","field_index":65,"field":"image super-resolution","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Robust real-world and weak-supervision SR","description":"Super-resolution whose primary objective is to work under unknown real-world degradations, domain shift, missing paired supervision, or to evaluate robustness under such mismatches.","inclusion":"Include atoms focused on blind degradation modeling or estimation, real-world degradation robustness, unpaired/self-supervised/source-free/domain-adaptive SR, test-time adaptation driven by mismatch, and systematic robustness or benchmark evaluation of real-world SR.","exclusion":"If the main mechanism is a generative prior for perceptual synthesis, assign to generative perceptual SR; if the defining constraint is a medical, scientific, sensor, or semantic domain, assign to domain-constrained SR; if the main goal is architecture quality under standard paired SISR, assign to representation architectures.","nearest_boundary":"Generative perceptual SR","representative_topics":["real-world single image super-resolution under rainy weather degradation","blind super-resolution via degradation-guided meta-restoration with joint consistency loss","blind super-resolution via meta-learning and implicit degradation representation extraction without groundtruth supervision"]},{"direction_index":1,"label":"D1","name":"Efficient deployable SR","description":"Super-resolution whose primary objective is reducing computation, memory, latency, energy, bit-width, or model size while preserving reconstruction quality.","inclusion":"Include lightweight networks, pruning, quantization, distillation, sparse or dynamic computation, mobile/NPU/edge deployment, and real-time SR when resource reduction is the dominant research program.","exclusion":"If efficiency is only incidental and the main contribution is feature modeling or reconstruction quality, assign to representation architectures; if temporal modeling dominates in video despite efficient components, assign to temporal sequence SR.","nearest_boundary":"Representation and scaling architectures","representative_topics":["structured sparsity pruning for efficient video super-resolution on resource-constrained devices","lightweight transformer-cnn hybrid for contextual-aware single image super-resolution","lightweight super-resolution via dual regression constraints and channel-level model compression"]},{"direction_index":2,"label":"D2","name":"Representation and scaling architectures","description":"Super-resolution whose primary objective is improving single-image reconstruction through architectures, feature representations, priors, optimization, or continuous/arbitrary-scale image representations.","inclusion":"Include general SISR architectures, attention or transformer/CNN hybrids, dictionary or sparse priors, self-similarity, frequency or wavelet modeling, loss/training designs, implicit neural representations, invertible rescaling, and arbitrary-scale or coordinate-continuous SR when the central goal is better image representation or scale generalization.","exclusion":"If reduced resource use is the explicit objective, assign to efficient deployable SR; if unknown real-world degradation or supervision scarcity defines the problem, assign to robust real-world and weak-supervision SR; if GAN/diffusion perceptual synthesis dominates, assign to generative perceptual SR.","nearest_boundary":"Efficient deployable SR","representative_topics":["patch-based super resolution via generalized gaussian mixture model and mmse reconstruction","hierarchical self-similarity learning for aliasing suppression in single image super-resolution","explicit dictionary learning for out-of-domain robustness in single-image super-resolution"]},{"direction_index":3,"label":"D3","name":"Generative perceptual SR","description":"Super-resolution whose primary objective is synthesizing perceptually plausible high-frequency detail or managing the fidelity-perception tradeoff through GAN, diffusion, flow, or other generative priors.","inclusion":"Include pretrained GAN prior inversion or latent banks, diffusion/score-based SR, conditional generative sampling, adversarially trained perceptual SR, GAN artifact handling, and real-world SR when the dominant contribution is generative synthesis rather than explicit degradation estimation.","exclusion":"If explicit degradation estimation or adaptation is the core objective, assign to robust real-world and weak-supervision SR; if medical or other domain constraints dominate over the generative mechanism, assign to the corresponding domain program.","nearest_boundary":"Robust real-world and weak-supervision SR","representative_topics":["pre-trained gan latent bank for perceptual image super-resolution without runtime inversion","gan-prior large-scale super-resolution via latent space graph-based multi-relation disentanglement and code generation","latent diffusion models for single image super-resolution with multimodal adversarial denoising"]},{"direction_index":4,"label":"D4","name":"Temporal sequence SR","description":"Super-resolution whose primary objective is recovering high-resolution images or videos from temporally related frames by modeling motion, alignment, propagation, fusion, and temporal consistency.","inclusion":"Include video SR, burst SR, recurrent or transformer sequence fusion, multi-frame alignment, temporal consistency learning, motion compensation, raw or exposure-bracketed burst reconstruction, and compressed video SR when temporal modeling is central.","exclusion":"If auxiliary views are non-temporal references or stereo/light-field geometry, assign to auxiliary-view and pipeline-conditioned SR; if resource reduction is the dominant goal, assign to efficient deployable SR.","nearest_boundary":"Auxiliary-view and pipeline-conditioned SR","representative_topics":["parametric 3d spatiotemporal filters for joint video super-resolution, denoising, and illumination enhancement","patch alignment for efficient multi-frame attention in video super-resolution transformers","super-resolution and dynamic range reconstruction from raw multi-exposure bursts using physics-based inverse optimization"]},{"direction_index":5,"label":"D5","name":"Biomedical SR","description":"Super-resolution whose primary objective is reconstructing medical or biomedical imagery while respecting modality-specific acquisition, anatomy, volumetric structure, contrast, biological texture, or clinical utility.","inclusion":"Include MRI, CT, ultrasound, OCT/OCTA, photoacoustic, microscopy, histopathology, medical volumes, multi-contrast medical fusion, unpaired medical SR, and medical-domain GAN/diffusion/transformer SR when biomedical constraints define the problem.","exclusion":"If the target is non-biomedical scientific, remote-sensing, thermal, depth, or semantic content, assign to domain-constrained non-medical SR; if a general method is merely evaluated on medical data without biomedical constraints, assign to representation architectures.","nearest_boundary":"Domain-constrained non-medical SR","representative_topics":["transformer-based volumetric super-resolution for anisotropic ct with real-paired datasets","generative adversarial network super-resolution for 2d brain mri enhancement","hybrid cnn-vision transformer super-resolution of mri with structure and texture constraints"]},{"direction_index":6,"label":"D6","name":"Domain-constrained non-medical SR","description":"Super-resolution whose primary objective is satisfying the physics, sensor geometry, spectral structure, semantic structure, or validation criteria of a non-medical specialized domain.","inclusion":"Include remote sensing, satellite, hyperspectral/multispectral, SAR, thermal, depth/time-of-flight, scientific simulations, dynamical systems, calorimetry, face, text, screen content, and other structured semantic or sensor domains when domain constraints are primary.","exclusion":"If the modality is medical or biomedical, assign to biomedical SR; if an external reference image or graphics/codec pipeline is the defining source of structure, assign to auxiliary-view and pipeline-conditioned SR; if domain shift or missing pairs is the main objective independent of domain constraints, assign to robust real-world and weak-supervision SR.","nearest_boundary":"Biomedical SR","representative_topics":["thermal image super-resolution via bilateral network with channel splitting and transformer fusion","stochastic attribute modeling for uncertainty-aware face super-resolution","real-world benchmark for multi-image super-resolution of satellite remote sensing imagery"]},{"direction_index":7,"label":"D7","name":"Auxiliary-view and pipeline-conditioned SR","description":"Super-resolution whose primary objective is exploiting external views, reference images, geometry, or production-pipeline structure such as codecs, tone mapping, rasterization, rendering buffers, or graphics metadata.","inclusion":"Include reference-based, multi-reference, stereo, light-field, multi-view, deformable correspondence or texture-transfer SR, codec-assisted or compressed-content SR, joint SR with inverse tone mapping, rasterized/game/rendered-content upscaling, and g-buffer or motion-vector guided rendering SR when these auxiliary sources or pipeline structures define the problem.","exclusion":"If temporally adjacent frames are fused primarily for video consistency, assign to temporal sequence SR; if compression is treated only as an unknown degradation without exploiting codec or pipeline structure, assign to robust real-world and weak-supervision SR; if the auxiliary source is a domain sensor whose physics dominates, assign to domain-constrained non-medical SR.","nearest_boundary":"Temporal sequence SR","representative_topics":["joint super-resolution and inverse tone-mapping via hierarchical feature decomposition aggregation","geometry-aware reference synthesis for multi-view image super-resolution with large-angle transformations","reference-based super-resolution via deformable attention and multi-scale texture feature transfer"]}],"split":"dev"}
{"episode_id":"F069_T2024-01","field_index":69,"field":"reinforcement learning from human feedback","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Foundations and field synthesis","description":"Research whose primary objective is to establish broad conceptual, theoretical, socio-technical, survey, benchmarking, or infrastructure-level understanding of RLHF as a field.","inclusion":"Include formal models and reductions for preference-based learning, convergence and sample-complexity analyses, field surveys, open-problem syntheses, broad socio-ethical analyses, and general RLHF systems or tooling papers whose contribution is field-level enablement rather than a specific optimizer, reward model, or elicitation method.","exclusion":"Concrete methods for collecting labels, representing rewards, optimizing policies, or mitigating specific failures belong to the corresponding technical direction when the main contribution is algorithmic or diagnostic rather than field-level synthesis.","nearest_boundary":"Offline and statistical preference learning","representative_topics":["regret-based reward function learning from human preferences in rlhf","social and ethical impact assessment of reinforcement learning with human feedback","surveying open problems, limitations, and auditing standards for rlhf alignment of large language models"]},{"direction_index":1,"label":"D1","name":"Preference elicitation efficiency","description":"Research whose primary objective is to obtain human preference information more efficiently, reliably, or annotator-aware before downstream RLHF optimization.","inclusion":"Include active query selection, disagreement- or uncertainty-driven feedback collection, few-shot and meta-learned preference acquisition, heterogeneous teacher expertise modeling, noisy or diverse annotator modeling, and data-efficient feedback strategies focused on reducing or improving human annotation.","exclusion":"Formal query-design theory belongs to foundational or statistical preference learning; downstream policy optimization on ordinary preference datasets belongs to direct, offline, or reward-guided policy optimization.","nearest_boundary":"Reward model design and analysis","representative_topics":["rlhf with heterogeneous teacher expertise modeling and query selection","dynamics-aware reward encoding to reduce label burden in preference-based reinforcement learning","few-shot preference learning for robot reward shaping via meta-learning adaptation"]},{"direction_index":2,"label":"D2","name":"Reward model design and analysis","description":"Research whose primary objective is to design, train, interpret, or analyze explicit reward or preference models used in RLHF pipelines.","inclusion":"Include tree-based, neuro-symbolic, argumentative, transformer, contrastive, GFlowNet-style, interpretable, structured, non-Markovian, or fine-grained reward-model representations, as well as studies of reward-model scaling, biases, generalization, and overoptimization when the reward model itself is the object of study.","exclusion":"Reward-free objectives that bypass explicit reward modeling belong to direct preference optimization; empirical mitigations of pipeline failures such as length bias or catastrophic forgetting belong to failure and robustness management when the main objective is correction rather than representation design.","nearest_boundary":"Failure and robustness management","representative_topics":["argumentative reward learning via neuro-symbolic preference-based argumentation for robust reward models","interpretable tree-based reward learning from human preference labels for alignment verification","contrastive reward model fine-tuning for preference-controlled story generation via reinforcement learning"]},{"direction_index":3,"label":"D3","name":"Direct preference optimization","description":"Research whose primary objective is to align policies directly from preference, ranking, comparison, or implicit feedback data while avoiding a separately optimized explicit reward-model-plus-online-RL loop.","inclusion":"Include DPO-style methods, f-divergence or KL-generalized preference objectives, rank-based and contrastive alignment losses, sequence likelihood calibration, implicit-reward policy updates, joint policy-reward formulations without a separate reward model, and reward-free preference optimization objectives.","exclusion":"PPO-style or reward-guided iterative RLHF belongs to reward-guided policy optimization; methods where fixed logged data, pessimism, or off-policy statistical validity is the main contribution belong to offline and statistical preference learning.","nearest_boundary":"Reward-guided policy optimization","representative_topics":["f-divergence-based distributional policy gradient for flexible language model preference alignment","ranking-based reinforcement learning from human feedback for language model alignment","sequence likelihood calibration for human feedback learning without reward model training"]},{"direction_index":4,"label":"D4","name":"Reward-guided policy optimization","description":"Research whose primary objective is to improve the iterative RLHF policy-optimization loop that fine-tunes a policy against reward, preference, or teacher-guided signals.","inclusion":"Include PPO variants, constrained or KL-regularized policy optimization, advantage-estimation methods, mode-collapse and stability improvements in the RL phase, memory-efficient PPO, teacher-guided RL, bilevel policy-alignment frameworks, Pareto weight interpolation when framed as reward-guided policy improvement, and alternative reward-guided objectives such as GFlowNet-based policy learning.","exclusion":"Preference objectives that remove the RL optimization loop belong to direct preference optimization; infrastructure-only scalable training belongs to foundations and field synthesis unless the scientific contribution is the optimizer itself.","nearest_boundary":"Direct preference optimization","representative_topics":["gflownets for reward function fitting via human feedback for policy learning","advantage-based policy gradient loss for stable rlhf fine-tuning of language models","pareto-optimal alignment via interpolating weights of diverse reward-specialized fine-tuned networks"]},{"direction_index":5,"label":"D5","name":"Failure and robustness management","description":"Research whose primary objective is to diagnose, prevent, or correct failures in RLHF feedback loops, including reward hacking, bias, forgetting, adversarial feedback, data construction failures, and robustness under conflicting or personalized objectives.","inclusion":"Include length bias analysis and mitigation, objective mismatch, reward hacking, catastrophic forgetting, continual preference alignment, selective rehearsal, hidden annotator context, reward poisoning, adversarial RM-LLM co-training, automated preference-pair construction for safer alignment, personalized or multi-objective robustness, and practical pipeline curation when framed as failure mitigation or robustness improvement.","exclusion":"New reward architectures belong to reward model design; optimizer changes aimed mainly at PPO stability belong to reward-guided policy optimization; ordinary annotator diversity for label efficiency belongs to preference elicitation efficiency.","nearest_boundary":"Reward model design and analysis","representative_topics":["customized reward model learning for personalized and domain-specific human preference alignment","stabilizing rlhf training via advantage modeling and selective rehearsal against reward hacking and catastrophic forgetting","length bias in rlhf reward models and its effect on language model alignment"]},{"direction_index":6,"label":"D6","name":"Offline and statistical preference learning","description":"Research whose primary objective is to learn, evaluate, or justify RLHF policies from fixed, logged, batch, off-policy, or statistically modeled human feedback rather than an interactive online human-feedback loop.","inclusion":"Include offline RLHF, pessimistic policy optimization from human choices, off-policy evaluation, importance-sampled preference learning, batch self-training, dynamic discrete choice modeling, statistical sample-complexity or convergence guarantees for preference learning, and formal analyses of ranking or comparison losses when the central issue is learning validity under limited or logged feedback.","exclusion":"Direct preference objectives belong to direct preference optimization when offline data are incidental; field-wide theoretical surveys belong to foundations and field synthesis; online PPO-style learning belongs to reward-guided policy optimization.","nearest_boundary":"Direct preference optimization","representative_topics":["theoretical convergence analysis of rlhf reward learning under pairwise and k-wise comparison models","theoretical optimal query design for nonparametric reward learning via rkhs bandit scaling laws","offline advantage-based policy gradient for language model alignment using pre-existing data"]},{"direction_index":7,"label":"D7","name":"Embodied and multimodal adaptation","description":"Research whose primary objective is to adapt RLHF or preference-based alignment to embodied agents, robotics, control, multimodal interaction, games, or non-text generative modalities where the domain changes the feedback, reward, or policy-learning problem.","inclusion":"Include RLHF for embodied multimodal agents, robotic manipulation, continuous control, skill discovery, safety-oriented physical behavior learning, hard-to-specify game tasks, interactive 3D environments, and text-to-image or other multimodal preference learning when the modality or embodiment is central to the research objective.","exclusion":"Generic reward-model, elicitation, or optimizer innovations that merely evaluate on an embodied or multimodal benchmark belong to their algorithmic direction; purely text-only language-model alignment belongs to direct, offline, reward-guided, or robustness directions as appropriate.","nearest_boundary":"Preference elicitation efficiency","representative_topics":["reinforcement learning from human feedback for embodied multimodal agents using inter-temporal reward modelling","preference-constrained autonomous skill discovery for diverse human-aligned behaviors","benchmarking reinforcement learning from human feedback for fine-tuning foundation models in game settings"]}],"split":"dev"}
{"episode_id":"F069_T2024-07","field_index":69,"field":"reinforcement learning from human feedback","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Foundations and field synthesis","description":"Research whose primary objective is to establish broad conceptual, theoretical, socio-technical, survey, benchmarking, or infrastructure-level understanding of RLHF as a field.","inclusion":"Include formal models and reductions for preference-based learning, convergence and sample-complexity analyses, field surveys, open-problem syntheses, broad socio-ethical analyses, and general RLHF systems or tooling papers whose contribution is field-level enablement rather than a specific optimizer, reward model, or elicitation method.","exclusion":"Concrete methods for collecting labels, representing rewards, optimizing policies, or mitigating specific failures belong to the corresponding technical direction when the main contribution is algorithmic or diagnostic rather than field-level synthesis.","nearest_boundary":"Offline and statistical preference learning","representative_topics":["regret-based reward function learning from human preferences in rlhf","social and ethical impact assessment of reinforcement learning with human feedback","surveying open problems, limitations, and auditing standards for rlhf alignment of large language models"]},{"direction_index":1,"label":"D1","name":"Preference elicitation efficiency","description":"Research whose primary objective is to obtain human preference information more efficiently, reliably, or annotator-aware before downstream RLHF optimization.","inclusion":"Include active query selection, disagreement- or uncertainty-driven feedback collection, few-shot and meta-learned preference acquisition, heterogeneous teacher expertise modeling, noisy or diverse annotator modeling, and data-efficient feedback strategies focused on reducing or improving human annotation.","exclusion":"Formal query-design theory belongs to foundational or statistical preference learning; downstream policy optimization on ordinary preference datasets belongs to direct, offline, or reward-guided policy optimization.","nearest_boundary":"Reward model design and analysis","representative_topics":["rlhf with heterogeneous teacher expertise modeling and query selection","dynamics-aware reward encoding to reduce label burden in preference-based reinforcement learning","few-shot preference learning for robot reward shaping via meta-learning adaptation"]},{"direction_index":2,"label":"D2","name":"Reward model design and analysis","description":"Research whose primary objective is to design, train, interpret, or analyze explicit reward or preference models used in RLHF pipelines.","inclusion":"Include tree-based, neuro-symbolic, argumentative, transformer, contrastive, GFlowNet-style, interpretable, structured, non-Markovian, or fine-grained reward-model representations, as well as studies of reward-model scaling, biases, generalization, and overoptimization when the reward model itself is the object of study.","exclusion":"Reward-free objectives that bypass explicit reward modeling belong to direct preference optimization; empirical mitigations of pipeline failures such as length bias or catastrophic forgetting belong to failure and robustness management when the main objective is correction rather than representation design.","nearest_boundary":"Failure and robustness management","representative_topics":["argumentative reward learning via neuro-symbolic preference-based argumentation for robust reward models","interpretable tree-based reward learning from human preference labels for alignment verification","contrastive reward model fine-tuning for preference-controlled story generation via reinforcement learning"]},{"direction_index":3,"label":"D3","name":"Direct preference optimization","description":"Research whose primary objective is to align policies directly from preference, ranking, comparison, or implicit feedback data while avoiding a separately optimized explicit reward-model-plus-online-RL loop.","inclusion":"Include DPO-style methods, f-divergence or KL-generalized preference objectives, rank-based and contrastive alignment losses, sequence likelihood calibration, implicit-reward policy updates, joint policy-reward formulations without a separate reward model, and reward-free preference optimization objectives.","exclusion":"PPO-style or reward-guided iterative RLHF belongs to reward-guided policy optimization; methods where fixed logged data, pessimism, or off-policy statistical validity is the main contribution belong to offline and statistical preference learning.","nearest_boundary":"Reward-guided policy optimization","representative_topics":["f-divergence-based distributional policy gradient for flexible language model preference alignment","ranking-based reinforcement learning from human feedback for language model alignment","sequence likelihood calibration for human feedback learning without reward model training"]},{"direction_index":4,"label":"D4","name":"Reward-guided policy optimization","description":"Research whose primary objective is to improve the iterative RLHF policy-optimization loop that fine-tunes a policy against reward, preference, or teacher-guided signals.","inclusion":"Include PPO variants, constrained or KL-regularized policy optimization, advantage-estimation methods, mode-collapse and stability improvements in the RL phase, memory-efficient PPO, teacher-guided RL, bilevel policy-alignment frameworks, Pareto weight interpolation when framed as reward-guided policy improvement, and alternative reward-guided objectives such as GFlowNet-based policy learning.","exclusion":"Preference objectives that remove the RL optimization loop belong to direct preference optimization; infrastructure-only scalable training belongs to foundations and field synthesis unless the scientific contribution is the optimizer itself.","nearest_boundary":"Direct preference optimization","representative_topics":["gflownets for reward function fitting via human feedback for policy learning","advantage-based policy gradient loss for stable rlhf fine-tuning of language models","pareto-optimal alignment via interpolating weights of diverse reward-specialized fine-tuned networks"]},{"direction_index":5,"label":"D5","name":"Failure and robustness management","description":"Research whose primary objective is to diagnose, prevent, or correct failures in RLHF feedback loops, including reward hacking, bias, forgetting, adversarial feedback, data construction failures, and robustness under conflicting or personalized objectives.","inclusion":"Include length bias analysis and mitigation, objective mismatch, reward hacking, catastrophic forgetting, continual preference alignment, selective rehearsal, hidden annotator context, reward poisoning, adversarial RM-LLM co-training, automated preference-pair construction for safer alignment, personalized or multi-objective robustness, and practical pipeline curation when framed as failure mitigation or robustness improvement.","exclusion":"New reward architectures belong to reward model design; optimizer changes aimed mainly at PPO stability belong to reward-guided policy optimization; ordinary annotator diversity for label efficiency belongs to preference elicitation efficiency.","nearest_boundary":"Reward model design and analysis","representative_topics":["customized reward model learning for personalized and domain-specific human preference alignment","stabilizing rlhf training via advantage modeling and selective rehearsal against reward hacking and catastrophic forgetting","length bias in rlhf reward models and its effect on language model alignment"]},{"direction_index":6,"label":"D6","name":"Offline and statistical preference learning","description":"Research whose primary objective is to learn, evaluate, or justify RLHF policies from fixed, logged, batch, off-policy, or statistically modeled human feedback rather than an interactive online human-feedback loop.","inclusion":"Include offline RLHF, pessimistic policy optimization from human choices, off-policy evaluation, importance-sampled preference learning, batch self-training, dynamic discrete choice modeling, statistical sample-complexity or convergence guarantees for preference learning, and formal analyses of ranking or comparison losses when the central issue is learning validity under limited or logged feedback.","exclusion":"Direct preference objectives belong to direct preference optimization when offline data are incidental; field-wide theoretical surveys belong to foundations and field synthesis; online PPO-style learning belongs to reward-guided policy optimization.","nearest_boundary":"Direct preference optimization","representative_topics":["theoretical convergence analysis of rlhf reward learning under pairwise and k-wise comparison models","theoretical optimal query design for nonparametric reward learning via rkhs bandit scaling laws","offline advantage-based policy gradient for language model alignment using pre-existing data"]},{"direction_index":7,"label":"D7","name":"Embodied and multimodal adaptation","description":"Research whose primary objective is to adapt RLHF or preference-based alignment to embodied agents, robotics, control, multimodal interaction, games, or non-text generative modalities where the domain changes the feedback, reward, or policy-learning problem.","inclusion":"Include RLHF for embodied multimodal agents, robotic manipulation, continuous control, skill discovery, safety-oriented physical behavior learning, hard-to-specify game tasks, interactive 3D environments, and text-to-image or other multimodal preference learning when the modality or embodiment is central to the research objective.","exclusion":"Generic reward-model, elicitation, or optimizer innovations that merely evaluate on an embodied or multimodal benchmark belong to their algorithmic direction; purely text-only language-model alignment belongs to direct, offline, reward-guided, or robustness directions as appropriate.","nearest_boundary":"Preference elicitation efficiency","representative_topics":["reinforcement learning from human feedback for embodied multimodal agents using inter-temporal reward modelling","preference-constrained autonomous skill discovery for diverse human-aligned behaviors","benchmarking reinforcement learning from human feedback for fine-tuning foundation models in game settings"]}],"split":"dev"}
{"episode_id":"F069_T2025-01","field_index":69,"field":"reinforcement learning from human feedback","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Foundations and field synthesis","description":"Research whose primary objective is to establish broad conceptual, theoretical, socio-technical, survey, benchmarking, or infrastructure-level understanding of RLHF as a field.","inclusion":"Include formal models and reductions for preference-based learning, convergence and sample-complexity analyses, field surveys, open-problem syntheses, broad socio-ethical analyses, and general RLHF systems or tooling papers whose contribution is field-level enablement rather than a specific optimizer, reward model, or elicitation method.","exclusion":"Concrete methods for collecting labels, representing rewards, optimizing policies, or mitigating specific failures belong to the corresponding technical direction when the main contribution is algorithmic or diagnostic rather than field-level synthesis.","nearest_boundary":"Offline and statistical preference learning","representative_topics":["regret-based reward function learning from human preferences in rlhf","social and ethical impact assessment of reinforcement learning with human feedback","surveying open problems, limitations, and auditing standards for rlhf alignment of large language models"]},{"direction_index":1,"label":"D1","name":"Preference elicitation efficiency","description":"Research whose primary objective is to obtain human preference information more efficiently, reliably, or annotator-aware before downstream RLHF optimization.","inclusion":"Include active query selection, disagreement- or uncertainty-driven feedback collection, few-shot and meta-learned preference acquisition, heterogeneous teacher expertise modeling, noisy or diverse annotator modeling, and data-efficient feedback strategies focused on reducing or improving human annotation.","exclusion":"Formal query-design theory belongs to foundational or statistical preference learning; downstream policy optimization on ordinary preference datasets belongs to direct, offline, or reward-guided policy optimization.","nearest_boundary":"Reward model design and analysis","representative_topics":["rlhf with heterogeneous teacher expertise modeling and query selection","dynamics-aware reward encoding to reduce label burden in preference-based reinforcement learning","few-shot preference learning for robot reward shaping via meta-learning adaptation"]},{"direction_index":2,"label":"D2","name":"Reward model design and analysis","description":"Research whose primary objective is to design, train, interpret, or analyze explicit reward or preference models used in RLHF pipelines.","inclusion":"Include tree-based, neuro-symbolic, argumentative, transformer, contrastive, GFlowNet-style, interpretable, structured, non-Markovian, or fine-grained reward-model representations, as well as studies of reward-model scaling, biases, generalization, and overoptimization when the reward model itself is the object of study.","exclusion":"Reward-free objectives that bypass explicit reward modeling belong to direct preference optimization; empirical mitigations of pipeline failures such as length bias or catastrophic forgetting belong to failure and robustness management when the main objective is correction rather than representation design.","nearest_boundary":"Failure and robustness management","representative_topics":["argumentative reward learning via neuro-symbolic preference-based argumentation for robust reward models","interpretable tree-based reward learning from human preference labels for alignment verification","contrastive reward model fine-tuning for preference-controlled story generation via reinforcement learning"]},{"direction_index":3,"label":"D3","name":"Direct preference optimization","description":"Research whose primary objective is to align policies directly from preference, ranking, comparison, or implicit feedback data while avoiding a separately optimized explicit reward-model-plus-online-RL loop.","inclusion":"Include DPO-style methods, f-divergence or KL-generalized preference objectives, rank-based and contrastive alignment losses, sequence likelihood calibration, implicit-reward policy updates, joint policy-reward formulations without a separate reward model, and reward-free preference optimization objectives.","exclusion":"PPO-style or reward-guided iterative RLHF belongs to reward-guided policy optimization; methods where fixed logged data, pessimism, or off-policy statistical validity is the main contribution belong to offline and statistical preference learning.","nearest_boundary":"Reward-guided policy optimization","representative_topics":["f-divergence-based distributional policy gradient for flexible language model preference alignment","ranking-based reinforcement learning from human feedback for language model alignment","sequence likelihood calibration for human feedback learning without reward model training"]},{"direction_index":4,"label":"D4","name":"Reward-guided policy optimization","description":"Research whose primary objective is to improve the iterative RLHF policy-optimization loop that fine-tunes a policy against reward, preference, or teacher-guided signals.","inclusion":"Include PPO variants, constrained or KL-regularized policy optimization, advantage-estimation methods, mode-collapse and stability improvements in the RL phase, memory-efficient PPO, teacher-guided RL, bilevel policy-alignment frameworks, Pareto weight interpolation when framed as reward-guided policy improvement, and alternative reward-guided objectives such as GFlowNet-based policy learning.","exclusion":"Preference objectives that remove the RL optimization loop belong to direct preference optimization; infrastructure-only scalable training belongs to foundations and field synthesis unless the scientific contribution is the optimizer itself.","nearest_boundary":"Direct preference optimization","representative_topics":["gflownets for reward function fitting via human feedback for policy learning","advantage-based policy gradient loss for stable rlhf fine-tuning of language models","pareto-optimal alignment via interpolating weights of diverse reward-specialized fine-tuned networks"]},{"direction_index":5,"label":"D5","name":"Failure and robustness management","description":"Research whose primary objective is to diagnose, prevent, or correct failures in RLHF feedback loops, including reward hacking, bias, forgetting, adversarial feedback, data construction failures, and robustness under conflicting or personalized objectives.","inclusion":"Include length bias analysis and mitigation, objective mismatch, reward hacking, catastrophic forgetting, continual preference alignment, selective rehearsal, hidden annotator context, reward poisoning, adversarial RM-LLM co-training, automated preference-pair construction for safer alignment, personalized or multi-objective robustness, and practical pipeline curation when framed as failure mitigation or robustness improvement.","exclusion":"New reward architectures belong to reward model design; optimizer changes aimed mainly at PPO stability belong to reward-guided policy optimization; ordinary annotator diversity for label efficiency belongs to preference elicitation efficiency.","nearest_boundary":"Reward model design and analysis","representative_topics":["customized reward model learning for personalized and domain-specific human preference alignment","stabilizing rlhf training via advantage modeling and selective rehearsal against reward hacking and catastrophic forgetting","length bias in rlhf reward models and its effect on language model alignment"]},{"direction_index":6,"label":"D6","name":"Offline and statistical preference learning","description":"Research whose primary objective is to learn, evaluate, or justify RLHF policies from fixed, logged, batch, off-policy, or statistically modeled human feedback rather than an interactive online human-feedback loop.","inclusion":"Include offline RLHF, pessimistic policy optimization from human choices, off-policy evaluation, importance-sampled preference learning, batch self-training, dynamic discrete choice modeling, statistical sample-complexity or convergence guarantees for preference learning, and formal analyses of ranking or comparison losses when the central issue is learning validity under limited or logged feedback.","exclusion":"Direct preference objectives belong to direct preference optimization when offline data are incidental; field-wide theoretical surveys belong to foundations and field synthesis; online PPO-style learning belongs to reward-guided policy optimization.","nearest_boundary":"Direct preference optimization","representative_topics":["theoretical convergence analysis of rlhf reward learning under pairwise and k-wise comparison models","theoretical optimal query design for nonparametric reward learning via rkhs bandit scaling laws","offline advantage-based policy gradient for language model alignment using pre-existing data"]},{"direction_index":7,"label":"D7","name":"Embodied and multimodal adaptation","description":"Research whose primary objective is to adapt RLHF or preference-based alignment to embodied agents, robotics, control, multimodal interaction, games, or non-text generative modalities where the domain changes the feedback, reward, or policy-learning problem.","inclusion":"Include RLHF for embodied multimodal agents, robotic manipulation, continuous control, skill discovery, safety-oriented physical behavior learning, hard-to-specify game tasks, interactive 3D environments, and text-to-image or other multimodal preference learning when the modality or embodiment is central to the research objective.","exclusion":"Generic reward-model, elicitation, or optimizer innovations that merely evaluate on an embodied or multimodal benchmark belong to their algorithmic direction; purely text-only language-model alignment belongs to direct, offline, reward-guided, or robustness directions as appropriate.","nearest_boundary":"Preference elicitation efficiency","representative_topics":["reinforcement learning from human feedback for embodied multimodal agents using inter-temporal reward modelling","preference-constrained autonomous skill discovery for diverse human-aligned behaviors","benchmarking reinforcement learning from human feedback for fine-tuning foundation models in game settings"]}],"split":"dev"}
{"episode_id":"F069_T2025-07","field_index":69,"field":"reinforcement learning from human feedback","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Foundations and field synthesis","description":"Research whose primary objective is to establish broad conceptual, theoretical, socio-technical, survey, benchmarking, or infrastructure-level understanding of RLHF as a field.","inclusion":"Include formal models and reductions for preference-based learning, convergence and sample-complexity analyses, field surveys, open-problem syntheses, broad socio-ethical analyses, and general RLHF systems or tooling papers whose contribution is field-level enablement rather than a specific optimizer, reward model, or elicitation method.","exclusion":"Concrete methods for collecting labels, representing rewards, optimizing policies, or mitigating specific failures belong to the corresponding technical direction when the main contribution is algorithmic or diagnostic rather than field-level synthesis.","nearest_boundary":"Offline and statistical preference learning","representative_topics":["regret-based reward function learning from human preferences in rlhf","social and ethical impact assessment of reinforcement learning with human feedback","surveying open problems, limitations, and auditing standards for rlhf alignment of large language models"]},{"direction_index":1,"label":"D1","name":"Preference elicitation efficiency","description":"Research whose primary objective is to obtain human preference information more efficiently, reliably, or annotator-aware before downstream RLHF optimization.","inclusion":"Include active query selection, disagreement- or uncertainty-driven feedback collection, few-shot and meta-learned preference acquisition, heterogeneous teacher expertise modeling, noisy or diverse annotator modeling, and data-efficient feedback strategies focused on reducing or improving human annotation.","exclusion":"Formal query-design theory belongs to foundational or statistical preference learning; downstream policy optimization on ordinary preference datasets belongs to direct, offline, or reward-guided policy optimization.","nearest_boundary":"Reward model design and analysis","representative_topics":["rlhf with heterogeneous teacher expertise modeling and query selection","dynamics-aware reward encoding to reduce label burden in preference-based reinforcement learning","few-shot preference learning for robot reward shaping via meta-learning adaptation"]},{"direction_index":2,"label":"D2","name":"Reward model design and analysis","description":"Research whose primary objective is to design, train, interpret, or analyze explicit reward or preference models used in RLHF pipelines.","inclusion":"Include tree-based, neuro-symbolic, argumentative, transformer, contrastive, GFlowNet-style, interpretable, structured, non-Markovian, or fine-grained reward-model representations, as well as studies of reward-model scaling, biases, generalization, and overoptimization when the reward model itself is the object of study.","exclusion":"Reward-free objectives that bypass explicit reward modeling belong to direct preference optimization; empirical mitigations of pipeline failures such as length bias or catastrophic forgetting belong to failure and robustness management when the main objective is correction rather than representation design.","nearest_boundary":"Failure and robustness management","representative_topics":["argumentative reward learning via neuro-symbolic preference-based argumentation for robust reward models","interpretable tree-based reward learning from human preference labels for alignment verification","contrastive reward model fine-tuning for preference-controlled story generation via reinforcement learning"]},{"direction_index":3,"label":"D3","name":"Direct preference optimization","description":"Research whose primary objective is to align policies directly from preference, ranking, comparison, or implicit feedback data while avoiding a separately optimized explicit reward-model-plus-online-RL loop.","inclusion":"Include DPO-style methods, f-divergence or KL-generalized preference objectives, rank-based and contrastive alignment losses, sequence likelihood calibration, implicit-reward policy updates, joint policy-reward formulations without a separate reward model, and reward-free preference optimization objectives.","exclusion":"PPO-style or reward-guided iterative RLHF belongs to reward-guided policy optimization; methods where fixed logged data, pessimism, or off-policy statistical validity is the main contribution belong to offline and statistical preference learning.","nearest_boundary":"Reward-guided policy optimization","representative_topics":["f-divergence-based distributional policy gradient for flexible language model preference alignment","ranking-based reinforcement learning from human feedback for language model alignment","sequence likelihood calibration for human feedback learning without reward model training"]},{"direction_index":4,"label":"D4","name":"Reward-guided policy optimization","description":"Research whose primary objective is to improve the iterative RLHF policy-optimization loop that fine-tunes a policy against reward, preference, or teacher-guided signals.","inclusion":"Include PPO variants, constrained or KL-regularized policy optimization, advantage-estimation methods, mode-collapse and stability improvements in the RL phase, memory-efficient PPO, teacher-guided RL, bilevel policy-alignment frameworks, Pareto weight interpolation when framed as reward-guided policy improvement, and alternative reward-guided objectives such as GFlowNet-based policy learning.","exclusion":"Preference objectives that remove the RL optimization loop belong to direct preference optimization; infrastructure-only scalable training belongs to foundations and field synthesis unless the scientific contribution is the optimizer itself.","nearest_boundary":"Direct preference optimization","representative_topics":["gflownets for reward function fitting via human feedback for policy learning","advantage-based policy gradient loss for stable rlhf fine-tuning of language models","pareto-optimal alignment via interpolating weights of diverse reward-specialized fine-tuned networks"]},{"direction_index":5,"label":"D5","name":"Failure and robustness management","description":"Research whose primary objective is to diagnose, prevent, or correct failures in RLHF feedback loops, including reward hacking, bias, forgetting, adversarial feedback, data construction failures, and robustness under conflicting or personalized objectives.","inclusion":"Include length bias analysis and mitigation, objective mismatch, reward hacking, catastrophic forgetting, continual preference alignment, selective rehearsal, hidden annotator context, reward poisoning, adversarial RM-LLM co-training, automated preference-pair construction for safer alignment, personalized or multi-objective robustness, and practical pipeline curation when framed as failure mitigation or robustness improvement.","exclusion":"New reward architectures belong to reward model design; optimizer changes aimed mainly at PPO stability belong to reward-guided policy optimization; ordinary annotator diversity for label efficiency belongs to preference elicitation efficiency.","nearest_boundary":"Reward model design and analysis","representative_topics":["customized reward model learning for personalized and domain-specific human preference alignment","stabilizing rlhf training via advantage modeling and selective rehearsal against reward hacking and catastrophic forgetting","length bias in rlhf reward models and its effect on language model alignment"]},{"direction_index":6,"label":"D6","name":"Offline and statistical preference learning","description":"Research whose primary objective is to learn, evaluate, or justify RLHF policies from fixed, logged, batch, off-policy, or statistically modeled human feedback rather than an interactive online human-feedback loop.","inclusion":"Include offline RLHF, pessimistic policy optimization from human choices, off-policy evaluation, importance-sampled preference learning, batch self-training, dynamic discrete choice modeling, statistical sample-complexity or convergence guarantees for preference learning, and formal analyses of ranking or comparison losses when the central issue is learning validity under limited or logged feedback.","exclusion":"Direct preference objectives belong to direct preference optimization when offline data are incidental; field-wide theoretical surveys belong to foundations and field synthesis; online PPO-style learning belongs to reward-guided policy optimization.","nearest_boundary":"Direct preference optimization","representative_topics":["theoretical convergence analysis of rlhf reward learning under pairwise and k-wise comparison models","theoretical optimal query design for nonparametric reward learning via rkhs bandit scaling laws","offline advantage-based policy gradient for language model alignment using pre-existing data"]},{"direction_index":7,"label":"D7","name":"Embodied and multimodal adaptation","description":"Research whose primary objective is to adapt RLHF or preference-based alignment to embodied agents, robotics, control, multimodal interaction, games, or non-text generative modalities where the domain changes the feedback, reward, or policy-learning problem.","inclusion":"Include RLHF for embodied multimodal agents, robotic manipulation, continuous control, skill discovery, safety-oriented physical behavior learning, hard-to-specify game tasks, interactive 3D environments, and text-to-image or other multimodal preference learning when the modality or embodiment is central to the research objective.","exclusion":"Generic reward-model, elicitation, or optimizer innovations that merely evaluate on an embodied or multimodal benchmark belong to their algorithmic direction; purely text-only language-model alignment belongs to direct, offline, reward-guided, or robustness directions as appropriate.","nearest_boundary":"Preference elicitation efficiency","representative_topics":["reinforcement learning from human feedback for embodied multimodal agents using inter-temporal reward modelling","preference-constrained autonomous skill discovery for diverse human-aligned behaviors","benchmarking reinforcement learning from human feedback for fine-tuning foundation models in game settings"]}],"split":"dev"}
{"episode_id":"F069_T2026-01","field_index":69,"field":"reinforcement learning from human feedback","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Foundations and field synthesis","description":"Research whose primary objective is to establish broad conceptual, theoretical, socio-technical, survey, benchmarking, or infrastructure-level understanding of RLHF as a field.","inclusion":"Include formal models and reductions for preference-based learning, convergence and sample-complexity analyses, field surveys, open-problem syntheses, broad socio-ethical analyses, and general RLHF systems or tooling papers whose contribution is field-level enablement rather than a specific optimizer, reward model, or elicitation method.","exclusion":"Concrete methods for collecting labels, representing rewards, optimizing policies, or mitigating specific failures belong to the corresponding technical direction when the main contribution is algorithmic or diagnostic rather than field-level synthesis.","nearest_boundary":"Offline and statistical preference learning","representative_topics":["regret-based reward function learning from human preferences in rlhf","social and ethical impact assessment of reinforcement learning with human feedback","surveying open problems, limitations, and auditing standards for rlhf alignment of large language models"]},{"direction_index":1,"label":"D1","name":"Preference elicitation efficiency","description":"Research whose primary objective is to obtain human preference information more efficiently, reliably, or annotator-aware before downstream RLHF optimization.","inclusion":"Include active query selection, disagreement- or uncertainty-driven feedback collection, few-shot and meta-learned preference acquisition, heterogeneous teacher expertise modeling, noisy or diverse annotator modeling, and data-efficient feedback strategies focused on reducing or improving human annotation.","exclusion":"Formal query-design theory belongs to foundational or statistical preference learning; downstream policy optimization on ordinary preference datasets belongs to direct, offline, or reward-guided policy optimization.","nearest_boundary":"Reward model design and analysis","representative_topics":["rlhf with heterogeneous teacher expertise modeling and query selection","dynamics-aware reward encoding to reduce label burden in preference-based reinforcement learning","few-shot preference learning for robot reward shaping via meta-learning adaptation"]},{"direction_index":2,"label":"D2","name":"Reward model design and analysis","description":"Research whose primary objective is to design, train, interpret, or analyze explicit reward or preference models used in RLHF pipelines.","inclusion":"Include tree-based, neuro-symbolic, argumentative, transformer, contrastive, GFlowNet-style, interpretable, structured, non-Markovian, or fine-grained reward-model representations, as well as studies of reward-model scaling, biases, generalization, and overoptimization when the reward model itself is the object of study.","exclusion":"Reward-free objectives that bypass explicit reward modeling belong to direct preference optimization; empirical mitigations of pipeline failures such as length bias or catastrophic forgetting belong to failure and robustness management when the main objective is correction rather than representation design.","nearest_boundary":"Failure and robustness management","representative_topics":["argumentative reward learning via neuro-symbolic preference-based argumentation for robust reward models","interpretable tree-based reward learning from human preference labels for alignment verification","contrastive reward model fine-tuning for preference-controlled story generation via reinforcement learning"]},{"direction_index":3,"label":"D3","name":"Direct preference optimization","description":"Research whose primary objective is to align policies directly from preference, ranking, comparison, or implicit feedback data while avoiding a separately optimized explicit reward-model-plus-online-RL loop.","inclusion":"Include DPO-style methods, f-divergence or KL-generalized preference objectives, rank-based and contrastive alignment losses, sequence likelihood calibration, implicit-reward policy updates, joint policy-reward formulations without a separate reward model, and reward-free preference optimization objectives.","exclusion":"PPO-style or reward-guided iterative RLHF belongs to reward-guided policy optimization; methods where fixed logged data, pessimism, or off-policy statistical validity is the main contribution belong to offline and statistical preference learning.","nearest_boundary":"Reward-guided policy optimization","representative_topics":["f-divergence-based distributional policy gradient for flexible language model preference alignment","ranking-based reinforcement learning from human feedback for language model alignment","sequence likelihood calibration for human feedback learning without reward model training"]},{"direction_index":4,"label":"D4","name":"Reward-guided policy optimization","description":"Research whose primary objective is to improve the iterative RLHF policy-optimization loop that fine-tunes a policy against reward, preference, or teacher-guided signals.","inclusion":"Include PPO variants, constrained or KL-regularized policy optimization, advantage-estimation methods, mode-collapse and stability improvements in the RL phase, memory-efficient PPO, teacher-guided RL, bilevel policy-alignment frameworks, Pareto weight interpolation when framed as reward-guided policy improvement, and alternative reward-guided objectives such as GFlowNet-based policy learning.","exclusion":"Preference objectives that remove the RL optimization loop belong to direct preference optimization; infrastructure-only scalable training belongs to foundations and field synthesis unless the scientific contribution is the optimizer itself.","nearest_boundary":"Direct preference optimization","representative_topics":["gflownets for reward function fitting via human feedback for policy learning","advantage-based policy gradient loss for stable rlhf fine-tuning of language models","pareto-optimal alignment via interpolating weights of diverse reward-specialized fine-tuned networks"]},{"direction_index":5,"label":"D5","name":"Failure and robustness management","description":"Research whose primary objective is to diagnose, prevent, or correct failures in RLHF feedback loops, including reward hacking, bias, forgetting, adversarial feedback, data construction failures, and robustness under conflicting or personalized objectives.","inclusion":"Include length bias analysis and mitigation, objective mismatch, reward hacking, catastrophic forgetting, continual preference alignment, selective rehearsal, hidden annotator context, reward poisoning, adversarial RM-LLM co-training, automated preference-pair construction for safer alignment, personalized or multi-objective robustness, and practical pipeline curation when framed as failure mitigation or robustness improvement.","exclusion":"New reward architectures belong to reward model design; optimizer changes aimed mainly at PPO stability belong to reward-guided policy optimization; ordinary annotator diversity for label efficiency belongs to preference elicitation efficiency.","nearest_boundary":"Reward model design and analysis","representative_topics":["customized reward model learning for personalized and domain-specific human preference alignment","stabilizing rlhf training via advantage modeling and selective rehearsal against reward hacking and catastrophic forgetting","length bias in rlhf reward models and its effect on language model alignment"]},{"direction_index":6,"label":"D6","name":"Offline and statistical preference learning","description":"Research whose primary objective is to learn, evaluate, or justify RLHF policies from fixed, logged, batch, off-policy, or statistically modeled human feedback rather than an interactive online human-feedback loop.","inclusion":"Include offline RLHF, pessimistic policy optimization from human choices, off-policy evaluation, importance-sampled preference learning, batch self-training, dynamic discrete choice modeling, statistical sample-complexity or convergence guarantees for preference learning, and formal analyses of ranking or comparison losses when the central issue is learning validity under limited or logged feedback.","exclusion":"Direct preference objectives belong to direct preference optimization when offline data are incidental; field-wide theoretical surveys belong to foundations and field synthesis; online PPO-style learning belongs to reward-guided policy optimization.","nearest_boundary":"Direct preference optimization","representative_topics":["theoretical convergence analysis of rlhf reward learning under pairwise and k-wise comparison models","theoretical optimal query design for nonparametric reward learning via rkhs bandit scaling laws","offline advantage-based policy gradient for language model alignment using pre-existing data"]},{"direction_index":7,"label":"D7","name":"Embodied and multimodal adaptation","description":"Research whose primary objective is to adapt RLHF or preference-based alignment to embodied agents, robotics, control, multimodal interaction, games, or non-text generative modalities where the domain changes the feedback, reward, or policy-learning problem.","inclusion":"Include RLHF for embodied multimodal agents, robotic manipulation, continuous control, skill discovery, safety-oriented physical behavior learning, hard-to-specify game tasks, interactive 3D environments, and text-to-image or other multimodal preference learning when the modality or embodiment is central to the research objective.","exclusion":"Generic reward-model, elicitation, or optimizer innovations that merely evaluate on an embodied or multimodal benchmark belong to their algorithmic direction; purely text-only language-model alignment belongs to direct, offline, reward-guided, or robustness directions as appropriate.","nearest_boundary":"Preference elicitation efficiency","representative_topics":["reinforcement learning from human feedback for embodied multimodal agents using inter-temporal reward modelling","preference-constrained autonomous skill discovery for diverse human-aligned behaviors","benchmarking reinforcement learning from human feedback for fine-tuning foundation models in game settings"]}],"split":"dev"}
{"episode_id":"F081_T2024-01","field_index":81,"field":"robot learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Sim-to-real deployment and evaluation infrastructure","description":"Develop transfer mechanisms, simulators, benchmarks, datasets, and evaluation protocols that make robot-learning systems trainable, comparable, scalable, and deployable across simulated and physical settings.","inclusion":"Include work whose primary objective is closing or measuring the simulation-reality gap, building scalable simulation/training infrastructure, standardizing benchmarks, generating tasks or environments, or creating evaluation frameworks for robot learning.","exclusion":"If simulation, benchmarks, or datasets are used mainly as tools for a specific policy-learning, navigation, manipulation, locomotion, or planning objective, assign to that objective instead.","nearest_boundary":"Generalizable representations and cross-domain priors","representative_topics":["real-to-sim policy adaptation for visual reinforcement learning insertion tasks","hybrid offline-online reinforcement learning with dynamics-aware policy evaluation for sim-to-real transfer","benchmarking offline and online adaptive domain randomization for sim-to-real robot policy transfer"]},{"direction_index":1,"label":"D1","name":"Objective discovery, exploration, and constraints","description":"Specify what robots should optimize and regulate how they discover or satisfy objectives under sparse feedback, human preferences, intrinsic motivation, risk, or explicit constraints.","inclusion":"Include reward learning, preference learning, inverse or adversarial reward inference, self-supervised goal discovery, intrinsic motivation, sparse-reward exploration, constrained RL, safe imitation, barrier or Lyapunov methods, risk prediction, and formal or physical constraint satisfaction when the objective or constraint structure is primary.","exclusion":"If demonstrations are used primarily to clone actions, use data-driven policy acquisition; if safety or reward mechanisms are incidental within a concrete navigation, manipulation, locomotion, or planning task, use that task program.","nearest_boundary":"Data-driven policy acquisition and RL algorithms","representative_topics":["heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations","scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","generative adversarial reward learning from partial demonstrations for agile robot skill acquisition"]},{"direction_index":2,"label":"D2","name":"Data-driven policy acquisition and RL algorithms","description":"Learn robot policies from demonstrations, fixed datasets, generated trajectories, or general-purpose policy-optimization improvements when the central objective is acquiring or improving policies from data rather than solving a specific embodied capability.","inclusion":"Include imitation learning, behavioral cloning, offline RL, offline-to-online adaptation, diffusion or other direct action-generation policies, generic continuous-control RL improvements, policy optimization, architecture and interpretability methods, and meta-methods for robotic control.","exclusion":"If the main contribution is reward recovery, safety enforcement, learned dynamics for planning, task hierarchy, or a concrete robot capability such as locomotion or manipulation, assign to that more specific program.","nearest_boundary":"Objective discovery, exploration, and constraints","representative_topics":["demonstration-based and rule-based 3d object manipulation policy learning for robotic arms","latent policy learning for stable adversarial imitation in high-dimensional robot tasks","offline prior data adaptation for autonomous robotic reinforcement learning and environment reset"]},{"direction_index":3,"label":"D3","name":"Predictive, hierarchical, and foundation-model planning","description":"Enable long-horizon robot behavior by learning predictive models, abstractions, skills, symbolic operators, or foundation-model reasoning modules for planning and control.","inclusion":"Include world models, learned dynamics, differentiable physics, MPC with learned models, visual foresight, hierarchical RL, skill discovery and composition, task-and-motion planning with learned operators, language or vision-language planning, and foundation-model embodied reasoning when planning or prediction is primary.","exclusion":"If a language or model component mainly supplies rewards, use objective discovery; if hierarchy is only a controller architecture for locomotion or manipulation, use the corresponding capability program.","nearest_boundary":"Data-driven policy acquisition and RL algorithms","representative_topics":["neurosymbolic vae with symbolic operators for autonomous battery disassembly task and motion planning","physics-constrained temporal convolutional learning of quadrotor dynamics for model predictive control","world model-based planning for online robot learning without simulators"]},{"direction_index":4,"label":"D4","name":"Navigation, exploration, and multi-agent coordination","description":"Learn mobile, embodied, human-interactive, or multi-robot behaviors for reaching goals, gathering spatial information, allocating tasks, coordinating with others, and operating in unknown or dynamic environments.","inclusion":"Include goal-directed navigation, object-goal or semantic navigation, visual and topological navigation, active sensing and exploration, help-seeking navigation, coverage, multi-robot task allocation, swarms, fleet learning, human-robot coordination, and collaborative navigation when spatial decision-making or coordination is primary.","exclusion":"If the main issue is body-level gait or flight stabilization, use locomotion and platform-specific control; if the main issue is LLM/VLM reasoning rather than navigation or coordination, use predictive, hierarchical, and foundation-model planning.","nearest_boundary":"Locomotion and platform-specific control","representative_topics":["monocular rgb-based autonomous exploration with directional pixel encoding for unexplored area coverage","decomposed reinforcement learning for mobile manipulation navigation in dynamic environments","uncertainty-driven help-asking in embodied navigation with intermittent feedback"]},{"direction_index":5,"label":"D5","name":"Locomotion and platform-specific control","description":"Learn movement and control policies for robot bodies or specialized platforms where morphology, actuation, balance, terrain interaction, or platform physics is the primary scientific challenge.","inclusion":"Include legged, bipedal, humanoid, quadruped, aerial, balancing, soft, musculoskeletal, space, surgical, endoscopic, and other non-standard platform control when the focus is body motion, actuation, or domain-specific physical control.","exclusion":"If the robot is moving primarily to reach semantic or spatial goals, use navigation and coordination; if the platform is only an evaluation setting for a general RL algorithm, use data-driven policy acquisition and RL algorithms.","nearest_boundary":"Navigation, exploration, and multi-agent coordination","representative_topics":["open-ended learning with neuroevolution-generated terrains for complex robot locomotion skill acquisition","semantics-aware quadrupedal locomotion skill learning from human demonstration for off-road terrain","model-free reinforcement learning for fast quadrupedal locomotion with symmetry-aware mirror-world control"]},{"direction_index":6,"label":"D6","name":"Manipulation skill control","description":"Acquire closed-loop manipulation behaviors for rigid, dexterous, contact-rich, deformable, soft, cloth, cable, food, or other physically interactive object manipulation tasks.","inclusion":"Include reinforcement learning or control learning for grasping, insertion, pushing, reorientation, bimanual and dexterous hands, tactile or force-guided manipulation, free-floating manipulation, deformable-object manipulation, cloth and garment handling, visuo-haptic food handling, and contact-rich skill acquisition when execution of the manipulation skill is primary.","exclusion":"If the main contribution is an object, affordance, visual, or video representation rather than the closed-loop skill, use generalizable representations and cross-domain priors; if the contribution is long-horizon symbolic sequencing of skills, use predictive, hierarchical, and foundation-model planning.","nearest_boundary":"Generalizable representations and cross-domain priors","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","reinforcement learning benchmark for bimanual dexterous manipulation across skill levels and task settings","reinforcement learning motion planning for free-float dual-arm manipulators tracking non-cooperative spinning objects"]},{"direction_index":7,"label":"D7","name":"Generalizable representations and cross-domain priors","description":"Learn reusable visual, geometric, affordance, semantic, video-derived, object-centric, or cross-embodiment representations that enable policies to generalize across objects, scenes, tasks, viewpoints, hands, robots, and embodiments.","inclusion":"Include object-centric manipulation representations, grasp and affordance perception, implicit geometry, neural fields, equivariant descriptors, human-video affordance extraction, visual or video pretraining, multimodal priors, diffusion-generated experience for representation or data scaling, cross-task transfer, cross-robot transfer, morphology-aware generalization, and broad representation learning when generalization is primary.","exclusion":"If the representation is mainly a learned dynamics model for planning, use predictive, hierarchical, and foundation-model planning; if it is only an implementation detail inside a specific manipulation or locomotion controller, use that capability program.","nearest_boundary":"Manipulation skill control","representative_topics":["se(3)-equivariant energy-based models for sample-efficient end-to-end visual robotic manipulation","dual-gripper collaborative affordance learning via disentangled subtask decomposition","implicit representation learning for cross-gripper grasp transfer via signed distance functions"]}],"split":"dev"}
{"episode_id":"F081_T2024-07","field_index":81,"field":"robot learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Sim-to-real deployment and evaluation infrastructure","description":"Develop transfer mechanisms, simulators, benchmarks, datasets, and evaluation protocols that make robot-learning systems trainable, comparable, scalable, and deployable across simulated and physical settings.","inclusion":"Include work whose primary objective is closing or measuring the simulation-reality gap, building scalable simulation/training infrastructure, standardizing benchmarks, generating tasks or environments, or creating evaluation frameworks for robot learning.","exclusion":"If simulation, benchmarks, or datasets are used mainly as tools for a specific policy-learning, navigation, manipulation, locomotion, or planning objective, assign to that objective instead.","nearest_boundary":"Generalizable representations and cross-domain priors","representative_topics":["real-to-sim policy adaptation for visual reinforcement learning insertion tasks","hybrid offline-online reinforcement learning with dynamics-aware policy evaluation for sim-to-real transfer","benchmarking offline and online adaptive domain randomization for sim-to-real robot policy transfer"]},{"direction_index":1,"label":"D1","name":"Objective discovery, exploration, and constraints","description":"Specify what robots should optimize and regulate how they discover or satisfy objectives under sparse feedback, human preferences, intrinsic motivation, risk, or explicit constraints.","inclusion":"Include reward learning, preference learning, inverse or adversarial reward inference, self-supervised goal discovery, intrinsic motivation, sparse-reward exploration, constrained RL, safe imitation, barrier or Lyapunov methods, risk prediction, and formal or physical constraint satisfaction when the objective or constraint structure is primary.","exclusion":"If demonstrations are used primarily to clone actions, use data-driven policy acquisition; if safety or reward mechanisms are incidental within a concrete navigation, manipulation, locomotion, or planning task, use that task program.","nearest_boundary":"Data-driven policy acquisition and RL algorithms","representative_topics":["heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations","scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","generative adversarial reward learning from partial demonstrations for agile robot skill acquisition"]},{"direction_index":2,"label":"D2","name":"Data-driven policy acquisition and RL algorithms","description":"Learn robot policies from demonstrations, fixed datasets, generated trajectories, or general-purpose policy-optimization improvements when the central objective is acquiring or improving policies from data rather than solving a specific embodied capability.","inclusion":"Include imitation learning, behavioral cloning, offline RL, offline-to-online adaptation, diffusion or other direct action-generation policies, generic continuous-control RL improvements, policy optimization, architecture and interpretability methods, and meta-methods for robotic control.","exclusion":"If the main contribution is reward recovery, safety enforcement, learned dynamics for planning, task hierarchy, or a concrete robot capability such as locomotion or manipulation, assign to that more specific program.","nearest_boundary":"Objective discovery, exploration, and constraints","representative_topics":["demonstration-based and rule-based 3d object manipulation policy learning for robotic arms","latent policy learning for stable adversarial imitation in high-dimensional robot tasks","offline prior data adaptation for autonomous robotic reinforcement learning and environment reset"]},{"direction_index":3,"label":"D3","name":"Predictive, hierarchical, and foundation-model planning","description":"Enable long-horizon robot behavior by learning predictive models, abstractions, skills, symbolic operators, or foundation-model reasoning modules for planning and control.","inclusion":"Include world models, learned dynamics, differentiable physics, MPC with learned models, visual foresight, hierarchical RL, skill discovery and composition, task-and-motion planning with learned operators, language or vision-language planning, and foundation-model embodied reasoning when planning or prediction is primary.","exclusion":"If a language or model component mainly supplies rewards, use objective discovery; if hierarchy is only a controller architecture for locomotion or manipulation, use the corresponding capability program.","nearest_boundary":"Data-driven policy acquisition and RL algorithms","representative_topics":["neurosymbolic vae with symbolic operators for autonomous battery disassembly task and motion planning","physics-constrained temporal convolutional learning of quadrotor dynamics for model predictive control","world model-based planning for online robot learning without simulators"]},{"direction_index":4,"label":"D4","name":"Navigation, exploration, and multi-agent coordination","description":"Learn mobile, embodied, human-interactive, or multi-robot behaviors for reaching goals, gathering spatial information, allocating tasks, coordinating with others, and operating in unknown or dynamic environments.","inclusion":"Include goal-directed navigation, object-goal or semantic navigation, visual and topological navigation, active sensing and exploration, help-seeking navigation, coverage, multi-robot task allocation, swarms, fleet learning, human-robot coordination, and collaborative navigation when spatial decision-making or coordination is primary.","exclusion":"If the main issue is body-level gait or flight stabilization, use locomotion and platform-specific control; if the main issue is LLM/VLM reasoning rather than navigation or coordination, use predictive, hierarchical, and foundation-model planning.","nearest_boundary":"Locomotion and platform-specific control","representative_topics":["monocular rgb-based autonomous exploration with directional pixel encoding for unexplored area coverage","decomposed reinforcement learning for mobile manipulation navigation in dynamic environments","uncertainty-driven help-asking in embodied navigation with intermittent feedback"]},{"direction_index":5,"label":"D5","name":"Locomotion and platform-specific control","description":"Learn movement and control policies for robot bodies or specialized platforms where morphology, actuation, balance, terrain interaction, or platform physics is the primary scientific challenge.","inclusion":"Include legged, bipedal, humanoid, quadruped, aerial, balancing, soft, musculoskeletal, space, surgical, endoscopic, and other non-standard platform control when the focus is body motion, actuation, or domain-specific physical control.","exclusion":"If the robot is moving primarily to reach semantic or spatial goals, use navigation and coordination; if the platform is only an evaluation setting for a general RL algorithm, use data-driven policy acquisition and RL algorithms.","nearest_boundary":"Navigation, exploration, and multi-agent coordination","representative_topics":["open-ended learning with neuroevolution-generated terrains for complex robot locomotion skill acquisition","semantics-aware quadrupedal locomotion skill learning from human demonstration for off-road terrain","model-free reinforcement learning for fast quadrupedal locomotion with symmetry-aware mirror-world control"]},{"direction_index":6,"label":"D6","name":"Manipulation skill control","description":"Acquire closed-loop manipulation behaviors for rigid, dexterous, contact-rich, deformable, soft, cloth, cable, food, or other physically interactive object manipulation tasks.","inclusion":"Include reinforcement learning or control learning for grasping, insertion, pushing, reorientation, bimanual and dexterous hands, tactile or force-guided manipulation, free-floating manipulation, deformable-object manipulation, cloth and garment handling, visuo-haptic food handling, and contact-rich skill acquisition when execution of the manipulation skill is primary.","exclusion":"If the main contribution is an object, affordance, visual, or video representation rather than the closed-loop skill, use generalizable representations and cross-domain priors; if the contribution is long-horizon symbolic sequencing of skills, use predictive, hierarchical, and foundation-model planning.","nearest_boundary":"Generalizable representations and cross-domain priors","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","reinforcement learning benchmark for bimanual dexterous manipulation across skill levels and task settings","reinforcement learning motion planning for free-float dual-arm manipulators tracking non-cooperative spinning objects"]},{"direction_index":7,"label":"D7","name":"Generalizable representations and cross-domain priors","description":"Learn reusable visual, geometric, affordance, semantic, video-derived, object-centric, or cross-embodiment representations that enable policies to generalize across objects, scenes, tasks, viewpoints, hands, robots, and embodiments.","inclusion":"Include object-centric manipulation representations, grasp and affordance perception, implicit geometry, neural fields, equivariant descriptors, human-video affordance extraction, visual or video pretraining, multimodal priors, diffusion-generated experience for representation or data scaling, cross-task transfer, cross-robot transfer, morphology-aware generalization, and broad representation learning when generalization is primary.","exclusion":"If the representation is mainly a learned dynamics model for planning, use predictive, hierarchical, and foundation-model planning; if it is only an implementation detail inside a specific manipulation or locomotion controller, use that capability program.","nearest_boundary":"Manipulation skill control","representative_topics":["se(3)-equivariant energy-based models for sample-efficient end-to-end visual robotic manipulation","dual-gripper collaborative affordance learning via disentangled subtask decomposition","implicit representation learning for cross-gripper grasp transfer via signed distance functions"]}],"split":"dev"}
{"episode_id":"F081_T2025-01","field_index":81,"field":"robot learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Sim-to-real deployment and evaluation infrastructure","description":"Develop transfer mechanisms, simulators, benchmarks, datasets, and evaluation protocols that make robot-learning systems trainable, comparable, scalable, and deployable across simulated and physical settings.","inclusion":"Include work whose primary objective is closing or measuring the simulation-reality gap, building scalable simulation/training infrastructure, standardizing benchmarks, generating tasks or environments, or creating evaluation frameworks for robot learning.","exclusion":"If simulation, benchmarks, or datasets are used mainly as tools for a specific policy-learning, navigation, manipulation, locomotion, or planning objective, assign to that objective instead.","nearest_boundary":"Generalizable representations and cross-domain priors","representative_topics":["real-to-sim policy adaptation for visual reinforcement learning insertion tasks","hybrid offline-online reinforcement learning with dynamics-aware policy evaluation for sim-to-real transfer","benchmarking offline and online adaptive domain randomization for sim-to-real robot policy transfer"]},{"direction_index":1,"label":"D1","name":"Objective discovery, exploration, and constraints","description":"Specify what robots should optimize and regulate how they discover or satisfy objectives under sparse feedback, human preferences, intrinsic motivation, risk, or explicit constraints.","inclusion":"Include reward learning, preference learning, inverse or adversarial reward inference, self-supervised goal discovery, intrinsic motivation, sparse-reward exploration, constrained RL, safe imitation, barrier or Lyapunov methods, risk prediction, and formal or physical constraint satisfaction when the objective or constraint structure is primary.","exclusion":"If demonstrations are used primarily to clone actions, use data-driven policy acquisition; if safety or reward mechanisms are incidental within a concrete navigation, manipulation, locomotion, or planning task, use that task program.","nearest_boundary":"Data-driven policy acquisition and RL algorithms","representative_topics":["heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations","scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","generative adversarial reward learning from partial demonstrations for agile robot skill acquisition"]},{"direction_index":2,"label":"D2","name":"Data-driven policy acquisition and RL algorithms","description":"Learn robot policies from demonstrations, fixed datasets, generated trajectories, or general-purpose policy-optimization improvements when the central objective is acquiring or improving policies from data rather than solving a specific embodied capability.","inclusion":"Include imitation learning, behavioral cloning, offline RL, offline-to-online adaptation, diffusion or other direct action-generation policies, generic continuous-control RL improvements, policy optimization, architecture and interpretability methods, and meta-methods for robotic control.","exclusion":"If the main contribution is reward recovery, safety enforcement, learned dynamics for planning, task hierarchy, or a concrete robot capability such as locomotion or manipulation, assign to that more specific program.","nearest_boundary":"Objective discovery, exploration, and constraints","representative_topics":["demonstration-based and rule-based 3d object manipulation policy learning for robotic arms","latent policy learning for stable adversarial imitation in high-dimensional robot tasks","offline prior data adaptation for autonomous robotic reinforcement learning and environment reset"]},{"direction_index":3,"label":"D3","name":"Predictive, hierarchical, and foundation-model planning","description":"Enable long-horizon robot behavior by learning predictive models, abstractions, skills, symbolic operators, or foundation-model reasoning modules for planning and control.","inclusion":"Include world models, learned dynamics, differentiable physics, MPC with learned models, visual foresight, hierarchical RL, skill discovery and composition, task-and-motion planning with learned operators, language or vision-language planning, and foundation-model embodied reasoning when planning or prediction is primary.","exclusion":"If a language or model component mainly supplies rewards, use objective discovery; if hierarchy is only a controller architecture for locomotion or manipulation, use the corresponding capability program.","nearest_boundary":"Data-driven policy acquisition and RL algorithms","representative_topics":["neurosymbolic vae with symbolic operators for autonomous battery disassembly task and motion planning","physics-constrained temporal convolutional learning of quadrotor dynamics for model predictive control","world model-based planning for online robot learning without simulators"]},{"direction_index":4,"label":"D4","name":"Navigation, exploration, and multi-agent coordination","description":"Learn mobile, embodied, human-interactive, or multi-robot behaviors for reaching goals, gathering spatial information, allocating tasks, coordinating with others, and operating in unknown or dynamic environments.","inclusion":"Include goal-directed navigation, object-goal or semantic navigation, visual and topological navigation, active sensing and exploration, help-seeking navigation, coverage, multi-robot task allocation, swarms, fleet learning, human-robot coordination, and collaborative navigation when spatial decision-making or coordination is primary.","exclusion":"If the main issue is body-level gait or flight stabilization, use locomotion and platform-specific control; if the main issue is LLM/VLM reasoning rather than navigation or coordination, use predictive, hierarchical, and foundation-model planning.","nearest_boundary":"Locomotion and platform-specific control","representative_topics":["monocular rgb-based autonomous exploration with directional pixel encoding for unexplored area coverage","decomposed reinforcement learning for mobile manipulation navigation in dynamic environments","uncertainty-driven help-asking in embodied navigation with intermittent feedback"]},{"direction_index":5,"label":"D5","name":"Locomotion and platform-specific control","description":"Learn movement and control policies for robot bodies or specialized platforms where morphology, actuation, balance, terrain interaction, or platform physics is the primary scientific challenge.","inclusion":"Include legged, bipedal, humanoid, quadruped, aerial, balancing, soft, musculoskeletal, space, surgical, endoscopic, and other non-standard platform control when the focus is body motion, actuation, or domain-specific physical control.","exclusion":"If the robot is moving primarily to reach semantic or spatial goals, use navigation and coordination; if the platform is only an evaluation setting for a general RL algorithm, use data-driven policy acquisition and RL algorithms.","nearest_boundary":"Navigation, exploration, and multi-agent coordination","representative_topics":["open-ended learning with neuroevolution-generated terrains for complex robot locomotion skill acquisition","semantics-aware quadrupedal locomotion skill learning from human demonstration for off-road terrain","model-free reinforcement learning for fast quadrupedal locomotion with symmetry-aware mirror-world control"]},{"direction_index":6,"label":"D6","name":"Manipulation skill control","description":"Acquire closed-loop manipulation behaviors for rigid, dexterous, contact-rich, deformable, soft, cloth, cable, food, or other physically interactive object manipulation tasks.","inclusion":"Include reinforcement learning or control learning for grasping, insertion, pushing, reorientation, bimanual and dexterous hands, tactile or force-guided manipulation, free-floating manipulation, deformable-object manipulation, cloth and garment handling, visuo-haptic food handling, and contact-rich skill acquisition when execution of the manipulation skill is primary.","exclusion":"If the main contribution is an object, affordance, visual, or video representation rather than the closed-loop skill, use generalizable representations and cross-domain priors; if the contribution is long-horizon symbolic sequencing of skills, use predictive, hierarchical, and foundation-model planning.","nearest_boundary":"Generalizable representations and cross-domain priors","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","reinforcement learning benchmark for bimanual dexterous manipulation across skill levels and task settings","reinforcement learning motion planning for free-float dual-arm manipulators tracking non-cooperative spinning objects"]},{"direction_index":7,"label":"D7","name":"Generalizable representations and cross-domain priors","description":"Learn reusable visual, geometric, affordance, semantic, video-derived, object-centric, or cross-embodiment representations that enable policies to generalize across objects, scenes, tasks, viewpoints, hands, robots, and embodiments.","inclusion":"Include object-centric manipulation representations, grasp and affordance perception, implicit geometry, neural fields, equivariant descriptors, human-video affordance extraction, visual or video pretraining, multimodal priors, diffusion-generated experience for representation or data scaling, cross-task transfer, cross-robot transfer, morphology-aware generalization, and broad representation learning when generalization is primary.","exclusion":"If the representation is mainly a learned dynamics model for planning, use predictive, hierarchical, and foundation-model planning; if it is only an implementation detail inside a specific manipulation or locomotion controller, use that capability program.","nearest_boundary":"Manipulation skill control","representative_topics":["se(3)-equivariant energy-based models for sample-efficient end-to-end visual robotic manipulation","dual-gripper collaborative affordance learning via disentangled subtask decomposition","implicit representation learning for cross-gripper grasp transfer via signed distance functions"]}],"split":"dev"}
{"episode_id":"F081_T2025-07","field_index":81,"field":"robot learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Sim-to-real deployment and evaluation infrastructure","description":"Develop transfer mechanisms, simulators, benchmarks, datasets, and evaluation protocols that make robot-learning systems trainable, comparable, scalable, and deployable across simulated and physical settings.","inclusion":"Include work whose primary objective is closing or measuring the simulation-reality gap, building scalable simulation/training infrastructure, standardizing benchmarks, generating tasks or environments, or creating evaluation frameworks for robot learning.","exclusion":"If simulation, benchmarks, or datasets are used mainly as tools for a specific policy-learning, navigation, manipulation, locomotion, or planning objective, assign to that objective instead.","nearest_boundary":"Generalizable representations and cross-domain priors","representative_topics":["real-to-sim policy adaptation for visual reinforcement learning insertion tasks","hybrid offline-online reinforcement learning with dynamics-aware policy evaluation for sim-to-real transfer","benchmarking offline and online adaptive domain randomization for sim-to-real robot policy transfer"]},{"direction_index":1,"label":"D1","name":"Objective discovery, exploration, and constraints","description":"Specify what robots should optimize and regulate how they discover or satisfy objectives under sparse feedback, human preferences, intrinsic motivation, risk, or explicit constraints.","inclusion":"Include reward learning, preference learning, inverse or adversarial reward inference, self-supervised goal discovery, intrinsic motivation, sparse-reward exploration, constrained RL, safe imitation, barrier or Lyapunov methods, risk prediction, and formal or physical constraint satisfaction when the objective or constraint structure is primary.","exclusion":"If demonstrations are used primarily to clone actions, use data-driven policy acquisition; if safety or reward mechanisms are incidental within a concrete navigation, manipulation, locomotion, or planning task, use that task program.","nearest_boundary":"Data-driven policy acquisition and RL algorithms","representative_topics":["heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations","scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","generative adversarial reward learning from partial demonstrations for agile robot skill acquisition"]},{"direction_index":2,"label":"D2","name":"Data-driven policy acquisition and RL algorithms","description":"Learn robot policies from demonstrations, fixed datasets, generated trajectories, or general-purpose policy-optimization improvements when the central objective is acquiring or improving policies from data rather than solving a specific embodied capability.","inclusion":"Include imitation learning, behavioral cloning, offline RL, offline-to-online adaptation, diffusion or other direct action-generation policies, generic continuous-control RL improvements, policy optimization, architecture and interpretability methods, and meta-methods for robotic control.","exclusion":"If the main contribution is reward recovery, safety enforcement, learned dynamics for planning, task hierarchy, or a concrete robot capability such as locomotion or manipulation, assign to that more specific program.","nearest_boundary":"Objective discovery, exploration, and constraints","representative_topics":["demonstration-based and rule-based 3d object manipulation policy learning for robotic arms","latent policy learning for stable adversarial imitation in high-dimensional robot tasks","offline prior data adaptation for autonomous robotic reinforcement learning and environment reset"]},{"direction_index":3,"label":"D3","name":"Predictive, hierarchical, and foundation-model planning","description":"Enable long-horizon robot behavior by learning predictive models, abstractions, skills, symbolic operators, or foundation-model reasoning modules for planning and control.","inclusion":"Include world models, learned dynamics, differentiable physics, MPC with learned models, visual foresight, hierarchical RL, skill discovery and composition, task-and-motion planning with learned operators, language or vision-language planning, and foundation-model embodied reasoning when planning or prediction is primary.","exclusion":"If a language or model component mainly supplies rewards, use objective discovery; if hierarchy is only a controller architecture for locomotion or manipulation, use the corresponding capability program.","nearest_boundary":"Data-driven policy acquisition and RL algorithms","representative_topics":["neurosymbolic vae with symbolic operators for autonomous battery disassembly task and motion planning","physics-constrained temporal convolutional learning of quadrotor dynamics for model predictive control","world model-based planning for online robot learning without simulators"]},{"direction_index":4,"label":"D4","name":"Navigation, exploration, and multi-agent coordination","description":"Learn mobile, embodied, human-interactive, or multi-robot behaviors for reaching goals, gathering spatial information, allocating tasks, coordinating with others, and operating in unknown or dynamic environments.","inclusion":"Include goal-directed navigation, object-goal or semantic navigation, visual and topological navigation, active sensing and exploration, help-seeking navigation, coverage, multi-robot task allocation, swarms, fleet learning, human-robot coordination, and collaborative navigation when spatial decision-making or coordination is primary.","exclusion":"If the main issue is body-level gait or flight stabilization, use locomotion and platform-specific control; if the main issue is LLM/VLM reasoning rather than navigation or coordination, use predictive, hierarchical, and foundation-model planning.","nearest_boundary":"Locomotion and platform-specific control","representative_topics":["monocular rgb-based autonomous exploration with directional pixel encoding for unexplored area coverage","decomposed reinforcement learning for mobile manipulation navigation in dynamic environments","uncertainty-driven help-asking in embodied navigation with intermittent feedback"]},{"direction_index":5,"label":"D5","name":"Locomotion and platform-specific control","description":"Learn movement and control policies for robot bodies or specialized platforms where morphology, actuation, balance, terrain interaction, or platform physics is the primary scientific challenge.","inclusion":"Include legged, bipedal, humanoid, quadruped, aerial, balancing, soft, musculoskeletal, space, surgical, endoscopic, and other non-standard platform control when the focus is body motion, actuation, or domain-specific physical control.","exclusion":"If the robot is moving primarily to reach semantic or spatial goals, use navigation and coordination; if the platform is only an evaluation setting for a general RL algorithm, use data-driven policy acquisition and RL algorithms.","nearest_boundary":"Navigation, exploration, and multi-agent coordination","representative_topics":["open-ended learning with neuroevolution-generated terrains for complex robot locomotion skill acquisition","semantics-aware quadrupedal locomotion skill learning from human demonstration for off-road terrain","model-free reinforcement learning for fast quadrupedal locomotion with symmetry-aware mirror-world control"]},{"direction_index":6,"label":"D6","name":"Manipulation skill control","description":"Acquire closed-loop manipulation behaviors for rigid, dexterous, contact-rich, deformable, soft, cloth, cable, food, or other physically interactive object manipulation tasks.","inclusion":"Include reinforcement learning or control learning for grasping, insertion, pushing, reorientation, bimanual and dexterous hands, tactile or force-guided manipulation, free-floating manipulation, deformable-object manipulation, cloth and garment handling, visuo-haptic food handling, and contact-rich skill acquisition when execution of the manipulation skill is primary.","exclusion":"If the main contribution is an object, affordance, visual, or video representation rather than the closed-loop skill, use generalizable representations and cross-domain priors; if the contribution is long-horizon symbolic sequencing of skills, use predictive, hierarchical, and foundation-model planning.","nearest_boundary":"Generalizable representations and cross-domain priors","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","reinforcement learning benchmark for bimanual dexterous manipulation across skill levels and task settings","reinforcement learning motion planning for free-float dual-arm manipulators tracking non-cooperative spinning objects"]},{"direction_index":7,"label":"D7","name":"Generalizable representations and cross-domain priors","description":"Learn reusable visual, geometric, affordance, semantic, video-derived, object-centric, or cross-embodiment representations that enable policies to generalize across objects, scenes, tasks, viewpoints, hands, robots, and embodiments.","inclusion":"Include object-centric manipulation representations, grasp and affordance perception, implicit geometry, neural fields, equivariant descriptors, human-video affordance extraction, visual or video pretraining, multimodal priors, diffusion-generated experience for representation or data scaling, cross-task transfer, cross-robot transfer, morphology-aware generalization, and broad representation learning when generalization is primary.","exclusion":"If the representation is mainly a learned dynamics model for planning, use predictive, hierarchical, and foundation-model planning; if it is only an implementation detail inside a specific manipulation or locomotion controller, use that capability program.","nearest_boundary":"Manipulation skill control","representative_topics":["se(3)-equivariant energy-based models for sample-efficient end-to-end visual robotic manipulation","dual-gripper collaborative affordance learning via disentangled subtask decomposition","implicit representation learning for cross-gripper grasp transfer via signed distance functions"]}],"split":"dev"}
{"episode_id":"F081_T2026-01","field_index":81,"field":"robot learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Sim-to-real deployment and evaluation infrastructure","description":"Develop transfer mechanisms, simulators, benchmarks, datasets, and evaluation protocols that make robot-learning systems trainable, comparable, scalable, and deployable across simulated and physical settings.","inclusion":"Include work whose primary objective is closing or measuring the simulation-reality gap, building scalable simulation/training infrastructure, standardizing benchmarks, generating tasks or environments, or creating evaluation frameworks for robot learning.","exclusion":"If simulation, benchmarks, or datasets are used mainly as tools for a specific policy-learning, navigation, manipulation, locomotion, or planning objective, assign to that objective instead.","nearest_boundary":"Generalizable representations and cross-domain priors","representative_topics":["real-to-sim policy adaptation for visual reinforcement learning insertion tasks","hybrid offline-online reinforcement learning with dynamics-aware policy evaluation for sim-to-real transfer","benchmarking offline and online adaptive domain randomization for sim-to-real robot policy transfer"]},{"direction_index":1,"label":"D1","name":"Objective discovery, exploration, and constraints","description":"Specify what robots should optimize and regulate how they discover or satisfy objectives under sparse feedback, human preferences, intrinsic motivation, risk, or explicit constraints.","inclusion":"Include reward learning, preference learning, inverse or adversarial reward inference, self-supervised goal discovery, intrinsic motivation, sparse-reward exploration, constrained RL, safe imitation, barrier or Lyapunov methods, risk prediction, and formal or physical constraint satisfaction when the objective or constraint structure is primary.","exclusion":"If demonstrations are used primarily to clone actions, use data-driven policy acquisition; if safety or reward mechanisms are incidental within a concrete navigation, manipulation, locomotion, or planning task, use that task program.","nearest_boundary":"Data-driven policy acquisition and RL algorithms","representative_topics":["heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations","scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","generative adversarial reward learning from partial demonstrations for agile robot skill acquisition"]},{"direction_index":2,"label":"D2","name":"Data-driven policy acquisition and RL algorithms","description":"Learn robot policies from demonstrations, fixed datasets, generated trajectories, or general-purpose policy-optimization improvements when the central objective is acquiring or improving policies from data rather than solving a specific embodied capability.","inclusion":"Include imitation learning, behavioral cloning, offline RL, offline-to-online adaptation, diffusion or other direct action-generation policies, generic continuous-control RL improvements, policy optimization, architecture and interpretability methods, and meta-methods for robotic control.","exclusion":"If the main contribution is reward recovery, safety enforcement, learned dynamics for planning, task hierarchy, or a concrete robot capability such as locomotion or manipulation, assign to that more specific program.","nearest_boundary":"Objective discovery, exploration, and constraints","representative_topics":["demonstration-based and rule-based 3d object manipulation policy learning for robotic arms","latent policy learning for stable adversarial imitation in high-dimensional robot tasks","offline prior data adaptation for autonomous robotic reinforcement learning and environment reset"]},{"direction_index":3,"label":"D3","name":"Predictive, hierarchical, and foundation-model planning","description":"Enable long-horizon robot behavior by learning predictive models, abstractions, skills, symbolic operators, or foundation-model reasoning modules for planning and control.","inclusion":"Include world models, learned dynamics, differentiable physics, MPC with learned models, visual foresight, hierarchical RL, skill discovery and composition, task-and-motion planning with learned operators, language or vision-language planning, and foundation-model embodied reasoning when planning or prediction is primary.","exclusion":"If a language or model component mainly supplies rewards, use objective discovery; if hierarchy is only a controller architecture for locomotion or manipulation, use the corresponding capability program.","nearest_boundary":"Data-driven policy acquisition and RL algorithms","representative_topics":["neurosymbolic vae with symbolic operators for autonomous battery disassembly task and motion planning","physics-constrained temporal convolutional learning of quadrotor dynamics for model predictive control","world model-based planning for online robot learning without simulators"]},{"direction_index":4,"label":"D4","name":"Navigation, exploration, and multi-agent coordination","description":"Learn mobile, embodied, human-interactive, or multi-robot behaviors for reaching goals, gathering spatial information, allocating tasks, coordinating with others, and operating in unknown or dynamic environments.","inclusion":"Include goal-directed navigation, object-goal or semantic navigation, visual and topological navigation, active sensing and exploration, help-seeking navigation, coverage, multi-robot task allocation, swarms, fleet learning, human-robot coordination, and collaborative navigation when spatial decision-making or coordination is primary.","exclusion":"If the main issue is body-level gait or flight stabilization, use locomotion and platform-specific control; if the main issue is LLM/VLM reasoning rather than navigation or coordination, use predictive, hierarchical, and foundation-model planning.","nearest_boundary":"Locomotion and platform-specific control","representative_topics":["monocular rgb-based autonomous exploration with directional pixel encoding for unexplored area coverage","decomposed reinforcement learning for mobile manipulation navigation in dynamic environments","uncertainty-driven help-asking in embodied navigation with intermittent feedback"]},{"direction_index":5,"label":"D5","name":"Locomotion and platform-specific control","description":"Learn movement and control policies for robot bodies or specialized platforms where morphology, actuation, balance, terrain interaction, or platform physics is the primary scientific challenge.","inclusion":"Include legged, bipedal, humanoid, quadruped, aerial, balancing, soft, musculoskeletal, space, surgical, endoscopic, and other non-standard platform control when the focus is body motion, actuation, or domain-specific physical control.","exclusion":"If the robot is moving primarily to reach semantic or spatial goals, use navigation and coordination; if the platform is only an evaluation setting for a general RL algorithm, use data-driven policy acquisition and RL algorithms.","nearest_boundary":"Navigation, exploration, and multi-agent coordination","representative_topics":["open-ended learning with neuroevolution-generated terrains for complex robot locomotion skill acquisition","semantics-aware quadrupedal locomotion skill learning from human demonstration for off-road terrain","model-free reinforcement learning for fast quadrupedal locomotion with symmetry-aware mirror-world control"]},{"direction_index":6,"label":"D6","name":"Manipulation skill control","description":"Acquire closed-loop manipulation behaviors for rigid, dexterous, contact-rich, deformable, soft, cloth, cable, food, or other physically interactive object manipulation tasks.","inclusion":"Include reinforcement learning or control learning for grasping, insertion, pushing, reorientation, bimanual and dexterous hands, tactile or force-guided manipulation, free-floating manipulation, deformable-object manipulation, cloth and garment handling, visuo-haptic food handling, and contact-rich skill acquisition when execution of the manipulation skill is primary.","exclusion":"If the main contribution is an object, affordance, visual, or video representation rather than the closed-loop skill, use generalizable representations and cross-domain priors; if the contribution is long-horizon symbolic sequencing of skills, use predictive, hierarchical, and foundation-model planning.","nearest_boundary":"Generalizable representations and cross-domain priors","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","reinforcement learning benchmark for bimanual dexterous manipulation across skill levels and task settings","reinforcement learning motion planning for free-float dual-arm manipulators tracking non-cooperative spinning objects"]},{"direction_index":7,"label":"D7","name":"Generalizable representations and cross-domain priors","description":"Learn reusable visual, geometric, affordance, semantic, video-derived, object-centric, or cross-embodiment representations that enable policies to generalize across objects, scenes, tasks, viewpoints, hands, robots, and embodiments.","inclusion":"Include object-centric manipulation representations, grasp and affordance perception, implicit geometry, neural fields, equivariant descriptors, human-video affordance extraction, visual or video pretraining, multimodal priors, diffusion-generated experience for representation or data scaling, cross-task transfer, cross-robot transfer, morphology-aware generalization, and broad representation learning when generalization is primary.","exclusion":"If the representation is mainly a learned dynamics model for planning, use predictive, hierarchical, and foundation-model planning; if it is only an implementation detail inside a specific manipulation or locomotion controller, use that capability program.","nearest_boundary":"Manipulation skill control","representative_topics":["se(3)-equivariant energy-based models for sample-efficient end-to-end visual robotic manipulation","dual-gripper collaborative affordance learning via disentangled subtask decomposition","implicit representation learning for cross-gripper grasp transfer via signed distance functions"]}],"split":"dev"}
{"episode_id":"F084_T2024-01","field_index":84,"field":"deep learning theory","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Supervised generalization phenomena","description":"Statistical theory of test performance in supervised deep learning, including risk bounds, excess-risk guarantees, sample complexity, structured-network generalization, interpolation-regime behavior, benign overfitting, and double descent.","inclusion":"Include atoms whose primary objective is to explain or bound supervised test risk, whether through PAC-Bayes, stability, complexity measures, sparsity, transfer/meta/active-learning settings, or interpolation-specific phenomena such as benign overfitting and double descent.","exclusion":"If the main objective is optimizer convergence or solution selection rather than test-risk behavior, assign to Training dynamics and regularization; if the objective is long-run scaling laws or task-level impossibility, assign to Scaling and learning limits.","nearest_boundary":"Training dynamics and regularization","representative_topics":["decision boundary variability as a generalization measure in deep learning theory","sgd-dependent generalization error bounds via parameter norm dynamics and rademacher complexity","theoretical analysis of benign overfitting failure under label noise in underparameterized regimes"]},{"direction_index":1,"label":"D1","name":"Training dynamics and regularization","description":"Theory of how training procedures, optimizer dynamics, initialization, normalization, sharpness, implicit bias, explicit interventions, and alternative learning rules make neural networks trainable and shape learned solutions.","inclusion":"Include atoms whose primary objective is convergence, trainability, signal or gradient propagation, optimizer-induced solution selection, flatness or edge-of-stability dynamics, explicit regularizers or augmentations, or nonstandard credit-assignment/training paradigms.","exclusion":"If the optimizer or regularizer is mainly a tool for proving test-risk bounds, assign to Supervised generalization phenomena; if the main object is a kernel, infinite-width, or tractable limiting model rather than training mechanisms themselves, assign to Kernel, continuum, and tractable limits.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["early and global convergence of gradient descent via neuron activation partition analysis","theoretical analysis of global convergence in over-parameterized relu nets via activation matrix angular geometry","stochastic gradient descent noise as flatness-favoring regularization via loss landscape anisotropy"]},{"direction_index":2,"label":"D2","name":"Kernel, continuum, and tractable limits","description":"Asymptotic, linearized, kernel-regime, infinite-width/depth, mean-field, Gaussian-process, stochastic-process, and tractable-model analyses used to obtain analytically controlled descriptions of deep networks.","inclusion":"Include atoms whose primary objective is to characterize NTK or lazy-training regimes, GP or mean-field limits, continuum ODE/SDE/PDE descriptions, large-width asymptotics, or deep linear/gated linear models as mathematically tractable proxies.","exclusion":"If the main target is finite-network optimization behavior, assign to Training dynamics and regularization; if the target is raw representational capacity rather than limiting dynamics, assign to Expressivity and architectural structure.","nearest_boundary":"Training dynamics and regularization","representative_topics":["generalization bounds for overparameterized 3-layer neural tangent kernel models with hidden-layer interactions","spectral bias bounds for deep networks via fisher information low effective rank in kernel regime","convergence guarantees for empirical neural tangent kernel approximation in wide readout networks"]},{"direction_index":3,"label":"D3","name":"Expressivity and architectural structure","description":"Theory of what neural architectures and parameterizations can represent, approximate, or structurally bias, including approximation rates, universality, depth-width tradeoffs, architectural inductive priors, equivariance, attention, convolution, MoE, and ReLU combinatorial geometry.","inclusion":"Include atoms whose primary objective is representational capacity, approximation complexity, architectural inductive bias, invariance/equivariance, architecture-specific learnability advantages or costs, or geometric/combinatorial structure of network function classes.","exclusion":"If the main target is learned feature evolution during training, assign to Feature and representation learning; if architecture is used only as the setting for a generalization bound, assign to Supervised generalization phenomena.","nearest_boundary":"Feature and representation learning","representative_topics":["l^p norm approximation lower bounds for feed-forward neural networks via fat-shattering dimension","smoothness-constrained function approximation in overparameterized convolutional residual networks","depth and size bounds for monotone neural networks via interpolation and approximation"]},{"direction_index":4,"label":"D4","name":"Feature and representation learning","description":"Theory of how trained networks form, transform, organize, and geometrize internal representations, including feature learning beyond kernels, representation multiplicity, low-rank or low-dimensional features, and neural-collapse geometry.","inclusion":"Include atoms whose primary objective is the emergence, dynamics, or final geometry of internal features or representations, including kernel-to-feature-learning transitions, subspace recovery, representational multiplicity, rank structure, and class-mean/classifier collapse.","exclusion":"If the main object is an architectural prior rather than learned representation dynamics, assign to Expressivity and architectural structure; if the main object is optimizer solution selection in parameter space, assign to Training dynamics and regularization.","nearest_boundary":"Expressivity and architectural structure","representative_topics":["theoretical analysis of feature learning overfitting via sparse representations in fully-connected networks","gradient descent representations for efficient learning of sparse high-dimensional polynomials","sgd training of two-layer networks for low-dimensional subspace recovery in high-dimensional settings"]},{"direction_index":5,"label":"D5","name":"Probabilistic and generative models","description":"Theoretical foundations of Bayesian, probabilistic, variational, ensemble, diffusion, score-based, normalizing-flow, and other deep generative modeling approaches.","inclusion":"Include atoms whose primary objective is posterior inference, uncertainty, Bayesian neural-network theory, variational free energy, posterior geometry, score estimation, diffusion sampling, distribution recovery, or convergence/statistical theory of generative models.","exclusion":"If Bayesian or posterior quantities are mainly used to derive frequentist test-risk bounds, assign to Supervised generalization phenomena; if a generative architecture is studied only for representational capacity, assign to Expressivity and architectural structure.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["repriorisation for accelerated mcmc sampling in wide bayesian neural networks","variational inference of overparameterized bayesian neural networks via elbo term trade-off analysis","theoretical convergence analysis of score-based generative models with l2-score accuracy and non-log-concave distributions"]},{"direction_index":6,"label":"D6","name":"Robustness and perturbation stability","description":"Theory of neural-network behavior under adversarial perturbations, input corruptions, sensitivity, Lipschitz constraints, distributional shifts, and robustness-relevant classifier or metric geometry.","inclusion":"Include atoms whose primary objective is to quantify, certify, or explain robustness, adversarial vulnerability, transferability, perturbation stability, Lipschitz behavior, or robustness-relevant geometry of trained networks.","exclusion":"If perturbations are introduced primarily as a training regularizer or augmentation, assign to Training dynamics and regularization; if robustness quantities are only proof devices for ordinary test-risk bounds, assign to Supervised generalization phenomena.","nearest_boundary":"Training dynamics and regularization","representative_topics":["theoretical framework for explaining neural network generalization and adversarial example transfer","lipschitz bound tightening for cnns via convolutional layer unrolling and toeplitz decomposition","theoretical analysis of bayesian neural network robustness to adversarial attacks via posterior geometry"]},{"direction_index":7,"label":"D7","name":"Scaling and learning limits","description":"Theory of performance scaling, task learnability, grokking and phase transitions, computational-statistical gaps, hardness, impossibility, and fundamental limits of neural-network learning on structured task families.","inclusion":"Include atoms whose primary objective is to explain scaling laws, delayed generalization, phase transitions, dataset/model/compute scaling, task-level learnability limits, computational hardness, or impossibility of learning particular structures.","exclusion":"If the target is a conventional finite-sample generalization guarantee, assign to Supervised generalization phenomena; if the target is feature emergence on a learnable task rather than a scaling law or limit, assign to Feature and representation learning.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["formal language theory grounded prediction of neural network generalization limits via chomsky hierarchy","theoretical analysis of neural network learnability of physical concepts via recursion and active perception","statistical model of neural scaling laws via random feature maps and spectral power laws"]}],"split":"dev"}
{"episode_id":"F084_T2024-07","field_index":84,"field":"deep learning theory","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Supervised generalization phenomena","description":"Statistical theory of test performance in supervised deep learning, including risk bounds, excess-risk guarantees, sample complexity, structured-network generalization, interpolation-regime behavior, benign overfitting, and double descent.","inclusion":"Include atoms whose primary objective is to explain or bound supervised test risk, whether through PAC-Bayes, stability, complexity measures, sparsity, transfer/meta/active-learning settings, or interpolation-specific phenomena such as benign overfitting and double descent.","exclusion":"If the main objective is optimizer convergence or solution selection rather than test-risk behavior, assign to Training dynamics and regularization; if the objective is long-run scaling laws or task-level impossibility, assign to Scaling and learning limits.","nearest_boundary":"Training dynamics and regularization","representative_topics":["decision boundary variability as a generalization measure in deep learning theory","sgd-dependent generalization error bounds via parameter norm dynamics and rademacher complexity","theoretical analysis of benign overfitting failure under label noise in underparameterized regimes"]},{"direction_index":1,"label":"D1","name":"Training dynamics and regularization","description":"Theory of how training procedures, optimizer dynamics, initialization, normalization, sharpness, implicit bias, explicit interventions, and alternative learning rules make neural networks trainable and shape learned solutions.","inclusion":"Include atoms whose primary objective is convergence, trainability, signal or gradient propagation, optimizer-induced solution selection, flatness or edge-of-stability dynamics, explicit regularizers or augmentations, or nonstandard credit-assignment/training paradigms.","exclusion":"If the optimizer or regularizer is mainly a tool for proving test-risk bounds, assign to Supervised generalization phenomena; if the main object is a kernel, infinite-width, or tractable limiting model rather than training mechanisms themselves, assign to Kernel, continuum, and tractable limits.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["early and global convergence of gradient descent via neuron activation partition analysis","theoretical analysis of global convergence in over-parameterized relu nets via activation matrix angular geometry","stochastic gradient descent noise as flatness-favoring regularization via loss landscape anisotropy"]},{"direction_index":2,"label":"D2","name":"Kernel, continuum, and tractable limits","description":"Asymptotic, linearized, kernel-regime, infinite-width/depth, mean-field, Gaussian-process, stochastic-process, and tractable-model analyses used to obtain analytically controlled descriptions of deep networks.","inclusion":"Include atoms whose primary objective is to characterize NTK or lazy-training regimes, GP or mean-field limits, continuum ODE/SDE/PDE descriptions, large-width asymptotics, or deep linear/gated linear models as mathematically tractable proxies.","exclusion":"If the main target is finite-network optimization behavior, assign to Training dynamics and regularization; if the target is raw representational capacity rather than limiting dynamics, assign to Expressivity and architectural structure.","nearest_boundary":"Training dynamics and regularization","representative_topics":["generalization bounds for overparameterized 3-layer neural tangent kernel models with hidden-layer interactions","spectral bias bounds for deep networks via fisher information low effective rank in kernel regime","convergence guarantees for empirical neural tangent kernel approximation in wide readout networks"]},{"direction_index":3,"label":"D3","name":"Expressivity and architectural structure","description":"Theory of what neural architectures and parameterizations can represent, approximate, or structurally bias, including approximation rates, universality, depth-width tradeoffs, architectural inductive priors, equivariance, attention, convolution, MoE, and ReLU combinatorial geometry.","inclusion":"Include atoms whose primary objective is representational capacity, approximation complexity, architectural inductive bias, invariance/equivariance, architecture-specific learnability advantages or costs, or geometric/combinatorial structure of network function classes.","exclusion":"If the main target is learned feature evolution during training, assign to Feature and representation learning; if architecture is used only as the setting for a generalization bound, assign to Supervised generalization phenomena.","nearest_boundary":"Feature and representation learning","representative_topics":["l^p norm approximation lower bounds for feed-forward neural networks via fat-shattering dimension","smoothness-constrained function approximation in overparameterized convolutional residual networks","depth and size bounds for monotone neural networks via interpolation and approximation"]},{"direction_index":4,"label":"D4","name":"Feature and representation learning","description":"Theory of how trained networks form, transform, organize, and geometrize internal representations, including feature learning beyond kernels, representation multiplicity, low-rank or low-dimensional features, and neural-collapse geometry.","inclusion":"Include atoms whose primary objective is the emergence, dynamics, or final geometry of internal features or representations, including kernel-to-feature-learning transitions, subspace recovery, representational multiplicity, rank structure, and class-mean/classifier collapse.","exclusion":"If the main object is an architectural prior rather than learned representation dynamics, assign to Expressivity and architectural structure; if the main object is optimizer solution selection in parameter space, assign to Training dynamics and regularization.","nearest_boundary":"Expressivity and architectural structure","representative_topics":["theoretical analysis of feature learning overfitting via sparse representations in fully-connected networks","gradient descent representations for efficient learning of sparse high-dimensional polynomials","sgd training of two-layer networks for low-dimensional subspace recovery in high-dimensional settings"]},{"direction_index":5,"label":"D5","name":"Probabilistic and generative models","description":"Theoretical foundations of Bayesian, probabilistic, variational, ensemble, diffusion, score-based, normalizing-flow, and other deep generative modeling approaches.","inclusion":"Include atoms whose primary objective is posterior inference, uncertainty, Bayesian neural-network theory, variational free energy, posterior geometry, score estimation, diffusion sampling, distribution recovery, or convergence/statistical theory of generative models.","exclusion":"If Bayesian or posterior quantities are mainly used to derive frequentist test-risk bounds, assign to Supervised generalization phenomena; if a generative architecture is studied only for representational capacity, assign to Expressivity and architectural structure.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["repriorisation for accelerated mcmc sampling in wide bayesian neural networks","variational inference of overparameterized bayesian neural networks via elbo term trade-off analysis","theoretical convergence analysis of score-based generative models with l2-score accuracy and non-log-concave distributions"]},{"direction_index":6,"label":"D6","name":"Robustness and perturbation stability","description":"Theory of neural-network behavior under adversarial perturbations, input corruptions, sensitivity, Lipschitz constraints, distributional shifts, and robustness-relevant classifier or metric geometry.","inclusion":"Include atoms whose primary objective is to quantify, certify, or explain robustness, adversarial vulnerability, transferability, perturbation stability, Lipschitz behavior, or robustness-relevant geometry of trained networks.","exclusion":"If perturbations are introduced primarily as a training regularizer or augmentation, assign to Training dynamics and regularization; if robustness quantities are only proof devices for ordinary test-risk bounds, assign to Supervised generalization phenomena.","nearest_boundary":"Training dynamics and regularization","representative_topics":["theoretical framework for explaining neural network generalization and adversarial example transfer","lipschitz bound tightening for cnns via convolutional layer unrolling and toeplitz decomposition","theoretical analysis of bayesian neural network robustness to adversarial attacks via posterior geometry"]},{"direction_index":7,"label":"D7","name":"Scaling and learning limits","description":"Theory of performance scaling, task learnability, grokking and phase transitions, computational-statistical gaps, hardness, impossibility, and fundamental limits of neural-network learning on structured task families.","inclusion":"Include atoms whose primary objective is to explain scaling laws, delayed generalization, phase transitions, dataset/model/compute scaling, task-level learnability limits, computational hardness, or impossibility of learning particular structures.","exclusion":"If the target is a conventional finite-sample generalization guarantee, assign to Supervised generalization phenomena; if the target is feature emergence on a learnable task rather than a scaling law or limit, assign to Feature and representation learning.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["formal language theory grounded prediction of neural network generalization limits via chomsky hierarchy","theoretical analysis of neural network learnability of physical concepts via recursion and active perception","statistical model of neural scaling laws via random feature maps and spectral power laws"]}],"split":"dev"}
{"episode_id":"F084_T2025-01","field_index":84,"field":"deep learning theory","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Supervised generalization phenomena","description":"Statistical theory of test performance in supervised deep learning, including risk bounds, excess-risk guarantees, sample complexity, structured-network generalization, interpolation-regime behavior, benign overfitting, and double descent.","inclusion":"Include atoms whose primary objective is to explain or bound supervised test risk, whether through PAC-Bayes, stability, complexity measures, sparsity, transfer/meta/active-learning settings, or interpolation-specific phenomena such as benign overfitting and double descent.","exclusion":"If the main objective is optimizer convergence or solution selection rather than test-risk behavior, assign to Training dynamics and regularization; if the objective is long-run scaling laws or task-level impossibility, assign to Scaling and learning limits.","nearest_boundary":"Training dynamics and regularization","representative_topics":["decision boundary variability as a generalization measure in deep learning theory","sgd-dependent generalization error bounds via parameter norm dynamics and rademacher complexity","theoretical analysis of benign overfitting failure under label noise in underparameterized regimes"]},{"direction_index":1,"label":"D1","name":"Training dynamics and regularization","description":"Theory of how training procedures, optimizer dynamics, initialization, normalization, sharpness, implicit bias, explicit interventions, and alternative learning rules make neural networks trainable and shape learned solutions.","inclusion":"Include atoms whose primary objective is convergence, trainability, signal or gradient propagation, optimizer-induced solution selection, flatness or edge-of-stability dynamics, explicit regularizers or augmentations, or nonstandard credit-assignment/training paradigms.","exclusion":"If the optimizer or regularizer is mainly a tool for proving test-risk bounds, assign to Supervised generalization phenomena; if the main object is a kernel, infinite-width, or tractable limiting model rather than training mechanisms themselves, assign to Kernel, continuum, and tractable limits.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["early and global convergence of gradient descent via neuron activation partition analysis","theoretical analysis of global convergence in over-parameterized relu nets via activation matrix angular geometry","stochastic gradient descent noise as flatness-favoring regularization via loss landscape anisotropy"]},{"direction_index":2,"label":"D2","name":"Kernel, continuum, and tractable limits","description":"Asymptotic, linearized, kernel-regime, infinite-width/depth, mean-field, Gaussian-process, stochastic-process, and tractable-model analyses used to obtain analytically controlled descriptions of deep networks.","inclusion":"Include atoms whose primary objective is to characterize NTK or lazy-training regimes, GP or mean-field limits, continuum ODE/SDE/PDE descriptions, large-width asymptotics, or deep linear/gated linear models as mathematically tractable proxies.","exclusion":"If the main target is finite-network optimization behavior, assign to Training dynamics and regularization; if the target is raw representational capacity rather than limiting dynamics, assign to Expressivity and architectural structure.","nearest_boundary":"Training dynamics and regularization","representative_topics":["generalization bounds for overparameterized 3-layer neural tangent kernel models with hidden-layer interactions","spectral bias bounds for deep networks via fisher information low effective rank in kernel regime","convergence guarantees for empirical neural tangent kernel approximation in wide readout networks"]},{"direction_index":3,"label":"D3","name":"Expressivity and architectural structure","description":"Theory of what neural architectures and parameterizations can represent, approximate, or structurally bias, including approximation rates, universality, depth-width tradeoffs, architectural inductive priors, equivariance, attention, convolution, MoE, and ReLU combinatorial geometry.","inclusion":"Include atoms whose primary objective is representational capacity, approximation complexity, architectural inductive bias, invariance/equivariance, architecture-specific learnability advantages or costs, or geometric/combinatorial structure of network function classes.","exclusion":"If the main target is learned feature evolution during training, assign to Feature and representation learning; if architecture is used only as the setting for a generalization bound, assign to Supervised generalization phenomena.","nearest_boundary":"Feature and representation learning","representative_topics":["l^p norm approximation lower bounds for feed-forward neural networks via fat-shattering dimension","smoothness-constrained function approximation in overparameterized convolutional residual networks","depth and size bounds for monotone neural networks via interpolation and approximation"]},{"direction_index":4,"label":"D4","name":"Feature and representation learning","description":"Theory of how trained networks form, transform, organize, and geometrize internal representations, including feature learning beyond kernels, representation multiplicity, low-rank or low-dimensional features, and neural-collapse geometry.","inclusion":"Include atoms whose primary objective is the emergence, dynamics, or final geometry of internal features or representations, including kernel-to-feature-learning transitions, subspace recovery, representational multiplicity, rank structure, and class-mean/classifier collapse.","exclusion":"If the main object is an architectural prior rather than learned representation dynamics, assign to Expressivity and architectural structure; if the main object is optimizer solution selection in parameter space, assign to Training dynamics and regularization.","nearest_boundary":"Expressivity and architectural structure","representative_topics":["theoretical analysis of feature learning overfitting via sparse representations in fully-connected networks","gradient descent representations for efficient learning of sparse high-dimensional polynomials","sgd training of two-layer networks for low-dimensional subspace recovery in high-dimensional settings"]},{"direction_index":5,"label":"D5","name":"Probabilistic and generative models","description":"Theoretical foundations of Bayesian, probabilistic, variational, ensemble, diffusion, score-based, normalizing-flow, and other deep generative modeling approaches.","inclusion":"Include atoms whose primary objective is posterior inference, uncertainty, Bayesian neural-network theory, variational free energy, posterior geometry, score estimation, diffusion sampling, distribution recovery, or convergence/statistical theory of generative models.","exclusion":"If Bayesian or posterior quantities are mainly used to derive frequentist test-risk bounds, assign to Supervised generalization phenomena; if a generative architecture is studied only for representational capacity, assign to Expressivity and architectural structure.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["repriorisation for accelerated mcmc sampling in wide bayesian neural networks","variational inference of overparameterized bayesian neural networks via elbo term trade-off analysis","theoretical convergence analysis of score-based generative models with l2-score accuracy and non-log-concave distributions"]},{"direction_index":6,"label":"D6","name":"Robustness and perturbation stability","description":"Theory of neural-network behavior under adversarial perturbations, input corruptions, sensitivity, Lipschitz constraints, distributional shifts, and robustness-relevant classifier or metric geometry.","inclusion":"Include atoms whose primary objective is to quantify, certify, or explain robustness, adversarial vulnerability, transferability, perturbation stability, Lipschitz behavior, or robustness-relevant geometry of trained networks.","exclusion":"If perturbations are introduced primarily as a training regularizer or augmentation, assign to Training dynamics and regularization; if robustness quantities are only proof devices for ordinary test-risk bounds, assign to Supervised generalization phenomena.","nearest_boundary":"Training dynamics and regularization","representative_topics":["theoretical framework for explaining neural network generalization and adversarial example transfer","lipschitz bound tightening for cnns via convolutional layer unrolling and toeplitz decomposition","theoretical analysis of bayesian neural network robustness to adversarial attacks via posterior geometry"]},{"direction_index":7,"label":"D7","name":"Scaling and learning limits","description":"Theory of performance scaling, task learnability, grokking and phase transitions, computational-statistical gaps, hardness, impossibility, and fundamental limits of neural-network learning on structured task families.","inclusion":"Include atoms whose primary objective is to explain scaling laws, delayed generalization, phase transitions, dataset/model/compute scaling, task-level learnability limits, computational hardness, or impossibility of learning particular structures.","exclusion":"If the target is a conventional finite-sample generalization guarantee, assign to Supervised generalization phenomena; if the target is feature emergence on a learnable task rather than a scaling law or limit, assign to Feature and representation learning.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["formal language theory grounded prediction of neural network generalization limits via chomsky hierarchy","theoretical analysis of neural network learnability of physical concepts via recursion and active perception","statistical model of neural scaling laws via random feature maps and spectral power laws"]}],"split":"dev"}
{"episode_id":"F084_T2025-07","field_index":84,"field":"deep learning theory","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Supervised generalization phenomena","description":"Statistical theory of test performance in supervised deep learning, including risk bounds, excess-risk guarantees, sample complexity, structured-network generalization, interpolation-regime behavior, benign overfitting, and double descent.","inclusion":"Include atoms whose primary objective is to explain or bound supervised test risk, whether through PAC-Bayes, stability, complexity measures, sparsity, transfer/meta/active-learning settings, or interpolation-specific phenomena such as benign overfitting and double descent.","exclusion":"If the main objective is optimizer convergence or solution selection rather than test-risk behavior, assign to Training dynamics and regularization; if the objective is long-run scaling laws or task-level impossibility, assign to Scaling and learning limits.","nearest_boundary":"Training dynamics and regularization","representative_topics":["decision boundary variability as a generalization measure in deep learning theory","sgd-dependent generalization error bounds via parameter norm dynamics and rademacher complexity","theoretical analysis of benign overfitting failure under label noise in underparameterized regimes"]},{"direction_index":1,"label":"D1","name":"Training dynamics and regularization","description":"Theory of how training procedures, optimizer dynamics, initialization, normalization, sharpness, implicit bias, explicit interventions, and alternative learning rules make neural networks trainable and shape learned solutions.","inclusion":"Include atoms whose primary objective is convergence, trainability, signal or gradient propagation, optimizer-induced solution selection, flatness or edge-of-stability dynamics, explicit regularizers or augmentations, or nonstandard credit-assignment/training paradigms.","exclusion":"If the optimizer or regularizer is mainly a tool for proving test-risk bounds, assign to Supervised generalization phenomena; if the main object is a kernel, infinite-width, or tractable limiting model rather than training mechanisms themselves, assign to Kernel, continuum, and tractable limits.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["early and global convergence of gradient descent via neuron activation partition analysis","theoretical analysis of global convergence in over-parameterized relu nets via activation matrix angular geometry","stochastic gradient descent noise as flatness-favoring regularization via loss landscape anisotropy"]},{"direction_index":2,"label":"D2","name":"Kernel, continuum, and tractable limits","description":"Asymptotic, linearized, kernel-regime, infinite-width/depth, mean-field, Gaussian-process, stochastic-process, and tractable-model analyses used to obtain analytically controlled descriptions of deep networks.","inclusion":"Include atoms whose primary objective is to characterize NTK or lazy-training regimes, GP or mean-field limits, continuum ODE/SDE/PDE descriptions, large-width asymptotics, or deep linear/gated linear models as mathematically tractable proxies.","exclusion":"If the main target is finite-network optimization behavior, assign to Training dynamics and regularization; if the target is raw representational capacity rather than limiting dynamics, assign to Expressivity and architectural structure.","nearest_boundary":"Training dynamics and regularization","representative_topics":["generalization bounds for overparameterized 3-layer neural tangent kernel models with hidden-layer interactions","spectral bias bounds for deep networks via fisher information low effective rank in kernel regime","convergence guarantees for empirical neural tangent kernel approximation in wide readout networks"]},{"direction_index":3,"label":"D3","name":"Expressivity and architectural structure","description":"Theory of what neural architectures and parameterizations can represent, approximate, or structurally bias, including approximation rates, universality, depth-width tradeoffs, architectural inductive priors, equivariance, attention, convolution, MoE, and ReLU combinatorial geometry.","inclusion":"Include atoms whose primary objective is representational capacity, approximation complexity, architectural inductive bias, invariance/equivariance, architecture-specific learnability advantages or costs, or geometric/combinatorial structure of network function classes.","exclusion":"If the main target is learned feature evolution during training, assign to Feature and representation learning; if architecture is used only as the setting for a generalization bound, assign to Supervised generalization phenomena.","nearest_boundary":"Feature and representation learning","representative_topics":["l^p norm approximation lower bounds for feed-forward neural networks via fat-shattering dimension","smoothness-constrained function approximation in overparameterized convolutional residual networks","depth and size bounds for monotone neural networks via interpolation and approximation"]},{"direction_index":4,"label":"D4","name":"Feature and representation learning","description":"Theory of how trained networks form, transform, organize, and geometrize internal representations, including feature learning beyond kernels, representation multiplicity, low-rank or low-dimensional features, and neural-collapse geometry.","inclusion":"Include atoms whose primary objective is the emergence, dynamics, or final geometry of internal features or representations, including kernel-to-feature-learning transitions, subspace recovery, representational multiplicity, rank structure, and class-mean/classifier collapse.","exclusion":"If the main object is an architectural prior rather than learned representation dynamics, assign to Expressivity and architectural structure; if the main object is optimizer solution selection in parameter space, assign to Training dynamics and regularization.","nearest_boundary":"Expressivity and architectural structure","representative_topics":["theoretical analysis of feature learning overfitting via sparse representations in fully-connected networks","gradient descent representations for efficient learning of sparse high-dimensional polynomials","sgd training of two-layer networks for low-dimensional subspace recovery in high-dimensional settings"]},{"direction_index":5,"label":"D5","name":"Probabilistic and generative models","description":"Theoretical foundations of Bayesian, probabilistic, variational, ensemble, diffusion, score-based, normalizing-flow, and other deep generative modeling approaches.","inclusion":"Include atoms whose primary objective is posterior inference, uncertainty, Bayesian neural-network theory, variational free energy, posterior geometry, score estimation, diffusion sampling, distribution recovery, or convergence/statistical theory of generative models.","exclusion":"If Bayesian or posterior quantities are mainly used to derive frequentist test-risk bounds, assign to Supervised generalization phenomena; if a generative architecture is studied only for representational capacity, assign to Expressivity and architectural structure.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["repriorisation for accelerated mcmc sampling in wide bayesian neural networks","variational inference of overparameterized bayesian neural networks via elbo term trade-off analysis","theoretical convergence analysis of score-based generative models with l2-score accuracy and non-log-concave distributions"]},{"direction_index":6,"label":"D6","name":"Robustness and perturbation stability","description":"Theory of neural-network behavior under adversarial perturbations, input corruptions, sensitivity, Lipschitz constraints, distributional shifts, and robustness-relevant classifier or metric geometry.","inclusion":"Include atoms whose primary objective is to quantify, certify, or explain robustness, adversarial vulnerability, transferability, perturbation stability, Lipschitz behavior, or robustness-relevant geometry of trained networks.","exclusion":"If perturbations are introduced primarily as a training regularizer or augmentation, assign to Training dynamics and regularization; if robustness quantities are only proof devices for ordinary test-risk bounds, assign to Supervised generalization phenomena.","nearest_boundary":"Training dynamics and regularization","representative_topics":["theoretical framework for explaining neural network generalization and adversarial example transfer","lipschitz bound tightening for cnns via convolutional layer unrolling and toeplitz decomposition","theoretical analysis of bayesian neural network robustness to adversarial attacks via posterior geometry"]},{"direction_index":7,"label":"D7","name":"Scaling and learning limits","description":"Theory of performance scaling, task learnability, grokking and phase transitions, computational-statistical gaps, hardness, impossibility, and fundamental limits of neural-network learning on structured task families.","inclusion":"Include atoms whose primary objective is to explain scaling laws, delayed generalization, phase transitions, dataset/model/compute scaling, task-level learnability limits, computational hardness, or impossibility of learning particular structures.","exclusion":"If the target is a conventional finite-sample generalization guarantee, assign to Supervised generalization phenomena; if the target is feature emergence on a learnable task rather than a scaling law or limit, assign to Feature and representation learning.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["formal language theory grounded prediction of neural network generalization limits via chomsky hierarchy","theoretical analysis of neural network learnability of physical concepts via recursion and active perception","statistical model of neural scaling laws via random feature maps and spectral power laws"]}],"split":"dev"}
{"episode_id":"F084_T2026-01","field_index":84,"field":"deep learning theory","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Supervised generalization phenomena","description":"Statistical theory of test performance in supervised deep learning, including risk bounds, excess-risk guarantees, sample complexity, structured-network generalization, interpolation-regime behavior, benign overfitting, and double descent.","inclusion":"Include atoms whose primary objective is to explain or bound supervised test risk, whether through PAC-Bayes, stability, complexity measures, sparsity, transfer/meta/active-learning settings, or interpolation-specific phenomena such as benign overfitting and double descent.","exclusion":"If the main objective is optimizer convergence or solution selection rather than test-risk behavior, assign to Training dynamics and regularization; if the objective is long-run scaling laws or task-level impossibility, assign to Scaling and learning limits.","nearest_boundary":"Training dynamics and regularization","representative_topics":["decision boundary variability as a generalization measure in deep learning theory","sgd-dependent generalization error bounds via parameter norm dynamics and rademacher complexity","theoretical analysis of benign overfitting failure under label noise in underparameterized regimes"]},{"direction_index":1,"label":"D1","name":"Training dynamics and regularization","description":"Theory of how training procedures, optimizer dynamics, initialization, normalization, sharpness, implicit bias, explicit interventions, and alternative learning rules make neural networks trainable and shape learned solutions.","inclusion":"Include atoms whose primary objective is convergence, trainability, signal or gradient propagation, optimizer-induced solution selection, flatness or edge-of-stability dynamics, explicit regularizers or augmentations, or nonstandard credit-assignment/training paradigms.","exclusion":"If the optimizer or regularizer is mainly a tool for proving test-risk bounds, assign to Supervised generalization phenomena; if the main object is a kernel, infinite-width, or tractable limiting model rather than training mechanisms themselves, assign to Kernel, continuum, and tractable limits.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["early and global convergence of gradient descent via neuron activation partition analysis","theoretical analysis of global convergence in over-parameterized relu nets via activation matrix angular geometry","stochastic gradient descent noise as flatness-favoring regularization via loss landscape anisotropy"]},{"direction_index":2,"label":"D2","name":"Kernel, continuum, and tractable limits","description":"Asymptotic, linearized, kernel-regime, infinite-width/depth, mean-field, Gaussian-process, stochastic-process, and tractable-model analyses used to obtain analytically controlled descriptions of deep networks.","inclusion":"Include atoms whose primary objective is to characterize NTK or lazy-training regimes, GP or mean-field limits, continuum ODE/SDE/PDE descriptions, large-width asymptotics, or deep linear/gated linear models as mathematically tractable proxies.","exclusion":"If the main target is finite-network optimization behavior, assign to Training dynamics and regularization; if the target is raw representational capacity rather than limiting dynamics, assign to Expressivity and architectural structure.","nearest_boundary":"Training dynamics and regularization","representative_topics":["generalization bounds for overparameterized 3-layer neural tangent kernel models with hidden-layer interactions","spectral bias bounds for deep networks via fisher information low effective rank in kernel regime","convergence guarantees for empirical neural tangent kernel approximation in wide readout networks"]},{"direction_index":3,"label":"D3","name":"Expressivity and architectural structure","description":"Theory of what neural architectures and parameterizations can represent, approximate, or structurally bias, including approximation rates, universality, depth-width tradeoffs, architectural inductive priors, equivariance, attention, convolution, MoE, and ReLU combinatorial geometry.","inclusion":"Include atoms whose primary objective is representational capacity, approximation complexity, architectural inductive bias, invariance/equivariance, architecture-specific learnability advantages or costs, or geometric/combinatorial structure of network function classes.","exclusion":"If the main target is learned feature evolution during training, assign to Feature and representation learning; if architecture is used only as the setting for a generalization bound, assign to Supervised generalization phenomena.","nearest_boundary":"Feature and representation learning","representative_topics":["l^p norm approximation lower bounds for feed-forward neural networks via fat-shattering dimension","smoothness-constrained function approximation in overparameterized convolutional residual networks","depth and size bounds for monotone neural networks via interpolation and approximation"]},{"direction_index":4,"label":"D4","name":"Feature and representation learning","description":"Theory of how trained networks form, transform, organize, and geometrize internal representations, including feature learning beyond kernels, representation multiplicity, low-rank or low-dimensional features, and neural-collapse geometry.","inclusion":"Include atoms whose primary objective is the emergence, dynamics, or final geometry of internal features or representations, including kernel-to-feature-learning transitions, subspace recovery, representational multiplicity, rank structure, and class-mean/classifier collapse.","exclusion":"If the main object is an architectural prior rather than learned representation dynamics, assign to Expressivity and architectural structure; if the main object is optimizer solution selection in parameter space, assign to Training dynamics and regularization.","nearest_boundary":"Expressivity and architectural structure","representative_topics":["theoretical analysis of feature learning overfitting via sparse representations in fully-connected networks","gradient descent representations for efficient learning of sparse high-dimensional polynomials","sgd training of two-layer networks for low-dimensional subspace recovery in high-dimensional settings"]},{"direction_index":5,"label":"D5","name":"Probabilistic and generative models","description":"Theoretical foundations of Bayesian, probabilistic, variational, ensemble, diffusion, score-based, normalizing-flow, and other deep generative modeling approaches.","inclusion":"Include atoms whose primary objective is posterior inference, uncertainty, Bayesian neural-network theory, variational free energy, posterior geometry, score estimation, diffusion sampling, distribution recovery, or convergence/statistical theory of generative models.","exclusion":"If Bayesian or posterior quantities are mainly used to derive frequentist test-risk bounds, assign to Supervised generalization phenomena; if a generative architecture is studied only for representational capacity, assign to Expressivity and architectural structure.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["repriorisation for accelerated mcmc sampling in wide bayesian neural networks","variational inference of overparameterized bayesian neural networks via elbo term trade-off analysis","theoretical convergence analysis of score-based generative models with l2-score accuracy and non-log-concave distributions"]},{"direction_index":6,"label":"D6","name":"Robustness and perturbation stability","description":"Theory of neural-network behavior under adversarial perturbations, input corruptions, sensitivity, Lipschitz constraints, distributional shifts, and robustness-relevant classifier or metric geometry.","inclusion":"Include atoms whose primary objective is to quantify, certify, or explain robustness, adversarial vulnerability, transferability, perturbation stability, Lipschitz behavior, or robustness-relevant geometry of trained networks.","exclusion":"If perturbations are introduced primarily as a training regularizer or augmentation, assign to Training dynamics and regularization; if robustness quantities are only proof devices for ordinary test-risk bounds, assign to Supervised generalization phenomena.","nearest_boundary":"Training dynamics and regularization","representative_topics":["theoretical framework for explaining neural network generalization and adversarial example transfer","lipschitz bound tightening for cnns via convolutional layer unrolling and toeplitz decomposition","theoretical analysis of bayesian neural network robustness to adversarial attacks via posterior geometry"]},{"direction_index":7,"label":"D7","name":"Scaling and learning limits","description":"Theory of performance scaling, task learnability, grokking and phase transitions, computational-statistical gaps, hardness, impossibility, and fundamental limits of neural-network learning on structured task families.","inclusion":"Include atoms whose primary objective is to explain scaling laws, delayed generalization, phase transitions, dataset/model/compute scaling, task-level learnability limits, computational hardness, or impossibility of learning particular structures.","exclusion":"If the target is a conventional finite-sample generalization guarantee, assign to Supervised generalization phenomena; if the target is feature emergence on a learnable task rather than a scaling law or limit, assign to Feature and representation learning.","nearest_boundary":"Supervised generalization phenomena","representative_topics":["formal language theory grounded prediction of neural network generalization limits via chomsky hierarchy","theoretical analysis of neural network learnability of physical concepts via recursion and active perception","statistical model of neural scaling laws via random feature maps and spectral power laws"]}],"split":"dev"}
{"episode_id":"F091_T2024-01","field_index":91,"field":"uncertainty quantification","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Validity Guarantees and UQ Theory","description":"Develop formally justified uncertainty outputs and theoretical foundations for reliable uncertainty quantification, including distribution-free prediction sets, finite-sample confidence objects, non-asymptotic bounds, risk/coherence analyses, and theory of uncertainty behavior.","inclusion":"Include work whose primary objective is a provable coverage, confidence, risk, coherence, robustness, misspecification, or non-asymptotic property of an uncertainty procedure, especially conformal and other distribution-free methods.","exclusion":"If the main objective is empirical probability calibration without a formal validity guarantee, assign to Calibration and Reliability Correction; if the main objective is a posterior or predictive estimator with theory only as support, assign to Probabilistic Predictive and Posterior Modeling.","nearest_boundary":"Calibration and Reliability Correction","representative_topics":["modular conformal calibration framework for transforming non-probabilistic regression models into calibrated probabilistic models","split conformal prediction with kernel-based local conditional coverage approximation","nonparametric non-asymptotic confidence bands for band-limited functions via paley-wiener kernels"]},{"direction_index":1,"label":"D1","name":"Calibration and Reliability Correction","description":"Improve, measure, visualize, or correct the agreement between predicted probabilities, confidence scores, uncertainty magnitudes, and empirical outcomes or losses.","inclusion":"Include post-hoc recalibration, training-time calibration losses, calibration-error estimators, reliability diagrams, subgroup or instance-level calibration analysis, and multivariate or spatial recalibration when reliability correction is the primary objective.","exclusion":"If the central deliverable is a coverage-valid prediction set or interval, assign to Validity Guarantees and UQ Theory; if calibration is only an evaluation criterion for a broader estimator, assign to that estimator’s primary program.","nearest_boundary":"Validity Guarantees and UQ Theory","representative_topics":["loss-calibrated bayesian uncertainty estimation via expected utility maximisation for high-risk prediction","geometric-based post-hoc calibration for real-time classifier uncertainty estimation","parametric gaussian process recalibration for multivariate spatial uncertainty in object detection"]},{"direction_index":2,"label":"D2","name":"Probabilistic Predictive and Posterior Modeling","description":"Model predictive distributions, intervals, aleatoric/epistemic components, Bayesian posteriors, Gaussian-process surrogates, evidential or second-order uncertainty representations, and regression uncertainty when the core objective is estimating uncertainty rather than proving coverage or calibrating scores.","inclusion":"Include Bayesian neural-network inference, variational/MCMC/Laplace/dropout posterior approximations, GP and probabilistic surrogate uncertainty, heteroscedastic or quantile regression, non-conformal prediction intervals, evidential and belief-function methods, Dirichlet or subjective-logic second-order uncertainty, and predictive-density estimators.","exclusion":"If the primary contribution is an efficient ensemble mechanism, assign to Ensemble and Algorithmic UQ Mechanisms; if a scientific inverse problem or physics-informed structure defines the objective, assign to Scientific and Inverse-Problem UQ.","nearest_boundary":"Ensemble and Algorithmic UQ Mechanisms","representative_topics":["bayesian dropout for uncertainty estimation in transformer-based models","bayesian deep noise networks for predictive density estimation and uncertainty quantification in regression","epistemic deep learning using belief function random sets for set-valued uncertainty quantification"]},{"direction_index":3,"label":"D3","name":"Ensemble and Algorithmic UQ Mechanisms","description":"Design computational mechanisms, architectures, losses, wrappers, or approximations that produce uncertainty estimates efficiently or improve uncertainty quality beyond standard posterior, conformal, or calibration pipelines.","inclusion":"Include deep ensembles, bootstrapped or prior-function ensembles, layer or architecture-weight ensembles, ensemble distillation or compression, single-model approximations to ensemble behavior, generative or density-aware mechanisms, sum-product or rule-based uncertainty estimators, normalizing-flow UQ, and other non-standard algorithmic innovations whose mechanism is the main contribution.","exclusion":"If ensemble members are explicitly posterior samples and posterior inference is the research target, assign to Probabilistic Predictive and Posterior Modeling; if the mechanism is primarily for rejecting OOD inputs, assign to OOD Detection and Selective Prediction.","nearest_boundary":"Probabilistic Predictive and Posterior Modeling","representative_topics":["ensemble uncertainty estimation with prior functions and bootstrapping for joint prediction benefits","single-model epistemic uncertainty estimation via stochastic input bias centering and ntk analysis","infinitesimal jackknife covariance estimation for model combinations and ensemble uncertainty"]},{"direction_index":4,"label":"D4","name":"OOD Detection and Selective Prediction","description":"Use uncertainty or confidence to decide when an input, prediction, generated output, or operating condition should be flagged, rejected, abstained from, deferred, or treated as outside the reliable domain.","inclusion":"Include OOD detection, epistemic boundary estimation, distance-aware extrapolation uncertainty, selective classification or regression, failure prediction, reject-option methods, and confidence measures designed primarily for trust, abstention, or screening under shift.","exclusion":"If the primary objective is calibrated probabilities for all cases rather than a trust/reject decision, assign to Calibration and Reliability Correction; if the focus is explaining why uncertainty is high rather than acting on it, assign to Assessment, Interpretation, and Human-Facing UQ.","nearest_boundary":"Calibration and Reliability Correction","representative_topics":["single-label one-vs-all classification for uncertainty estimation and out-of-distribution detection","distance-aware prior calibration for bayesian neural network uncertainty outside training domain","uncertainty estimation and ood detection in on-policy deep reinforcement learning via pareto-optimized ensembles"]},{"direction_index":5,"label":"D5","name":"Scientific and Inverse-Problem UQ","description":"Quantify uncertainty in scientific machine learning, physical simulations, mechanistic models, differential equations, neural operators, surrogate models, data assimilation, and scientific or engineering inverse problems.","inclusion":"Include UQ for physics-informed models, neural PDEs, neural operators, scientific surrogates, multi-fidelity workflows, meteorological or cosmological inverse problems, process monitoring, nuclear or engineering prediction, and domain-structured parameter inference when the scientific or mechanistic objective defines the problem.","exclusion":"If the work is a general Bayesian, GP, or regression uncertainty method with only illustrative scientific examples, assign to Probabilistic Predictive and Posterior Modeling; if the core output is dense visual perception or reconstruction without dominant mechanistic structure, assign to Structured, Sequential, and Perception UQ.","nearest_boundary":"Probabilistic Predictive and Posterior Modeling","representative_topics":["physics-informed generative adversarial networks for uncertainty quantification in traffic state estimation","chilled sampling for laplace approximation in high-dimensional meteorological inverse problems","bayesian neural network inference of galaxy structural parameters with calibrated uncertainty quantification"]},{"direction_index":6,"label":"D6","name":"Structured, Sequential, and Perception UQ","description":"Quantify and propagate uncertainty for structured outputs, temporally dependent data, graph or multi-view data, perception systems, reconstruction pipelines, segmentation, localization, radiance fields, trajectories, and safety-critical visual or embodied systems.","inclusion":"Include uncertainty in dense prediction, medical or visual segmentation, object detection, image or MRI reconstruction, computational imaging, NeRFs, pose or gaze estimation, face verification pipelines, time-series and spatiotemporal forecasting, graph neural networks, multi-view fusion, trajectory prediction, and cascaded or autonomous-system uncertainty propagation.","exclusion":"If physical laws or scientific inverse modeling dominate the formulation, assign to Scientific and Inverse-Problem UQ; if the main objective is OOD rejection rather than structured-output uncertainty itself, assign to OOD Detection and Selective Prediction.","nearest_boundary":"Scientific and Inverse-Problem UQ","representative_topics":["trusted brain tumor segmentation via subjective logic uncertainty modeling and evidence aggregation","bayesian deep unrolling for epistemic and aleatoric uncertainty quantification in computational imaging","probabilistic medical image segmentation with generalized latent space distributions for uncertainty"]},{"direction_index":7,"label":"D7","name":"Assessment, Interpretation, and Human-Facing UQ","description":"Evaluate, benchmark, explain, communicate, elicit, or use uncertainty estimates in settings where understanding, auditing, comparing, or human interaction with uncertainty is the primary objective, including language-generation uncertainty.","inclusion":"Include UQ benchmarks, surveys, evaluation protocols, metric studies, uncertainty visualization, attribution and interpretability of uncertainty, fairness audits of uncertainty estimates, atypicality analysis, human-AI decision studies, uncertainty communication, reproducibility workflows, LLM confidence elicitation, semantic uncertainty, known-unknown expression, and LLM-specific UQ evaluation.","exclusion":"If the primary contribution is a new estimator whose benchmarks only validate it, assign to that estimator’s program; if language uncertainty is used mainly for a generic reject decision rather than LLM-specific elicitation or semantic uncertainty, assign to OOD Detection and Selective Prediction.","nearest_boundary":"OOD Detection and Selective Prediction","representative_topics":["architectural and training regime effects on uncertainty estimation and selective prediction in image classifiers","benchmarking bayesian neural network posterior approximation methods for regression uncertainty quantification","semantic entropy for uncertainty estimation in large language models via linguistic invariances"]}],"split":"dev"}
{"episode_id":"F091_T2024-07","field_index":91,"field":"uncertainty quantification","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Validity Guarantees and UQ Theory","description":"Develop formally justified uncertainty outputs and theoretical foundations for reliable uncertainty quantification, including distribution-free prediction sets, finite-sample confidence objects, non-asymptotic bounds, risk/coherence analyses, and theory of uncertainty behavior.","inclusion":"Include work whose primary objective is a provable coverage, confidence, risk, coherence, robustness, misspecification, or non-asymptotic property of an uncertainty procedure, especially conformal and other distribution-free methods.","exclusion":"If the main objective is empirical probability calibration without a formal validity guarantee, assign to Calibration and Reliability Correction; if the main objective is a posterior or predictive estimator with theory only as support, assign to Probabilistic Predictive and Posterior Modeling.","nearest_boundary":"Calibration and Reliability Correction","representative_topics":["modular conformal calibration framework for transforming non-probabilistic regression models into calibrated probabilistic models","split conformal prediction with kernel-based local conditional coverage approximation","nonparametric non-asymptotic confidence bands for band-limited functions via paley-wiener kernels"]},{"direction_index":1,"label":"D1","name":"Calibration and Reliability Correction","description":"Improve, measure, visualize, or correct the agreement between predicted probabilities, confidence scores, uncertainty magnitudes, and empirical outcomes or losses.","inclusion":"Include post-hoc recalibration, training-time calibration losses, calibration-error estimators, reliability diagrams, subgroup or instance-level calibration analysis, and multivariate or spatial recalibration when reliability correction is the primary objective.","exclusion":"If the central deliverable is a coverage-valid prediction set or interval, assign to Validity Guarantees and UQ Theory; if calibration is only an evaluation criterion for a broader estimator, assign to that estimator’s primary program.","nearest_boundary":"Validity Guarantees and UQ Theory","representative_topics":["loss-calibrated bayesian uncertainty estimation via expected utility maximisation for high-risk prediction","geometric-based post-hoc calibration for real-time classifier uncertainty estimation","parametric gaussian process recalibration for multivariate spatial uncertainty in object detection"]},{"direction_index":2,"label":"D2","name":"Probabilistic Predictive and Posterior Modeling","description":"Model predictive distributions, intervals, aleatoric/epistemic components, Bayesian posteriors, Gaussian-process surrogates, evidential or second-order uncertainty representations, and regression uncertainty when the core objective is estimating uncertainty rather than proving coverage or calibrating scores.","inclusion":"Include Bayesian neural-network inference, variational/MCMC/Laplace/dropout posterior approximations, GP and probabilistic surrogate uncertainty, heteroscedastic or quantile regression, non-conformal prediction intervals, evidential and belief-function methods, Dirichlet or subjective-logic second-order uncertainty, and predictive-density estimators.","exclusion":"If the primary contribution is an efficient ensemble mechanism, assign to Ensemble and Algorithmic UQ Mechanisms; if a scientific inverse problem or physics-informed structure defines the objective, assign to Scientific and Inverse-Problem UQ.","nearest_boundary":"Ensemble and Algorithmic UQ Mechanisms","representative_topics":["bayesian dropout for uncertainty estimation in transformer-based models","bayesian deep noise networks for predictive density estimation and uncertainty quantification in regression","epistemic deep learning using belief function random sets for set-valued uncertainty quantification"]},{"direction_index":3,"label":"D3","name":"Ensemble and Algorithmic UQ Mechanisms","description":"Design computational mechanisms, architectures, losses, wrappers, or approximations that produce uncertainty estimates efficiently or improve uncertainty quality beyond standard posterior, conformal, or calibration pipelines.","inclusion":"Include deep ensembles, bootstrapped or prior-function ensembles, layer or architecture-weight ensembles, ensemble distillation or compression, single-model approximations to ensemble behavior, generative or density-aware mechanisms, sum-product or rule-based uncertainty estimators, normalizing-flow UQ, and other non-standard algorithmic innovations whose mechanism is the main contribution.","exclusion":"If ensemble members are explicitly posterior samples and posterior inference is the research target, assign to Probabilistic Predictive and Posterior Modeling; if the mechanism is primarily for rejecting OOD inputs, assign to OOD Detection and Selective Prediction.","nearest_boundary":"Probabilistic Predictive and Posterior Modeling","representative_topics":["ensemble uncertainty estimation with prior functions and bootstrapping for joint prediction benefits","single-model epistemic uncertainty estimation via stochastic input bias centering and ntk analysis","infinitesimal jackknife covariance estimation for model combinations and ensemble uncertainty"]},{"direction_index":4,"label":"D4","name":"OOD Detection and Selective Prediction","description":"Use uncertainty or confidence to decide when an input, prediction, generated output, or operating condition should be flagged, rejected, abstained from, deferred, or treated as outside the reliable domain.","inclusion":"Include OOD detection, epistemic boundary estimation, distance-aware extrapolation uncertainty, selective classification or regression, failure prediction, reject-option methods, and confidence measures designed primarily for trust, abstention, or screening under shift.","exclusion":"If the primary objective is calibrated probabilities for all cases rather than a trust/reject decision, assign to Calibration and Reliability Correction; if the focus is explaining why uncertainty is high rather than acting on it, assign to Assessment, Interpretation, and Human-Facing UQ.","nearest_boundary":"Calibration and Reliability Correction","representative_topics":["single-label one-vs-all classification for uncertainty estimation and out-of-distribution detection","distance-aware prior calibration for bayesian neural network uncertainty outside training domain","uncertainty estimation and ood detection in on-policy deep reinforcement learning via pareto-optimized ensembles"]},{"direction_index":5,"label":"D5","name":"Scientific and Inverse-Problem UQ","description":"Quantify uncertainty in scientific machine learning, physical simulations, mechanistic models, differential equations, neural operators, surrogate models, data assimilation, and scientific or engineering inverse problems.","inclusion":"Include UQ for physics-informed models, neural PDEs, neural operators, scientific surrogates, multi-fidelity workflows, meteorological or cosmological inverse problems, process monitoring, nuclear or engineering prediction, and domain-structured parameter inference when the scientific or mechanistic objective defines the problem.","exclusion":"If the work is a general Bayesian, GP, or regression uncertainty method with only illustrative scientific examples, assign to Probabilistic Predictive and Posterior Modeling; if the core output is dense visual perception or reconstruction without dominant mechanistic structure, assign to Structured, Sequential, and Perception UQ.","nearest_boundary":"Probabilistic Predictive and Posterior Modeling","representative_topics":["physics-informed generative adversarial networks for uncertainty quantification in traffic state estimation","chilled sampling for laplace approximation in high-dimensional meteorological inverse problems","bayesian neural network inference of galaxy structural parameters with calibrated uncertainty quantification"]},{"direction_index":6,"label":"D6","name":"Structured, Sequential, and Perception UQ","description":"Quantify and propagate uncertainty for structured outputs, temporally dependent data, graph or multi-view data, perception systems, reconstruction pipelines, segmentation, localization, radiance fields, trajectories, and safety-critical visual or embodied systems.","inclusion":"Include uncertainty in dense prediction, medical or visual segmentation, object detection, image or MRI reconstruction, computational imaging, NeRFs, pose or gaze estimation, face verification pipelines, time-series and spatiotemporal forecasting, graph neural networks, multi-view fusion, trajectory prediction, and cascaded or autonomous-system uncertainty propagation.","exclusion":"If physical laws or scientific inverse modeling dominate the formulation, assign to Scientific and Inverse-Problem UQ; if the main objective is OOD rejection rather than structured-output uncertainty itself, assign to OOD Detection and Selective Prediction.","nearest_boundary":"Scientific and Inverse-Problem UQ","representative_topics":["trusted brain tumor segmentation via subjective logic uncertainty modeling and evidence aggregation","bayesian deep unrolling for epistemic and aleatoric uncertainty quantification in computational imaging","probabilistic medical image segmentation with generalized latent space distributions for uncertainty"]},{"direction_index":7,"label":"D7","name":"Assessment, Interpretation, and Human-Facing UQ","description":"Evaluate, benchmark, explain, communicate, elicit, or use uncertainty estimates in settings where understanding, auditing, comparing, or human interaction with uncertainty is the primary objective, including language-generation uncertainty.","inclusion":"Include UQ benchmarks, surveys, evaluation protocols, metric studies, uncertainty visualization, attribution and interpretability of uncertainty, fairness audits of uncertainty estimates, atypicality analysis, human-AI decision studies, uncertainty communication, reproducibility workflows, LLM confidence elicitation, semantic uncertainty, known-unknown expression, and LLM-specific UQ evaluation.","exclusion":"If the primary contribution is a new estimator whose benchmarks only validate it, assign to that estimator’s program; if language uncertainty is used mainly for a generic reject decision rather than LLM-specific elicitation or semantic uncertainty, assign to OOD Detection and Selective Prediction.","nearest_boundary":"OOD Detection and Selective Prediction","representative_topics":["architectural and training regime effects on uncertainty estimation and selective prediction in image classifiers","benchmarking bayesian neural network posterior approximation methods for regression uncertainty quantification","semantic entropy for uncertainty estimation in large language models via linguistic invariances"]}],"split":"dev"}
{"episode_id":"F091_T2025-01","field_index":91,"field":"uncertainty quantification","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Validity Guarantees and UQ Theory","description":"Develop formally justified uncertainty outputs and theoretical foundations for reliable uncertainty quantification, including distribution-free prediction sets, finite-sample confidence objects, non-asymptotic bounds, risk/coherence analyses, and theory of uncertainty behavior.","inclusion":"Include work whose primary objective is a provable coverage, confidence, risk, coherence, robustness, misspecification, or non-asymptotic property of an uncertainty procedure, especially conformal and other distribution-free methods.","exclusion":"If the main objective is empirical probability calibration without a formal validity guarantee, assign to Calibration and Reliability Correction; if the main objective is a posterior or predictive estimator with theory only as support, assign to Probabilistic Predictive and Posterior Modeling.","nearest_boundary":"Calibration and Reliability Correction","representative_topics":["modular conformal calibration framework for transforming non-probabilistic regression models into calibrated probabilistic models","split conformal prediction with kernel-based local conditional coverage approximation","nonparametric non-asymptotic confidence bands for band-limited functions via paley-wiener kernels"]},{"direction_index":1,"label":"D1","name":"Calibration and Reliability Correction","description":"Improve, measure, visualize, or correct the agreement between predicted probabilities, confidence scores, uncertainty magnitudes, and empirical outcomes or losses.","inclusion":"Include post-hoc recalibration, training-time calibration losses, calibration-error estimators, reliability diagrams, subgroup or instance-level calibration analysis, and multivariate or spatial recalibration when reliability correction is the primary objective.","exclusion":"If the central deliverable is a coverage-valid prediction set or interval, assign to Validity Guarantees and UQ Theory; if calibration is only an evaluation criterion for a broader estimator, assign to that estimator’s primary program.","nearest_boundary":"Validity Guarantees and UQ Theory","representative_topics":["loss-calibrated bayesian uncertainty estimation via expected utility maximisation for high-risk prediction","geometric-based post-hoc calibration for real-time classifier uncertainty estimation","parametric gaussian process recalibration for multivariate spatial uncertainty in object detection"]},{"direction_index":2,"label":"D2","name":"Probabilistic Predictive and Posterior Modeling","description":"Model predictive distributions, intervals, aleatoric/epistemic components, Bayesian posteriors, Gaussian-process surrogates, evidential or second-order uncertainty representations, and regression uncertainty when the core objective is estimating uncertainty rather than proving coverage or calibrating scores.","inclusion":"Include Bayesian neural-network inference, variational/MCMC/Laplace/dropout posterior approximations, GP and probabilistic surrogate uncertainty, heteroscedastic or quantile regression, non-conformal prediction intervals, evidential and belief-function methods, Dirichlet or subjective-logic second-order uncertainty, and predictive-density estimators.","exclusion":"If the primary contribution is an efficient ensemble mechanism, assign to Ensemble and Algorithmic UQ Mechanisms; if a scientific inverse problem or physics-informed structure defines the objective, assign to Scientific and Inverse-Problem UQ.","nearest_boundary":"Ensemble and Algorithmic UQ Mechanisms","representative_topics":["bayesian dropout for uncertainty estimation in transformer-based models","bayesian deep noise networks for predictive density estimation and uncertainty quantification in regression","epistemic deep learning using belief function random sets for set-valued uncertainty quantification"]},{"direction_index":3,"label":"D3","name":"Ensemble and Algorithmic UQ Mechanisms","description":"Design computational mechanisms, architectures, losses, wrappers, or approximations that produce uncertainty estimates efficiently or improve uncertainty quality beyond standard posterior, conformal, or calibration pipelines.","inclusion":"Include deep ensembles, bootstrapped or prior-function ensembles, layer or architecture-weight ensembles, ensemble distillation or compression, single-model approximations to ensemble behavior, generative or density-aware mechanisms, sum-product or rule-based uncertainty estimators, normalizing-flow UQ, and other non-standard algorithmic innovations whose mechanism is the main contribution.","exclusion":"If ensemble members are explicitly posterior samples and posterior inference is the research target, assign to Probabilistic Predictive and Posterior Modeling; if the mechanism is primarily for rejecting OOD inputs, assign to OOD Detection and Selective Prediction.","nearest_boundary":"Probabilistic Predictive and Posterior Modeling","representative_topics":["ensemble uncertainty estimation with prior functions and bootstrapping for joint prediction benefits","single-model epistemic uncertainty estimation via stochastic input bias centering and ntk analysis","infinitesimal jackknife covariance estimation for model combinations and ensemble uncertainty"]},{"direction_index":4,"label":"D4","name":"OOD Detection and Selective Prediction","description":"Use uncertainty or confidence to decide when an input, prediction, generated output, or operating condition should be flagged, rejected, abstained from, deferred, or treated as outside the reliable domain.","inclusion":"Include OOD detection, epistemic boundary estimation, distance-aware extrapolation uncertainty, selective classification or regression, failure prediction, reject-option methods, and confidence measures designed primarily for trust, abstention, or screening under shift.","exclusion":"If the primary objective is calibrated probabilities for all cases rather than a trust/reject decision, assign to Calibration and Reliability Correction; if the focus is explaining why uncertainty is high rather than acting on it, assign to Assessment, Interpretation, and Human-Facing UQ.","nearest_boundary":"Calibration and Reliability Correction","representative_topics":["single-label one-vs-all classification for uncertainty estimation and out-of-distribution detection","distance-aware prior calibration for bayesian neural network uncertainty outside training domain","uncertainty estimation and ood detection in on-policy deep reinforcement learning via pareto-optimized ensembles"]},{"direction_index":5,"label":"D5","name":"Scientific and Inverse-Problem UQ","description":"Quantify uncertainty in scientific machine learning, physical simulations, mechanistic models, differential equations, neural operators, surrogate models, data assimilation, and scientific or engineering inverse problems.","inclusion":"Include UQ for physics-informed models, neural PDEs, neural operators, scientific surrogates, multi-fidelity workflows, meteorological or cosmological inverse problems, process monitoring, nuclear or engineering prediction, and domain-structured parameter inference when the scientific or mechanistic objective defines the problem.","exclusion":"If the work is a general Bayesian, GP, or regression uncertainty method with only illustrative scientific examples, assign to Probabilistic Predictive and Posterior Modeling; if the core output is dense visual perception or reconstruction without dominant mechanistic structure, assign to Structured, Sequential, and Perception UQ.","nearest_boundary":"Probabilistic Predictive and Posterior Modeling","representative_topics":["physics-informed generative adversarial networks for uncertainty quantification in traffic state estimation","chilled sampling for laplace approximation in high-dimensional meteorological inverse problems","bayesian neural network inference of galaxy structural parameters with calibrated uncertainty quantification"]},{"direction_index":6,"label":"D6","name":"Structured, Sequential, and Perception UQ","description":"Quantify and propagate uncertainty for structured outputs, temporally dependent data, graph or multi-view data, perception systems, reconstruction pipelines, segmentation, localization, radiance fields, trajectories, and safety-critical visual or embodied systems.","inclusion":"Include uncertainty in dense prediction, medical or visual segmentation, object detection, image or MRI reconstruction, computational imaging, NeRFs, pose or gaze estimation, face verification pipelines, time-series and spatiotemporal forecasting, graph neural networks, multi-view fusion, trajectory prediction, and cascaded or autonomous-system uncertainty propagation.","exclusion":"If physical laws or scientific inverse modeling dominate the formulation, assign to Scientific and Inverse-Problem UQ; if the main objective is OOD rejection rather than structured-output uncertainty itself, assign to OOD Detection and Selective Prediction.","nearest_boundary":"Scientific and Inverse-Problem UQ","representative_topics":["trusted brain tumor segmentation via subjective logic uncertainty modeling and evidence aggregation","bayesian deep unrolling for epistemic and aleatoric uncertainty quantification in computational imaging","probabilistic medical image segmentation with generalized latent space distributions for uncertainty"]},{"direction_index":7,"label":"D7","name":"Assessment, Interpretation, and Human-Facing UQ","description":"Evaluate, benchmark, explain, communicate, elicit, or use uncertainty estimates in settings where understanding, auditing, comparing, or human interaction with uncertainty is the primary objective, including language-generation uncertainty.","inclusion":"Include UQ benchmarks, surveys, evaluation protocols, metric studies, uncertainty visualization, attribution and interpretability of uncertainty, fairness audits of uncertainty estimates, atypicality analysis, human-AI decision studies, uncertainty communication, reproducibility workflows, LLM confidence elicitation, semantic uncertainty, known-unknown expression, and LLM-specific UQ evaluation.","exclusion":"If the primary contribution is a new estimator whose benchmarks only validate it, assign to that estimator’s program; if language uncertainty is used mainly for a generic reject decision rather than LLM-specific elicitation or semantic uncertainty, assign to OOD Detection and Selective Prediction.","nearest_boundary":"OOD Detection and Selective Prediction","representative_topics":["architectural and training regime effects on uncertainty estimation and selective prediction in image classifiers","benchmarking bayesian neural network posterior approximation methods for regression uncertainty quantification","semantic entropy for uncertainty estimation in large language models via linguistic invariances"]}],"split":"dev"}
{"episode_id":"F091_T2025-07","field_index":91,"field":"uncertainty quantification","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Validity Guarantees and UQ Theory","description":"Develop formally justified uncertainty outputs and theoretical foundations for reliable uncertainty quantification, including distribution-free prediction sets, finite-sample confidence objects, non-asymptotic bounds, risk/coherence analyses, and theory of uncertainty behavior.","inclusion":"Include work whose primary objective is a provable coverage, confidence, risk, coherence, robustness, misspecification, or non-asymptotic property of an uncertainty procedure, especially conformal and other distribution-free methods.","exclusion":"If the main objective is empirical probability calibration without a formal validity guarantee, assign to Calibration and Reliability Correction; if the main objective is a posterior or predictive estimator with theory only as support, assign to Probabilistic Predictive and Posterior Modeling.","nearest_boundary":"Calibration and Reliability Correction","representative_topics":["modular conformal calibration framework for transforming non-probabilistic regression models into calibrated probabilistic models","split conformal prediction with kernel-based local conditional coverage approximation","nonparametric non-asymptotic confidence bands for band-limited functions via paley-wiener kernels"]},{"direction_index":1,"label":"D1","name":"Calibration and Reliability Correction","description":"Improve, measure, visualize, or correct the agreement between predicted probabilities, confidence scores, uncertainty magnitudes, and empirical outcomes or losses.","inclusion":"Include post-hoc recalibration, training-time calibration losses, calibration-error estimators, reliability diagrams, subgroup or instance-level calibration analysis, and multivariate or spatial recalibration when reliability correction is the primary objective.","exclusion":"If the central deliverable is a coverage-valid prediction set or interval, assign to Validity Guarantees and UQ Theory; if calibration is only an evaluation criterion for a broader estimator, assign to that estimator’s primary program.","nearest_boundary":"Validity Guarantees and UQ Theory","representative_topics":["loss-calibrated bayesian uncertainty estimation via expected utility maximisation for high-risk prediction","geometric-based post-hoc calibration for real-time classifier uncertainty estimation","parametric gaussian process recalibration for multivariate spatial uncertainty in object detection"]},{"direction_index":2,"label":"D2","name":"Probabilistic Predictive and Posterior Modeling","description":"Model predictive distributions, intervals, aleatoric/epistemic components, Bayesian posteriors, Gaussian-process surrogates, evidential or second-order uncertainty representations, and regression uncertainty when the core objective is estimating uncertainty rather than proving coverage or calibrating scores.","inclusion":"Include Bayesian neural-network inference, variational/MCMC/Laplace/dropout posterior approximations, GP and probabilistic surrogate uncertainty, heteroscedastic or quantile regression, non-conformal prediction intervals, evidential and belief-function methods, Dirichlet or subjective-logic second-order uncertainty, and predictive-density estimators.","exclusion":"If the primary contribution is an efficient ensemble mechanism, assign to Ensemble and Algorithmic UQ Mechanisms; if a scientific inverse problem or physics-informed structure defines the objective, assign to Scientific and Inverse-Problem UQ.","nearest_boundary":"Ensemble and Algorithmic UQ Mechanisms","representative_topics":["bayesian dropout for uncertainty estimation in transformer-based models","bayesian deep noise networks for predictive density estimation and uncertainty quantification in regression","epistemic deep learning using belief function random sets for set-valued uncertainty quantification"]},{"direction_index":3,"label":"D3","name":"Ensemble and Algorithmic UQ Mechanisms","description":"Design computational mechanisms, architectures, losses, wrappers, or approximations that produce uncertainty estimates efficiently or improve uncertainty quality beyond standard posterior, conformal, or calibration pipelines.","inclusion":"Include deep ensembles, bootstrapped or prior-function ensembles, layer or architecture-weight ensembles, ensemble distillation or compression, single-model approximations to ensemble behavior, generative or density-aware mechanisms, sum-product or rule-based uncertainty estimators, normalizing-flow UQ, and other non-standard algorithmic innovations whose mechanism is the main contribution.","exclusion":"If ensemble members are explicitly posterior samples and posterior inference is the research target, assign to Probabilistic Predictive and Posterior Modeling; if the mechanism is primarily for rejecting OOD inputs, assign to OOD Detection and Selective Prediction.","nearest_boundary":"Probabilistic Predictive and Posterior Modeling","representative_topics":["ensemble uncertainty estimation with prior functions and bootstrapping for joint prediction benefits","single-model epistemic uncertainty estimation via stochastic input bias centering and ntk analysis","infinitesimal jackknife covariance estimation for model combinations and ensemble uncertainty"]},{"direction_index":4,"label":"D4","name":"OOD Detection and Selective Prediction","description":"Use uncertainty or confidence to decide when an input, prediction, generated output, or operating condition should be flagged, rejected, abstained from, deferred, or treated as outside the reliable domain.","inclusion":"Include OOD detection, epistemic boundary estimation, distance-aware extrapolation uncertainty, selective classification or regression, failure prediction, reject-option methods, and confidence measures designed primarily for trust, abstention, or screening under shift.","exclusion":"If the primary objective is calibrated probabilities for all cases rather than a trust/reject decision, assign to Calibration and Reliability Correction; if the focus is explaining why uncertainty is high rather than acting on it, assign to Assessment, Interpretation, and Human-Facing UQ.","nearest_boundary":"Calibration and Reliability Correction","representative_topics":["single-label one-vs-all classification for uncertainty estimation and out-of-distribution detection","distance-aware prior calibration for bayesian neural network uncertainty outside training domain","uncertainty estimation and ood detection in on-policy deep reinforcement learning via pareto-optimized ensembles"]},{"direction_index":5,"label":"D5","name":"Scientific and Inverse-Problem UQ","description":"Quantify uncertainty in scientific machine learning, physical simulations, mechanistic models, differential equations, neural operators, surrogate models, data assimilation, and scientific or engineering inverse problems.","inclusion":"Include UQ for physics-informed models, neural PDEs, neural operators, scientific surrogates, multi-fidelity workflows, meteorological or cosmological inverse problems, process monitoring, nuclear or engineering prediction, and domain-structured parameter inference when the scientific or mechanistic objective defines the problem.","exclusion":"If the work is a general Bayesian, GP, or regression uncertainty method with only illustrative scientific examples, assign to Probabilistic Predictive and Posterior Modeling; if the core output is dense visual perception or reconstruction without dominant mechanistic structure, assign to Structured, Sequential, and Perception UQ.","nearest_boundary":"Probabilistic Predictive and Posterior Modeling","representative_topics":["physics-informed generative adversarial networks for uncertainty quantification in traffic state estimation","chilled sampling for laplace approximation in high-dimensional meteorological inverse problems","bayesian neural network inference of galaxy structural parameters with calibrated uncertainty quantification"]},{"direction_index":6,"label":"D6","name":"Structured, Sequential, and Perception UQ","description":"Quantify and propagate uncertainty for structured outputs, temporally dependent data, graph or multi-view data, perception systems, reconstruction pipelines, segmentation, localization, radiance fields, trajectories, and safety-critical visual or embodied systems.","inclusion":"Include uncertainty in dense prediction, medical or visual segmentation, object detection, image or MRI reconstruction, computational imaging, NeRFs, pose or gaze estimation, face verification pipelines, time-series and spatiotemporal forecasting, graph neural networks, multi-view fusion, trajectory prediction, and cascaded or autonomous-system uncertainty propagation.","exclusion":"If physical laws or scientific inverse modeling dominate the formulation, assign to Scientific and Inverse-Problem UQ; if the main objective is OOD rejection rather than structured-output uncertainty itself, assign to OOD Detection and Selective Prediction.","nearest_boundary":"Scientific and Inverse-Problem UQ","representative_topics":["trusted brain tumor segmentation via subjective logic uncertainty modeling and evidence aggregation","bayesian deep unrolling for epistemic and aleatoric uncertainty quantification in computational imaging","probabilistic medical image segmentation with generalized latent space distributions for uncertainty"]},{"direction_index":7,"label":"D7","name":"Assessment, Interpretation, and Human-Facing UQ","description":"Evaluate, benchmark, explain, communicate, elicit, or use uncertainty estimates in settings where understanding, auditing, comparing, or human interaction with uncertainty is the primary objective, including language-generation uncertainty.","inclusion":"Include UQ benchmarks, surveys, evaluation protocols, metric studies, uncertainty visualization, attribution and interpretability of uncertainty, fairness audits of uncertainty estimates, atypicality analysis, human-AI decision studies, uncertainty communication, reproducibility workflows, LLM confidence elicitation, semantic uncertainty, known-unknown expression, and LLM-specific UQ evaluation.","exclusion":"If the primary contribution is a new estimator whose benchmarks only validate it, assign to that estimator’s program; if language uncertainty is used mainly for a generic reject decision rather than LLM-specific elicitation or semantic uncertainty, assign to OOD Detection and Selective Prediction.","nearest_boundary":"OOD Detection and Selective Prediction","representative_topics":["architectural and training regime effects on uncertainty estimation and selective prediction in image classifiers","benchmarking bayesian neural network posterior approximation methods for regression uncertainty quantification","semantic entropy for uncertainty estimation in large language models via linguistic invariances"]}],"split":"dev"}
{"episode_id":"F091_T2026-01","field_index":91,"field":"uncertainty quantification","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Validity Guarantees and UQ Theory","description":"Develop formally justified uncertainty outputs and theoretical foundations for reliable uncertainty quantification, including distribution-free prediction sets, finite-sample confidence objects, non-asymptotic bounds, risk/coherence analyses, and theory of uncertainty behavior.","inclusion":"Include work whose primary objective is a provable coverage, confidence, risk, coherence, robustness, misspecification, or non-asymptotic property of an uncertainty procedure, especially conformal and other distribution-free methods.","exclusion":"If the main objective is empirical probability calibration without a formal validity guarantee, assign to Calibration and Reliability Correction; if the main objective is a posterior or predictive estimator with theory only as support, assign to Probabilistic Predictive and Posterior Modeling.","nearest_boundary":"Calibration and Reliability Correction","representative_topics":["modular conformal calibration framework for transforming non-probabilistic regression models into calibrated probabilistic models","split conformal prediction with kernel-based local conditional coverage approximation","nonparametric non-asymptotic confidence bands for band-limited functions via paley-wiener kernels"]},{"direction_index":1,"label":"D1","name":"Calibration and Reliability Correction","description":"Improve, measure, visualize, or correct the agreement between predicted probabilities, confidence scores, uncertainty magnitudes, and empirical outcomes or losses.","inclusion":"Include post-hoc recalibration, training-time calibration losses, calibration-error estimators, reliability diagrams, subgroup or instance-level calibration analysis, and multivariate or spatial recalibration when reliability correction is the primary objective.","exclusion":"If the central deliverable is a coverage-valid prediction set or interval, assign to Validity Guarantees and UQ Theory; if calibration is only an evaluation criterion for a broader estimator, assign to that estimator’s primary program.","nearest_boundary":"Validity Guarantees and UQ Theory","representative_topics":["loss-calibrated bayesian uncertainty estimation via expected utility maximisation for high-risk prediction","geometric-based post-hoc calibration for real-time classifier uncertainty estimation","parametric gaussian process recalibration for multivariate spatial uncertainty in object detection"]},{"direction_index":2,"label":"D2","name":"Probabilistic Predictive and Posterior Modeling","description":"Model predictive distributions, intervals, aleatoric/epistemic components, Bayesian posteriors, Gaussian-process surrogates, evidential or second-order uncertainty representations, and regression uncertainty when the core objective is estimating uncertainty rather than proving coverage or calibrating scores.","inclusion":"Include Bayesian neural-network inference, variational/MCMC/Laplace/dropout posterior approximations, GP and probabilistic surrogate uncertainty, heteroscedastic or quantile regression, non-conformal prediction intervals, evidential and belief-function methods, Dirichlet or subjective-logic second-order uncertainty, and predictive-density estimators.","exclusion":"If the primary contribution is an efficient ensemble mechanism, assign to Ensemble and Algorithmic UQ Mechanisms; if a scientific inverse problem or physics-informed structure defines the objective, assign to Scientific and Inverse-Problem UQ.","nearest_boundary":"Ensemble and Algorithmic UQ Mechanisms","representative_topics":["bayesian dropout for uncertainty estimation in transformer-based models","bayesian deep noise networks for predictive density estimation and uncertainty quantification in regression","epistemic deep learning using belief function random sets for set-valued uncertainty quantification"]},{"direction_index":3,"label":"D3","name":"Ensemble and Algorithmic UQ Mechanisms","description":"Design computational mechanisms, architectures, losses, wrappers, or approximations that produce uncertainty estimates efficiently or improve uncertainty quality beyond standard posterior, conformal, or calibration pipelines.","inclusion":"Include deep ensembles, bootstrapped or prior-function ensembles, layer or architecture-weight ensembles, ensemble distillation or compression, single-model approximations to ensemble behavior, generative or density-aware mechanisms, sum-product or rule-based uncertainty estimators, normalizing-flow UQ, and other non-standard algorithmic innovations whose mechanism is the main contribution.","exclusion":"If ensemble members are explicitly posterior samples and posterior inference is the research target, assign to Probabilistic Predictive and Posterior Modeling; if the mechanism is primarily for rejecting OOD inputs, assign to OOD Detection and Selective Prediction.","nearest_boundary":"Probabilistic Predictive and Posterior Modeling","representative_topics":["ensemble uncertainty estimation with prior functions and bootstrapping for joint prediction benefits","single-model epistemic uncertainty estimation via stochastic input bias centering and ntk analysis","infinitesimal jackknife covariance estimation for model combinations and ensemble uncertainty"]},{"direction_index":4,"label":"D4","name":"OOD Detection and Selective Prediction","description":"Use uncertainty or confidence to decide when an input, prediction, generated output, or operating condition should be flagged, rejected, abstained from, deferred, or treated as outside the reliable domain.","inclusion":"Include OOD detection, epistemic boundary estimation, distance-aware extrapolation uncertainty, selective classification or regression, failure prediction, reject-option methods, and confidence measures designed primarily for trust, abstention, or screening under shift.","exclusion":"If the primary objective is calibrated probabilities for all cases rather than a trust/reject decision, assign to Calibration and Reliability Correction; if the focus is explaining why uncertainty is high rather than acting on it, assign to Assessment, Interpretation, and Human-Facing UQ.","nearest_boundary":"Calibration and Reliability Correction","representative_topics":["single-label one-vs-all classification for uncertainty estimation and out-of-distribution detection","distance-aware prior calibration for bayesian neural network uncertainty outside training domain","uncertainty estimation and ood detection in on-policy deep reinforcement learning via pareto-optimized ensembles"]},{"direction_index":5,"label":"D5","name":"Scientific and Inverse-Problem UQ","description":"Quantify uncertainty in scientific machine learning, physical simulations, mechanistic models, differential equations, neural operators, surrogate models, data assimilation, and scientific or engineering inverse problems.","inclusion":"Include UQ for physics-informed models, neural PDEs, neural operators, scientific surrogates, multi-fidelity workflows, meteorological or cosmological inverse problems, process monitoring, nuclear or engineering prediction, and domain-structured parameter inference when the scientific or mechanistic objective defines the problem.","exclusion":"If the work is a general Bayesian, GP, or regression uncertainty method with only illustrative scientific examples, assign to Probabilistic Predictive and Posterior Modeling; if the core output is dense visual perception or reconstruction without dominant mechanistic structure, assign to Structured, Sequential, and Perception UQ.","nearest_boundary":"Probabilistic Predictive and Posterior Modeling","representative_topics":["physics-informed generative adversarial networks for uncertainty quantification in traffic state estimation","chilled sampling for laplace approximation in high-dimensional meteorological inverse problems","bayesian neural network inference of galaxy structural parameters with calibrated uncertainty quantification"]},{"direction_index":6,"label":"D6","name":"Structured, Sequential, and Perception UQ","description":"Quantify and propagate uncertainty for structured outputs, temporally dependent data, graph or multi-view data, perception systems, reconstruction pipelines, segmentation, localization, radiance fields, trajectories, and safety-critical visual or embodied systems.","inclusion":"Include uncertainty in dense prediction, medical or visual segmentation, object detection, image or MRI reconstruction, computational imaging, NeRFs, pose or gaze estimation, face verification pipelines, time-series and spatiotemporal forecasting, graph neural networks, multi-view fusion, trajectory prediction, and cascaded or autonomous-system uncertainty propagation.","exclusion":"If physical laws or scientific inverse modeling dominate the formulation, assign to Scientific and Inverse-Problem UQ; if the main objective is OOD rejection rather than structured-output uncertainty itself, assign to OOD Detection and Selective Prediction.","nearest_boundary":"Scientific and Inverse-Problem UQ","representative_topics":["trusted brain tumor segmentation via subjective logic uncertainty modeling and evidence aggregation","bayesian deep unrolling for epistemic and aleatoric uncertainty quantification in computational imaging","probabilistic medical image segmentation with generalized latent space distributions for uncertainty"]},{"direction_index":7,"label":"D7","name":"Assessment, Interpretation, and Human-Facing UQ","description":"Evaluate, benchmark, explain, communicate, elicit, or use uncertainty estimates in settings where understanding, auditing, comparing, or human interaction with uncertainty is the primary objective, including language-generation uncertainty.","inclusion":"Include UQ benchmarks, surveys, evaluation protocols, metric studies, uncertainty visualization, attribution and interpretability of uncertainty, fairness audits of uncertainty estimates, atypicality analysis, human-AI decision studies, uncertainty communication, reproducibility workflows, LLM confidence elicitation, semantic uncertainty, known-unknown expression, and LLM-specific UQ evaluation.","exclusion":"If the primary contribution is a new estimator whose benchmarks only validate it, assign to that estimator’s program; if language uncertainty is used mainly for a generic reject decision rather than LLM-specific elicitation or semantic uncertainty, assign to OOD Detection and Selective Prediction.","nearest_boundary":"OOD Detection and Selective Prediction","representative_topics":["architectural and training regime effects on uncertainty estimation and selective prediction in image classifiers","benchmarking bayesian neural network posterior approximation methods for regression uncertainty quantification","semantic entropy for uncertainty estimation in large language models via linguistic invariances"]}],"split":"dev"}
{"episode_id":"F093_T2024-01","field_index":93,"field":"reinforcement learning theory","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Policy and Value Algorithm Convergence","description":"Theoretical guarantees for core model-free RL optimization and evaluation algorithms, including policy-gradient, natural-gradient, actor-critic, TD, Q-learning, value iteration, replay, and function-approximation dynamics.","inclusion":"Include atoms whose primary objective is proving convergence, stability, finite-time rates, sample complexity, global optimality, or landscape/generalization properties of policy-optimization or Bellman/value-learning algorithms.","exclusion":"If the same algorithms are analyzed mainly to satisfy constraints, handle fixed offline data, learn equilibria, achieve robustness, or adapt to non-stationarity, assign to that primary objective instead.","nearest_boundary":"D1","representative_topics":["stabilizing q-learning via modular algorithmic tools for provably efficient linear mdp learning","finite-time complexity analysis of entropy-regularized neural natural actor-critic algorithms","temporal difference learning dynamics and generalization in deep reinforcement learning"]},{"direction_index":1,"label":"D1","name":"Online Sample-Efficient Learning","description":"Regret, PAC, minimax, Bayesian, structural, reward-free, and exploration theory for learning efficiently through online interaction with single-agent MDPs and related function-approximation classes.","inclusion":"Include atoms whose primary objective is characterizing or improving online sample efficiency, exploration, regret, learnability, structural complexity, or function-approximation guarantees in stationary single-agent RL.","exclusion":"Offline fixed-dataset limitations belong to D2; safety/robustness/risk objectives belong to D3; non-stationary or adversarial environments belong to D4; representation learning as the main object belongs to D6.","nearest_boundary":"D6","representative_topics":["low-rank structure assumptions for sample-efficient policy learning in long-horizon mdps","optimistic posterior sampling for model-based rl with hellinger-distance regret analysis","near-minimax optimal reinforcement learning via linear function approximation and bernstein self-normalized exploration"]},{"direction_index":2,"label":"D2","name":"Offline and Off-Policy Inference","description":"Theory for policy learning, evaluation, selection, and estimation from fixed or previously collected data, including pessimism, conservatism, coverage limits, off-policy evaluation, and hybrid offline-online efficiency.","inclusion":"Include atoms whose primary objective is overcoming distribution shift, coverage, estimation, or policy-selection limitations caused by learning from non-interactive or logged data.","exclusion":"If the primary objective is explicit worst-case robustness under uncertainty sets, assign to D3; if equilibrium learning in games is central, assign to D5; if hidden-state inference is central, assign to D6.","nearest_boundary":"D3","representative_topics":["near-optimal primal-dual offline learning in constrained markov decision processes","theoretical efficiency of offline reinforcement learning with differentiable function class approximation","hybrid q-learning exploiting offline dataset quality and bilinear rank for coverage-free efficient policy learning"]},{"direction_index":3,"label":"D3","name":"Reliable RL under Constraints and Uncertainty","description":"Theory for RL objectives requiring safety, feasibility, robustness, distributional robustness, adversarial-corruption tolerance, or risk-sensitive control under uncertainty in costs, transitions, rewards, or returns.","inclusion":"Include atoms whose primary objective is satisfying constraints, bounding violations, optimizing under explicit uncertainty sets or misspecification, learning robust policies, or optimizing risk/return-distribution criteria.","exclusion":"Non-stationary/adversarial changes over time belong to D4; fixed-dataset pessimism without a robustness formulation belongs to D2; reward/specification design without reliability as the main objective belongs to D7.","nearest_boundary":"D2","representative_topics":["dual methods for constrained markov decision processes with linear convergence guarantees","near-minimax optimal sample complexity bounds for constrained mdp policy learning","uncertainty set regularization for robust policy generalization in continuous control reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Adaptation across Shifts and Tasks","description":"Theory for adapting policies, representations, or knowledge across changing environments, adversarial/non-stationary feedback, continual streams, task distributions, lifelong learning, transfer, meta-RL, and multi-task settings.","inclusion":"Include atoms whose primary objective is adaptation across time, contexts, tasks, or task distributions, including dynamic regret, detection of changes, transfer guarantees, policy reuse, and meta-learning sample complexity.","exclusion":"Single stationary-task online sample efficiency belongs to D1; single-task representation learning without transfer belongs to D6; strategic multi-agent interaction belongs to D5.","nearest_boundary":"D1","representative_topics":["lifelong reinforcement learning with provably efficient multi-task policy learning via value distillation","theoretical analysis of multitask representation learning benefits in low-rank reinforcement learning","pac bounds for meta rl via density estimation and manifold-aware task distribution learning"]},{"direction_index":5,"label":"D5","name":"Multi-Agent and Game-Theoretic RL","description":"Theory for learning, optimizing, or computing equilibria and decentralized/cooperative policies in Markov games, stochastic games, mean-field settings, competitive RL, and distributed multi-agent RL.","inclusion":"Include atoms whose primary objective is strategic interaction, equilibrium learning, multi-agent regret/sample complexity, decentralized coordination, opponent interaction, or game-theoretic computation.","exclusion":"Single-agent robust minimax formulations without strategic learners belong to D3; single-agent adversarial losses belong to D4; offline data issues are secondary when equilibrium learning is central.","nearest_boundary":"D3","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","byzantine-robust distributed reinforcement learning via robust mean estimation and pessimistic/optimistic value iteration","offline nash equilibrium learning in zero-sum markov games via pessimistic model-based confidence bounds"]},{"direction_index":6,"label":"D6","name":"State, Representation, and Model Structure","description":"Foundations for discovering, compressing, representing, or planning with internal structure in RL, including latent state, partial observability, abstractions, causal/invariant representations, learned models, world models, options, and hierarchical decomposition.","inclusion":"Include atoms whose primary objective is overcoming hidden or high-dimensional information structure, learning useful representations or abstractions, constructing value-equivalent or model-based representations, or using hierarchical/model-based structure for generalization and planning.","exclusion":"If representation is mainly used for online regret under a specified structural class, assign to D1; if it is mainly shared across tasks, assign to D4; if formal task syntax is central, assign to D7.","nearest_boundary":"D1","representative_topics":["rate-distortion-based approximate value equivalence for tractable model-based reinforcement learning","value-equivalent model compression via rate-distortion for capacity-constrained model-based reinforcement learning","oracle-free reinforcement learning in observable pomdps via barycentric spanner policy covers"]},{"direction_index":7,"label":"D7","name":"Objective and Specification Design","description":"Theory for specifying, transforming, learning, or synthesizing RL objectives, including reward shaping, preference and demonstration objectives, utility axioms, nonstandard discounting, generalized sequential objectives, temporal logic, automata, and formal task specifications.","inclusion":"Include atoms whose primary objective is defining what the agent should optimize or satisfy, learning that objective from feedback, proving equivalence or validity of reward/objective transformations, or synthesizing policies from formal/hierarchical specifications.","exclusion":"Risk measures over return distributions belong to D3; algorithmic convergence for a fixed objective belongs to D0; ordinary CMDP constraints without formal objective syntax belong to D3.","nearest_boundary":"D3","representative_topics":["theoretical analysis of outcome-conditioned imitation learning and its relation to reward maximization","axiomatic utility characterization for sequential decision making and reward hypothesis","discrete calculus on mdps formalizing potential shaping as gradient-based reward equivalence"]}],"split":"dev"}
{"episode_id":"F093_T2024-07","field_index":93,"field":"reinforcement learning theory","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Policy and Value Algorithm Convergence","description":"Theoretical guarantees for core model-free RL optimization and evaluation algorithms, including policy-gradient, natural-gradient, actor-critic, TD, Q-learning, value iteration, replay, and function-approximation dynamics.","inclusion":"Include atoms whose primary objective is proving convergence, stability, finite-time rates, sample complexity, global optimality, or landscape/generalization properties of policy-optimization or Bellman/value-learning algorithms.","exclusion":"If the same algorithms are analyzed mainly to satisfy constraints, handle fixed offline data, learn equilibria, achieve robustness, or adapt to non-stationarity, assign to that primary objective instead.","nearest_boundary":"D1","representative_topics":["stabilizing q-learning via modular algorithmic tools for provably efficient linear mdp learning","finite-time complexity analysis of entropy-regularized neural natural actor-critic algorithms","temporal difference learning dynamics and generalization in deep reinforcement learning"]},{"direction_index":1,"label":"D1","name":"Online Sample-Efficient Learning","description":"Regret, PAC, minimax, Bayesian, structural, reward-free, and exploration theory for learning efficiently through online interaction with single-agent MDPs and related function-approximation classes.","inclusion":"Include atoms whose primary objective is characterizing or improving online sample efficiency, exploration, regret, learnability, structural complexity, or function-approximation guarantees in stationary single-agent RL.","exclusion":"Offline fixed-dataset limitations belong to D2; safety/robustness/risk objectives belong to D3; non-stationary or adversarial environments belong to D4; representation learning as the main object belongs to D6.","nearest_boundary":"D6","representative_topics":["low-rank structure assumptions for sample-efficient policy learning in long-horizon mdps","optimistic posterior sampling for model-based rl with hellinger-distance regret analysis","near-minimax optimal reinforcement learning via linear function approximation and bernstein self-normalized exploration"]},{"direction_index":2,"label":"D2","name":"Offline and Off-Policy Inference","description":"Theory for policy learning, evaluation, selection, and estimation from fixed or previously collected data, including pessimism, conservatism, coverage limits, off-policy evaluation, and hybrid offline-online efficiency.","inclusion":"Include atoms whose primary objective is overcoming distribution shift, coverage, estimation, or policy-selection limitations caused by learning from non-interactive or logged data.","exclusion":"If the primary objective is explicit worst-case robustness under uncertainty sets, assign to D3; if equilibrium learning in games is central, assign to D5; if hidden-state inference is central, assign to D6.","nearest_boundary":"D3","representative_topics":["near-optimal primal-dual offline learning in constrained markov decision processes","theoretical efficiency of offline reinforcement learning with differentiable function class approximation","hybrid q-learning exploiting offline dataset quality and bilinear rank for coverage-free efficient policy learning"]},{"direction_index":3,"label":"D3","name":"Reliable RL under Constraints and Uncertainty","description":"Theory for RL objectives requiring safety, feasibility, robustness, distributional robustness, adversarial-corruption tolerance, or risk-sensitive control under uncertainty in costs, transitions, rewards, or returns.","inclusion":"Include atoms whose primary objective is satisfying constraints, bounding violations, optimizing under explicit uncertainty sets or misspecification, learning robust policies, or optimizing risk/return-distribution criteria.","exclusion":"Non-stationary/adversarial changes over time belong to D4; fixed-dataset pessimism without a robustness formulation belongs to D2; reward/specification design without reliability as the main objective belongs to D7.","nearest_boundary":"D2","representative_topics":["dual methods for constrained markov decision processes with linear convergence guarantees","near-minimax optimal sample complexity bounds for constrained mdp policy learning","uncertainty set regularization for robust policy generalization in continuous control reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Adaptation across Shifts and Tasks","description":"Theory for adapting policies, representations, or knowledge across changing environments, adversarial/non-stationary feedback, continual streams, task distributions, lifelong learning, transfer, meta-RL, and multi-task settings.","inclusion":"Include atoms whose primary objective is adaptation across time, contexts, tasks, or task distributions, including dynamic regret, detection of changes, transfer guarantees, policy reuse, and meta-learning sample complexity.","exclusion":"Single stationary-task online sample efficiency belongs to D1; single-task representation learning without transfer belongs to D6; strategic multi-agent interaction belongs to D5.","nearest_boundary":"D1","representative_topics":["lifelong reinforcement learning with provably efficient multi-task policy learning via value distillation","theoretical analysis of multitask representation learning benefits in low-rank reinforcement learning","pac bounds for meta rl via density estimation and manifold-aware task distribution learning"]},{"direction_index":5,"label":"D5","name":"Multi-Agent and Game-Theoretic RL","description":"Theory for learning, optimizing, or computing equilibria and decentralized/cooperative policies in Markov games, stochastic games, mean-field settings, competitive RL, and distributed multi-agent RL.","inclusion":"Include atoms whose primary objective is strategic interaction, equilibrium learning, multi-agent regret/sample complexity, decentralized coordination, opponent interaction, or game-theoretic computation.","exclusion":"Single-agent robust minimax formulations without strategic learners belong to D3; single-agent adversarial losses belong to D4; offline data issues are secondary when equilibrium learning is central.","nearest_boundary":"D3","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","byzantine-robust distributed reinforcement learning via robust mean estimation and pessimistic/optimistic value iteration","offline nash equilibrium learning in zero-sum markov games via pessimistic model-based confidence bounds"]},{"direction_index":6,"label":"D6","name":"State, Representation, and Model Structure","description":"Foundations for discovering, compressing, representing, or planning with internal structure in RL, including latent state, partial observability, abstractions, causal/invariant representations, learned models, world models, options, and hierarchical decomposition.","inclusion":"Include atoms whose primary objective is overcoming hidden or high-dimensional information structure, learning useful representations or abstractions, constructing value-equivalent or model-based representations, or using hierarchical/model-based structure for generalization and planning.","exclusion":"If representation is mainly used for online regret under a specified structural class, assign to D1; if it is mainly shared across tasks, assign to D4; if formal task syntax is central, assign to D7.","nearest_boundary":"D1","representative_topics":["rate-distortion-based approximate value equivalence for tractable model-based reinforcement learning","value-equivalent model compression via rate-distortion for capacity-constrained model-based reinforcement learning","oracle-free reinforcement learning in observable pomdps via barycentric spanner policy covers"]},{"direction_index":7,"label":"D7","name":"Objective and Specification Design","description":"Theory for specifying, transforming, learning, or synthesizing RL objectives, including reward shaping, preference and demonstration objectives, utility axioms, nonstandard discounting, generalized sequential objectives, temporal logic, automata, and formal task specifications.","inclusion":"Include atoms whose primary objective is defining what the agent should optimize or satisfy, learning that objective from feedback, proving equivalence or validity of reward/objective transformations, or synthesizing policies from formal/hierarchical specifications.","exclusion":"Risk measures over return distributions belong to D3; algorithmic convergence for a fixed objective belongs to D0; ordinary CMDP constraints without formal objective syntax belong to D3.","nearest_boundary":"D3","representative_topics":["theoretical analysis of outcome-conditioned imitation learning and its relation to reward maximization","axiomatic utility characterization for sequential decision making and reward hypothesis","discrete calculus on mdps formalizing potential shaping as gradient-based reward equivalence"]}],"split":"dev"}
{"episode_id":"F093_T2025-01","field_index":93,"field":"reinforcement learning theory","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Policy and Value Algorithm Convergence","description":"Theoretical guarantees for core model-free RL optimization and evaluation algorithms, including policy-gradient, natural-gradient, actor-critic, TD, Q-learning, value iteration, replay, and function-approximation dynamics.","inclusion":"Include atoms whose primary objective is proving convergence, stability, finite-time rates, sample complexity, global optimality, or landscape/generalization properties of policy-optimization or Bellman/value-learning algorithms.","exclusion":"If the same algorithms are analyzed mainly to satisfy constraints, handle fixed offline data, learn equilibria, achieve robustness, or adapt to non-stationarity, assign to that primary objective instead.","nearest_boundary":"D1","representative_topics":["stabilizing q-learning via modular algorithmic tools for provably efficient linear mdp learning","finite-time complexity analysis of entropy-regularized neural natural actor-critic algorithms","temporal difference learning dynamics and generalization in deep reinforcement learning"]},{"direction_index":1,"label":"D1","name":"Online Sample-Efficient Learning","description":"Regret, PAC, minimax, Bayesian, structural, reward-free, and exploration theory for learning efficiently through online interaction with single-agent MDPs and related function-approximation classes.","inclusion":"Include atoms whose primary objective is characterizing or improving online sample efficiency, exploration, regret, learnability, structural complexity, or function-approximation guarantees in stationary single-agent RL.","exclusion":"Offline fixed-dataset limitations belong to D2; safety/robustness/risk objectives belong to D3; non-stationary or adversarial environments belong to D4; representation learning as the main object belongs to D6.","nearest_boundary":"D6","representative_topics":["low-rank structure assumptions for sample-efficient policy learning in long-horizon mdps","optimistic posterior sampling for model-based rl with hellinger-distance regret analysis","near-minimax optimal reinforcement learning via linear function approximation and bernstein self-normalized exploration"]},{"direction_index":2,"label":"D2","name":"Offline and Off-Policy Inference","description":"Theory for policy learning, evaluation, selection, and estimation from fixed or previously collected data, including pessimism, conservatism, coverage limits, off-policy evaluation, and hybrid offline-online efficiency.","inclusion":"Include atoms whose primary objective is overcoming distribution shift, coverage, estimation, or policy-selection limitations caused by learning from non-interactive or logged data.","exclusion":"If the primary objective is explicit worst-case robustness under uncertainty sets, assign to D3; if equilibrium learning in games is central, assign to D5; if hidden-state inference is central, assign to D6.","nearest_boundary":"D3","representative_topics":["near-optimal primal-dual offline learning in constrained markov decision processes","theoretical efficiency of offline reinforcement learning with differentiable function class approximation","hybrid q-learning exploiting offline dataset quality and bilinear rank for coverage-free efficient policy learning"]},{"direction_index":3,"label":"D3","name":"Reliable RL under Constraints and Uncertainty","description":"Theory for RL objectives requiring safety, feasibility, robustness, distributional robustness, adversarial-corruption tolerance, or risk-sensitive control under uncertainty in costs, transitions, rewards, or returns.","inclusion":"Include atoms whose primary objective is satisfying constraints, bounding violations, optimizing under explicit uncertainty sets or misspecification, learning robust policies, or optimizing risk/return-distribution criteria.","exclusion":"Non-stationary/adversarial changes over time belong to D4; fixed-dataset pessimism without a robustness formulation belongs to D2; reward/specification design without reliability as the main objective belongs to D7.","nearest_boundary":"D2","representative_topics":["dual methods for constrained markov decision processes with linear convergence guarantees","near-minimax optimal sample complexity bounds for constrained mdp policy learning","uncertainty set regularization for robust policy generalization in continuous control reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Adaptation across Shifts and Tasks","description":"Theory for adapting policies, representations, or knowledge across changing environments, adversarial/non-stationary feedback, continual streams, task distributions, lifelong learning, transfer, meta-RL, and multi-task settings.","inclusion":"Include atoms whose primary objective is adaptation across time, contexts, tasks, or task distributions, including dynamic regret, detection of changes, transfer guarantees, policy reuse, and meta-learning sample complexity.","exclusion":"Single stationary-task online sample efficiency belongs to D1; single-task representation learning without transfer belongs to D6; strategic multi-agent interaction belongs to D5.","nearest_boundary":"D1","representative_topics":["lifelong reinforcement learning with provably efficient multi-task policy learning via value distillation","theoretical analysis of multitask representation learning benefits in low-rank reinforcement learning","pac bounds for meta rl via density estimation and manifold-aware task distribution learning"]},{"direction_index":5,"label":"D5","name":"Multi-Agent and Game-Theoretic RL","description":"Theory for learning, optimizing, or computing equilibria and decentralized/cooperative policies in Markov games, stochastic games, mean-field settings, competitive RL, and distributed multi-agent RL.","inclusion":"Include atoms whose primary objective is strategic interaction, equilibrium learning, multi-agent regret/sample complexity, decentralized coordination, opponent interaction, or game-theoretic computation.","exclusion":"Single-agent robust minimax formulations without strategic learners belong to D3; single-agent adversarial losses belong to D4; offline data issues are secondary when equilibrium learning is central.","nearest_boundary":"D3","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","byzantine-robust distributed reinforcement learning via robust mean estimation and pessimistic/optimistic value iteration","offline nash equilibrium learning in zero-sum markov games via pessimistic model-based confidence bounds"]},{"direction_index":6,"label":"D6","name":"State, Representation, and Model Structure","description":"Foundations for discovering, compressing, representing, or planning with internal structure in RL, including latent state, partial observability, abstractions, causal/invariant representations, learned models, world models, options, and hierarchical decomposition.","inclusion":"Include atoms whose primary objective is overcoming hidden or high-dimensional information structure, learning useful representations or abstractions, constructing value-equivalent or model-based representations, or using hierarchical/model-based structure for generalization and planning.","exclusion":"If representation is mainly used for online regret under a specified structural class, assign to D1; if it is mainly shared across tasks, assign to D4; if formal task syntax is central, assign to D7.","nearest_boundary":"D1","representative_topics":["rate-distortion-based approximate value equivalence for tractable model-based reinforcement learning","value-equivalent model compression via rate-distortion for capacity-constrained model-based reinforcement learning","oracle-free reinforcement learning in observable pomdps via barycentric spanner policy covers"]},{"direction_index":7,"label":"D7","name":"Objective and Specification Design","description":"Theory for specifying, transforming, learning, or synthesizing RL objectives, including reward shaping, preference and demonstration objectives, utility axioms, nonstandard discounting, generalized sequential objectives, temporal logic, automata, and formal task specifications.","inclusion":"Include atoms whose primary objective is defining what the agent should optimize or satisfy, learning that objective from feedback, proving equivalence or validity of reward/objective transformations, or synthesizing policies from formal/hierarchical specifications.","exclusion":"Risk measures over return distributions belong to D3; algorithmic convergence for a fixed objective belongs to D0; ordinary CMDP constraints without formal objective syntax belong to D3.","nearest_boundary":"D3","representative_topics":["theoretical analysis of outcome-conditioned imitation learning and its relation to reward maximization","axiomatic utility characterization for sequential decision making and reward hypothesis","discrete calculus on mdps formalizing potential shaping as gradient-based reward equivalence"]}],"split":"dev"}
{"episode_id":"F093_T2025-07","field_index":93,"field":"reinforcement learning theory","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Policy and Value Algorithm Convergence","description":"Theoretical guarantees for core model-free RL optimization and evaluation algorithms, including policy-gradient, natural-gradient, actor-critic, TD, Q-learning, value iteration, replay, and function-approximation dynamics.","inclusion":"Include atoms whose primary objective is proving convergence, stability, finite-time rates, sample complexity, global optimality, or landscape/generalization properties of policy-optimization or Bellman/value-learning algorithms.","exclusion":"If the same algorithms are analyzed mainly to satisfy constraints, handle fixed offline data, learn equilibria, achieve robustness, or adapt to non-stationarity, assign to that primary objective instead.","nearest_boundary":"D1","representative_topics":["stabilizing q-learning via modular algorithmic tools for provably efficient linear mdp learning","finite-time complexity analysis of entropy-regularized neural natural actor-critic algorithms","temporal difference learning dynamics and generalization in deep reinforcement learning"]},{"direction_index":1,"label":"D1","name":"Online Sample-Efficient Learning","description":"Regret, PAC, minimax, Bayesian, structural, reward-free, and exploration theory for learning efficiently through online interaction with single-agent MDPs and related function-approximation classes.","inclusion":"Include atoms whose primary objective is characterizing or improving online sample efficiency, exploration, regret, learnability, structural complexity, or function-approximation guarantees in stationary single-agent RL.","exclusion":"Offline fixed-dataset limitations belong to D2; safety/robustness/risk objectives belong to D3; non-stationary or adversarial environments belong to D4; representation learning as the main object belongs to D6.","nearest_boundary":"D6","representative_topics":["low-rank structure assumptions for sample-efficient policy learning in long-horizon mdps","optimistic posterior sampling for model-based rl with hellinger-distance regret analysis","near-minimax optimal reinforcement learning via linear function approximation and bernstein self-normalized exploration"]},{"direction_index":2,"label":"D2","name":"Offline and Off-Policy Inference","description":"Theory for policy learning, evaluation, selection, and estimation from fixed or previously collected data, including pessimism, conservatism, coverage limits, off-policy evaluation, and hybrid offline-online efficiency.","inclusion":"Include atoms whose primary objective is overcoming distribution shift, coverage, estimation, or policy-selection limitations caused by learning from non-interactive or logged data.","exclusion":"If the primary objective is explicit worst-case robustness under uncertainty sets, assign to D3; if equilibrium learning in games is central, assign to D5; if hidden-state inference is central, assign to D6.","nearest_boundary":"D3","representative_topics":["near-optimal primal-dual offline learning in constrained markov decision processes","theoretical efficiency of offline reinforcement learning with differentiable function class approximation","hybrid q-learning exploiting offline dataset quality and bilinear rank for coverage-free efficient policy learning"]},{"direction_index":3,"label":"D3","name":"Reliable RL under Constraints and Uncertainty","description":"Theory for RL objectives requiring safety, feasibility, robustness, distributional robustness, adversarial-corruption tolerance, or risk-sensitive control under uncertainty in costs, transitions, rewards, or returns.","inclusion":"Include atoms whose primary objective is satisfying constraints, bounding violations, optimizing under explicit uncertainty sets or misspecification, learning robust policies, or optimizing risk/return-distribution criteria.","exclusion":"Non-stationary/adversarial changes over time belong to D4; fixed-dataset pessimism without a robustness formulation belongs to D2; reward/specification design without reliability as the main objective belongs to D7.","nearest_boundary":"D2","representative_topics":["dual methods for constrained markov decision processes with linear convergence guarantees","near-minimax optimal sample complexity bounds for constrained mdp policy learning","uncertainty set regularization for robust policy generalization in continuous control reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Adaptation across Shifts and Tasks","description":"Theory for adapting policies, representations, or knowledge across changing environments, adversarial/non-stationary feedback, continual streams, task distributions, lifelong learning, transfer, meta-RL, and multi-task settings.","inclusion":"Include atoms whose primary objective is adaptation across time, contexts, tasks, or task distributions, including dynamic regret, detection of changes, transfer guarantees, policy reuse, and meta-learning sample complexity.","exclusion":"Single stationary-task online sample efficiency belongs to D1; single-task representation learning without transfer belongs to D6; strategic multi-agent interaction belongs to D5.","nearest_boundary":"D1","representative_topics":["lifelong reinforcement learning with provably efficient multi-task policy learning via value distillation","theoretical analysis of multitask representation learning benefits in low-rank reinforcement learning","pac bounds for meta rl via density estimation and manifold-aware task distribution learning"]},{"direction_index":5,"label":"D5","name":"Multi-Agent and Game-Theoretic RL","description":"Theory for learning, optimizing, or computing equilibria and decentralized/cooperative policies in Markov games, stochastic games, mean-field settings, competitive RL, and distributed multi-agent RL.","inclusion":"Include atoms whose primary objective is strategic interaction, equilibrium learning, multi-agent regret/sample complexity, decentralized coordination, opponent interaction, or game-theoretic computation.","exclusion":"Single-agent robust minimax formulations without strategic learners belong to D3; single-agent adversarial losses belong to D4; offline data issues are secondary when equilibrium learning is central.","nearest_boundary":"D3","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","byzantine-robust distributed reinforcement learning via robust mean estimation and pessimistic/optimistic value iteration","offline nash equilibrium learning in zero-sum markov games via pessimistic model-based confidence bounds"]},{"direction_index":6,"label":"D6","name":"State, Representation, and Model Structure","description":"Foundations for discovering, compressing, representing, or planning with internal structure in RL, including latent state, partial observability, abstractions, causal/invariant representations, learned models, world models, options, and hierarchical decomposition.","inclusion":"Include atoms whose primary objective is overcoming hidden or high-dimensional information structure, learning useful representations or abstractions, constructing value-equivalent or model-based representations, or using hierarchical/model-based structure for generalization and planning.","exclusion":"If representation is mainly used for online regret under a specified structural class, assign to D1; if it is mainly shared across tasks, assign to D4; if formal task syntax is central, assign to D7.","nearest_boundary":"D1","representative_topics":["rate-distortion-based approximate value equivalence for tractable model-based reinforcement learning","value-equivalent model compression via rate-distortion for capacity-constrained model-based reinforcement learning","oracle-free reinforcement learning in observable pomdps via barycentric spanner policy covers"]},{"direction_index":7,"label":"D7","name":"Objective and Specification Design","description":"Theory for specifying, transforming, learning, or synthesizing RL objectives, including reward shaping, preference and demonstration objectives, utility axioms, nonstandard discounting, generalized sequential objectives, temporal logic, automata, and formal task specifications.","inclusion":"Include atoms whose primary objective is defining what the agent should optimize or satisfy, learning that objective from feedback, proving equivalence or validity of reward/objective transformations, or synthesizing policies from formal/hierarchical specifications.","exclusion":"Risk measures over return distributions belong to D3; algorithmic convergence for a fixed objective belongs to D0; ordinary CMDP constraints without formal objective syntax belong to D3.","nearest_boundary":"D3","representative_topics":["theoretical analysis of outcome-conditioned imitation learning and its relation to reward maximization","axiomatic utility characterization for sequential decision making and reward hypothesis","discrete calculus on mdps formalizing potential shaping as gradient-based reward equivalence"]}],"split":"dev"}
{"episode_id":"F093_T2026-01","field_index":93,"field":"reinforcement learning theory","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Policy and Value Algorithm Convergence","description":"Theoretical guarantees for core model-free RL optimization and evaluation algorithms, including policy-gradient, natural-gradient, actor-critic, TD, Q-learning, value iteration, replay, and function-approximation dynamics.","inclusion":"Include atoms whose primary objective is proving convergence, stability, finite-time rates, sample complexity, global optimality, or landscape/generalization properties of policy-optimization or Bellman/value-learning algorithms.","exclusion":"If the same algorithms are analyzed mainly to satisfy constraints, handle fixed offline data, learn equilibria, achieve robustness, or adapt to non-stationarity, assign to that primary objective instead.","nearest_boundary":"D1","representative_topics":["stabilizing q-learning via modular algorithmic tools for provably efficient linear mdp learning","finite-time complexity analysis of entropy-regularized neural natural actor-critic algorithms","temporal difference learning dynamics and generalization in deep reinforcement learning"]},{"direction_index":1,"label":"D1","name":"Online Sample-Efficient Learning","description":"Regret, PAC, minimax, Bayesian, structural, reward-free, and exploration theory for learning efficiently through online interaction with single-agent MDPs and related function-approximation classes.","inclusion":"Include atoms whose primary objective is characterizing or improving online sample efficiency, exploration, regret, learnability, structural complexity, or function-approximation guarantees in stationary single-agent RL.","exclusion":"Offline fixed-dataset limitations belong to D2; safety/robustness/risk objectives belong to D3; non-stationary or adversarial environments belong to D4; representation learning as the main object belongs to D6.","nearest_boundary":"D6","representative_topics":["low-rank structure assumptions for sample-efficient policy learning in long-horizon mdps","optimistic posterior sampling for model-based rl with hellinger-distance regret analysis","near-minimax optimal reinforcement learning via linear function approximation and bernstein self-normalized exploration"]},{"direction_index":2,"label":"D2","name":"Offline and Off-Policy Inference","description":"Theory for policy learning, evaluation, selection, and estimation from fixed or previously collected data, including pessimism, conservatism, coverage limits, off-policy evaluation, and hybrid offline-online efficiency.","inclusion":"Include atoms whose primary objective is overcoming distribution shift, coverage, estimation, or policy-selection limitations caused by learning from non-interactive or logged data.","exclusion":"If the primary objective is explicit worst-case robustness under uncertainty sets, assign to D3; if equilibrium learning in games is central, assign to D5; if hidden-state inference is central, assign to D6.","nearest_boundary":"D3","representative_topics":["near-optimal primal-dual offline learning in constrained markov decision processes","theoretical efficiency of offline reinforcement learning with differentiable function class approximation","hybrid q-learning exploiting offline dataset quality and bilinear rank for coverage-free efficient policy learning"]},{"direction_index":3,"label":"D3","name":"Reliable RL under Constraints and Uncertainty","description":"Theory for RL objectives requiring safety, feasibility, robustness, distributional robustness, adversarial-corruption tolerance, or risk-sensitive control under uncertainty in costs, transitions, rewards, or returns.","inclusion":"Include atoms whose primary objective is satisfying constraints, bounding violations, optimizing under explicit uncertainty sets or misspecification, learning robust policies, or optimizing risk/return-distribution criteria.","exclusion":"Non-stationary/adversarial changes over time belong to D4; fixed-dataset pessimism without a robustness formulation belongs to D2; reward/specification design without reliability as the main objective belongs to D7.","nearest_boundary":"D2","representative_topics":["dual methods for constrained markov decision processes with linear convergence guarantees","near-minimax optimal sample complexity bounds for constrained mdp policy learning","uncertainty set regularization for robust policy generalization in continuous control reinforcement learning"]},{"direction_index":4,"label":"D4","name":"Adaptation across Shifts and Tasks","description":"Theory for adapting policies, representations, or knowledge across changing environments, adversarial/non-stationary feedback, continual streams, task distributions, lifelong learning, transfer, meta-RL, and multi-task settings.","inclusion":"Include atoms whose primary objective is adaptation across time, contexts, tasks, or task distributions, including dynamic regret, detection of changes, transfer guarantees, policy reuse, and meta-learning sample complexity.","exclusion":"Single stationary-task online sample efficiency belongs to D1; single-task representation learning without transfer belongs to D6; strategic multi-agent interaction belongs to D5.","nearest_boundary":"D1","representative_topics":["lifelong reinforcement learning with provably efficient multi-task policy learning via value distillation","theoretical analysis of multitask representation learning benefits in low-rank reinforcement learning","pac bounds for meta rl via density estimation and manifold-aware task distribution learning"]},{"direction_index":5,"label":"D5","name":"Multi-Agent and Game-Theoretic RL","description":"Theory for learning, optimizing, or computing equilibria and decentralized/cooperative policies in Markov games, stochastic games, mean-field settings, competitive RL, and distributed multi-agent RL.","inclusion":"Include atoms whose primary objective is strategic interaction, equilibrium learning, multi-agent regret/sample complexity, decentralized coordination, opponent interaction, or game-theoretic computation.","exclusion":"Single-agent robust minimax formulations without strategic learners belong to D3; single-agent adversarial losses belong to D4; offline data issues are secondary when equilibrium learning is central.","nearest_boundary":"D3","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","byzantine-robust distributed reinforcement learning via robust mean estimation and pessimistic/optimistic value iteration","offline nash equilibrium learning in zero-sum markov games via pessimistic model-based confidence bounds"]},{"direction_index":6,"label":"D6","name":"State, Representation, and Model Structure","description":"Foundations for discovering, compressing, representing, or planning with internal structure in RL, including latent state, partial observability, abstractions, causal/invariant representations, learned models, world models, options, and hierarchical decomposition.","inclusion":"Include atoms whose primary objective is overcoming hidden or high-dimensional information structure, learning useful representations or abstractions, constructing value-equivalent or model-based representations, or using hierarchical/model-based structure for generalization and planning.","exclusion":"If representation is mainly used for online regret under a specified structural class, assign to D1; if it is mainly shared across tasks, assign to D4; if formal task syntax is central, assign to D7.","nearest_boundary":"D1","representative_topics":["rate-distortion-based approximate value equivalence for tractable model-based reinforcement learning","value-equivalent model compression via rate-distortion for capacity-constrained model-based reinforcement learning","oracle-free reinforcement learning in observable pomdps via barycentric spanner policy covers"]},{"direction_index":7,"label":"D7","name":"Objective and Specification Design","description":"Theory for specifying, transforming, learning, or synthesizing RL objectives, including reward shaping, preference and demonstration objectives, utility axioms, nonstandard discounting, generalized sequential objectives, temporal logic, automata, and formal task specifications.","inclusion":"Include atoms whose primary objective is defining what the agent should optimize or satisfy, learning that objective from feedback, proving equivalence or validity of reward/objective transformations, or synthesizing policies from formal/hierarchical specifications.","exclusion":"Risk measures over return distributions belong to D3; algorithmic convergence for a fixed objective belongs to D0; ordinary CMDP constraints without formal objective syntax belong to D3.","nearest_boundary":"D3","representative_topics":["theoretical analysis of outcome-conditioned imitation learning and its relation to reward maximization","axiomatic utility characterization for sequential decision making and reward hypothesis","discrete calculus on mdps formalizing potential shaping as gradient-based reward equivalence"]}],"split":"dev"}
{"episode_id":"F102_T2024-01","field_index":102,"field":"multimodal representation learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cross-modal alignment and foundations","description":"Learning, characterizing, and evaluating shared multimodal embedding spaces in which heterogeneous modalities become semantically comparable, together with formal, empirical, and benchmark-level analyses of such representation learning.","inclusion":"Include work whose primary objective is global cross-modal semantic alignment, contrastive or embedding-space design, theoretical analysis of multimodal objectives or transfer, controlled empirical characterization, or survey/benchmark infrastructure for multimodal representation learning.","exclusion":"If the main objective is a unified generative architecture, masked/generative pretraining objective, fine-grained region/phrase grounding, domain-specific modeling, or lightweight adaptation of an existing pretrained model, assign to the corresponding specialized direction.","nearest_boundary":"D2","representative_topics":["sparse mixture of experts for multimodal contrastive image-text learning with balanced expert activation","prototypical contrastive pretraining with back translation for robust image-text representation learning","controlled study of caption supervision effects on transferable visual representations"]},{"direction_index":1,"label":"D1","name":"Unified generalist architectures","description":"Designing single multimodal architectures or interfaces that unify multiple modalities, tasks, and input-output formats for broad multimodal understanding, generation, or instruction-following.","inclusion":"Include work whose primary objective is architectural unification through language-model interfaces, autoregressive or encoder-decoder input-output frameworks, discrete token homogenization, multimodal instruction systems, or generalist multimodal foundation architectures.","exclusion":"If the work mainly proposes a non-generative fusion block, a contrastive alignment loss, or a masked/generative objective without a generalist architecture, assign to D3, D0, or D2 respectively.","nearest_boundary":"D3","representative_topics":["language model as universal interface for modular multimodal encoder integration via semi-causal pretraining","unified transformer model via discrete token homogenization for diverse vision and language tasks","multimodal foundation model pretraining unifying vision and vision-language tasks via masked language modeling"]},{"direction_index":2,"label":"D2","name":"Masked and generative pretraining","description":"Developing transferable multimodal representations through self-supervised, masked, reconstruction, prediction, or hybrid generative-discriminative pretraining objectives.","inclusion":"Include work whose primary objective is a multimodal pretraining objective or recipe such as masked multimodal modeling, masked autoencoding, encoder-decoder reconstruction, cross-modal prediction, or weakly/self-supervised generative representation learning.","exclusion":"If contrastive shared-space comparability is the dominant contribution, assign to D0; if the objective is subordinate to a generalist unified architecture, assign to D1; if the objective is mainly domain-specific, assign to D7.","nearest_boundary":"D0","representative_topics":["masked encoder-decoder pre-training unifying vision and language for text recognition","unified masked autoencoding for joint image and video vision transformer pretraining","bidirectional vision-language pre-training for visually-rich document understanding and information extraction"]},{"direction_index":3,"label":"D3","name":"Modality interaction architectures","description":"Designing architectural mechanisms that determine how modality streams exchange, fuse, attend to, bridge, or jointly structure information.","inclusion":"Include work whose primary objective is a fusion architecture or modality-interaction mechanism, including cross-attention variants, bridge layers, token exchange, progressive fusion, hierarchical modality attention, graph-based multimodal interaction, or transformer fusion design.","exclusion":"If the architecture is primarily a generalist LLM or unified generative interface, assign to D1; if the main problem is missing/noisy modality robustness rather than interaction design, assign to D4.","nearest_boundary":"D1","representative_topics":["two-tower vision-language models with multi-layer bridge connections for cross-modal alignment","unified audio-visual model architecture for joint event classification","orthogonal projection and branched fusion for multilingual face-voice association"]},{"direction_index":4,"label":"D4","name":"Robust and balanced latent structure","description":"Learning multimodal representations that remain reliable under missing, noisy, imbalanced, or conflicting modalities while structuring shared and modality-specific latent information.","inclusion":"Include work whose primary objective is robustness to missing or corrupted modalities, modality imbalance, modality collapse mitigation, gradient harmonization, common-private decomposition, uncertainty-aware or information-theoretic latent structure, or multi-view factorization.","exclusion":"If robustness is only incidental to a general alignment objective, assign to D0; if the contribution is mainly a fusion architecture without explicit robustness or latent-factor balancing, assign to D3.","nearest_boundary":"D0","representative_topics":["mitigating modality collapse in multimodal vae training via impartial gradient optimization","hierarchical variational autoencoders with factor analysis latent space for heterogeneous multimodal data","optimal transport alignment dynamics for missing-modality inference in multimodal sequences"]},{"direction_index":5,"label":"D5","name":"Fine-grained structured grounding","description":"Learning multimodal representations that preserve local, part-level, region-level, phrase-level, relation-level, multi-scale, or otherwise structured cross-modal correspondences.","inclusion":"Include work whose primary objective is explicit fine-grained grounding or structured correspondence beyond global modality-level similarity, including part-aware retrieval, region-word alignment, spatial or relation-aware image-text retrieval, multi-granularity representations, and multi-scale cross-modal correspondence learning.","exclusion":"If the work only learns global embeddings without explicit local or structured correspondence, assign to D0; if fine-grained modeling is mainly a domain-specific adaptation device, assign to D7 or D6 according to the primary objective.","nearest_boundary":"D0","representative_topics":["granularity-unified cross-modal representation learning for text-to-image person re-identification","multi-scale detail injection into clip feature space for fine-grained remote sensing retrieval","masked multimodal modeling for fine-grained vision-language representation learning"]},{"direction_index":6,"label":"D6","name":"Adaptation and efficient transfer","description":"Adapting, compressing, transferring, or deploying multimodal representations under limited compute, data, parameters, communication, modality acquisition, or annotation budgets.","inclusion":"Include work whose primary objective is transfer of pretrained multimodal or vision-language models to new languages, modalities, tasks, or low-resource settings; prompt/adaptor/parameter-efficient tuning; multimodal-to-unimodal distillation; modality selection; efficient training; federated learning; active learning; or resource-constrained multimodal representation learning.","exclusion":"If the main contribution is new full pretraining from scratch, assign to D0 or D2; if the central novelty is domain-specific structure rather than transfer or efficiency, assign to D7; if missing-modality robustness rather than resource-aware deployment is primary, assign to D4.","nearest_boundary":"D7","representative_topics":["multilingual multimodal retrieval via cross-lingual pseudo-label transfer from pretrained multilingual models","post-pretraining image-text models for video-language representation alignment via omnisource cross-modal learning","parameter-free attention enhancement of contrastive vision-language pre-training for zero-shot transfer"]},{"direction_index":7,"label":"D7","name":"Domain and human-centered representations","description":"Developing multimodal representations whose primary objective is to capture specialized domain structures or human-centered affective, behavioral, physiological, motion, music, or sentiment signal structure.","inclusion":"Include work where the scientific objective is modeling domain-specific multimodal semantics, such as biomedical, neuroimaging, molecular, document, remote-sensing, product, commodity, or registration structure, or human-centered affect, emotion, sentiment, physiology, music, dance, motion, and behavior representations.","exclusion":"Do not include general alignment, fusion, or pretraining methods merely evaluated on a domain benchmark; assign those to D0, D2, or D3. If the main objective is lightweight adaptation of an existing pretrained model rather than domain or human-signal structure, assign to D6.","nearest_boundary":"D6","representative_topics":["entity graph enhanced cross-modal pretraining for weakly-supervised instance-level product retrieval","unsupervised contrastive representation learning for non-rigid multimodal deformable image registration","audio-visual music-dance representation learning via explicit-implicit rhythm synchronization"]}],"split":"dev"}
{"episode_id":"F102_T2024-07","field_index":102,"field":"multimodal representation learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cross-modal alignment and foundations","description":"Learning, characterizing, and evaluating shared multimodal embedding spaces in which heterogeneous modalities become semantically comparable, together with formal, empirical, and benchmark-level analyses of such representation learning.","inclusion":"Include work whose primary objective is global cross-modal semantic alignment, contrastive or embedding-space design, theoretical analysis of multimodal objectives or transfer, controlled empirical characterization, or survey/benchmark infrastructure for multimodal representation learning.","exclusion":"If the main objective is a unified generative architecture, masked/generative pretraining objective, fine-grained region/phrase grounding, domain-specific modeling, or lightweight adaptation of an existing pretrained model, assign to the corresponding specialized direction.","nearest_boundary":"D2","representative_topics":["sparse mixture of experts for multimodal contrastive image-text learning with balanced expert activation","prototypical contrastive pretraining with back translation for robust image-text representation learning","controlled study of caption supervision effects on transferable visual representations"]},{"direction_index":1,"label":"D1","name":"Unified generalist architectures","description":"Designing single multimodal architectures or interfaces that unify multiple modalities, tasks, and input-output formats for broad multimodal understanding, generation, or instruction-following.","inclusion":"Include work whose primary objective is architectural unification through language-model interfaces, autoregressive or encoder-decoder input-output frameworks, discrete token homogenization, multimodal instruction systems, or generalist multimodal foundation architectures.","exclusion":"If the work mainly proposes a non-generative fusion block, a contrastive alignment loss, or a masked/generative objective without a generalist architecture, assign to D3, D0, or D2 respectively.","nearest_boundary":"D3","representative_topics":["language model as universal interface for modular multimodal encoder integration via semi-causal pretraining","unified transformer model via discrete token homogenization for diverse vision and language tasks","multimodal foundation model pretraining unifying vision and vision-language tasks via masked language modeling"]},{"direction_index":2,"label":"D2","name":"Masked and generative pretraining","description":"Developing transferable multimodal representations through self-supervised, masked, reconstruction, prediction, or hybrid generative-discriminative pretraining objectives.","inclusion":"Include work whose primary objective is a multimodal pretraining objective or recipe such as masked multimodal modeling, masked autoencoding, encoder-decoder reconstruction, cross-modal prediction, or weakly/self-supervised generative representation learning.","exclusion":"If contrastive shared-space comparability is the dominant contribution, assign to D0; if the objective is subordinate to a generalist unified architecture, assign to D1; if the objective is mainly domain-specific, assign to D7.","nearest_boundary":"D0","representative_topics":["masked encoder-decoder pre-training unifying vision and language for text recognition","unified masked autoencoding for joint image and video vision transformer pretraining","bidirectional vision-language pre-training for visually-rich document understanding and information extraction"]},{"direction_index":3,"label":"D3","name":"Modality interaction architectures","description":"Designing architectural mechanisms that determine how modality streams exchange, fuse, attend to, bridge, or jointly structure information.","inclusion":"Include work whose primary objective is a fusion architecture or modality-interaction mechanism, including cross-attention variants, bridge layers, token exchange, progressive fusion, hierarchical modality attention, graph-based multimodal interaction, or transformer fusion design.","exclusion":"If the architecture is primarily a generalist LLM or unified generative interface, assign to D1; if the main problem is missing/noisy modality robustness rather than interaction design, assign to D4.","nearest_boundary":"D1","representative_topics":["two-tower vision-language models with multi-layer bridge connections for cross-modal alignment","unified audio-visual model architecture for joint event classification","orthogonal projection and branched fusion for multilingual face-voice association"]},{"direction_index":4,"label":"D4","name":"Robust and balanced latent structure","description":"Learning multimodal representations that remain reliable under missing, noisy, imbalanced, or conflicting modalities while structuring shared and modality-specific latent information.","inclusion":"Include work whose primary objective is robustness to missing or corrupted modalities, modality imbalance, modality collapse mitigation, gradient harmonization, common-private decomposition, uncertainty-aware or information-theoretic latent structure, or multi-view factorization.","exclusion":"If robustness is only incidental to a general alignment objective, assign to D0; if the contribution is mainly a fusion architecture without explicit robustness or latent-factor balancing, assign to D3.","nearest_boundary":"D0","representative_topics":["mitigating modality collapse in multimodal vae training via impartial gradient optimization","hierarchical variational autoencoders with factor analysis latent space for heterogeneous multimodal data","optimal transport alignment dynamics for missing-modality inference in multimodal sequences"]},{"direction_index":5,"label":"D5","name":"Fine-grained structured grounding","description":"Learning multimodal representations that preserve local, part-level, region-level, phrase-level, relation-level, multi-scale, or otherwise structured cross-modal correspondences.","inclusion":"Include work whose primary objective is explicit fine-grained grounding or structured correspondence beyond global modality-level similarity, including part-aware retrieval, region-word alignment, spatial or relation-aware image-text retrieval, multi-granularity representations, and multi-scale cross-modal correspondence learning.","exclusion":"If the work only learns global embeddings without explicit local or structured correspondence, assign to D0; if fine-grained modeling is mainly a domain-specific adaptation device, assign to D7 or D6 according to the primary objective.","nearest_boundary":"D0","representative_topics":["granularity-unified cross-modal representation learning for text-to-image person re-identification","multi-scale detail injection into clip feature space for fine-grained remote sensing retrieval","masked multimodal modeling for fine-grained vision-language representation learning"]},{"direction_index":6,"label":"D6","name":"Adaptation and efficient transfer","description":"Adapting, compressing, transferring, or deploying multimodal representations under limited compute, data, parameters, communication, modality acquisition, or annotation budgets.","inclusion":"Include work whose primary objective is transfer of pretrained multimodal or vision-language models to new languages, modalities, tasks, or low-resource settings; prompt/adaptor/parameter-efficient tuning; multimodal-to-unimodal distillation; modality selection; efficient training; federated learning; active learning; or resource-constrained multimodal representation learning.","exclusion":"If the main contribution is new full pretraining from scratch, assign to D0 or D2; if the central novelty is domain-specific structure rather than transfer or efficiency, assign to D7; if missing-modality robustness rather than resource-aware deployment is primary, assign to D4.","nearest_boundary":"D7","representative_topics":["multilingual multimodal retrieval via cross-lingual pseudo-label transfer from pretrained multilingual models","post-pretraining image-text models for video-language representation alignment via omnisource cross-modal learning","parameter-free attention enhancement of contrastive vision-language pre-training for zero-shot transfer"]},{"direction_index":7,"label":"D7","name":"Domain and human-centered representations","description":"Developing multimodal representations whose primary objective is to capture specialized domain structures or human-centered affective, behavioral, physiological, motion, music, or sentiment signal structure.","inclusion":"Include work where the scientific objective is modeling domain-specific multimodal semantics, such as biomedical, neuroimaging, molecular, document, remote-sensing, product, commodity, or registration structure, or human-centered affect, emotion, sentiment, physiology, music, dance, motion, and behavior representations.","exclusion":"Do not include general alignment, fusion, or pretraining methods merely evaluated on a domain benchmark; assign those to D0, D2, or D3. If the main objective is lightweight adaptation of an existing pretrained model rather than domain or human-signal structure, assign to D6.","nearest_boundary":"D6","representative_topics":["entity graph enhanced cross-modal pretraining for weakly-supervised instance-level product retrieval","unsupervised contrastive representation learning for non-rigid multimodal deformable image registration","audio-visual music-dance representation learning via explicit-implicit rhythm synchronization"]}],"split":"dev"}
{"episode_id":"F102_T2025-01","field_index":102,"field":"multimodal representation learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cross-modal alignment and foundations","description":"Learning, characterizing, and evaluating shared multimodal embedding spaces in which heterogeneous modalities become semantically comparable, together with formal, empirical, and benchmark-level analyses of such representation learning.","inclusion":"Include work whose primary objective is global cross-modal semantic alignment, contrastive or embedding-space design, theoretical analysis of multimodal objectives or transfer, controlled empirical characterization, or survey/benchmark infrastructure for multimodal representation learning.","exclusion":"If the main objective is a unified generative architecture, masked/generative pretraining objective, fine-grained region/phrase grounding, domain-specific modeling, or lightweight adaptation of an existing pretrained model, assign to the corresponding specialized direction.","nearest_boundary":"D2","representative_topics":["sparse mixture of experts for multimodal contrastive image-text learning with balanced expert activation","prototypical contrastive pretraining with back translation for robust image-text representation learning","controlled study of caption supervision effects on transferable visual representations"]},{"direction_index":1,"label":"D1","name":"Unified generalist architectures","description":"Designing single multimodal architectures or interfaces that unify multiple modalities, tasks, and input-output formats for broad multimodal understanding, generation, or instruction-following.","inclusion":"Include work whose primary objective is architectural unification through language-model interfaces, autoregressive or encoder-decoder input-output frameworks, discrete token homogenization, multimodal instruction systems, or generalist multimodal foundation architectures.","exclusion":"If the work mainly proposes a non-generative fusion block, a contrastive alignment loss, or a masked/generative objective without a generalist architecture, assign to D3, D0, or D2 respectively.","nearest_boundary":"D3","representative_topics":["language model as universal interface for modular multimodal encoder integration via semi-causal pretraining","unified transformer model via discrete token homogenization for diverse vision and language tasks","multimodal foundation model pretraining unifying vision and vision-language tasks via masked language modeling"]},{"direction_index":2,"label":"D2","name":"Masked and generative pretraining","description":"Developing transferable multimodal representations through self-supervised, masked, reconstruction, prediction, or hybrid generative-discriminative pretraining objectives.","inclusion":"Include work whose primary objective is a multimodal pretraining objective or recipe such as masked multimodal modeling, masked autoencoding, encoder-decoder reconstruction, cross-modal prediction, or weakly/self-supervised generative representation learning.","exclusion":"If contrastive shared-space comparability is the dominant contribution, assign to D0; if the objective is subordinate to a generalist unified architecture, assign to D1; if the objective is mainly domain-specific, assign to D7.","nearest_boundary":"D0","representative_topics":["masked encoder-decoder pre-training unifying vision and language for text recognition","unified masked autoencoding for joint image and video vision transformer pretraining","bidirectional vision-language pre-training for visually-rich document understanding and information extraction"]},{"direction_index":3,"label":"D3","name":"Modality interaction architectures","description":"Designing architectural mechanisms that determine how modality streams exchange, fuse, attend to, bridge, or jointly structure information.","inclusion":"Include work whose primary objective is a fusion architecture or modality-interaction mechanism, including cross-attention variants, bridge layers, token exchange, progressive fusion, hierarchical modality attention, graph-based multimodal interaction, or transformer fusion design.","exclusion":"If the architecture is primarily a generalist LLM or unified generative interface, assign to D1; if the main problem is missing/noisy modality robustness rather than interaction design, assign to D4.","nearest_boundary":"D1","representative_topics":["two-tower vision-language models with multi-layer bridge connections for cross-modal alignment","unified audio-visual model architecture for joint event classification","orthogonal projection and branched fusion for multilingual face-voice association"]},{"direction_index":4,"label":"D4","name":"Robust and balanced latent structure","description":"Learning multimodal representations that remain reliable under missing, noisy, imbalanced, or conflicting modalities while structuring shared and modality-specific latent information.","inclusion":"Include work whose primary objective is robustness to missing or corrupted modalities, modality imbalance, modality collapse mitigation, gradient harmonization, common-private decomposition, uncertainty-aware or information-theoretic latent structure, or multi-view factorization.","exclusion":"If robustness is only incidental to a general alignment objective, assign to D0; if the contribution is mainly a fusion architecture without explicit robustness or latent-factor balancing, assign to D3.","nearest_boundary":"D0","representative_topics":["mitigating modality collapse in multimodal vae training via impartial gradient optimization","hierarchical variational autoencoders with factor analysis latent space for heterogeneous multimodal data","optimal transport alignment dynamics for missing-modality inference in multimodal sequences"]},{"direction_index":5,"label":"D5","name":"Fine-grained structured grounding","description":"Learning multimodal representations that preserve local, part-level, region-level, phrase-level, relation-level, multi-scale, or otherwise structured cross-modal correspondences.","inclusion":"Include work whose primary objective is explicit fine-grained grounding or structured correspondence beyond global modality-level similarity, including part-aware retrieval, region-word alignment, spatial or relation-aware image-text retrieval, multi-granularity representations, and multi-scale cross-modal correspondence learning.","exclusion":"If the work only learns global embeddings without explicit local or structured correspondence, assign to D0; if fine-grained modeling is mainly a domain-specific adaptation device, assign to D7 or D6 according to the primary objective.","nearest_boundary":"D0","representative_topics":["granularity-unified cross-modal representation learning for text-to-image person re-identification","multi-scale detail injection into clip feature space for fine-grained remote sensing retrieval","masked multimodal modeling for fine-grained vision-language representation learning"]},{"direction_index":6,"label":"D6","name":"Adaptation and efficient transfer","description":"Adapting, compressing, transferring, or deploying multimodal representations under limited compute, data, parameters, communication, modality acquisition, or annotation budgets.","inclusion":"Include work whose primary objective is transfer of pretrained multimodal or vision-language models to new languages, modalities, tasks, or low-resource settings; prompt/adaptor/parameter-efficient tuning; multimodal-to-unimodal distillation; modality selection; efficient training; federated learning; active learning; or resource-constrained multimodal representation learning.","exclusion":"If the main contribution is new full pretraining from scratch, assign to D0 or D2; if the central novelty is domain-specific structure rather than transfer or efficiency, assign to D7; if missing-modality robustness rather than resource-aware deployment is primary, assign to D4.","nearest_boundary":"D7","representative_topics":["multilingual multimodal retrieval via cross-lingual pseudo-label transfer from pretrained multilingual models","post-pretraining image-text models for video-language representation alignment via omnisource cross-modal learning","parameter-free attention enhancement of contrastive vision-language pre-training for zero-shot transfer"]},{"direction_index":7,"label":"D7","name":"Domain and human-centered representations","description":"Developing multimodal representations whose primary objective is to capture specialized domain structures or human-centered affective, behavioral, physiological, motion, music, or sentiment signal structure.","inclusion":"Include work where the scientific objective is modeling domain-specific multimodal semantics, such as biomedical, neuroimaging, molecular, document, remote-sensing, product, commodity, or registration structure, or human-centered affect, emotion, sentiment, physiology, music, dance, motion, and behavior representations.","exclusion":"Do not include general alignment, fusion, or pretraining methods merely evaluated on a domain benchmark; assign those to D0, D2, or D3. If the main objective is lightweight adaptation of an existing pretrained model rather than domain or human-signal structure, assign to D6.","nearest_boundary":"D6","representative_topics":["entity graph enhanced cross-modal pretraining for weakly-supervised instance-level product retrieval","unsupervised contrastive representation learning for non-rigid multimodal deformable image registration","audio-visual music-dance representation learning via explicit-implicit rhythm synchronization"]}],"split":"dev"}
{"episode_id":"F102_T2025-07","field_index":102,"field":"multimodal representation learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cross-modal alignment and foundations","description":"Learning, characterizing, and evaluating shared multimodal embedding spaces in which heterogeneous modalities become semantically comparable, together with formal, empirical, and benchmark-level analyses of such representation learning.","inclusion":"Include work whose primary objective is global cross-modal semantic alignment, contrastive or embedding-space design, theoretical analysis of multimodal objectives or transfer, controlled empirical characterization, or survey/benchmark infrastructure for multimodal representation learning.","exclusion":"If the main objective is a unified generative architecture, masked/generative pretraining objective, fine-grained region/phrase grounding, domain-specific modeling, or lightweight adaptation of an existing pretrained model, assign to the corresponding specialized direction.","nearest_boundary":"D2","representative_topics":["sparse mixture of experts for multimodal contrastive image-text learning with balanced expert activation","prototypical contrastive pretraining with back translation for robust image-text representation learning","controlled study of caption supervision effects on transferable visual representations"]},{"direction_index":1,"label":"D1","name":"Unified generalist architectures","description":"Designing single multimodal architectures or interfaces that unify multiple modalities, tasks, and input-output formats for broad multimodal understanding, generation, or instruction-following.","inclusion":"Include work whose primary objective is architectural unification through language-model interfaces, autoregressive or encoder-decoder input-output frameworks, discrete token homogenization, multimodal instruction systems, or generalist multimodal foundation architectures.","exclusion":"If the work mainly proposes a non-generative fusion block, a contrastive alignment loss, or a masked/generative objective without a generalist architecture, assign to D3, D0, or D2 respectively.","nearest_boundary":"D3","representative_topics":["language model as universal interface for modular multimodal encoder integration via semi-causal pretraining","unified transformer model via discrete token homogenization for diverse vision and language tasks","multimodal foundation model pretraining unifying vision and vision-language tasks via masked language modeling"]},{"direction_index":2,"label":"D2","name":"Masked and generative pretraining","description":"Developing transferable multimodal representations through self-supervised, masked, reconstruction, prediction, or hybrid generative-discriminative pretraining objectives.","inclusion":"Include work whose primary objective is a multimodal pretraining objective or recipe such as masked multimodal modeling, masked autoencoding, encoder-decoder reconstruction, cross-modal prediction, or weakly/self-supervised generative representation learning.","exclusion":"If contrastive shared-space comparability is the dominant contribution, assign to D0; if the objective is subordinate to a generalist unified architecture, assign to D1; if the objective is mainly domain-specific, assign to D7.","nearest_boundary":"D0","representative_topics":["masked encoder-decoder pre-training unifying vision and language for text recognition","unified masked autoencoding for joint image and video vision transformer pretraining","bidirectional vision-language pre-training for visually-rich document understanding and information extraction"]},{"direction_index":3,"label":"D3","name":"Modality interaction architectures","description":"Designing architectural mechanisms that determine how modality streams exchange, fuse, attend to, bridge, or jointly structure information.","inclusion":"Include work whose primary objective is a fusion architecture or modality-interaction mechanism, including cross-attention variants, bridge layers, token exchange, progressive fusion, hierarchical modality attention, graph-based multimodal interaction, or transformer fusion design.","exclusion":"If the architecture is primarily a generalist LLM or unified generative interface, assign to D1; if the main problem is missing/noisy modality robustness rather than interaction design, assign to D4.","nearest_boundary":"D1","representative_topics":["two-tower vision-language models with multi-layer bridge connections for cross-modal alignment","unified audio-visual model architecture for joint event classification","orthogonal projection and branched fusion for multilingual face-voice association"]},{"direction_index":4,"label":"D4","name":"Robust and balanced latent structure","description":"Learning multimodal representations that remain reliable under missing, noisy, imbalanced, or conflicting modalities while structuring shared and modality-specific latent information.","inclusion":"Include work whose primary objective is robustness to missing or corrupted modalities, modality imbalance, modality collapse mitigation, gradient harmonization, common-private decomposition, uncertainty-aware or information-theoretic latent structure, or multi-view factorization.","exclusion":"If robustness is only incidental to a general alignment objective, assign to D0; if the contribution is mainly a fusion architecture without explicit robustness or latent-factor balancing, assign to D3.","nearest_boundary":"D0","representative_topics":["mitigating modality collapse in multimodal vae training via impartial gradient optimization","hierarchical variational autoencoders with factor analysis latent space for heterogeneous multimodal data","optimal transport alignment dynamics for missing-modality inference in multimodal sequences"]},{"direction_index":5,"label":"D5","name":"Fine-grained structured grounding","description":"Learning multimodal representations that preserve local, part-level, region-level, phrase-level, relation-level, multi-scale, or otherwise structured cross-modal correspondences.","inclusion":"Include work whose primary objective is explicit fine-grained grounding or structured correspondence beyond global modality-level similarity, including part-aware retrieval, region-word alignment, spatial or relation-aware image-text retrieval, multi-granularity representations, and multi-scale cross-modal correspondence learning.","exclusion":"If the work only learns global embeddings without explicit local or structured correspondence, assign to D0; if fine-grained modeling is mainly a domain-specific adaptation device, assign to D7 or D6 according to the primary objective.","nearest_boundary":"D0","representative_topics":["granularity-unified cross-modal representation learning for text-to-image person re-identification","multi-scale detail injection into clip feature space for fine-grained remote sensing retrieval","masked multimodal modeling for fine-grained vision-language representation learning"]},{"direction_index":6,"label":"D6","name":"Adaptation and efficient transfer","description":"Adapting, compressing, transferring, or deploying multimodal representations under limited compute, data, parameters, communication, modality acquisition, or annotation budgets.","inclusion":"Include work whose primary objective is transfer of pretrained multimodal or vision-language models to new languages, modalities, tasks, or low-resource settings; prompt/adaptor/parameter-efficient tuning; multimodal-to-unimodal distillation; modality selection; efficient training; federated learning; active learning; or resource-constrained multimodal representation learning.","exclusion":"If the main contribution is new full pretraining from scratch, assign to D0 or D2; if the central novelty is domain-specific structure rather than transfer or efficiency, assign to D7; if missing-modality robustness rather than resource-aware deployment is primary, assign to D4.","nearest_boundary":"D7","representative_topics":["multilingual multimodal retrieval via cross-lingual pseudo-label transfer from pretrained multilingual models","post-pretraining image-text models for video-language representation alignment via omnisource cross-modal learning","parameter-free attention enhancement of contrastive vision-language pre-training for zero-shot transfer"]},{"direction_index":7,"label":"D7","name":"Domain and human-centered representations","description":"Developing multimodal representations whose primary objective is to capture specialized domain structures or human-centered affective, behavioral, physiological, motion, music, or sentiment signal structure.","inclusion":"Include work where the scientific objective is modeling domain-specific multimodal semantics, such as biomedical, neuroimaging, molecular, document, remote-sensing, product, commodity, or registration structure, or human-centered affect, emotion, sentiment, physiology, music, dance, motion, and behavior representations.","exclusion":"Do not include general alignment, fusion, or pretraining methods merely evaluated on a domain benchmark; assign those to D0, D2, or D3. If the main objective is lightweight adaptation of an existing pretrained model rather than domain or human-signal structure, assign to D6.","nearest_boundary":"D6","representative_topics":["entity graph enhanced cross-modal pretraining for weakly-supervised instance-level product retrieval","unsupervised contrastive representation learning for non-rigid multimodal deformable image registration","audio-visual music-dance representation learning via explicit-implicit rhythm synchronization"]}],"split":"dev"}
{"episode_id":"F102_T2026-01","field_index":102,"field":"multimodal representation learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cross-modal alignment and foundations","description":"Learning, characterizing, and evaluating shared multimodal embedding spaces in which heterogeneous modalities become semantically comparable, together with formal, empirical, and benchmark-level analyses of such representation learning.","inclusion":"Include work whose primary objective is global cross-modal semantic alignment, contrastive or embedding-space design, theoretical analysis of multimodal objectives or transfer, controlled empirical characterization, or survey/benchmark infrastructure for multimodal representation learning.","exclusion":"If the main objective is a unified generative architecture, masked/generative pretraining objective, fine-grained region/phrase grounding, domain-specific modeling, or lightweight adaptation of an existing pretrained model, assign to the corresponding specialized direction.","nearest_boundary":"D2","representative_topics":["sparse mixture of experts for multimodal contrastive image-text learning with balanced expert activation","prototypical contrastive pretraining with back translation for robust image-text representation learning","controlled study of caption supervision effects on transferable visual representations"]},{"direction_index":1,"label":"D1","name":"Unified generalist architectures","description":"Designing single multimodal architectures or interfaces that unify multiple modalities, tasks, and input-output formats for broad multimodal understanding, generation, or instruction-following.","inclusion":"Include work whose primary objective is architectural unification through language-model interfaces, autoregressive or encoder-decoder input-output frameworks, discrete token homogenization, multimodal instruction systems, or generalist multimodal foundation architectures.","exclusion":"If the work mainly proposes a non-generative fusion block, a contrastive alignment loss, or a masked/generative objective without a generalist architecture, assign to D3, D0, or D2 respectively.","nearest_boundary":"D3","representative_topics":["language model as universal interface for modular multimodal encoder integration via semi-causal pretraining","unified transformer model via discrete token homogenization for diverse vision and language tasks","multimodal foundation model pretraining unifying vision and vision-language tasks via masked language modeling"]},{"direction_index":2,"label":"D2","name":"Masked and generative pretraining","description":"Developing transferable multimodal representations through self-supervised, masked, reconstruction, prediction, or hybrid generative-discriminative pretraining objectives.","inclusion":"Include work whose primary objective is a multimodal pretraining objective or recipe such as masked multimodal modeling, masked autoencoding, encoder-decoder reconstruction, cross-modal prediction, or weakly/self-supervised generative representation learning.","exclusion":"If contrastive shared-space comparability is the dominant contribution, assign to D0; if the objective is subordinate to a generalist unified architecture, assign to D1; if the objective is mainly domain-specific, assign to D7.","nearest_boundary":"D0","representative_topics":["masked encoder-decoder pre-training unifying vision and language for text recognition","unified masked autoencoding for joint image and video vision transformer pretraining","bidirectional vision-language pre-training for visually-rich document understanding and information extraction"]},{"direction_index":3,"label":"D3","name":"Modality interaction architectures","description":"Designing architectural mechanisms that determine how modality streams exchange, fuse, attend to, bridge, or jointly structure information.","inclusion":"Include work whose primary objective is a fusion architecture or modality-interaction mechanism, including cross-attention variants, bridge layers, token exchange, progressive fusion, hierarchical modality attention, graph-based multimodal interaction, or transformer fusion design.","exclusion":"If the architecture is primarily a generalist LLM or unified generative interface, assign to D1; if the main problem is missing/noisy modality robustness rather than interaction design, assign to D4.","nearest_boundary":"D1","representative_topics":["two-tower vision-language models with multi-layer bridge connections for cross-modal alignment","unified audio-visual model architecture for joint event classification","orthogonal projection and branched fusion for multilingual face-voice association"]},{"direction_index":4,"label":"D4","name":"Robust and balanced latent structure","description":"Learning multimodal representations that remain reliable under missing, noisy, imbalanced, or conflicting modalities while structuring shared and modality-specific latent information.","inclusion":"Include work whose primary objective is robustness to missing or corrupted modalities, modality imbalance, modality collapse mitigation, gradient harmonization, common-private decomposition, uncertainty-aware or information-theoretic latent structure, or multi-view factorization.","exclusion":"If robustness is only incidental to a general alignment objective, assign to D0; if the contribution is mainly a fusion architecture without explicit robustness or latent-factor balancing, assign to D3.","nearest_boundary":"D0","representative_topics":["mitigating modality collapse in multimodal vae training via impartial gradient optimization","hierarchical variational autoencoders with factor analysis latent space for heterogeneous multimodal data","optimal transport alignment dynamics for missing-modality inference in multimodal sequences"]},{"direction_index":5,"label":"D5","name":"Fine-grained structured grounding","description":"Learning multimodal representations that preserve local, part-level, region-level, phrase-level, relation-level, multi-scale, or otherwise structured cross-modal correspondences.","inclusion":"Include work whose primary objective is explicit fine-grained grounding or structured correspondence beyond global modality-level similarity, including part-aware retrieval, region-word alignment, spatial or relation-aware image-text retrieval, multi-granularity representations, and multi-scale cross-modal correspondence learning.","exclusion":"If the work only learns global embeddings without explicit local or structured correspondence, assign to D0; if fine-grained modeling is mainly a domain-specific adaptation device, assign to D7 or D6 according to the primary objective.","nearest_boundary":"D0","representative_topics":["granularity-unified cross-modal representation learning for text-to-image person re-identification","multi-scale detail injection into clip feature space for fine-grained remote sensing retrieval","masked multimodal modeling for fine-grained vision-language representation learning"]},{"direction_index":6,"label":"D6","name":"Adaptation and efficient transfer","description":"Adapting, compressing, transferring, or deploying multimodal representations under limited compute, data, parameters, communication, modality acquisition, or annotation budgets.","inclusion":"Include work whose primary objective is transfer of pretrained multimodal or vision-language models to new languages, modalities, tasks, or low-resource settings; prompt/adaptor/parameter-efficient tuning; multimodal-to-unimodal distillation; modality selection; efficient training; federated learning; active learning; or resource-constrained multimodal representation learning.","exclusion":"If the main contribution is new full pretraining from scratch, assign to D0 or D2; if the central novelty is domain-specific structure rather than transfer or efficiency, assign to D7; if missing-modality robustness rather than resource-aware deployment is primary, assign to D4.","nearest_boundary":"D7","representative_topics":["multilingual multimodal retrieval via cross-lingual pseudo-label transfer from pretrained multilingual models","post-pretraining image-text models for video-language representation alignment via omnisource cross-modal learning","parameter-free attention enhancement of contrastive vision-language pre-training for zero-shot transfer"]},{"direction_index":7,"label":"D7","name":"Domain and human-centered representations","description":"Developing multimodal representations whose primary objective is to capture specialized domain structures or human-centered affective, behavioral, physiological, motion, music, or sentiment signal structure.","inclusion":"Include work where the scientific objective is modeling domain-specific multimodal semantics, such as biomedical, neuroimaging, molecular, document, remote-sensing, product, commodity, or registration structure, or human-centered affect, emotion, sentiment, physiology, music, dance, motion, and behavior representations.","exclusion":"Do not include general alignment, fusion, or pretraining methods merely evaluated on a domain benchmark; assign those to D0, D2, or D3. If the main objective is lightweight adaptation of an existing pretrained model rather than domain or human-signal structure, assign to D6.","nearest_boundary":"D6","representative_topics":["entity graph enhanced cross-modal pretraining for weakly-supervised instance-level product retrieval","unsupervised contrastive representation learning for non-rigid multimodal deformable image registration","audio-visual music-dance representation learning via explicit-implicit rhythm synchronization"]}],"split":"dev"}
{"episode_id":"F105_T2024-01","field_index":105,"field":"neural network optimization theory","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Implicit bias and induced regularization","description":"Characterize the solution structures selected by training dynamics beyond explicit empirical-risk minimization, including margin, norm, sparsity, rank, balancing, and optimizer- or reparametrization-induced regularization.","inclusion":"Include work whose primary objective is to identify or prove optimizer-induced selection of particular final solutions, representations of solutions, or structural regularizers in overparameterized neural-network-like parameterizations.","exclusion":"Use D1 for results mainly proving that training converges; use D3 when the central phenomenon is sharpness or edge-of-stability behavior; use D7 when the main object is feature acquisition rather than selected final-solution structure.","nearest_boundary":"D3 Sharpness and stability-boundary dynamics","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","layer-wise noise injection for explicit regularization in overparametrized neural networks","stochastic gradient descent and weight decay bias toward low-rank weight matrix learning"]},{"direction_index":1,"label":"D1","name":"Global convergence of neural training","description":"Prove that standard gradient descent, gradient flow, or plain SGD-style first-order training reaches interpolation, global optimality, stationarity, or learning targets for specified neural-network classes.","inclusion":"Include work whose primary objective is a neural-network-specific convergence, rate, escape, or reachability theorem tied to architecture, initialization, width, distribution, or loss assumptions.","exclusion":"Use D2 when the loss landscape or solution set is the main object rather than a training trajectory; use D4 for stochastic-protocol or broad stochastic-approximation rate theory; use D5 for adaptive, momentum, proximal, natural-gradient, or other nonstandard optimizer families as the main object.","nearest_boundary":"D2 Landscape geometry and computational tractability","representative_topics":["early and global convergence of gradient descent via neuron activation partition analysis","theoretical analysis of global convergence in over-parameterized relu nets via activation matrix angular geometry","coupled weight initialization for tighter width bounds in two-layer relu network convergence"]},{"direction_index":2,"label":"D2","name":"Landscape geometry and tractability","description":"Characterize the static structure, topology, geometry, combinatorics, solution sets, critical points, symmetries, convexifications, and computational tractability or hardness of neural-network optimization objectives.","inclusion":"Include work whose primary objective is to describe the objective landscape, minimizer or critical-point structure, connectivity, activation-region geometry, architectural landscape effects, or complexity barriers for training problems.","exclusion":"Use D1 when landscape facts are mainly tools for proving a particular gradient-training convergence result; use D3 when the focus is sharpness evolution under training; use D5 when the contribution is chiefly a new optimizer or training framework rather than landscape or hardness characterization.","nearest_boundary":"D1 Global convergence of neural training","representative_topics":["parameter convex neural networks using exponential activations for improved generalization","depth flattening of optimization landscape in robust over-parameterized deep linear models","intrinsic loss landscape analysis of scale-invariant neural network training via spherical optimization regimes"]},{"direction_index":3,"label":"D3","name":"Sharpness and stability-boundary dynamics","description":"Explain optimization behavior governed by sharpness, flatness, large learning rates, edge-of-stability thresholds, catapult phases, progressive sharpening, and sharpness-aware objectives such as SAM.","inclusion":"Include work whose primary objective is to analyze sharpness dynamics, flat-minima selection through stability mechanisms, edge-of-stability training, or SAM/SAM-like optimization and regularization mechanisms.","exclusion":"Use D0 when sharpness is secondary to a broader implicit-bias claim about margin, norm, rank, or sparsity; use D4 when stochastic noise is modeled as a process without sharpness as the central objective; use D2 for static curvature or flatness properties not tied to training dynamics.","nearest_boundary":"D0 Implicit bias and induced regularization","representative_topics":["pde-based theoretical framework for edge of stability and numerical instability in deep network training","normalization-induced sharpness reduction in gradient descent for generalization","theoretical analysis of edge-of-stability gradient descent dynamics via minimalist counterexample"]},{"direction_index":4,"label":"D4","name":"Stochastic-gradient dynamics and rates","description":"Analyze stochastic-gradient optimization through noise models, diffusion or statistical-mechanics limits, sampling protocols, distributed/asynchronous execution, and broad stochastic-approximation convergence or efficiency theory.","inclusion":"Include work whose primary objective is to understand SGD or stochastic approximation as a stochastic process, to prove convergence or complexity under stochastic noise, or to quantify effects of reshuffling, client sampling, minibatching, delays, low precision, or weak stochastic assumptions.","exclusion":"Use D3 when stochastic effects are studied mainly to explain sharpness or edge-of-stability behavior; use D5 for adaptive, momentum, sign, proximal, natural-gradient, or engineered optimizer rules as the main object; use D6 when the main deliverable is a generalization or stability bound rather than optimization dynamics or rates.","nearest_boundary":"D5 Optimizer and training-framework design","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","sharp sgd convergence rates under flexible client and data sampling in nonconvex optimization","unified convergence theorem for stochastic gradient and proximal gradient methods in nonsmooth nonconvex settings"]},{"direction_index":5,"label":"D5","name":"Optimizer and training-framework design","description":"Develop or analyze optimizer families and training mechanisms beyond plain GD/SGD, including adaptive methods, momentum, sign methods, natural-gradient, proximal, alternating-minimization, control-theoretic, scheduling, normalization, feedback, and reparameterization-aware frameworks.","inclusion":"Include work whose primary objective is the theory, mechanism, convergence, robustness, or design of a modified update rule or practical training-control framework rather than the baseline neural-network objective itself.","exclusion":"Use D1 for convergence of ordinary gradient methods in neural networks; use D4 when the main variable is stochastic sampling or distributed execution rather than the update formula; use D3 for SAM or sharpness-aware objectives as the central object.","nearest_boundary":"D4 Stochastic-gradient dynamics and rates","representative_topics":["mini-batch momentum convergence dynamics via volterra approximation and implicit conditioning ratio in high dimensions","control-theoretic framework for adaptive gradient optimizers via state-space transfer functions","adaptive optimizer slingshot mechanism as cause of grokking via cyclic weight norm transitions"]},{"direction_index":6,"label":"D6","name":"Optimization-driven generalization and stability","description":"Derive generalization, excess-risk, robustness, or algorithmic-stability guarantees from properties of gradient-based training algorithms, stochastic trajectories, or optimization-induced controls.","inclusion":"Include work whose primary objective is to connect a training algorithm or trajectory to out-of-sample error, stability, robustness, excess risk, consistency, or generalization bounds.","exclusion":"Use D0 when the main claim is structural solution selection without a generalization bound; use D4 for convergence or stochastic-process results without generalization as the target; use D7 for representation-learning mechanisms where feature acquisition, not a bound, is central.","nearest_boundary":"D0 Implicit bias and induced regularization","representative_topics":["algorithmic stability bounds for heavy-tailed stochastic gradient descent generalization","stability and generalization bounds for stochastic gradient methods on nonconvex nonsmooth objectives","consistency theory for regularized neural networks via sieve and minimal network methods"]},{"direction_index":7,"label":"D7","name":"Feature learning and wide-regime limits","description":"Explain how optimization produces learned features or representations and how training behaves across kernel, mean-field, infinite-width, large-width, and finite-width feature-learning regimes.","inclusion":"Include work whose primary objective is representation formation, feature-learning departures from lazy kernels, NTK-to-mean-field transitions, width or particle limits of training, scaling regimes, spectral or hierarchical feature acquisition, or mean-field/kernel descriptions of learned dynamics.","exclusion":"Use D0 for final-solution implicit bias without feature-acquisition dynamics; use D1 for pure NTK-style convergence proofs without representation-learning analysis; use D4 for diffusion or SGD-noise limits not primarily arising from network width or mean-field structure.","nearest_boundary":"D1 Global convergence of neural training","representative_topics":["spectral ntk analysis for joint dense and sparse polynomial learning via gradient descent escape","large-width asymptotic analysis of relu networks with alpha-stable initialization and random kernel dynamics","theoretical analysis of feature learning overfitting via sparse representations in fully-connected networks"]}],"split":"dev"}
{"episode_id":"F105_T2024-07","field_index":105,"field":"neural network optimization theory","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Implicit bias and induced regularization","description":"Characterize the solution structures selected by training dynamics beyond explicit empirical-risk minimization, including margin, norm, sparsity, rank, balancing, and optimizer- or reparametrization-induced regularization.","inclusion":"Include work whose primary objective is to identify or prove optimizer-induced selection of particular final solutions, representations of solutions, or structural regularizers in overparameterized neural-network-like parameterizations.","exclusion":"Use D1 for results mainly proving that training converges; use D3 when the central phenomenon is sharpness or edge-of-stability behavior; use D7 when the main object is feature acquisition rather than selected final-solution structure.","nearest_boundary":"D3 Sharpness and stability-boundary dynamics","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","layer-wise noise injection for explicit regularization in overparametrized neural networks","stochastic gradient descent and weight decay bias toward low-rank weight matrix learning"]},{"direction_index":1,"label":"D1","name":"Global convergence of neural training","description":"Prove that standard gradient descent, gradient flow, or plain SGD-style first-order training reaches interpolation, global optimality, stationarity, or learning targets for specified neural-network classes.","inclusion":"Include work whose primary objective is a neural-network-specific convergence, rate, escape, or reachability theorem tied to architecture, initialization, width, distribution, or loss assumptions.","exclusion":"Use D2 when the loss landscape or solution set is the main object rather than a training trajectory; use D4 for stochastic-protocol or broad stochastic-approximation rate theory; use D5 for adaptive, momentum, proximal, natural-gradient, or other nonstandard optimizer families as the main object.","nearest_boundary":"D2 Landscape geometry and computational tractability","representative_topics":["early and global convergence of gradient descent via neuron activation partition analysis","theoretical analysis of global convergence in over-parameterized relu nets via activation matrix angular geometry","coupled weight initialization for tighter width bounds in two-layer relu network convergence"]},{"direction_index":2,"label":"D2","name":"Landscape geometry and tractability","description":"Characterize the static structure, topology, geometry, combinatorics, solution sets, critical points, symmetries, convexifications, and computational tractability or hardness of neural-network optimization objectives.","inclusion":"Include work whose primary objective is to describe the objective landscape, minimizer or critical-point structure, connectivity, activation-region geometry, architectural landscape effects, or complexity barriers for training problems.","exclusion":"Use D1 when landscape facts are mainly tools for proving a particular gradient-training convergence result; use D3 when the focus is sharpness evolution under training; use D5 when the contribution is chiefly a new optimizer or training framework rather than landscape or hardness characterization.","nearest_boundary":"D1 Global convergence of neural training","representative_topics":["parameter convex neural networks using exponential activations for improved generalization","depth flattening of optimization landscape in robust over-parameterized deep linear models","intrinsic loss landscape analysis of scale-invariant neural network training via spherical optimization regimes"]},{"direction_index":3,"label":"D3","name":"Sharpness and stability-boundary dynamics","description":"Explain optimization behavior governed by sharpness, flatness, large learning rates, edge-of-stability thresholds, catapult phases, progressive sharpening, and sharpness-aware objectives such as SAM.","inclusion":"Include work whose primary objective is to analyze sharpness dynamics, flat-minima selection through stability mechanisms, edge-of-stability training, or SAM/SAM-like optimization and regularization mechanisms.","exclusion":"Use D0 when sharpness is secondary to a broader implicit-bias claim about margin, norm, rank, or sparsity; use D4 when stochastic noise is modeled as a process without sharpness as the central objective; use D2 for static curvature or flatness properties not tied to training dynamics.","nearest_boundary":"D0 Implicit bias and induced regularization","representative_topics":["pde-based theoretical framework for edge of stability and numerical instability in deep network training","normalization-induced sharpness reduction in gradient descent for generalization","theoretical analysis of edge-of-stability gradient descent dynamics via minimalist counterexample"]},{"direction_index":4,"label":"D4","name":"Stochastic-gradient dynamics and rates","description":"Analyze stochastic-gradient optimization through noise models, diffusion or statistical-mechanics limits, sampling protocols, distributed/asynchronous execution, and broad stochastic-approximation convergence or efficiency theory.","inclusion":"Include work whose primary objective is to understand SGD or stochastic approximation as a stochastic process, to prove convergence or complexity under stochastic noise, or to quantify effects of reshuffling, client sampling, minibatching, delays, low precision, or weak stochastic assumptions.","exclusion":"Use D3 when stochastic effects are studied mainly to explain sharpness or edge-of-stability behavior; use D5 for adaptive, momentum, sign, proximal, natural-gradient, or engineered optimizer rules as the main object; use D6 when the main deliverable is a generalization or stability bound rather than optimization dynamics or rates.","nearest_boundary":"D5 Optimizer and training-framework design","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","sharp sgd convergence rates under flexible client and data sampling in nonconvex optimization","unified convergence theorem for stochastic gradient and proximal gradient methods in nonsmooth nonconvex settings"]},{"direction_index":5,"label":"D5","name":"Optimizer and training-framework design","description":"Develop or analyze optimizer families and training mechanisms beyond plain GD/SGD, including adaptive methods, momentum, sign methods, natural-gradient, proximal, alternating-minimization, control-theoretic, scheduling, normalization, feedback, and reparameterization-aware frameworks.","inclusion":"Include work whose primary objective is the theory, mechanism, convergence, robustness, or design of a modified update rule or practical training-control framework rather than the baseline neural-network objective itself.","exclusion":"Use D1 for convergence of ordinary gradient methods in neural networks; use D4 when the main variable is stochastic sampling or distributed execution rather than the update formula; use D3 for SAM or sharpness-aware objectives as the central object.","nearest_boundary":"D4 Stochastic-gradient dynamics and rates","representative_topics":["mini-batch momentum convergence dynamics via volterra approximation and implicit conditioning ratio in high dimensions","control-theoretic framework for adaptive gradient optimizers via state-space transfer functions","adaptive optimizer slingshot mechanism as cause of grokking via cyclic weight norm transitions"]},{"direction_index":6,"label":"D6","name":"Optimization-driven generalization and stability","description":"Derive generalization, excess-risk, robustness, or algorithmic-stability guarantees from properties of gradient-based training algorithms, stochastic trajectories, or optimization-induced controls.","inclusion":"Include work whose primary objective is to connect a training algorithm or trajectory to out-of-sample error, stability, robustness, excess risk, consistency, or generalization bounds.","exclusion":"Use D0 when the main claim is structural solution selection without a generalization bound; use D4 for convergence or stochastic-process results without generalization as the target; use D7 for representation-learning mechanisms where feature acquisition, not a bound, is central.","nearest_boundary":"D0 Implicit bias and induced regularization","representative_topics":["algorithmic stability bounds for heavy-tailed stochastic gradient descent generalization","stability and generalization bounds for stochastic gradient methods on nonconvex nonsmooth objectives","consistency theory for regularized neural networks via sieve and minimal network methods"]},{"direction_index":7,"label":"D7","name":"Feature learning and wide-regime limits","description":"Explain how optimization produces learned features or representations and how training behaves across kernel, mean-field, infinite-width, large-width, and finite-width feature-learning regimes.","inclusion":"Include work whose primary objective is representation formation, feature-learning departures from lazy kernels, NTK-to-mean-field transitions, width or particle limits of training, scaling regimes, spectral or hierarchical feature acquisition, or mean-field/kernel descriptions of learned dynamics.","exclusion":"Use D0 for final-solution implicit bias without feature-acquisition dynamics; use D1 for pure NTK-style convergence proofs without representation-learning analysis; use D4 for diffusion or SGD-noise limits not primarily arising from network width or mean-field structure.","nearest_boundary":"D1 Global convergence of neural training","representative_topics":["spectral ntk analysis for joint dense and sparse polynomial learning via gradient descent escape","large-width asymptotic analysis of relu networks with alpha-stable initialization and random kernel dynamics","theoretical analysis of feature learning overfitting via sparse representations in fully-connected networks"]}],"split":"dev"}
{"episode_id":"F105_T2025-01","field_index":105,"field":"neural network optimization theory","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Implicit bias and induced regularization","description":"Characterize the solution structures selected by training dynamics beyond explicit empirical-risk minimization, including margin, norm, sparsity, rank, balancing, and optimizer- or reparametrization-induced regularization.","inclusion":"Include work whose primary objective is to identify or prove optimizer-induced selection of particular final solutions, representations of solutions, or structural regularizers in overparameterized neural-network-like parameterizations.","exclusion":"Use D1 for results mainly proving that training converges; use D3 when the central phenomenon is sharpness or edge-of-stability behavior; use D7 when the main object is feature acquisition rather than selected final-solution structure.","nearest_boundary":"D3 Sharpness and stability-boundary dynamics","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","layer-wise noise injection for explicit regularization in overparametrized neural networks","stochastic gradient descent and weight decay bias toward low-rank weight matrix learning"]},{"direction_index":1,"label":"D1","name":"Global convergence of neural training","description":"Prove that standard gradient descent, gradient flow, or plain SGD-style first-order training reaches interpolation, global optimality, stationarity, or learning targets for specified neural-network classes.","inclusion":"Include work whose primary objective is a neural-network-specific convergence, rate, escape, or reachability theorem tied to architecture, initialization, width, distribution, or loss assumptions.","exclusion":"Use D2 when the loss landscape or solution set is the main object rather than a training trajectory; use D4 for stochastic-protocol or broad stochastic-approximation rate theory; use D5 for adaptive, momentum, proximal, natural-gradient, or other nonstandard optimizer families as the main object.","nearest_boundary":"D2 Landscape geometry and computational tractability","representative_topics":["early and global convergence of gradient descent via neuron activation partition analysis","theoretical analysis of global convergence in over-parameterized relu nets via activation matrix angular geometry","coupled weight initialization for tighter width bounds in two-layer relu network convergence"]},{"direction_index":2,"label":"D2","name":"Landscape geometry and tractability","description":"Characterize the static structure, topology, geometry, combinatorics, solution sets, critical points, symmetries, convexifications, and computational tractability or hardness of neural-network optimization objectives.","inclusion":"Include work whose primary objective is to describe the objective landscape, minimizer or critical-point structure, connectivity, activation-region geometry, architectural landscape effects, or complexity barriers for training problems.","exclusion":"Use D1 when landscape facts are mainly tools for proving a particular gradient-training convergence result; use D3 when the focus is sharpness evolution under training; use D5 when the contribution is chiefly a new optimizer or training framework rather than landscape or hardness characterization.","nearest_boundary":"D1 Global convergence of neural training","representative_topics":["parameter convex neural networks using exponential activations for improved generalization","depth flattening of optimization landscape in robust over-parameterized deep linear models","intrinsic loss landscape analysis of scale-invariant neural network training via spherical optimization regimes"]},{"direction_index":3,"label":"D3","name":"Sharpness and stability-boundary dynamics","description":"Explain optimization behavior governed by sharpness, flatness, large learning rates, edge-of-stability thresholds, catapult phases, progressive sharpening, and sharpness-aware objectives such as SAM.","inclusion":"Include work whose primary objective is to analyze sharpness dynamics, flat-minima selection through stability mechanisms, edge-of-stability training, or SAM/SAM-like optimization and regularization mechanisms.","exclusion":"Use D0 when sharpness is secondary to a broader implicit-bias claim about margin, norm, rank, or sparsity; use D4 when stochastic noise is modeled as a process without sharpness as the central objective; use D2 for static curvature or flatness properties not tied to training dynamics.","nearest_boundary":"D0 Implicit bias and induced regularization","representative_topics":["pde-based theoretical framework for edge of stability and numerical instability in deep network training","normalization-induced sharpness reduction in gradient descent for generalization","theoretical analysis of edge-of-stability gradient descent dynamics via minimalist counterexample"]},{"direction_index":4,"label":"D4","name":"Stochastic-gradient dynamics and rates","description":"Analyze stochastic-gradient optimization through noise models, diffusion or statistical-mechanics limits, sampling protocols, distributed/asynchronous execution, and broad stochastic-approximation convergence or efficiency theory.","inclusion":"Include work whose primary objective is to understand SGD or stochastic approximation as a stochastic process, to prove convergence or complexity under stochastic noise, or to quantify effects of reshuffling, client sampling, minibatching, delays, low precision, or weak stochastic assumptions.","exclusion":"Use D3 when stochastic effects are studied mainly to explain sharpness or edge-of-stability behavior; use D5 for adaptive, momentum, sign, proximal, natural-gradient, or engineered optimizer rules as the main object; use D6 when the main deliverable is a generalization or stability bound rather than optimization dynamics or rates.","nearest_boundary":"D5 Optimizer and training-framework design","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","sharp sgd convergence rates under flexible client and data sampling in nonconvex optimization","unified convergence theorem for stochastic gradient and proximal gradient methods in nonsmooth nonconvex settings"]},{"direction_index":5,"label":"D5","name":"Optimizer and training-framework design","description":"Develop or analyze optimizer families and training mechanisms beyond plain GD/SGD, including adaptive methods, momentum, sign methods, natural-gradient, proximal, alternating-minimization, control-theoretic, scheduling, normalization, feedback, and reparameterization-aware frameworks.","inclusion":"Include work whose primary objective is the theory, mechanism, convergence, robustness, or design of a modified update rule or practical training-control framework rather than the baseline neural-network objective itself.","exclusion":"Use D1 for convergence of ordinary gradient methods in neural networks; use D4 when the main variable is stochastic sampling or distributed execution rather than the update formula; use D3 for SAM or sharpness-aware objectives as the central object.","nearest_boundary":"D4 Stochastic-gradient dynamics and rates","representative_topics":["mini-batch momentum convergence dynamics via volterra approximation and implicit conditioning ratio in high dimensions","control-theoretic framework for adaptive gradient optimizers via state-space transfer functions","adaptive optimizer slingshot mechanism as cause of grokking via cyclic weight norm transitions"]},{"direction_index":6,"label":"D6","name":"Optimization-driven generalization and stability","description":"Derive generalization, excess-risk, robustness, or algorithmic-stability guarantees from properties of gradient-based training algorithms, stochastic trajectories, or optimization-induced controls.","inclusion":"Include work whose primary objective is to connect a training algorithm or trajectory to out-of-sample error, stability, robustness, excess risk, consistency, or generalization bounds.","exclusion":"Use D0 when the main claim is structural solution selection without a generalization bound; use D4 for convergence or stochastic-process results without generalization as the target; use D7 for representation-learning mechanisms where feature acquisition, not a bound, is central.","nearest_boundary":"D0 Implicit bias and induced regularization","representative_topics":["algorithmic stability bounds for heavy-tailed stochastic gradient descent generalization","stability and generalization bounds for stochastic gradient methods on nonconvex nonsmooth objectives","consistency theory for regularized neural networks via sieve and minimal network methods"]},{"direction_index":7,"label":"D7","name":"Feature learning and wide-regime limits","description":"Explain how optimization produces learned features or representations and how training behaves across kernel, mean-field, infinite-width, large-width, and finite-width feature-learning regimes.","inclusion":"Include work whose primary objective is representation formation, feature-learning departures from lazy kernels, NTK-to-mean-field transitions, width or particle limits of training, scaling regimes, spectral or hierarchical feature acquisition, or mean-field/kernel descriptions of learned dynamics.","exclusion":"Use D0 for final-solution implicit bias without feature-acquisition dynamics; use D1 for pure NTK-style convergence proofs without representation-learning analysis; use D4 for diffusion or SGD-noise limits not primarily arising from network width or mean-field structure.","nearest_boundary":"D1 Global convergence of neural training","representative_topics":["spectral ntk analysis for joint dense and sparse polynomial learning via gradient descent escape","large-width asymptotic analysis of relu networks with alpha-stable initialization and random kernel dynamics","theoretical analysis of feature learning overfitting via sparse representations in fully-connected networks"]}],"split":"dev"}
{"episode_id":"F105_T2025-07","field_index":105,"field":"neural network optimization theory","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Implicit bias and induced regularization","description":"Characterize the solution structures selected by training dynamics beyond explicit empirical-risk minimization, including margin, norm, sparsity, rank, balancing, and optimizer- or reparametrization-induced regularization.","inclusion":"Include work whose primary objective is to identify or prove optimizer-induced selection of particular final solutions, representations of solutions, or structural regularizers in overparameterized neural-network-like parameterizations.","exclusion":"Use D1 for results mainly proving that training converges; use D3 when the central phenomenon is sharpness or edge-of-stability behavior; use D7 when the main object is feature acquisition rather than selected final-solution structure.","nearest_boundary":"D3 Sharpness and stability-boundary dynamics","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","layer-wise noise injection for explicit regularization in overparametrized neural networks","stochastic gradient descent and weight decay bias toward low-rank weight matrix learning"]},{"direction_index":1,"label":"D1","name":"Global convergence of neural training","description":"Prove that standard gradient descent, gradient flow, or plain SGD-style first-order training reaches interpolation, global optimality, stationarity, or learning targets for specified neural-network classes.","inclusion":"Include work whose primary objective is a neural-network-specific convergence, rate, escape, or reachability theorem tied to architecture, initialization, width, distribution, or loss assumptions.","exclusion":"Use D2 when the loss landscape or solution set is the main object rather than a training trajectory; use D4 for stochastic-protocol or broad stochastic-approximation rate theory; use D5 for adaptive, momentum, proximal, natural-gradient, or other nonstandard optimizer families as the main object.","nearest_boundary":"D2 Landscape geometry and computational tractability","representative_topics":["early and global convergence of gradient descent via neuron activation partition analysis","theoretical analysis of global convergence in over-parameterized relu nets via activation matrix angular geometry","coupled weight initialization for tighter width bounds in two-layer relu network convergence"]},{"direction_index":2,"label":"D2","name":"Landscape geometry and tractability","description":"Characterize the static structure, topology, geometry, combinatorics, solution sets, critical points, symmetries, convexifications, and computational tractability or hardness of neural-network optimization objectives.","inclusion":"Include work whose primary objective is to describe the objective landscape, minimizer or critical-point structure, connectivity, activation-region geometry, architectural landscape effects, or complexity barriers for training problems.","exclusion":"Use D1 when landscape facts are mainly tools for proving a particular gradient-training convergence result; use D3 when the focus is sharpness evolution under training; use D5 when the contribution is chiefly a new optimizer or training framework rather than landscape or hardness characterization.","nearest_boundary":"D1 Global convergence of neural training","representative_topics":["parameter convex neural networks using exponential activations for improved generalization","depth flattening of optimization landscape in robust over-parameterized deep linear models","intrinsic loss landscape analysis of scale-invariant neural network training via spherical optimization regimes"]},{"direction_index":3,"label":"D3","name":"Sharpness and stability-boundary dynamics","description":"Explain optimization behavior governed by sharpness, flatness, large learning rates, edge-of-stability thresholds, catapult phases, progressive sharpening, and sharpness-aware objectives such as SAM.","inclusion":"Include work whose primary objective is to analyze sharpness dynamics, flat-minima selection through stability mechanisms, edge-of-stability training, or SAM/SAM-like optimization and regularization mechanisms.","exclusion":"Use D0 when sharpness is secondary to a broader implicit-bias claim about margin, norm, rank, or sparsity; use D4 when stochastic noise is modeled as a process without sharpness as the central objective; use D2 for static curvature or flatness properties not tied to training dynamics.","nearest_boundary":"D0 Implicit bias and induced regularization","representative_topics":["pde-based theoretical framework for edge of stability and numerical instability in deep network training","normalization-induced sharpness reduction in gradient descent for generalization","theoretical analysis of edge-of-stability gradient descent dynamics via minimalist counterexample"]},{"direction_index":4,"label":"D4","name":"Stochastic-gradient dynamics and rates","description":"Analyze stochastic-gradient optimization through noise models, diffusion or statistical-mechanics limits, sampling protocols, distributed/asynchronous execution, and broad stochastic-approximation convergence or efficiency theory.","inclusion":"Include work whose primary objective is to understand SGD or stochastic approximation as a stochastic process, to prove convergence or complexity under stochastic noise, or to quantify effects of reshuffling, client sampling, minibatching, delays, low precision, or weak stochastic assumptions.","exclusion":"Use D3 when stochastic effects are studied mainly to explain sharpness or edge-of-stability behavior; use D5 for adaptive, momentum, sign, proximal, natural-gradient, or engineered optimizer rules as the main object; use D6 when the main deliverable is a generalization or stability bound rather than optimization dynamics or rates.","nearest_boundary":"D5 Optimizer and training-framework design","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","sharp sgd convergence rates under flexible client and data sampling in nonconvex optimization","unified convergence theorem for stochastic gradient and proximal gradient methods in nonsmooth nonconvex settings"]},{"direction_index":5,"label":"D5","name":"Optimizer and training-framework design","description":"Develop or analyze optimizer families and training mechanisms beyond plain GD/SGD, including adaptive methods, momentum, sign methods, natural-gradient, proximal, alternating-minimization, control-theoretic, scheduling, normalization, feedback, and reparameterization-aware frameworks.","inclusion":"Include work whose primary objective is the theory, mechanism, convergence, robustness, or design of a modified update rule or practical training-control framework rather than the baseline neural-network objective itself.","exclusion":"Use D1 for convergence of ordinary gradient methods in neural networks; use D4 when the main variable is stochastic sampling or distributed execution rather than the update formula; use D3 for SAM or sharpness-aware objectives as the central object.","nearest_boundary":"D4 Stochastic-gradient dynamics and rates","representative_topics":["mini-batch momentum convergence dynamics via volterra approximation and implicit conditioning ratio in high dimensions","control-theoretic framework for adaptive gradient optimizers via state-space transfer functions","adaptive optimizer slingshot mechanism as cause of grokking via cyclic weight norm transitions"]},{"direction_index":6,"label":"D6","name":"Optimization-driven generalization and stability","description":"Derive generalization, excess-risk, robustness, or algorithmic-stability guarantees from properties of gradient-based training algorithms, stochastic trajectories, or optimization-induced controls.","inclusion":"Include work whose primary objective is to connect a training algorithm or trajectory to out-of-sample error, stability, robustness, excess risk, consistency, or generalization bounds.","exclusion":"Use D0 when the main claim is structural solution selection without a generalization bound; use D4 for convergence or stochastic-process results without generalization as the target; use D7 for representation-learning mechanisms where feature acquisition, not a bound, is central.","nearest_boundary":"D0 Implicit bias and induced regularization","representative_topics":["algorithmic stability bounds for heavy-tailed stochastic gradient descent generalization","stability and generalization bounds for stochastic gradient methods on nonconvex nonsmooth objectives","consistency theory for regularized neural networks via sieve and minimal network methods"]},{"direction_index":7,"label":"D7","name":"Feature learning and wide-regime limits","description":"Explain how optimization produces learned features or representations and how training behaves across kernel, mean-field, infinite-width, large-width, and finite-width feature-learning regimes.","inclusion":"Include work whose primary objective is representation formation, feature-learning departures from lazy kernels, NTK-to-mean-field transitions, width or particle limits of training, scaling regimes, spectral or hierarchical feature acquisition, or mean-field/kernel descriptions of learned dynamics.","exclusion":"Use D0 for final-solution implicit bias without feature-acquisition dynamics; use D1 for pure NTK-style convergence proofs without representation-learning analysis; use D4 for diffusion or SGD-noise limits not primarily arising from network width or mean-field structure.","nearest_boundary":"D1 Global convergence of neural training","representative_topics":["spectral ntk analysis for joint dense and sparse polynomial learning via gradient descent escape","large-width asymptotic analysis of relu networks with alpha-stable initialization and random kernel dynamics","theoretical analysis of feature learning overfitting via sparse representations in fully-connected networks"]}],"split":"dev"}
{"episode_id":"F105_T2026-01","field_index":105,"field":"neural network optimization theory","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Implicit bias and induced regularization","description":"Characterize the solution structures selected by training dynamics beyond explicit empirical-risk minimization, including margin, norm, sparsity, rank, balancing, and optimizer- or reparametrization-induced regularization.","inclusion":"Include work whose primary objective is to identify or prove optimizer-induced selection of particular final solutions, representations of solutions, or structural regularizers in overparameterized neural-network-like parameterizations.","exclusion":"Use D1 for results mainly proving that training converges; use D3 when the central phenomenon is sharpness or edge-of-stability behavior; use D7 when the main object is feature acquisition rather than selected final-solution structure.","nearest_boundary":"D3 Sharpness and stability-boundary dynamics","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","layer-wise noise injection for explicit regularization in overparametrized neural networks","stochastic gradient descent and weight decay bias toward low-rank weight matrix learning"]},{"direction_index":1,"label":"D1","name":"Global convergence of neural training","description":"Prove that standard gradient descent, gradient flow, or plain SGD-style first-order training reaches interpolation, global optimality, stationarity, or learning targets for specified neural-network classes.","inclusion":"Include work whose primary objective is a neural-network-specific convergence, rate, escape, or reachability theorem tied to architecture, initialization, width, distribution, or loss assumptions.","exclusion":"Use D2 when the loss landscape or solution set is the main object rather than a training trajectory; use D4 for stochastic-protocol or broad stochastic-approximation rate theory; use D5 for adaptive, momentum, proximal, natural-gradient, or other nonstandard optimizer families as the main object.","nearest_boundary":"D2 Landscape geometry and computational tractability","representative_topics":["early and global convergence of gradient descent via neuron activation partition analysis","theoretical analysis of global convergence in over-parameterized relu nets via activation matrix angular geometry","coupled weight initialization for tighter width bounds in two-layer relu network convergence"]},{"direction_index":2,"label":"D2","name":"Landscape geometry and tractability","description":"Characterize the static structure, topology, geometry, combinatorics, solution sets, critical points, symmetries, convexifications, and computational tractability or hardness of neural-network optimization objectives.","inclusion":"Include work whose primary objective is to describe the objective landscape, minimizer or critical-point structure, connectivity, activation-region geometry, architectural landscape effects, or complexity barriers for training problems.","exclusion":"Use D1 when landscape facts are mainly tools for proving a particular gradient-training convergence result; use D3 when the focus is sharpness evolution under training; use D5 when the contribution is chiefly a new optimizer or training framework rather than landscape or hardness characterization.","nearest_boundary":"D1 Global convergence of neural training","representative_topics":["parameter convex neural networks using exponential activations for improved generalization","depth flattening of optimization landscape in robust over-parameterized deep linear models","intrinsic loss landscape analysis of scale-invariant neural network training via spherical optimization regimes"]},{"direction_index":3,"label":"D3","name":"Sharpness and stability-boundary dynamics","description":"Explain optimization behavior governed by sharpness, flatness, large learning rates, edge-of-stability thresholds, catapult phases, progressive sharpening, and sharpness-aware objectives such as SAM.","inclusion":"Include work whose primary objective is to analyze sharpness dynamics, flat-minima selection through stability mechanisms, edge-of-stability training, or SAM/SAM-like optimization and regularization mechanisms.","exclusion":"Use D0 when sharpness is secondary to a broader implicit-bias claim about margin, norm, rank, or sparsity; use D4 when stochastic noise is modeled as a process without sharpness as the central objective; use D2 for static curvature or flatness properties not tied to training dynamics.","nearest_boundary":"D0 Implicit bias and induced regularization","representative_topics":["pde-based theoretical framework for edge of stability and numerical instability in deep network training","normalization-induced sharpness reduction in gradient descent for generalization","theoretical analysis of edge-of-stability gradient descent dynamics via minimalist counterexample"]},{"direction_index":4,"label":"D4","name":"Stochastic-gradient dynamics and rates","description":"Analyze stochastic-gradient optimization through noise models, diffusion or statistical-mechanics limits, sampling protocols, distributed/asynchronous execution, and broad stochastic-approximation convergence or efficiency theory.","inclusion":"Include work whose primary objective is to understand SGD or stochastic approximation as a stochastic process, to prove convergence or complexity under stochastic noise, or to quantify effects of reshuffling, client sampling, minibatching, delays, low precision, or weak stochastic assumptions.","exclusion":"Use D3 when stochastic effects are studied mainly to explain sharpness or edge-of-stability behavior; use D5 for adaptive, momentum, sign, proximal, natural-gradient, or engineered optimizer rules as the main object; use D6 when the main deliverable is a generalization or stability bound rather than optimization dynamics or rates.","nearest_boundary":"D5 Optimizer and training-framework design","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","sharp sgd convergence rates under flexible client and data sampling in nonconvex optimization","unified convergence theorem for stochastic gradient and proximal gradient methods in nonsmooth nonconvex settings"]},{"direction_index":5,"label":"D5","name":"Optimizer and training-framework design","description":"Develop or analyze optimizer families and training mechanisms beyond plain GD/SGD, including adaptive methods, momentum, sign methods, natural-gradient, proximal, alternating-minimization, control-theoretic, scheduling, normalization, feedback, and reparameterization-aware frameworks.","inclusion":"Include work whose primary objective is the theory, mechanism, convergence, robustness, or design of a modified update rule or practical training-control framework rather than the baseline neural-network objective itself.","exclusion":"Use D1 for convergence of ordinary gradient methods in neural networks; use D4 when the main variable is stochastic sampling or distributed execution rather than the update formula; use D3 for SAM or sharpness-aware objectives as the central object.","nearest_boundary":"D4 Stochastic-gradient dynamics and rates","representative_topics":["mini-batch momentum convergence dynamics via volterra approximation and implicit conditioning ratio in high dimensions","control-theoretic framework for adaptive gradient optimizers via state-space transfer functions","adaptive optimizer slingshot mechanism as cause of grokking via cyclic weight norm transitions"]},{"direction_index":6,"label":"D6","name":"Optimization-driven generalization and stability","description":"Derive generalization, excess-risk, robustness, or algorithmic-stability guarantees from properties of gradient-based training algorithms, stochastic trajectories, or optimization-induced controls.","inclusion":"Include work whose primary objective is to connect a training algorithm or trajectory to out-of-sample error, stability, robustness, excess risk, consistency, or generalization bounds.","exclusion":"Use D0 when the main claim is structural solution selection without a generalization bound; use D4 for convergence or stochastic-process results without generalization as the target; use D7 for representation-learning mechanisms where feature acquisition, not a bound, is central.","nearest_boundary":"D0 Implicit bias and induced regularization","representative_topics":["algorithmic stability bounds for heavy-tailed stochastic gradient descent generalization","stability and generalization bounds for stochastic gradient methods on nonconvex nonsmooth objectives","consistency theory for regularized neural networks via sieve and minimal network methods"]},{"direction_index":7,"label":"D7","name":"Feature learning and wide-regime limits","description":"Explain how optimization produces learned features or representations and how training behaves across kernel, mean-field, infinite-width, large-width, and finite-width feature-learning regimes.","inclusion":"Include work whose primary objective is representation formation, feature-learning departures from lazy kernels, NTK-to-mean-field transitions, width or particle limits of training, scaling regimes, spectral or hierarchical feature acquisition, or mean-field/kernel descriptions of learned dynamics.","exclusion":"Use D0 for final-solution implicit bias without feature-acquisition dynamics; use D1 for pure NTK-style convergence proofs without representation-learning analysis; use D4 for diffusion or SGD-noise limits not primarily arising from network width or mean-field structure.","nearest_boundary":"D1 Global convergence of neural training","representative_topics":["spectral ntk analysis for joint dense and sparse polynomial learning via gradient descent escape","large-width asymptotic analysis of relu networks with alpha-stable initialization and random kernel dynamics","theoretical analysis of feature learning overfitting via sparse representations in fully-connected networks"]}],"split":"dev"}
{"episode_id":"F113_T2024-01","field_index":113,"field":"large language model fairness","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"General bias measurement","description":"Developing, validating, critiquing, and applying general-purpose instruments for measuring demographic and social bias in language models, including gender-focused probing when the central objective is characterization rather than mitigation or deployment auditing.","inclusion":"Assign work whose primary objective is to quantify, benchmark, probe, or validate measurements of social or demographic bias in model representations or outputs across protected or socially salient groups, including studies of metric reliability, template sensitivity, benchmark design, persona probing, and gender-specific characterization.","exclusion":"Work belongs to the nearest alternative when the organizing objective is a concrete deployment domain, multilingual or cross-cultural transfer, a weight-changing or inference-time mitigation mechanism, stereotype-theory formalization, automated detector construction, or causal/mechanistic explanation.","nearest_boundary":"Stereotype detection and theory-grounded modeling","representative_topics":["robustness of template-based social bias quantification in large language models","hierarchical evaluation of regional bias in pre-trained language models via sub-region clustering","psychometric framework for construct validity and reliability of nlp bias measurement tools"]},{"direction_index":1,"label":"D1","name":"Multilingual cultural fairness","description":"Studying how bias, stereotypes, fairness disparities, and cultural norms appear, transfer, or require adaptation across languages, regions, cultures, and underrepresented linguistic communities.","inclusion":"Assign work whose primary objective is fairness analysis or method adaptation for non-English, multilingual, cross-lingual, regional, Indigenous, or culturally specific contexts, including non-English benchmark construction and cross-language comparisons.","exclusion":"Work belongs to political, moral, and value bias characterization when the central objective is ideology, values, moral reasoning, or geopolitical stance without a multilingual or cross-cultural transfer program.","nearest_boundary":"Political, moral, and value bias","representative_topics":["bias characterization methodology for word embeddings and language models in latin american nlp","cross-lingual group fairness evaluation of pretrained multilingual language models","cross-lingual moral bias analysis in pre-trained multilingual language models"]},{"direction_index":2,"label":"D2","name":"Domain fairness auditing","description":"Auditing and characterizing fairness when language models are used as components of concrete downstream applications, institutional workflows, or high-stakes decision settings.","inclusion":"Assign work whose primary objective is empirical fairness assessment in a specific application domain such as hiring, recommendation, credit, finance, healthcare, education, research instrumentation, social-science analysis, or other deployed decision-support contexts.","exclusion":"Work belongs to general bias measurement when the main contribution is a domain-agnostic benchmark or metric, and to mitigation directions when the main contribution is a transferable technical intervention rather than the domain audit.","nearest_boundary":"General bias measurement","representative_topics":["fairness evaluation of large language model recommendation across sensitive user attributes","bias detection in llms for automated resume screening in hiring contexts","fairness evaluation of large language models across high-stakes application domains"]},{"direction_index":3,"label":"D3","name":"Training-time mitigation","description":"Designing and evaluating fairness interventions that reduce bias by changing model parameters, training data, fine-tuning data, adapters, pruning, projections, or optimization procedures.","inclusion":"Assign work whose primary contribution is a debiasing or fairness intervention applied during pre-training, fine-tuning, model editing, parameter-efficient adaptation, corpus curation, or privacy-aware training.","exclusion":"Work belongs to inference-time control when it mitigates bias through prompts, decoding, calibration, self-critique, or in-context learning without weight updates; it belongs to mechanisms when the objective is explanation rather than intervention.","nearest_boundary":"Inference-time fairness control","representative_topics":["disentangling factual gender from stereotypical bias in language model representations","stereotype content model-guided fine-tuning to reduce minority group stereotyping in language models","parameter-efficient modular adapter-based debiasing for on-demand fairness in language models"]},{"direction_index":4,"label":"D4","name":"Inference-time fairness control","description":"Reducing, steering, calibrating, or testing biased behavior at runtime without changing model weights, using prompts, in-context examples, decoding, explanation-guided steering, self-critique, or controllable generation.","inclusion":"Assign work whose primary objective is an inference-time fairness method, including fairness-guided prompting, chain-of-thought debiasing, decoding choices, prompt search, self-improvement loops, runtime calibration, or LLM-assisted generation of probes when the runtime mechanism is central.","exclusion":"Work belongs to training-time mitigation when it changes weights or training data, and to general measurement or automated detection when the primary deliverable is an evaluation instrument rather than a runtime control method.","nearest_boundary":"Training-time mitigation","representative_topics":["llm-based counterfactual generation for probing fairness in text classifiers","decoding algorithm effects on fairness and diversity trade-offs in open-ended language generation","fairness-guided prompt search for reducing predictive bias in few-shot in-context learning"]},{"direction_index":5,"label":"D5","name":"Bias mechanisms and dynamics","description":"Explaining where language-model bias comes from, how it emerges, propagates, amplifies, and is represented internally across data, training trajectories, components, tasks, or behavioral perturbations.","inclusion":"Assign work whose primary objective is causal, mechanistic, developmental, dynamic, or attributional understanding of bias, including spurious correlations, training-checkpoint evolution, internal localization, propagation from pretraining data to downstream systems, or self-reinforcing amplification loops.","exclusion":"Work belongs to mitigation when the main contribution is a concrete debiasing intervention, and to measurement when the main contribution is a static benchmark or score rather than an explanatory account.","nearest_boundary":"Training-time mitigation","representative_topics":["selection bias induced spurious correlations in pretrained language model gender representations","training dynamics of gender bias formation in language model input embeddings","intrinsic-extrinsic bias evaluation gap in debiased masked language models for downstream tasks"]},{"direction_index":6,"label":"D6","name":"Political and value bias","description":"Characterizing language-model bias in political ideology, moral reasoning, human values, geopolitical stance, cultural dominance, affective polarization, and other value-laden representational skews.","inclusion":"Assign work whose primary objective is measuring, explaining, or tracking political leaning, moral mimicry, value bias, geopolitical perspective, cultural dominance, or ideology-conditioned generated behavior.","exclusion":"Work belongs to multilingual cultural fairness when language transfer, non-English adaptation, or cross-lingual comparison is the organizing objective; it belongs to domain auditing when the value-laden issue is studied mainly inside a concrete deployment workflow.","nearest_boundary":"Multilingual cultural fairness","representative_topics":["political identity-conditioned moral mimicry in large language model text generation","probing geographic and geopolitical bias in multilingual pretrained language model representations","value bias characterization in llms using psychological value theory and corpus analysis"]},{"direction_index":7,"label":"D7","name":"Stereotype and conceptual modeling","description":"Formalizing, detecting, and systematizing stereotype and bias constructs through theory-grounded models, automated detection systems, red-teaming frameworks, and field-level conceptual taxonomies.","inclusion":"Assign work whose primary objective is to operationalize stereotype or bias constructs using social-science theory, build stereotype or bias detection systems, automate bias testing or red-teaming, or synthesize conceptual taxonomies and lifecycle frameworks for the fairness field.","exclusion":"Work belongs to general bias measurement when it mainly reports demographic bias scores using standard benchmarks, and to mitigation directions when the central contribution is reducing bias rather than modeling, detecting, or conceptually organizing it.","nearest_boundary":"General bias measurement","representative_topics":["theory-grounded measurement of social stereotyping in english language models using psychological stereotype frameworks","bias analysis and mitigation in large language models across data, design, and deployment","prompt-based intersectional stereotype measurement in language models via markedness and persona generation"]}],"split":"dev"}
{"episode_id":"F113_T2024-07","field_index":113,"field":"large language model fairness","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"General bias measurement","description":"Developing, validating, critiquing, and applying general-purpose instruments for measuring demographic and social bias in language models, including gender-focused probing when the central objective is characterization rather than mitigation or deployment auditing.","inclusion":"Assign work whose primary objective is to quantify, benchmark, probe, or validate measurements of social or demographic bias in model representations or outputs across protected or socially salient groups, including studies of metric reliability, template sensitivity, benchmark design, persona probing, and gender-specific characterization.","exclusion":"Work belongs to the nearest alternative when the organizing objective is a concrete deployment domain, multilingual or cross-cultural transfer, a weight-changing or inference-time mitigation mechanism, stereotype-theory formalization, automated detector construction, or causal/mechanistic explanation.","nearest_boundary":"Stereotype detection and theory-grounded modeling","representative_topics":["robustness of template-based social bias quantification in large language models","hierarchical evaluation of regional bias in pre-trained language models via sub-region clustering","psychometric framework for construct validity and reliability of nlp bias measurement tools"]},{"direction_index":1,"label":"D1","name":"Multilingual cultural fairness","description":"Studying how bias, stereotypes, fairness disparities, and cultural norms appear, transfer, or require adaptation across languages, regions, cultures, and underrepresented linguistic communities.","inclusion":"Assign work whose primary objective is fairness analysis or method adaptation for non-English, multilingual, cross-lingual, regional, Indigenous, or culturally specific contexts, including non-English benchmark construction and cross-language comparisons.","exclusion":"Work belongs to political, moral, and value bias characterization when the central objective is ideology, values, moral reasoning, or geopolitical stance without a multilingual or cross-cultural transfer program.","nearest_boundary":"Political, moral, and value bias","representative_topics":["bias characterization methodology for word embeddings and language models in latin american nlp","cross-lingual group fairness evaluation of pretrained multilingual language models","cross-lingual moral bias analysis in pre-trained multilingual language models"]},{"direction_index":2,"label":"D2","name":"Domain fairness auditing","description":"Auditing and characterizing fairness when language models are used as components of concrete downstream applications, institutional workflows, or high-stakes decision settings.","inclusion":"Assign work whose primary objective is empirical fairness assessment in a specific application domain such as hiring, recommendation, credit, finance, healthcare, education, research instrumentation, social-science analysis, or other deployed decision-support contexts.","exclusion":"Work belongs to general bias measurement when the main contribution is a domain-agnostic benchmark or metric, and to mitigation directions when the main contribution is a transferable technical intervention rather than the domain audit.","nearest_boundary":"General bias measurement","representative_topics":["fairness evaluation of large language model recommendation across sensitive user attributes","bias detection in llms for automated resume screening in hiring contexts","fairness evaluation of large language models across high-stakes application domains"]},{"direction_index":3,"label":"D3","name":"Training-time mitigation","description":"Designing and evaluating fairness interventions that reduce bias by changing model parameters, training data, fine-tuning data, adapters, pruning, projections, or optimization procedures.","inclusion":"Assign work whose primary contribution is a debiasing or fairness intervention applied during pre-training, fine-tuning, model editing, parameter-efficient adaptation, corpus curation, or privacy-aware training.","exclusion":"Work belongs to inference-time control when it mitigates bias through prompts, decoding, calibration, self-critique, or in-context learning without weight updates; it belongs to mechanisms when the objective is explanation rather than intervention.","nearest_boundary":"Inference-time fairness control","representative_topics":["disentangling factual gender from stereotypical bias in language model representations","stereotype content model-guided fine-tuning to reduce minority group stereotyping in language models","parameter-efficient modular adapter-based debiasing for on-demand fairness in language models"]},{"direction_index":4,"label":"D4","name":"Inference-time fairness control","description":"Reducing, steering, calibrating, or testing biased behavior at runtime without changing model weights, using prompts, in-context examples, decoding, explanation-guided steering, self-critique, or controllable generation.","inclusion":"Assign work whose primary objective is an inference-time fairness method, including fairness-guided prompting, chain-of-thought debiasing, decoding choices, prompt search, self-improvement loops, runtime calibration, or LLM-assisted generation of probes when the runtime mechanism is central.","exclusion":"Work belongs to training-time mitigation when it changes weights or training data, and to general measurement or automated detection when the primary deliverable is an evaluation instrument rather than a runtime control method.","nearest_boundary":"Training-time mitigation","representative_topics":["llm-based counterfactual generation for probing fairness in text classifiers","decoding algorithm effects on fairness and diversity trade-offs in open-ended language generation","fairness-guided prompt search for reducing predictive bias in few-shot in-context learning"]},{"direction_index":5,"label":"D5","name":"Bias mechanisms and dynamics","description":"Explaining where language-model bias comes from, how it emerges, propagates, amplifies, and is represented internally across data, training trajectories, components, tasks, or behavioral perturbations.","inclusion":"Assign work whose primary objective is causal, mechanistic, developmental, dynamic, or attributional understanding of bias, including spurious correlations, training-checkpoint evolution, internal localization, propagation from pretraining data to downstream systems, or self-reinforcing amplification loops.","exclusion":"Work belongs to mitigation when the main contribution is a concrete debiasing intervention, and to measurement when the main contribution is a static benchmark or score rather than an explanatory account.","nearest_boundary":"Training-time mitigation","representative_topics":["selection bias induced spurious correlations in pretrained language model gender representations","training dynamics of gender bias formation in language model input embeddings","intrinsic-extrinsic bias evaluation gap in debiased masked language models for downstream tasks"]},{"direction_index":6,"label":"D6","name":"Political and value bias","description":"Characterizing language-model bias in political ideology, moral reasoning, human values, geopolitical stance, cultural dominance, affective polarization, and other value-laden representational skews.","inclusion":"Assign work whose primary objective is measuring, explaining, or tracking political leaning, moral mimicry, value bias, geopolitical perspective, cultural dominance, or ideology-conditioned generated behavior.","exclusion":"Work belongs to multilingual cultural fairness when language transfer, non-English adaptation, or cross-lingual comparison is the organizing objective; it belongs to domain auditing when the value-laden issue is studied mainly inside a concrete deployment workflow.","nearest_boundary":"Multilingual cultural fairness","representative_topics":["political identity-conditioned moral mimicry in large language model text generation","probing geographic and geopolitical bias in multilingual pretrained language model representations","value bias characterization in llms using psychological value theory and corpus analysis"]},{"direction_index":7,"label":"D7","name":"Stereotype and conceptual modeling","description":"Formalizing, detecting, and systematizing stereotype and bias constructs through theory-grounded models, automated detection systems, red-teaming frameworks, and field-level conceptual taxonomies.","inclusion":"Assign work whose primary objective is to operationalize stereotype or bias constructs using social-science theory, build stereotype or bias detection systems, automate bias testing or red-teaming, or synthesize conceptual taxonomies and lifecycle frameworks for the fairness field.","exclusion":"Work belongs to general bias measurement when it mainly reports demographic bias scores using standard benchmarks, and to mitigation directions when the central contribution is reducing bias rather than modeling, detecting, or conceptually organizing it.","nearest_boundary":"General bias measurement","representative_topics":["theory-grounded measurement of social stereotyping in english language models using psychological stereotype frameworks","bias analysis and mitigation in large language models across data, design, and deployment","prompt-based intersectional stereotype measurement in language models via markedness and persona generation"]}],"split":"dev"}
{"episode_id":"F113_T2025-01","field_index":113,"field":"large language model fairness","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"General bias measurement","description":"Developing, validating, critiquing, and applying general-purpose instruments for measuring demographic and social bias in language models, including gender-focused probing when the central objective is characterization rather than mitigation or deployment auditing.","inclusion":"Assign work whose primary objective is to quantify, benchmark, probe, or validate measurements of social or demographic bias in model representations or outputs across protected or socially salient groups, including studies of metric reliability, template sensitivity, benchmark design, persona probing, and gender-specific characterization.","exclusion":"Work belongs to the nearest alternative when the organizing objective is a concrete deployment domain, multilingual or cross-cultural transfer, a weight-changing or inference-time mitigation mechanism, stereotype-theory formalization, automated detector construction, or causal/mechanistic explanation.","nearest_boundary":"Stereotype detection and theory-grounded modeling","representative_topics":["robustness of template-based social bias quantification in large language models","hierarchical evaluation of regional bias in pre-trained language models via sub-region clustering","psychometric framework for construct validity and reliability of nlp bias measurement tools"]},{"direction_index":1,"label":"D1","name":"Multilingual cultural fairness","description":"Studying how bias, stereotypes, fairness disparities, and cultural norms appear, transfer, or require adaptation across languages, regions, cultures, and underrepresented linguistic communities.","inclusion":"Assign work whose primary objective is fairness analysis or method adaptation for non-English, multilingual, cross-lingual, regional, Indigenous, or culturally specific contexts, including non-English benchmark construction and cross-language comparisons.","exclusion":"Work belongs to political, moral, and value bias characterization when the central objective is ideology, values, moral reasoning, or geopolitical stance without a multilingual or cross-cultural transfer program.","nearest_boundary":"Political, moral, and value bias","representative_topics":["bias characterization methodology for word embeddings and language models in latin american nlp","cross-lingual group fairness evaluation of pretrained multilingual language models","cross-lingual moral bias analysis in pre-trained multilingual language models"]},{"direction_index":2,"label":"D2","name":"Domain fairness auditing","description":"Auditing and characterizing fairness when language models are used as components of concrete downstream applications, institutional workflows, or high-stakes decision settings.","inclusion":"Assign work whose primary objective is empirical fairness assessment in a specific application domain such as hiring, recommendation, credit, finance, healthcare, education, research instrumentation, social-science analysis, or other deployed decision-support contexts.","exclusion":"Work belongs to general bias measurement when the main contribution is a domain-agnostic benchmark or metric, and to mitigation directions when the main contribution is a transferable technical intervention rather than the domain audit.","nearest_boundary":"General bias measurement","representative_topics":["fairness evaluation of large language model recommendation across sensitive user attributes","bias detection in llms for automated resume screening in hiring contexts","fairness evaluation of large language models across high-stakes application domains"]},{"direction_index":3,"label":"D3","name":"Training-time mitigation","description":"Designing and evaluating fairness interventions that reduce bias by changing model parameters, training data, fine-tuning data, adapters, pruning, projections, or optimization procedures.","inclusion":"Assign work whose primary contribution is a debiasing or fairness intervention applied during pre-training, fine-tuning, model editing, parameter-efficient adaptation, corpus curation, or privacy-aware training.","exclusion":"Work belongs to inference-time control when it mitigates bias through prompts, decoding, calibration, self-critique, or in-context learning without weight updates; it belongs to mechanisms when the objective is explanation rather than intervention.","nearest_boundary":"Inference-time fairness control","representative_topics":["disentangling factual gender from stereotypical bias in language model representations","stereotype content model-guided fine-tuning to reduce minority group stereotyping in language models","parameter-efficient modular adapter-based debiasing for on-demand fairness in language models"]},{"direction_index":4,"label":"D4","name":"Inference-time fairness control","description":"Reducing, steering, calibrating, or testing biased behavior at runtime without changing model weights, using prompts, in-context examples, decoding, explanation-guided steering, self-critique, or controllable generation.","inclusion":"Assign work whose primary objective is an inference-time fairness method, including fairness-guided prompting, chain-of-thought debiasing, decoding choices, prompt search, self-improvement loops, runtime calibration, or LLM-assisted generation of probes when the runtime mechanism is central.","exclusion":"Work belongs to training-time mitigation when it changes weights or training data, and to general measurement or automated detection when the primary deliverable is an evaluation instrument rather than a runtime control method.","nearest_boundary":"Training-time mitigation","representative_topics":["llm-based counterfactual generation for probing fairness in text classifiers","decoding algorithm effects on fairness and diversity trade-offs in open-ended language generation","fairness-guided prompt search for reducing predictive bias in few-shot in-context learning"]},{"direction_index":5,"label":"D5","name":"Bias mechanisms and dynamics","description":"Explaining where language-model bias comes from, how it emerges, propagates, amplifies, and is represented internally across data, training trajectories, components, tasks, or behavioral perturbations.","inclusion":"Assign work whose primary objective is causal, mechanistic, developmental, dynamic, or attributional understanding of bias, including spurious correlations, training-checkpoint evolution, internal localization, propagation from pretraining data to downstream systems, or self-reinforcing amplification loops.","exclusion":"Work belongs to mitigation when the main contribution is a concrete debiasing intervention, and to measurement when the main contribution is a static benchmark or score rather than an explanatory account.","nearest_boundary":"Training-time mitigation","representative_topics":["selection bias induced spurious correlations in pretrained language model gender representations","training dynamics of gender bias formation in language model input embeddings","intrinsic-extrinsic bias evaluation gap in debiased masked language models for downstream tasks"]},{"direction_index":6,"label":"D6","name":"Political and value bias","description":"Characterizing language-model bias in political ideology, moral reasoning, human values, geopolitical stance, cultural dominance, affective polarization, and other value-laden representational skews.","inclusion":"Assign work whose primary objective is measuring, explaining, or tracking political leaning, moral mimicry, value bias, geopolitical perspective, cultural dominance, or ideology-conditioned generated behavior.","exclusion":"Work belongs to multilingual cultural fairness when language transfer, non-English adaptation, or cross-lingual comparison is the organizing objective; it belongs to domain auditing when the value-laden issue is studied mainly inside a concrete deployment workflow.","nearest_boundary":"Multilingual cultural fairness","representative_topics":["political identity-conditioned moral mimicry in large language model text generation","probing geographic and geopolitical bias in multilingual pretrained language model representations","value bias characterization in llms using psychological value theory and corpus analysis"]},{"direction_index":7,"label":"D7","name":"Stereotype and conceptual modeling","description":"Formalizing, detecting, and systematizing stereotype and bias constructs through theory-grounded models, automated detection systems, red-teaming frameworks, and field-level conceptual taxonomies.","inclusion":"Assign work whose primary objective is to operationalize stereotype or bias constructs using social-science theory, build stereotype or bias detection systems, automate bias testing or red-teaming, or synthesize conceptual taxonomies and lifecycle frameworks for the fairness field.","exclusion":"Work belongs to general bias measurement when it mainly reports demographic bias scores using standard benchmarks, and to mitigation directions when the central contribution is reducing bias rather than modeling, detecting, or conceptually organizing it.","nearest_boundary":"General bias measurement","representative_topics":["theory-grounded measurement of social stereotyping in english language models using psychological stereotype frameworks","bias analysis and mitigation in large language models across data, design, and deployment","prompt-based intersectional stereotype measurement in language models via markedness and persona generation"]}],"split":"dev"}
{"episode_id":"F113_T2025-07","field_index":113,"field":"large language model fairness","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"General bias measurement","description":"Developing, validating, critiquing, and applying general-purpose instruments for measuring demographic and social bias in language models, including gender-focused probing when the central objective is characterization rather than mitigation or deployment auditing.","inclusion":"Assign work whose primary objective is to quantify, benchmark, probe, or validate measurements of social or demographic bias in model representations or outputs across protected or socially salient groups, including studies of metric reliability, template sensitivity, benchmark design, persona probing, and gender-specific characterization.","exclusion":"Work belongs to the nearest alternative when the organizing objective is a concrete deployment domain, multilingual or cross-cultural transfer, a weight-changing or inference-time mitigation mechanism, stereotype-theory formalization, automated detector construction, or causal/mechanistic explanation.","nearest_boundary":"Stereotype detection and theory-grounded modeling","representative_topics":["robustness of template-based social bias quantification in large language models","hierarchical evaluation of regional bias in pre-trained language models via sub-region clustering","psychometric framework for construct validity and reliability of nlp bias measurement tools"]},{"direction_index":1,"label":"D1","name":"Multilingual cultural fairness","description":"Studying how bias, stereotypes, fairness disparities, and cultural norms appear, transfer, or require adaptation across languages, regions, cultures, and underrepresented linguistic communities.","inclusion":"Assign work whose primary objective is fairness analysis or method adaptation for non-English, multilingual, cross-lingual, regional, Indigenous, or culturally specific contexts, including non-English benchmark construction and cross-language comparisons.","exclusion":"Work belongs to political, moral, and value bias characterization when the central objective is ideology, values, moral reasoning, or geopolitical stance without a multilingual or cross-cultural transfer program.","nearest_boundary":"Political, moral, and value bias","representative_topics":["bias characterization methodology for word embeddings and language models in latin american nlp","cross-lingual group fairness evaluation of pretrained multilingual language models","cross-lingual moral bias analysis in pre-trained multilingual language models"]},{"direction_index":2,"label":"D2","name":"Domain fairness auditing","description":"Auditing and characterizing fairness when language models are used as components of concrete downstream applications, institutional workflows, or high-stakes decision settings.","inclusion":"Assign work whose primary objective is empirical fairness assessment in a specific application domain such as hiring, recommendation, credit, finance, healthcare, education, research instrumentation, social-science analysis, or other deployed decision-support contexts.","exclusion":"Work belongs to general bias measurement when the main contribution is a domain-agnostic benchmark or metric, and to mitigation directions when the main contribution is a transferable technical intervention rather than the domain audit.","nearest_boundary":"General bias measurement","representative_topics":["fairness evaluation of large language model recommendation across sensitive user attributes","bias detection in llms for automated resume screening in hiring contexts","fairness evaluation of large language models across high-stakes application domains"]},{"direction_index":3,"label":"D3","name":"Training-time mitigation","description":"Designing and evaluating fairness interventions that reduce bias by changing model parameters, training data, fine-tuning data, adapters, pruning, projections, or optimization procedures.","inclusion":"Assign work whose primary contribution is a debiasing or fairness intervention applied during pre-training, fine-tuning, model editing, parameter-efficient adaptation, corpus curation, or privacy-aware training.","exclusion":"Work belongs to inference-time control when it mitigates bias through prompts, decoding, calibration, self-critique, or in-context learning without weight updates; it belongs to mechanisms when the objective is explanation rather than intervention.","nearest_boundary":"Inference-time fairness control","representative_topics":["disentangling factual gender from stereotypical bias in language model representations","stereotype content model-guided fine-tuning to reduce minority group stereotyping in language models","parameter-efficient modular adapter-based debiasing for on-demand fairness in language models"]},{"direction_index":4,"label":"D4","name":"Inference-time fairness control","description":"Reducing, steering, calibrating, or testing biased behavior at runtime without changing model weights, using prompts, in-context examples, decoding, explanation-guided steering, self-critique, or controllable generation.","inclusion":"Assign work whose primary objective is an inference-time fairness method, including fairness-guided prompting, chain-of-thought debiasing, decoding choices, prompt search, self-improvement loops, runtime calibration, or LLM-assisted generation of probes when the runtime mechanism is central.","exclusion":"Work belongs to training-time mitigation when it changes weights or training data, and to general measurement or automated detection when the primary deliverable is an evaluation instrument rather than a runtime control method.","nearest_boundary":"Training-time mitigation","representative_topics":["llm-based counterfactual generation for probing fairness in text classifiers","decoding algorithm effects on fairness and diversity trade-offs in open-ended language generation","fairness-guided prompt search for reducing predictive bias in few-shot in-context learning"]},{"direction_index":5,"label":"D5","name":"Bias mechanisms and dynamics","description":"Explaining where language-model bias comes from, how it emerges, propagates, amplifies, and is represented internally across data, training trajectories, components, tasks, or behavioral perturbations.","inclusion":"Assign work whose primary objective is causal, mechanistic, developmental, dynamic, or attributional understanding of bias, including spurious correlations, training-checkpoint evolution, internal localization, propagation from pretraining data to downstream systems, or self-reinforcing amplification loops.","exclusion":"Work belongs to mitigation when the main contribution is a concrete debiasing intervention, and to measurement when the main contribution is a static benchmark or score rather than an explanatory account.","nearest_boundary":"Training-time mitigation","representative_topics":["selection bias induced spurious correlations in pretrained language model gender representations","training dynamics of gender bias formation in language model input embeddings","intrinsic-extrinsic bias evaluation gap in debiased masked language models for downstream tasks"]},{"direction_index":6,"label":"D6","name":"Political and value bias","description":"Characterizing language-model bias in political ideology, moral reasoning, human values, geopolitical stance, cultural dominance, affective polarization, and other value-laden representational skews.","inclusion":"Assign work whose primary objective is measuring, explaining, or tracking political leaning, moral mimicry, value bias, geopolitical perspective, cultural dominance, or ideology-conditioned generated behavior.","exclusion":"Work belongs to multilingual cultural fairness when language transfer, non-English adaptation, or cross-lingual comparison is the organizing objective; it belongs to domain auditing when the value-laden issue is studied mainly inside a concrete deployment workflow.","nearest_boundary":"Multilingual cultural fairness","representative_topics":["political identity-conditioned moral mimicry in large language model text generation","probing geographic and geopolitical bias in multilingual pretrained language model representations","value bias characterization in llms using psychological value theory and corpus analysis"]},{"direction_index":7,"label":"D7","name":"Stereotype and conceptual modeling","description":"Formalizing, detecting, and systematizing stereotype and bias constructs through theory-grounded models, automated detection systems, red-teaming frameworks, and field-level conceptual taxonomies.","inclusion":"Assign work whose primary objective is to operationalize stereotype or bias constructs using social-science theory, build stereotype or bias detection systems, automate bias testing or red-teaming, or synthesize conceptual taxonomies and lifecycle frameworks for the fairness field.","exclusion":"Work belongs to general bias measurement when it mainly reports demographic bias scores using standard benchmarks, and to mitigation directions when the central contribution is reducing bias rather than modeling, detecting, or conceptually organizing it.","nearest_boundary":"General bias measurement","representative_topics":["theory-grounded measurement of social stereotyping in english language models using psychological stereotype frameworks","bias analysis and mitigation in large language models across data, design, and deployment","prompt-based intersectional stereotype measurement in language models via markedness and persona generation"]}],"split":"dev"}
{"episode_id":"F113_T2026-01","field_index":113,"field":"large language model fairness","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"General bias measurement","description":"Developing, validating, critiquing, and applying general-purpose instruments for measuring demographic and social bias in language models, including gender-focused probing when the central objective is characterization rather than mitigation or deployment auditing.","inclusion":"Assign work whose primary objective is to quantify, benchmark, probe, or validate measurements of social or demographic bias in model representations or outputs across protected or socially salient groups, including studies of metric reliability, template sensitivity, benchmark design, persona probing, and gender-specific characterization.","exclusion":"Work belongs to the nearest alternative when the organizing objective is a concrete deployment domain, multilingual or cross-cultural transfer, a weight-changing or inference-time mitigation mechanism, stereotype-theory formalization, automated detector construction, or causal/mechanistic explanation.","nearest_boundary":"Stereotype detection and theory-grounded modeling","representative_topics":["robustness of template-based social bias quantification in large language models","hierarchical evaluation of regional bias in pre-trained language models via sub-region clustering","psychometric framework for construct validity and reliability of nlp bias measurement tools"]},{"direction_index":1,"label":"D1","name":"Multilingual cultural fairness","description":"Studying how bias, stereotypes, fairness disparities, and cultural norms appear, transfer, or require adaptation across languages, regions, cultures, and underrepresented linguistic communities.","inclusion":"Assign work whose primary objective is fairness analysis or method adaptation for non-English, multilingual, cross-lingual, regional, Indigenous, or culturally specific contexts, including non-English benchmark construction and cross-language comparisons.","exclusion":"Work belongs to political, moral, and value bias characterization when the central objective is ideology, values, moral reasoning, or geopolitical stance without a multilingual or cross-cultural transfer program.","nearest_boundary":"Political, moral, and value bias","representative_topics":["bias characterization methodology for word embeddings and language models in latin american nlp","cross-lingual group fairness evaluation of pretrained multilingual language models","cross-lingual moral bias analysis in pre-trained multilingual language models"]},{"direction_index":2,"label":"D2","name":"Domain fairness auditing","description":"Auditing and characterizing fairness when language models are used as components of concrete downstream applications, institutional workflows, or high-stakes decision settings.","inclusion":"Assign work whose primary objective is empirical fairness assessment in a specific application domain such as hiring, recommendation, credit, finance, healthcare, education, research instrumentation, social-science analysis, or other deployed decision-support contexts.","exclusion":"Work belongs to general bias measurement when the main contribution is a domain-agnostic benchmark or metric, and to mitigation directions when the main contribution is a transferable technical intervention rather than the domain audit.","nearest_boundary":"General bias measurement","representative_topics":["fairness evaluation of large language model recommendation across sensitive user attributes","bias detection in llms for automated resume screening in hiring contexts","fairness evaluation of large language models across high-stakes application domains"]},{"direction_index":3,"label":"D3","name":"Training-time mitigation","description":"Designing and evaluating fairness interventions that reduce bias by changing model parameters, training data, fine-tuning data, adapters, pruning, projections, or optimization procedures.","inclusion":"Assign work whose primary contribution is a debiasing or fairness intervention applied during pre-training, fine-tuning, model editing, parameter-efficient adaptation, corpus curation, or privacy-aware training.","exclusion":"Work belongs to inference-time control when it mitigates bias through prompts, decoding, calibration, self-critique, or in-context learning without weight updates; it belongs to mechanisms when the objective is explanation rather than intervention.","nearest_boundary":"Inference-time fairness control","representative_topics":["disentangling factual gender from stereotypical bias in language model representations","stereotype content model-guided fine-tuning to reduce minority group stereotyping in language models","parameter-efficient modular adapter-based debiasing for on-demand fairness in language models"]},{"direction_index":4,"label":"D4","name":"Inference-time fairness control","description":"Reducing, steering, calibrating, or testing biased behavior at runtime without changing model weights, using prompts, in-context examples, decoding, explanation-guided steering, self-critique, or controllable generation.","inclusion":"Assign work whose primary objective is an inference-time fairness method, including fairness-guided prompting, chain-of-thought debiasing, decoding choices, prompt search, self-improvement loops, runtime calibration, or LLM-assisted generation of probes when the runtime mechanism is central.","exclusion":"Work belongs to training-time mitigation when it changes weights or training data, and to general measurement or automated detection when the primary deliverable is an evaluation instrument rather than a runtime control method.","nearest_boundary":"Training-time mitigation","representative_topics":["llm-based counterfactual generation for probing fairness in text classifiers","decoding algorithm effects on fairness and diversity trade-offs in open-ended language generation","fairness-guided prompt search for reducing predictive bias in few-shot in-context learning"]},{"direction_index":5,"label":"D5","name":"Bias mechanisms and dynamics","description":"Explaining where language-model bias comes from, how it emerges, propagates, amplifies, and is represented internally across data, training trajectories, components, tasks, or behavioral perturbations.","inclusion":"Assign work whose primary objective is causal, mechanistic, developmental, dynamic, or attributional understanding of bias, including spurious correlations, training-checkpoint evolution, internal localization, propagation from pretraining data to downstream systems, or self-reinforcing amplification loops.","exclusion":"Work belongs to mitigation when the main contribution is a concrete debiasing intervention, and to measurement when the main contribution is a static benchmark or score rather than an explanatory account.","nearest_boundary":"Training-time mitigation","representative_topics":["selection bias induced spurious correlations in pretrained language model gender representations","training dynamics of gender bias formation in language model input embeddings","intrinsic-extrinsic bias evaluation gap in debiased masked language models for downstream tasks"]},{"direction_index":6,"label":"D6","name":"Political and value bias","description":"Characterizing language-model bias in political ideology, moral reasoning, human values, geopolitical stance, cultural dominance, affective polarization, and other value-laden representational skews.","inclusion":"Assign work whose primary objective is measuring, explaining, or tracking political leaning, moral mimicry, value bias, geopolitical perspective, cultural dominance, or ideology-conditioned generated behavior.","exclusion":"Work belongs to multilingual cultural fairness when language transfer, non-English adaptation, or cross-lingual comparison is the organizing objective; it belongs to domain auditing when the value-laden issue is studied mainly inside a concrete deployment workflow.","nearest_boundary":"Multilingual cultural fairness","representative_topics":["political identity-conditioned moral mimicry in large language model text generation","probing geographic and geopolitical bias in multilingual pretrained language model representations","value bias characterization in llms using psychological value theory and corpus analysis"]},{"direction_index":7,"label":"D7","name":"Stereotype and conceptual modeling","description":"Formalizing, detecting, and systematizing stereotype and bias constructs through theory-grounded models, automated detection systems, red-teaming frameworks, and field-level conceptual taxonomies.","inclusion":"Assign work whose primary objective is to operationalize stereotype or bias constructs using social-science theory, build stereotype or bias detection systems, automate bias testing or red-teaming, or synthesize conceptual taxonomies and lifecycle frameworks for the fairness field.","exclusion":"Work belongs to general bias measurement when it mainly reports demographic bias scores using standard benchmarks, and to mitigation directions when the central contribution is reducing bias rather than modeling, detecting, or conceptually organizing it.","nearest_boundary":"General bias measurement","representative_topics":["theory-grounded measurement of social stereotyping in english language models using psychological stereotype frameworks","bias analysis and mitigation in large language models across data, design, and deployment","prompt-based intersectional stereotype measurement in language models via markedness and persona generation"]}],"split":"dev"}
{"episode_id":"F116_T2024-01","field_index":116,"field":"neural operator learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Physics-constrained operator learning","description":"Learning neural solution operators by directly imposing governing equations, residuals, boundary conditions, weak forms, conservation laws, symmetries, or other explicit physical constraints in training or architecture.","inclusion":"Include work whose primary objective is data-free, low-label, semi-supervised, or physically consistent operator learning through PDE residuals, physics-informed losses, boundary/weak constraints, or embedded physical laws.","exclusion":"If equations only define a benchmark or application and the operator is mainly a black-box surrogate, assign to Physical-system surrogates; if the main novelty is a general architecture without physics integration, assign to General operator methods.","nearest_boundary":"Physical-system surrogates","representative_topics":["physics-constrained neural operators for parametric pdes via low-rank long-range entanglement modeling","physics-informed deep operator networks for surrogate modeling in computational mechanics","semi-supervised fourier neural operator learning of nonlinear pde evolution operators"]},{"direction_index":1,"label":"D1","name":"General operator methods","description":"Developing broadly applicable neural-operator architectures, representations, training procedures, and numerical-neural integrations for mappings between function spaces.","inclusion":"Include new operator layers, transforms, DeepONet variants, FNO/spectral/wavelet/Laplace/attention/latent/variational/multimodal representations, super-resolution or symbolic extensions, and solver-augmented operator methods when the main objective is a reusable method rather than a particular physical system.","exclusion":"If the central objective is handling arbitrary geometries or meshes, assign to Geometry-general operators; if the main objective is theory, physics constraints, data efficiency, or uncertainty/inversion, assign to those directions instead.","nearest_boundary":"Geometry-general operators","representative_topics":["multiscale neural operator learning of grid-independent non-local pde parametrizations","hybrid deep operator and relaxation methods for scalable spectral pde solving","clifford algebraic neural layers for multivector field pde surrogate modeling"]},{"direction_index":2,"label":"D2","name":"Geometry-general operators","description":"Learning operators that generalize across discretizations, meshes, point distributions, irregular or arbitrary domains, deformations, unbounded domains, and evolving geometries.","inclusion":"Include work whose primary objective is mesh-free, point-cloud, irregular-grid, arbitrary-domain, deformation-aware, domain-agnostic, or topology-robust operator learning.","exclusion":"If the geometry is fixed and the contribution is a general architecture or representation, assign to General operator methods; if geometry handling is incidental to a specific application, assign to Physical-system surrogates.","nearest_boundary":"General operator methods","representative_topics":["fourier neural operator with learned deformations for pdes on general geometries","mesh-invariant neural operator for wasserstein geodesic via coupled pde saddle point learning","transformer-based neural operator learning for pdes on irregular meshes with multi-scale gating"]},{"direction_index":3,"label":"D3","name":"Physical-system surrogates","description":"Using neural operators as fast forward surrogates, forecasters, downscalers, optimizers, or design tools for concrete physical, geophysical, and engineering systems.","inclusion":"Include domain-centered work in fluids, turbulence, weather, climate, geophysics, seismic waves, photonics, electromagnetics, MHD, infrastructure, aerodynamics, combustion, nuclear transport, acoustics, materials, or related systems where the operator primarily enables simulation, prediction, or design.","exclusion":"If explicit physics-informed residual enforcement is the main contribution, assign to Physics-constrained operator learning; if long-horizon temporal stability is central, assign to Stable dynamical and structure-preserving operators.","nearest_boundary":"Physics-constrained operator learning","representative_topics":["deep operator networks for efficient interdependent systems recovery modeling with sparse data","physics-agnostic neural operator framework for parametric photonic device electromagnetic simulation","fourier neural operator learning for seismic wave equation solving on variable velocity models"]},{"direction_index":4,"label":"D4","name":"Stable and structured dynamics","description":"Designing neural operators for reliable dynamical prediction, long-time integration, invariant or symmetry preservation, and control- or observer-oriented guarantees.","inclusion":"Include work focused on autoregressive stability, rollout error control, long-horizon chaotic or turbulent prediction, recurrent temporal integration, invariant-measure preservation, equivariance, variational or mimetic structure, conservation, and stabilizing controller or observer operator maps.","exclusion":"If time dependence is only the application setting without a stability, invariant, symmetry, or control objective, assign to Physical-system surrogates; if the structure is only a generic architectural device, assign to General operator methods.","nearest_boundary":"Physical-system surrogates","representative_topics":["recurrent neural operator learning for stable long-time integration in dynamical systems","neural operator learning of backstepping gain functions for parabolic reaction-diffusion pdes with stability proofs","neural operators trained to preserve chaotic attractor invariant measures via optimal transport and contrastive learning"]},{"direction_index":5,"label":"D5","name":"Data-efficient adaptation","description":"Reducing data requirements and improving transfer, adaptation, sampling, extrapolation, and continual or meta-learning behavior of neural operators.","inclusion":"Include active learning, critical sampling, sparse-data learning, multi-fidelity or multi-resolution training, transfer learning, zero-shot transfer, meta-learning, continual learning, extrapolation from limited support, and data augmentation when the main objective is improved learning efficiency or adaptation.","exclusion":"If multiscale or multiresolution behavior is mainly an architectural contribution without an explicit data-efficiency or transfer objective, assign to General operator methods; if sparse observations are used primarily for inverse reconstruction, assign to Inverse and uncertainty operators.","nearest_boundary":"General operator methods","representative_topics":["multi-fidelity wavelet neural operator learning for uncertainty quantification and surrogate modeling","zero-shot transferable deep graph operator networks for networked system dynamics approximation","physics-guided data augmentation for neural operator learning of linear differential equations"]},{"direction_index":6,"label":"D6","name":"Operator-learning theory","description":"Mathematical, statistical, and computational foundations for neural operator learning, including approximation, expressivity, complexity, generalization, risk, minimax optimality, and recovery theory.","inclusion":"Include work whose primary objective is a theorem, bound, rate, lower or upper complexity characterization, universality result, generalization analysis, statistical risk analysis, or principled comparison of operator-learning families.","exclusion":"If theory mainly supports a concrete new architecture or solver, assign to General operator methods; if guarantees concern control stability or invariant preservation of a designed operator, assign to Stable and structured dynamics.","nearest_boundary":"General operator methods","representative_topics":["rademacher complexity bounds for fourier neural operators via group norm capacity analysis","minimax optimal kernel operator learning via spectral bias-variance multilevel training in infinite-dimensional function spaces","out-of-distribution risk bounds for neural operators via stochastic depth and rademacher complexity"]},{"direction_index":7,"label":"D7","name":"Inverse and uncertainty operators","description":"Learning neural operators for inverse inference, posterior approximation, uncertainty quantification, probabilistic prediction, sparse-observation reconstruction, field completion, and super-resolution.","inclusion":"Include Bayesian neural operators, variational or approximate Bayesian operator learning, randomized priors, invertible operators for posterior sampling, neural inverse operators, uncertainty-aware prediction, and reconstruction or super-resolution from sparse or coarse observations.","exclusion":"If uncertainty or reconstruction is only a downstream evaluation of a deterministic forward surrogate, assign to Physical-system surrogates; if the central issue is limited training data rather than inverse or probabilistic inference, assign to Data-efficient adaptation.","nearest_boundary":"Physical-system surrogates","representative_topics":["variational bayesian deep operator networks for uncertainty-aware parametric differential equation solving","bayesian uncertainty quantification for deep neural operators solving parametric pdes","invertible deep operator networks for full posterior approximation in bayesian parametric pde inverse problems"]}],"split":"dev"}
{"episode_id":"F116_T2024-07","field_index":116,"field":"neural operator learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Physics-constrained operator learning","description":"Learning neural solution operators by directly imposing governing equations, residuals, boundary conditions, weak forms, conservation laws, symmetries, or other explicit physical constraints in training or architecture.","inclusion":"Include work whose primary objective is data-free, low-label, semi-supervised, or physically consistent operator learning through PDE residuals, physics-informed losses, boundary/weak constraints, or embedded physical laws.","exclusion":"If equations only define a benchmark or application and the operator is mainly a black-box surrogate, assign to Physical-system surrogates; if the main novelty is a general architecture without physics integration, assign to General operator methods.","nearest_boundary":"Physical-system surrogates","representative_topics":["physics-constrained neural operators for parametric pdes via low-rank long-range entanglement modeling","physics-informed deep operator networks for surrogate modeling in computational mechanics","semi-supervised fourier neural operator learning of nonlinear pde evolution operators"]},{"direction_index":1,"label":"D1","name":"General operator methods","description":"Developing broadly applicable neural-operator architectures, representations, training procedures, and numerical-neural integrations for mappings between function spaces.","inclusion":"Include new operator layers, transforms, DeepONet variants, FNO/spectral/wavelet/Laplace/attention/latent/variational/multimodal representations, super-resolution or symbolic extensions, and solver-augmented operator methods when the main objective is a reusable method rather than a particular physical system.","exclusion":"If the central objective is handling arbitrary geometries or meshes, assign to Geometry-general operators; if the main objective is theory, physics constraints, data efficiency, or uncertainty/inversion, assign to those directions instead.","nearest_boundary":"Geometry-general operators","representative_topics":["multiscale neural operator learning of grid-independent non-local pde parametrizations","hybrid deep operator and relaxation methods for scalable spectral pde solving","clifford algebraic neural layers for multivector field pde surrogate modeling"]},{"direction_index":2,"label":"D2","name":"Geometry-general operators","description":"Learning operators that generalize across discretizations, meshes, point distributions, irregular or arbitrary domains, deformations, unbounded domains, and evolving geometries.","inclusion":"Include work whose primary objective is mesh-free, point-cloud, irregular-grid, arbitrary-domain, deformation-aware, domain-agnostic, or topology-robust operator learning.","exclusion":"If the geometry is fixed and the contribution is a general architecture or representation, assign to General operator methods; if geometry handling is incidental to a specific application, assign to Physical-system surrogates.","nearest_boundary":"General operator methods","representative_topics":["fourier neural operator with learned deformations for pdes on general geometries","mesh-invariant neural operator for wasserstein geodesic via coupled pde saddle point learning","transformer-based neural operator learning for pdes on irregular meshes with multi-scale gating"]},{"direction_index":3,"label":"D3","name":"Physical-system surrogates","description":"Using neural operators as fast forward surrogates, forecasters, downscalers, optimizers, or design tools for concrete physical, geophysical, and engineering systems.","inclusion":"Include domain-centered work in fluids, turbulence, weather, climate, geophysics, seismic waves, photonics, electromagnetics, MHD, infrastructure, aerodynamics, combustion, nuclear transport, acoustics, materials, or related systems where the operator primarily enables simulation, prediction, or design.","exclusion":"If explicit physics-informed residual enforcement is the main contribution, assign to Physics-constrained operator learning; if long-horizon temporal stability is central, assign to Stable dynamical and structure-preserving operators.","nearest_boundary":"Physics-constrained operator learning","representative_topics":["deep operator networks for efficient interdependent systems recovery modeling with sparse data","physics-agnostic neural operator framework for parametric photonic device electromagnetic simulation","fourier neural operator learning for seismic wave equation solving on variable velocity models"]},{"direction_index":4,"label":"D4","name":"Stable and structured dynamics","description":"Designing neural operators for reliable dynamical prediction, long-time integration, invariant or symmetry preservation, and control- or observer-oriented guarantees.","inclusion":"Include work focused on autoregressive stability, rollout error control, long-horizon chaotic or turbulent prediction, recurrent temporal integration, invariant-measure preservation, equivariance, variational or mimetic structure, conservation, and stabilizing controller or observer operator maps.","exclusion":"If time dependence is only the application setting without a stability, invariant, symmetry, or control objective, assign to Physical-system surrogates; if the structure is only a generic architectural device, assign to General operator methods.","nearest_boundary":"Physical-system surrogates","representative_topics":["recurrent neural operator learning for stable long-time integration in dynamical systems","neural operator learning of backstepping gain functions for parabolic reaction-diffusion pdes with stability proofs","neural operators trained to preserve chaotic attractor invariant measures via optimal transport and contrastive learning"]},{"direction_index":5,"label":"D5","name":"Data-efficient adaptation","description":"Reducing data requirements and improving transfer, adaptation, sampling, extrapolation, and continual or meta-learning behavior of neural operators.","inclusion":"Include active learning, critical sampling, sparse-data learning, multi-fidelity or multi-resolution training, transfer learning, zero-shot transfer, meta-learning, continual learning, extrapolation from limited support, and data augmentation when the main objective is improved learning efficiency or adaptation.","exclusion":"If multiscale or multiresolution behavior is mainly an architectural contribution without an explicit data-efficiency or transfer objective, assign to General operator methods; if sparse observations are used primarily for inverse reconstruction, assign to Inverse and uncertainty operators.","nearest_boundary":"General operator methods","representative_topics":["multi-fidelity wavelet neural operator learning for uncertainty quantification and surrogate modeling","zero-shot transferable deep graph operator networks for networked system dynamics approximation","physics-guided data augmentation for neural operator learning of linear differential equations"]},{"direction_index":6,"label":"D6","name":"Operator-learning theory","description":"Mathematical, statistical, and computational foundations for neural operator learning, including approximation, expressivity, complexity, generalization, risk, minimax optimality, and recovery theory.","inclusion":"Include work whose primary objective is a theorem, bound, rate, lower or upper complexity characterization, universality result, generalization analysis, statistical risk analysis, or principled comparison of operator-learning families.","exclusion":"If theory mainly supports a concrete new architecture or solver, assign to General operator methods; if guarantees concern control stability or invariant preservation of a designed operator, assign to Stable and structured dynamics.","nearest_boundary":"General operator methods","representative_topics":["rademacher complexity bounds for fourier neural operators via group norm capacity analysis","minimax optimal kernel operator learning via spectral bias-variance multilevel training in infinite-dimensional function spaces","out-of-distribution risk bounds for neural operators via stochastic depth and rademacher complexity"]},{"direction_index":7,"label":"D7","name":"Inverse and uncertainty operators","description":"Learning neural operators for inverse inference, posterior approximation, uncertainty quantification, probabilistic prediction, sparse-observation reconstruction, field completion, and super-resolution.","inclusion":"Include Bayesian neural operators, variational or approximate Bayesian operator learning, randomized priors, invertible operators for posterior sampling, neural inverse operators, uncertainty-aware prediction, and reconstruction or super-resolution from sparse or coarse observations.","exclusion":"If uncertainty or reconstruction is only a downstream evaluation of a deterministic forward surrogate, assign to Physical-system surrogates; if the central issue is limited training data rather than inverse or probabilistic inference, assign to Data-efficient adaptation.","nearest_boundary":"Physical-system surrogates","representative_topics":["variational bayesian deep operator networks for uncertainty-aware parametric differential equation solving","bayesian uncertainty quantification for deep neural operators solving parametric pdes","invertible deep operator networks for full posterior approximation in bayesian parametric pde inverse problems"]}],"split":"dev"}
{"episode_id":"F116_T2025-01","field_index":116,"field":"neural operator learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Physics-constrained operator learning","description":"Learning neural solution operators by directly imposing governing equations, residuals, boundary conditions, weak forms, conservation laws, symmetries, or other explicit physical constraints in training or architecture.","inclusion":"Include work whose primary objective is data-free, low-label, semi-supervised, or physically consistent operator learning through PDE residuals, physics-informed losses, boundary/weak constraints, or embedded physical laws.","exclusion":"If equations only define a benchmark or application and the operator is mainly a black-box surrogate, assign to Physical-system surrogates; if the main novelty is a general architecture without physics integration, assign to General operator methods.","nearest_boundary":"Physical-system surrogates","representative_topics":["physics-constrained neural operators for parametric pdes via low-rank long-range entanglement modeling","physics-informed deep operator networks for surrogate modeling in computational mechanics","semi-supervised fourier neural operator learning of nonlinear pde evolution operators"]},{"direction_index":1,"label":"D1","name":"General operator methods","description":"Developing broadly applicable neural-operator architectures, representations, training procedures, and numerical-neural integrations for mappings between function spaces.","inclusion":"Include new operator layers, transforms, DeepONet variants, FNO/spectral/wavelet/Laplace/attention/latent/variational/multimodal representations, super-resolution or symbolic extensions, and solver-augmented operator methods when the main objective is a reusable method rather than a particular physical system.","exclusion":"If the central objective is handling arbitrary geometries or meshes, assign to Geometry-general operators; if the main objective is theory, physics constraints, data efficiency, or uncertainty/inversion, assign to those directions instead.","nearest_boundary":"Geometry-general operators","representative_topics":["multiscale neural operator learning of grid-independent non-local pde parametrizations","hybrid deep operator and relaxation methods for scalable spectral pde solving","clifford algebraic neural layers for multivector field pde surrogate modeling"]},{"direction_index":2,"label":"D2","name":"Geometry-general operators","description":"Learning operators that generalize across discretizations, meshes, point distributions, irregular or arbitrary domains, deformations, unbounded domains, and evolving geometries.","inclusion":"Include work whose primary objective is mesh-free, point-cloud, irregular-grid, arbitrary-domain, deformation-aware, domain-agnostic, or topology-robust operator learning.","exclusion":"If the geometry is fixed and the contribution is a general architecture or representation, assign to General operator methods; if geometry handling is incidental to a specific application, assign to Physical-system surrogates.","nearest_boundary":"General operator methods","representative_topics":["fourier neural operator with learned deformations for pdes on general geometries","mesh-invariant neural operator for wasserstein geodesic via coupled pde saddle point learning","transformer-based neural operator learning for pdes on irregular meshes with multi-scale gating"]},{"direction_index":3,"label":"D3","name":"Physical-system surrogates","description":"Using neural operators as fast forward surrogates, forecasters, downscalers, optimizers, or design tools for concrete physical, geophysical, and engineering systems.","inclusion":"Include domain-centered work in fluids, turbulence, weather, climate, geophysics, seismic waves, photonics, electromagnetics, MHD, infrastructure, aerodynamics, combustion, nuclear transport, acoustics, materials, or related systems where the operator primarily enables simulation, prediction, or design.","exclusion":"If explicit physics-informed residual enforcement is the main contribution, assign to Physics-constrained operator learning; if long-horizon temporal stability is central, assign to Stable dynamical and structure-preserving operators.","nearest_boundary":"Physics-constrained operator learning","representative_topics":["deep operator networks for efficient interdependent systems recovery modeling with sparse data","physics-agnostic neural operator framework for parametric photonic device electromagnetic simulation","fourier neural operator learning for seismic wave equation solving on variable velocity models"]},{"direction_index":4,"label":"D4","name":"Stable and structured dynamics","description":"Designing neural operators for reliable dynamical prediction, long-time integration, invariant or symmetry preservation, and control- or observer-oriented guarantees.","inclusion":"Include work focused on autoregressive stability, rollout error control, long-horizon chaotic or turbulent prediction, recurrent temporal integration, invariant-measure preservation, equivariance, variational or mimetic structure, conservation, and stabilizing controller or observer operator maps.","exclusion":"If time dependence is only the application setting without a stability, invariant, symmetry, or control objective, assign to Physical-system surrogates; if the structure is only a generic architectural device, assign to General operator methods.","nearest_boundary":"Physical-system surrogates","representative_topics":["recurrent neural operator learning for stable long-time integration in dynamical systems","neural operator learning of backstepping gain functions for parabolic reaction-diffusion pdes with stability proofs","neural operators trained to preserve chaotic attractor invariant measures via optimal transport and contrastive learning"]},{"direction_index":5,"label":"D5","name":"Data-efficient adaptation","description":"Reducing data requirements and improving transfer, adaptation, sampling, extrapolation, and continual or meta-learning behavior of neural operators.","inclusion":"Include active learning, critical sampling, sparse-data learning, multi-fidelity or multi-resolution training, transfer learning, zero-shot transfer, meta-learning, continual learning, extrapolation from limited support, and data augmentation when the main objective is improved learning efficiency or adaptation.","exclusion":"If multiscale or multiresolution behavior is mainly an architectural contribution without an explicit data-efficiency or transfer objective, assign to General operator methods; if sparse observations are used primarily for inverse reconstruction, assign to Inverse and uncertainty operators.","nearest_boundary":"General operator methods","representative_topics":["multi-fidelity wavelet neural operator learning for uncertainty quantification and surrogate modeling","zero-shot transferable deep graph operator networks for networked system dynamics approximation","physics-guided data augmentation for neural operator learning of linear differential equations"]},{"direction_index":6,"label":"D6","name":"Operator-learning theory","description":"Mathematical, statistical, and computational foundations for neural operator learning, including approximation, expressivity, complexity, generalization, risk, minimax optimality, and recovery theory.","inclusion":"Include work whose primary objective is a theorem, bound, rate, lower or upper complexity characterization, universality result, generalization analysis, statistical risk analysis, or principled comparison of operator-learning families.","exclusion":"If theory mainly supports a concrete new architecture or solver, assign to General operator methods; if guarantees concern control stability or invariant preservation of a designed operator, assign to Stable and structured dynamics.","nearest_boundary":"General operator methods","representative_topics":["rademacher complexity bounds for fourier neural operators via group norm capacity analysis","minimax optimal kernel operator learning via spectral bias-variance multilevel training in infinite-dimensional function spaces","out-of-distribution risk bounds for neural operators via stochastic depth and rademacher complexity"]},{"direction_index":7,"label":"D7","name":"Inverse and uncertainty operators","description":"Learning neural operators for inverse inference, posterior approximation, uncertainty quantification, probabilistic prediction, sparse-observation reconstruction, field completion, and super-resolution.","inclusion":"Include Bayesian neural operators, variational or approximate Bayesian operator learning, randomized priors, invertible operators for posterior sampling, neural inverse operators, uncertainty-aware prediction, and reconstruction or super-resolution from sparse or coarse observations.","exclusion":"If uncertainty or reconstruction is only a downstream evaluation of a deterministic forward surrogate, assign to Physical-system surrogates; if the central issue is limited training data rather than inverse or probabilistic inference, assign to Data-efficient adaptation.","nearest_boundary":"Physical-system surrogates","representative_topics":["variational bayesian deep operator networks for uncertainty-aware parametric differential equation solving","bayesian uncertainty quantification for deep neural operators solving parametric pdes","invertible deep operator networks for full posterior approximation in bayesian parametric pde inverse problems"]}],"split":"dev"}
{"episode_id":"F116_T2025-07","field_index":116,"field":"neural operator learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Physics-constrained operator learning","description":"Learning neural solution operators by directly imposing governing equations, residuals, boundary conditions, weak forms, conservation laws, symmetries, or other explicit physical constraints in training or architecture.","inclusion":"Include work whose primary objective is data-free, low-label, semi-supervised, or physically consistent operator learning through PDE residuals, physics-informed losses, boundary/weak constraints, or embedded physical laws.","exclusion":"If equations only define a benchmark or application and the operator is mainly a black-box surrogate, assign to Physical-system surrogates; if the main novelty is a general architecture without physics integration, assign to General operator methods.","nearest_boundary":"Physical-system surrogates","representative_topics":["physics-constrained neural operators for parametric pdes via low-rank long-range entanglement modeling","physics-informed deep operator networks for surrogate modeling in computational mechanics","semi-supervised fourier neural operator learning of nonlinear pde evolution operators"]},{"direction_index":1,"label":"D1","name":"General operator methods","description":"Developing broadly applicable neural-operator architectures, representations, training procedures, and numerical-neural integrations for mappings between function spaces.","inclusion":"Include new operator layers, transforms, DeepONet variants, FNO/spectral/wavelet/Laplace/attention/latent/variational/multimodal representations, super-resolution or symbolic extensions, and solver-augmented operator methods when the main objective is a reusable method rather than a particular physical system.","exclusion":"If the central objective is handling arbitrary geometries or meshes, assign to Geometry-general operators; if the main objective is theory, physics constraints, data efficiency, or uncertainty/inversion, assign to those directions instead.","nearest_boundary":"Geometry-general operators","representative_topics":["multiscale neural operator learning of grid-independent non-local pde parametrizations","hybrid deep operator and relaxation methods for scalable spectral pde solving","clifford algebraic neural layers for multivector field pde surrogate modeling"]},{"direction_index":2,"label":"D2","name":"Geometry-general operators","description":"Learning operators that generalize across discretizations, meshes, point distributions, irregular or arbitrary domains, deformations, unbounded domains, and evolving geometries.","inclusion":"Include work whose primary objective is mesh-free, point-cloud, irregular-grid, arbitrary-domain, deformation-aware, domain-agnostic, or topology-robust operator learning.","exclusion":"If the geometry is fixed and the contribution is a general architecture or representation, assign to General operator methods; if geometry handling is incidental to a specific application, assign to Physical-system surrogates.","nearest_boundary":"General operator methods","representative_topics":["fourier neural operator with learned deformations for pdes on general geometries","mesh-invariant neural operator for wasserstein geodesic via coupled pde saddle point learning","transformer-based neural operator learning for pdes on irregular meshes with multi-scale gating"]},{"direction_index":3,"label":"D3","name":"Physical-system surrogates","description":"Using neural operators as fast forward surrogates, forecasters, downscalers, optimizers, or design tools for concrete physical, geophysical, and engineering systems.","inclusion":"Include domain-centered work in fluids, turbulence, weather, climate, geophysics, seismic waves, photonics, electromagnetics, MHD, infrastructure, aerodynamics, combustion, nuclear transport, acoustics, materials, or related systems where the operator primarily enables simulation, prediction, or design.","exclusion":"If explicit physics-informed residual enforcement is the main contribution, assign to Physics-constrained operator learning; if long-horizon temporal stability is central, assign to Stable dynamical and structure-preserving operators.","nearest_boundary":"Physics-constrained operator learning","representative_topics":["deep operator networks for efficient interdependent systems recovery modeling with sparse data","physics-agnostic neural operator framework for parametric photonic device electromagnetic simulation","fourier neural operator learning for seismic wave equation solving on variable velocity models"]},{"direction_index":4,"label":"D4","name":"Stable and structured dynamics","description":"Designing neural operators for reliable dynamical prediction, long-time integration, invariant or symmetry preservation, and control- or observer-oriented guarantees.","inclusion":"Include work focused on autoregressive stability, rollout error control, long-horizon chaotic or turbulent prediction, recurrent temporal integration, invariant-measure preservation, equivariance, variational or mimetic structure, conservation, and stabilizing controller or observer operator maps.","exclusion":"If time dependence is only the application setting without a stability, invariant, symmetry, or control objective, assign to Physical-system surrogates; if the structure is only a generic architectural device, assign to General operator methods.","nearest_boundary":"Physical-system surrogates","representative_topics":["recurrent neural operator learning for stable long-time integration in dynamical systems","neural operator learning of backstepping gain functions for parabolic reaction-diffusion pdes with stability proofs","neural operators trained to preserve chaotic attractor invariant measures via optimal transport and contrastive learning"]},{"direction_index":5,"label":"D5","name":"Data-efficient adaptation","description":"Reducing data requirements and improving transfer, adaptation, sampling, extrapolation, and continual or meta-learning behavior of neural operators.","inclusion":"Include active learning, critical sampling, sparse-data learning, multi-fidelity or multi-resolution training, transfer learning, zero-shot transfer, meta-learning, continual learning, extrapolation from limited support, and data augmentation when the main objective is improved learning efficiency or adaptation.","exclusion":"If multiscale or multiresolution behavior is mainly an architectural contribution without an explicit data-efficiency or transfer objective, assign to General operator methods; if sparse observations are used primarily for inverse reconstruction, assign to Inverse and uncertainty operators.","nearest_boundary":"General operator methods","representative_topics":["multi-fidelity wavelet neural operator learning for uncertainty quantification and surrogate modeling","zero-shot transferable deep graph operator networks for networked system dynamics approximation","physics-guided data augmentation for neural operator learning of linear differential equations"]},{"direction_index":6,"label":"D6","name":"Operator-learning theory","description":"Mathematical, statistical, and computational foundations for neural operator learning, including approximation, expressivity, complexity, generalization, risk, minimax optimality, and recovery theory.","inclusion":"Include work whose primary objective is a theorem, bound, rate, lower or upper complexity characterization, universality result, generalization analysis, statistical risk analysis, or principled comparison of operator-learning families.","exclusion":"If theory mainly supports a concrete new architecture or solver, assign to General operator methods; if guarantees concern control stability or invariant preservation of a designed operator, assign to Stable and structured dynamics.","nearest_boundary":"General operator methods","representative_topics":["rademacher complexity bounds for fourier neural operators via group norm capacity analysis","minimax optimal kernel operator learning via spectral bias-variance multilevel training in infinite-dimensional function spaces","out-of-distribution risk bounds for neural operators via stochastic depth and rademacher complexity"]},{"direction_index":7,"label":"D7","name":"Inverse and uncertainty operators","description":"Learning neural operators for inverse inference, posterior approximation, uncertainty quantification, probabilistic prediction, sparse-observation reconstruction, field completion, and super-resolution.","inclusion":"Include Bayesian neural operators, variational or approximate Bayesian operator learning, randomized priors, invertible operators for posterior sampling, neural inverse operators, uncertainty-aware prediction, and reconstruction or super-resolution from sparse or coarse observations.","exclusion":"If uncertainty or reconstruction is only a downstream evaluation of a deterministic forward surrogate, assign to Physical-system surrogates; if the central issue is limited training data rather than inverse or probabilistic inference, assign to Data-efficient adaptation.","nearest_boundary":"Physical-system surrogates","representative_topics":["variational bayesian deep operator networks for uncertainty-aware parametric differential equation solving","bayesian uncertainty quantification for deep neural operators solving parametric pdes","invertible deep operator networks for full posterior approximation in bayesian parametric pde inverse problems"]}],"split":"dev"}
{"episode_id":"F116_T2026-01","field_index":116,"field":"neural operator learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Physics-constrained operator learning","description":"Learning neural solution operators by directly imposing governing equations, residuals, boundary conditions, weak forms, conservation laws, symmetries, or other explicit physical constraints in training or architecture.","inclusion":"Include work whose primary objective is data-free, low-label, semi-supervised, or physically consistent operator learning through PDE residuals, physics-informed losses, boundary/weak constraints, or embedded physical laws.","exclusion":"If equations only define a benchmark or application and the operator is mainly a black-box surrogate, assign to Physical-system surrogates; if the main novelty is a general architecture without physics integration, assign to General operator methods.","nearest_boundary":"Physical-system surrogates","representative_topics":["physics-constrained neural operators for parametric pdes via low-rank long-range entanglement modeling","physics-informed deep operator networks for surrogate modeling in computational mechanics","semi-supervised fourier neural operator learning of nonlinear pde evolution operators"]},{"direction_index":1,"label":"D1","name":"General operator methods","description":"Developing broadly applicable neural-operator architectures, representations, training procedures, and numerical-neural integrations for mappings between function spaces.","inclusion":"Include new operator layers, transforms, DeepONet variants, FNO/spectral/wavelet/Laplace/attention/latent/variational/multimodal representations, super-resolution or symbolic extensions, and solver-augmented operator methods when the main objective is a reusable method rather than a particular physical system.","exclusion":"If the central objective is handling arbitrary geometries or meshes, assign to Geometry-general operators; if the main objective is theory, physics constraints, data efficiency, or uncertainty/inversion, assign to those directions instead.","nearest_boundary":"Geometry-general operators","representative_topics":["multiscale neural operator learning of grid-independent non-local pde parametrizations","hybrid deep operator and relaxation methods for scalable spectral pde solving","clifford algebraic neural layers for multivector field pde surrogate modeling"]},{"direction_index":2,"label":"D2","name":"Geometry-general operators","description":"Learning operators that generalize across discretizations, meshes, point distributions, irregular or arbitrary domains, deformations, unbounded domains, and evolving geometries.","inclusion":"Include work whose primary objective is mesh-free, point-cloud, irregular-grid, arbitrary-domain, deformation-aware, domain-agnostic, or topology-robust operator learning.","exclusion":"If the geometry is fixed and the contribution is a general architecture or representation, assign to General operator methods; if geometry handling is incidental to a specific application, assign to Physical-system surrogates.","nearest_boundary":"General operator methods","representative_topics":["fourier neural operator with learned deformations for pdes on general geometries","mesh-invariant neural operator for wasserstein geodesic via coupled pde saddle point learning","transformer-based neural operator learning for pdes on irregular meshes with multi-scale gating"]},{"direction_index":3,"label":"D3","name":"Physical-system surrogates","description":"Using neural operators as fast forward surrogates, forecasters, downscalers, optimizers, or design tools for concrete physical, geophysical, and engineering systems.","inclusion":"Include domain-centered work in fluids, turbulence, weather, climate, geophysics, seismic waves, photonics, electromagnetics, MHD, infrastructure, aerodynamics, combustion, nuclear transport, acoustics, materials, or related systems where the operator primarily enables simulation, prediction, or design.","exclusion":"If explicit physics-informed residual enforcement is the main contribution, assign to Physics-constrained operator learning; if long-horizon temporal stability is central, assign to Stable dynamical and structure-preserving operators.","nearest_boundary":"Physics-constrained operator learning","representative_topics":["deep operator networks for efficient interdependent systems recovery modeling with sparse data","physics-agnostic neural operator framework for parametric photonic device electromagnetic simulation","fourier neural operator learning for seismic wave equation solving on variable velocity models"]},{"direction_index":4,"label":"D4","name":"Stable and structured dynamics","description":"Designing neural operators for reliable dynamical prediction, long-time integration, invariant or symmetry preservation, and control- or observer-oriented guarantees.","inclusion":"Include work focused on autoregressive stability, rollout error control, long-horizon chaotic or turbulent prediction, recurrent temporal integration, invariant-measure preservation, equivariance, variational or mimetic structure, conservation, and stabilizing controller or observer operator maps.","exclusion":"If time dependence is only the application setting without a stability, invariant, symmetry, or control objective, assign to Physical-system surrogates; if the structure is only a generic architectural device, assign to General operator methods.","nearest_boundary":"Physical-system surrogates","representative_topics":["recurrent neural operator learning for stable long-time integration in dynamical systems","neural operator learning of backstepping gain functions for parabolic reaction-diffusion pdes with stability proofs","neural operators trained to preserve chaotic attractor invariant measures via optimal transport and contrastive learning"]},{"direction_index":5,"label":"D5","name":"Data-efficient adaptation","description":"Reducing data requirements and improving transfer, adaptation, sampling, extrapolation, and continual or meta-learning behavior of neural operators.","inclusion":"Include active learning, critical sampling, sparse-data learning, multi-fidelity or multi-resolution training, transfer learning, zero-shot transfer, meta-learning, continual learning, extrapolation from limited support, and data augmentation when the main objective is improved learning efficiency or adaptation.","exclusion":"If multiscale or multiresolution behavior is mainly an architectural contribution without an explicit data-efficiency or transfer objective, assign to General operator methods; if sparse observations are used primarily for inverse reconstruction, assign to Inverse and uncertainty operators.","nearest_boundary":"General operator methods","representative_topics":["multi-fidelity wavelet neural operator learning for uncertainty quantification and surrogate modeling","zero-shot transferable deep graph operator networks for networked system dynamics approximation","physics-guided data augmentation for neural operator learning of linear differential equations"]},{"direction_index":6,"label":"D6","name":"Operator-learning theory","description":"Mathematical, statistical, and computational foundations for neural operator learning, including approximation, expressivity, complexity, generalization, risk, minimax optimality, and recovery theory.","inclusion":"Include work whose primary objective is a theorem, bound, rate, lower or upper complexity characterization, universality result, generalization analysis, statistical risk analysis, or principled comparison of operator-learning families.","exclusion":"If theory mainly supports a concrete new architecture or solver, assign to General operator methods; if guarantees concern control stability or invariant preservation of a designed operator, assign to Stable and structured dynamics.","nearest_boundary":"General operator methods","representative_topics":["rademacher complexity bounds for fourier neural operators via group norm capacity analysis","minimax optimal kernel operator learning via spectral bias-variance multilevel training in infinite-dimensional function spaces","out-of-distribution risk bounds for neural operators via stochastic depth and rademacher complexity"]},{"direction_index":7,"label":"D7","name":"Inverse and uncertainty operators","description":"Learning neural operators for inverse inference, posterior approximation, uncertainty quantification, probabilistic prediction, sparse-observation reconstruction, field completion, and super-resolution.","inclusion":"Include Bayesian neural operators, variational or approximate Bayesian operator learning, randomized priors, invertible operators for posterior sampling, neural inverse operators, uncertainty-aware prediction, and reconstruction or super-resolution from sparse or coarse observations.","exclusion":"If uncertainty or reconstruction is only a downstream evaluation of a deterministic forward surrogate, assign to Physical-system surrogates; if the central issue is limited training data rather than inverse or probabilistic inference, assign to Data-efficient adaptation.","nearest_boundary":"Physical-system surrogates","representative_topics":["variational bayesian deep operator networks for uncertainty-aware parametric differential equation solving","bayesian uncertainty quantification for deep neural operators solving parametric pdes","invertible deep operator networks for full posterior approximation in bayesian parametric pde inverse problems"]}],"split":"dev"}
{"episode_id":"F118_T2024-01","field_index":118,"field":"multimodal learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Robust and constrained multimodal transfer","description":"Methods whose primary objective is to make multimodal learning work under modality absence, limited labels or pairing, domain shift, frozen-model adaptation, or teacher-student transfer constraints.","inclusion":"Include papers centered on missing or corrupted modalities, modality hallucination or imputation, incomplete multi-view learning, low-resource or source-free multimodal transfer, prompt/adaptor-style lightweight adaptation, and cross-modal distillation or transfer when the constraint itself is the main scientific objective.","exclusion":"If the main objective is a general reusable pretraining or retrieval embedding rather than constrained transfer, assign to Cross-modal representation and retrieval; if the main objective is safety evaluation or adversarial/federated deployment, assign to Fusion, evaluation, and trustworthy mechanisms.","nearest_boundary":"Cross-modal representation and retrieval","representative_topics":["multimodal inertial sensor fusion with hallucination handling via privileged information learning","efficient multimodal transformer with dual-level feature restoration for robust sentiment analysis under missing modalities","generative adversarial clustering for incomplete image-text with modality-specific subspace learning"]},{"direction_index":1,"label":"D1","name":"Human affect and social content understanding","description":"Multimodal systems whose primary objective is to infer human affective states, subjective attitudes, or socially consequential meanings from multimodal human communication and social media.","inclusion":"Include sentiment, emotion, stress, pain, humor or affective-state recognition, as well as hate, misogyny, ideology, misinformation, fact verification, persuasion, memes, sarcasm when treated as social-pragmatic content, and other social-media integrity tasks.","exclusion":"If the target is a general caption, answer, document extraction, or assistant response rather than affective or social-content judgment, assign to Multimodal language, reasoning, and structured knowledge systems.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["uncertainty-aware audiovisual fusion via calibrated and ordinal latent distribution constraints","multi-scale cooperative transformer for unaligned multimodal video sentiment analysis","multimodal aspect-category sentiment analysis with cross-modal graph-based fine-grained fusion"]},{"direction_index":2,"label":"D2","name":"Cross-modal representation and retrieval","description":"Learning reusable aligned representations or shared embedding spaces across modalities for downstream transfer, semantic matching, grounding, search, and retrieval.","inclusion":"Include contrastive, masked, generative, cross-lingual, or self-supervised multimodal pretraining when the output is a general representation; image-text, video-text, audio-text, product, person, composed-query, referring-expression, or fine-grained retrieval and grounding when embedding alignment or matching is primary.","exclusion":"If the main objective is instruction-following with a generative language backbone, assign to Multimodal language, reasoning, and structured knowledge systems; if the main objective is non-text media synthesis, assign to Multimodal media generation.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["masked encoder-decoder pre-training unifying vision and language for text recognition","unified cross-lingual cross-modal pre-training via cross-view mutual information alignment","generative vision-language pretraining via bidirectional masked prediction on shared transformer"]},{"direction_index":3,"label":"D3","name":"Multimodal language, reasoning, and structured knowledge systems","description":"Systems whose primary objective is to use multimodal evidence for language-centered outputs, assistant behavior, explicit reasoning, dialogue, document understanding, QA, captioning, or structured knowledge inference.","inclusion":"Include multimodal instruction-following and large language/foundation models, multimodal chain-of-thought, visual chat, multimodal QA and dialogue, captioning, summarization, document information extraction, visually assisted language tasks, and multimodal knowledge-graph construction, alignment, completion, or reasoning.","exclusion":"If the main output is synthesized image, audio, video, 3D, or other non-text media, assign to Multimodal media generation; if the work is only learning a retrieval embedding without language generation or explicit reasoning, assign to Cross-modal representation and retrieval.","nearest_boundary":"Cross-modal representation and retrieval","representative_topics":["zero-shot multimodal instruction tuning with cross-modal transfer and instruction sensitivity evaluation","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","zero-shot multimodal large language model training from scratch on interleaved text-image corpora"]},{"direction_index":4,"label":"D4","name":"Multimodal media generation","description":"Generative modeling programs whose primary objective is to synthesize, translate, complete, edit, or jointly model non-text media or multimodal outputs across modalities.","inclusion":"Include multimodal VAEs, GANs, diffusion, autoregressive or latent generative models for image, audio, speech, video, 3D, motion, multimodal explanation media, cross-modal synthesis, audio-to-image, text-to-image with multimodal controls, and media completion or editing.","exclusion":"If generation is primarily a language answer, caption, summary, or document output, assign to Multimodal language, reasoning, and structured knowledge systems; if instruction-following alignment is primary, assign there even when generation tools are used.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["multimodal electropalatography and audio fusion for speech generation and enhancement","survey of variational autoencoder-based multimodal deep generative models for cross-modal generation","cross-modal 3d shape generation and manipulation via shared latent spaces for sketches and color images"]},{"direction_index":5,"label":"D5","name":"Sensorimotor and domain perception systems","description":"Multimodal learning for physical-world perception, sensing, tracking, action, planning, and domain-tailored decision systems outside clinical diagnosis.","inclusion":"Include autonomous-driving sensor fusion, RGB-depth/thermal/infrared/multispectral/event visual perception, audio-visual scene or event understanding, embodied agents, robotics, manipulation, multimodal planning, geospatial or disaster sensing, industrial perception, recommendation or tracking systems when the primary objective is a domain-tailored perception/action/decision capability.","exclusion":"If the primary objective is biomedical diagnosis or prognosis, assign to Biomedical multimodal prediction; if the work proposes a general fusion principle independent of the domain objective, assign to Fusion, evaluation, and trustworthy mechanisms.","nearest_boundary":"Fusion, evaluation, and trustworthy mechanisms","representative_topics":["multi-sensor bev perception: key training and sensor integration factors beyond lifting strategies","deformable cross-modal feature aggregation for efficient multi-modal 3d object detection","cross-modal visible-thermal person re-identification via local shortest path alignment and global feature enhancement"]},{"direction_index":6,"label":"D6","name":"Biomedical multimodal prediction","description":"Integrating biomedical, clinical, imaging, omics, pathology, EHR, neuroimaging, report, biomarker, or physiological modalities for health-state prediction and clinical decision support.","inclusion":"Include multimodal diagnosis, prognosis, survival modeling, risk stratification, disease severity prediction, medical decision support, medical image-report fusion, imaging-tabular or imaging-omics fusion, and biomedical reviews when clinical prediction or decision support is the primary objective.","exclusion":"Medical image registration or segmentation without clinical prediction belongs to Sensorimotor and domain perception systems; general medical representation pretraining without a specific clinical endpoint belongs to Cross-modal representation and retrieval.","nearest_boundary":"Sensorimotor and domain perception systems","representative_topics":["multimodal stacking of chest x-ray and clinical data for covid-19 mortality risk prediction","cross-modal attention fusion for multimodal alzheimer's disease and mild cognitive impairment diagnosis","multimodal survival modeling for radiation encephalopathy prediction using weighted concordance index loss"]},{"direction_index":7,"label":"D7","name":"Fusion, evaluation, and trustworthy mechanisms","description":"Research whose primary objective is to understand, design, benchmark, interpret, secure, or make reliable the general mechanisms by which modalities interact and are evaluated.","inclusion":"Include general fusion architectures, attention or graph fusion mechanisms, modality interaction theory, uncertainty-aware or imbalance-aware fusion, adversarial robustness and defenses, poisoning/backdoor studies, privacy-preserving or federated multimodal learning, compression or efficiency analysis, interpretability, surveys, diagnostic benchmarks, and standardized toolkits when mechanism or assessment is primary.","exclusion":"If a fusion module is mainly a contribution to a sustained task or domain such as affect, driving, medicine, retrieval, generation, or language QA, assign to that objective-specific direction; if missing-modality recovery is central, assign to Robust and constrained multimodal transfer.","nearest_boundary":"Robust and constrained multimodal transfer","representative_topics":["defending multimodal fusion models against single-source adversarial perturbations via cross-modal inconsistency detection","uncertainty-aware multimodal fusion via cross-modal random network prediction for feature density estimation","progressive fusion with cross-stage residual connections for multimodal representation refinement"]}],"split":"dev"}
{"episode_id":"F118_T2024-07","field_index":118,"field":"multimodal learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Robust and constrained multimodal transfer","description":"Methods whose primary objective is to make multimodal learning work under modality absence, limited labels or pairing, domain shift, frozen-model adaptation, or teacher-student transfer constraints.","inclusion":"Include papers centered on missing or corrupted modalities, modality hallucination or imputation, incomplete multi-view learning, low-resource or source-free multimodal transfer, prompt/adaptor-style lightweight adaptation, and cross-modal distillation or transfer when the constraint itself is the main scientific objective.","exclusion":"If the main objective is a general reusable pretraining or retrieval embedding rather than constrained transfer, assign to Cross-modal representation and retrieval; if the main objective is safety evaluation or adversarial/federated deployment, assign to Fusion, evaluation, and trustworthy mechanisms.","nearest_boundary":"Cross-modal representation and retrieval","representative_topics":["multimodal inertial sensor fusion with hallucination handling via privileged information learning","efficient multimodal transformer with dual-level feature restoration for robust sentiment analysis under missing modalities","generative adversarial clustering for incomplete image-text with modality-specific subspace learning"]},{"direction_index":1,"label":"D1","name":"Human affect and social content understanding","description":"Multimodal systems whose primary objective is to infer human affective states, subjective attitudes, or socially consequential meanings from multimodal human communication and social media.","inclusion":"Include sentiment, emotion, stress, pain, humor or affective-state recognition, as well as hate, misogyny, ideology, misinformation, fact verification, persuasion, memes, sarcasm when treated as social-pragmatic content, and other social-media integrity tasks.","exclusion":"If the target is a general caption, answer, document extraction, or assistant response rather than affective or social-content judgment, assign to Multimodal language, reasoning, and structured knowledge systems.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["uncertainty-aware audiovisual fusion via calibrated and ordinal latent distribution constraints","multi-scale cooperative transformer for unaligned multimodal video sentiment analysis","multimodal aspect-category sentiment analysis with cross-modal graph-based fine-grained fusion"]},{"direction_index":2,"label":"D2","name":"Cross-modal representation and retrieval","description":"Learning reusable aligned representations or shared embedding spaces across modalities for downstream transfer, semantic matching, grounding, search, and retrieval.","inclusion":"Include contrastive, masked, generative, cross-lingual, or self-supervised multimodal pretraining when the output is a general representation; image-text, video-text, audio-text, product, person, composed-query, referring-expression, or fine-grained retrieval and grounding when embedding alignment or matching is primary.","exclusion":"If the main objective is instruction-following with a generative language backbone, assign to Multimodal language, reasoning, and structured knowledge systems; if the main objective is non-text media synthesis, assign to Multimodal media generation.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["masked encoder-decoder pre-training unifying vision and language for text recognition","unified cross-lingual cross-modal pre-training via cross-view mutual information alignment","generative vision-language pretraining via bidirectional masked prediction on shared transformer"]},{"direction_index":3,"label":"D3","name":"Multimodal language, reasoning, and structured knowledge systems","description":"Systems whose primary objective is to use multimodal evidence for language-centered outputs, assistant behavior, explicit reasoning, dialogue, document understanding, QA, captioning, or structured knowledge inference.","inclusion":"Include multimodal instruction-following and large language/foundation models, multimodal chain-of-thought, visual chat, multimodal QA and dialogue, captioning, summarization, document information extraction, visually assisted language tasks, and multimodal knowledge-graph construction, alignment, completion, or reasoning.","exclusion":"If the main output is synthesized image, audio, video, 3D, or other non-text media, assign to Multimodal media generation; if the work is only learning a retrieval embedding without language generation or explicit reasoning, assign to Cross-modal representation and retrieval.","nearest_boundary":"Cross-modal representation and retrieval","representative_topics":["zero-shot multimodal instruction tuning with cross-modal transfer and instruction sensitivity evaluation","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","zero-shot multimodal large language model training from scratch on interleaved text-image corpora"]},{"direction_index":4,"label":"D4","name":"Multimodal media generation","description":"Generative modeling programs whose primary objective is to synthesize, translate, complete, edit, or jointly model non-text media or multimodal outputs across modalities.","inclusion":"Include multimodal VAEs, GANs, diffusion, autoregressive or latent generative models for image, audio, speech, video, 3D, motion, multimodal explanation media, cross-modal synthesis, audio-to-image, text-to-image with multimodal controls, and media completion or editing.","exclusion":"If generation is primarily a language answer, caption, summary, or document output, assign to Multimodal language, reasoning, and structured knowledge systems; if instruction-following alignment is primary, assign there even when generation tools are used.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["multimodal electropalatography and audio fusion for speech generation and enhancement","survey of variational autoencoder-based multimodal deep generative models for cross-modal generation","cross-modal 3d shape generation and manipulation via shared latent spaces for sketches and color images"]},{"direction_index":5,"label":"D5","name":"Sensorimotor and domain perception systems","description":"Multimodal learning for physical-world perception, sensing, tracking, action, planning, and domain-tailored decision systems outside clinical diagnosis.","inclusion":"Include autonomous-driving sensor fusion, RGB-depth/thermal/infrared/multispectral/event visual perception, audio-visual scene or event understanding, embodied agents, robotics, manipulation, multimodal planning, geospatial or disaster sensing, industrial perception, recommendation or tracking systems when the primary objective is a domain-tailored perception/action/decision capability.","exclusion":"If the primary objective is biomedical diagnosis or prognosis, assign to Biomedical multimodal prediction; if the work proposes a general fusion principle independent of the domain objective, assign to Fusion, evaluation, and trustworthy mechanisms.","nearest_boundary":"Fusion, evaluation, and trustworthy mechanisms","representative_topics":["multi-sensor bev perception: key training and sensor integration factors beyond lifting strategies","deformable cross-modal feature aggregation for efficient multi-modal 3d object detection","cross-modal visible-thermal person re-identification via local shortest path alignment and global feature enhancement"]},{"direction_index":6,"label":"D6","name":"Biomedical multimodal prediction","description":"Integrating biomedical, clinical, imaging, omics, pathology, EHR, neuroimaging, report, biomarker, or physiological modalities for health-state prediction and clinical decision support.","inclusion":"Include multimodal diagnosis, prognosis, survival modeling, risk stratification, disease severity prediction, medical decision support, medical image-report fusion, imaging-tabular or imaging-omics fusion, and biomedical reviews when clinical prediction or decision support is the primary objective.","exclusion":"Medical image registration or segmentation without clinical prediction belongs to Sensorimotor and domain perception systems; general medical representation pretraining without a specific clinical endpoint belongs to Cross-modal representation and retrieval.","nearest_boundary":"Sensorimotor and domain perception systems","representative_topics":["multimodal stacking of chest x-ray and clinical data for covid-19 mortality risk prediction","cross-modal attention fusion for multimodal alzheimer's disease and mild cognitive impairment diagnosis","multimodal survival modeling for radiation encephalopathy prediction using weighted concordance index loss"]},{"direction_index":7,"label":"D7","name":"Fusion, evaluation, and trustworthy mechanisms","description":"Research whose primary objective is to understand, design, benchmark, interpret, secure, or make reliable the general mechanisms by which modalities interact and are evaluated.","inclusion":"Include general fusion architectures, attention or graph fusion mechanisms, modality interaction theory, uncertainty-aware or imbalance-aware fusion, adversarial robustness and defenses, poisoning/backdoor studies, privacy-preserving or federated multimodal learning, compression or efficiency analysis, interpretability, surveys, diagnostic benchmarks, and standardized toolkits when mechanism or assessment is primary.","exclusion":"If a fusion module is mainly a contribution to a sustained task or domain such as affect, driving, medicine, retrieval, generation, or language QA, assign to that objective-specific direction; if missing-modality recovery is central, assign to Robust and constrained multimodal transfer.","nearest_boundary":"Robust and constrained multimodal transfer","representative_topics":["defending multimodal fusion models against single-source adversarial perturbations via cross-modal inconsistency detection","uncertainty-aware multimodal fusion via cross-modal random network prediction for feature density estimation","progressive fusion with cross-stage residual connections for multimodal representation refinement"]}],"split":"dev"}
{"episode_id":"F118_T2025-01","field_index":118,"field":"multimodal learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Robust and constrained multimodal transfer","description":"Methods whose primary objective is to make multimodal learning work under modality absence, limited labels or pairing, domain shift, frozen-model adaptation, or teacher-student transfer constraints.","inclusion":"Include papers centered on missing or corrupted modalities, modality hallucination or imputation, incomplete multi-view learning, low-resource or source-free multimodal transfer, prompt/adaptor-style lightweight adaptation, and cross-modal distillation or transfer when the constraint itself is the main scientific objective.","exclusion":"If the main objective is a general reusable pretraining or retrieval embedding rather than constrained transfer, assign to Cross-modal representation and retrieval; if the main objective is safety evaluation or adversarial/federated deployment, assign to Fusion, evaluation, and trustworthy mechanisms.","nearest_boundary":"Cross-modal representation and retrieval","representative_topics":["multimodal inertial sensor fusion with hallucination handling via privileged information learning","efficient multimodal transformer with dual-level feature restoration for robust sentiment analysis under missing modalities","generative adversarial clustering for incomplete image-text with modality-specific subspace learning"]},{"direction_index":1,"label":"D1","name":"Human affect and social content understanding","description":"Multimodal systems whose primary objective is to infer human affective states, subjective attitudes, or socially consequential meanings from multimodal human communication and social media.","inclusion":"Include sentiment, emotion, stress, pain, humor or affective-state recognition, as well as hate, misogyny, ideology, misinformation, fact verification, persuasion, memes, sarcasm when treated as social-pragmatic content, and other social-media integrity tasks.","exclusion":"If the target is a general caption, answer, document extraction, or assistant response rather than affective or social-content judgment, assign to Multimodal language, reasoning, and structured knowledge systems.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["uncertainty-aware audiovisual fusion via calibrated and ordinal latent distribution constraints","multi-scale cooperative transformer for unaligned multimodal video sentiment analysis","multimodal aspect-category sentiment analysis with cross-modal graph-based fine-grained fusion"]},{"direction_index":2,"label":"D2","name":"Cross-modal representation and retrieval","description":"Learning reusable aligned representations or shared embedding spaces across modalities for downstream transfer, semantic matching, grounding, search, and retrieval.","inclusion":"Include contrastive, masked, generative, cross-lingual, or self-supervised multimodal pretraining when the output is a general representation; image-text, video-text, audio-text, product, person, composed-query, referring-expression, or fine-grained retrieval and grounding when embedding alignment or matching is primary.","exclusion":"If the main objective is instruction-following with a generative language backbone, assign to Multimodal language, reasoning, and structured knowledge systems; if the main objective is non-text media synthesis, assign to Multimodal media generation.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["masked encoder-decoder pre-training unifying vision and language for text recognition","unified cross-lingual cross-modal pre-training via cross-view mutual information alignment","generative vision-language pretraining via bidirectional masked prediction on shared transformer"]},{"direction_index":3,"label":"D3","name":"Multimodal language, reasoning, and structured knowledge systems","description":"Systems whose primary objective is to use multimodal evidence for language-centered outputs, assistant behavior, explicit reasoning, dialogue, document understanding, QA, captioning, or structured knowledge inference.","inclusion":"Include multimodal instruction-following and large language/foundation models, multimodal chain-of-thought, visual chat, multimodal QA and dialogue, captioning, summarization, document information extraction, visually assisted language tasks, and multimodal knowledge-graph construction, alignment, completion, or reasoning.","exclusion":"If the main output is synthesized image, audio, video, 3D, or other non-text media, assign to Multimodal media generation; if the work is only learning a retrieval embedding without language generation or explicit reasoning, assign to Cross-modal representation and retrieval.","nearest_boundary":"Cross-modal representation and retrieval","representative_topics":["zero-shot multimodal instruction tuning with cross-modal transfer and instruction sensitivity evaluation","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","zero-shot multimodal large language model training from scratch on interleaved text-image corpora"]},{"direction_index":4,"label":"D4","name":"Multimodal media generation","description":"Generative modeling programs whose primary objective is to synthesize, translate, complete, edit, or jointly model non-text media or multimodal outputs across modalities.","inclusion":"Include multimodal VAEs, GANs, diffusion, autoregressive or latent generative models for image, audio, speech, video, 3D, motion, multimodal explanation media, cross-modal synthesis, audio-to-image, text-to-image with multimodal controls, and media completion or editing.","exclusion":"If generation is primarily a language answer, caption, summary, or document output, assign to Multimodal language, reasoning, and structured knowledge systems; if instruction-following alignment is primary, assign there even when generation tools are used.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["multimodal electropalatography and audio fusion for speech generation and enhancement","survey of variational autoencoder-based multimodal deep generative models for cross-modal generation","cross-modal 3d shape generation and manipulation via shared latent spaces for sketches and color images"]},{"direction_index":5,"label":"D5","name":"Sensorimotor and domain perception systems","description":"Multimodal learning for physical-world perception, sensing, tracking, action, planning, and domain-tailored decision systems outside clinical diagnosis.","inclusion":"Include autonomous-driving sensor fusion, RGB-depth/thermal/infrared/multispectral/event visual perception, audio-visual scene or event understanding, embodied agents, robotics, manipulation, multimodal planning, geospatial or disaster sensing, industrial perception, recommendation or tracking systems when the primary objective is a domain-tailored perception/action/decision capability.","exclusion":"If the primary objective is biomedical diagnosis or prognosis, assign to Biomedical multimodal prediction; if the work proposes a general fusion principle independent of the domain objective, assign to Fusion, evaluation, and trustworthy mechanisms.","nearest_boundary":"Fusion, evaluation, and trustworthy mechanisms","representative_topics":["multi-sensor bev perception: key training and sensor integration factors beyond lifting strategies","deformable cross-modal feature aggregation for efficient multi-modal 3d object detection","cross-modal visible-thermal person re-identification via local shortest path alignment and global feature enhancement"]},{"direction_index":6,"label":"D6","name":"Biomedical multimodal prediction","description":"Integrating biomedical, clinical, imaging, omics, pathology, EHR, neuroimaging, report, biomarker, or physiological modalities for health-state prediction and clinical decision support.","inclusion":"Include multimodal diagnosis, prognosis, survival modeling, risk stratification, disease severity prediction, medical decision support, medical image-report fusion, imaging-tabular or imaging-omics fusion, and biomedical reviews when clinical prediction or decision support is the primary objective.","exclusion":"Medical image registration or segmentation without clinical prediction belongs to Sensorimotor and domain perception systems; general medical representation pretraining without a specific clinical endpoint belongs to Cross-modal representation and retrieval.","nearest_boundary":"Sensorimotor and domain perception systems","representative_topics":["multimodal stacking of chest x-ray and clinical data for covid-19 mortality risk prediction","cross-modal attention fusion for multimodal alzheimer's disease and mild cognitive impairment diagnosis","multimodal survival modeling for radiation encephalopathy prediction using weighted concordance index loss"]},{"direction_index":7,"label":"D7","name":"Fusion, evaluation, and trustworthy mechanisms","description":"Research whose primary objective is to understand, design, benchmark, interpret, secure, or make reliable the general mechanisms by which modalities interact and are evaluated.","inclusion":"Include general fusion architectures, attention or graph fusion mechanisms, modality interaction theory, uncertainty-aware or imbalance-aware fusion, adversarial robustness and defenses, poisoning/backdoor studies, privacy-preserving or federated multimodal learning, compression or efficiency analysis, interpretability, surveys, diagnostic benchmarks, and standardized toolkits when mechanism or assessment is primary.","exclusion":"If a fusion module is mainly a contribution to a sustained task or domain such as affect, driving, medicine, retrieval, generation, or language QA, assign to that objective-specific direction; if missing-modality recovery is central, assign to Robust and constrained multimodal transfer.","nearest_boundary":"Robust and constrained multimodal transfer","representative_topics":["defending multimodal fusion models against single-source adversarial perturbations via cross-modal inconsistency detection","uncertainty-aware multimodal fusion via cross-modal random network prediction for feature density estimation","progressive fusion with cross-stage residual connections for multimodal representation refinement"]}],"split":"dev"}
{"episode_id":"F118_T2025-07","field_index":118,"field":"multimodal learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Robust and constrained multimodal transfer","description":"Methods whose primary objective is to make multimodal learning work under modality absence, limited labels or pairing, domain shift, frozen-model adaptation, or teacher-student transfer constraints.","inclusion":"Include papers centered on missing or corrupted modalities, modality hallucination or imputation, incomplete multi-view learning, low-resource or source-free multimodal transfer, prompt/adaptor-style lightweight adaptation, and cross-modal distillation or transfer when the constraint itself is the main scientific objective.","exclusion":"If the main objective is a general reusable pretraining or retrieval embedding rather than constrained transfer, assign to Cross-modal representation and retrieval; if the main objective is safety evaluation or adversarial/federated deployment, assign to Fusion, evaluation, and trustworthy mechanisms.","nearest_boundary":"Cross-modal representation and retrieval","representative_topics":["multimodal inertial sensor fusion with hallucination handling via privileged information learning","efficient multimodal transformer with dual-level feature restoration for robust sentiment analysis under missing modalities","generative adversarial clustering for incomplete image-text with modality-specific subspace learning"]},{"direction_index":1,"label":"D1","name":"Human affect and social content understanding","description":"Multimodal systems whose primary objective is to infer human affective states, subjective attitudes, or socially consequential meanings from multimodal human communication and social media.","inclusion":"Include sentiment, emotion, stress, pain, humor or affective-state recognition, as well as hate, misogyny, ideology, misinformation, fact verification, persuasion, memes, sarcasm when treated as social-pragmatic content, and other social-media integrity tasks.","exclusion":"If the target is a general caption, answer, document extraction, or assistant response rather than affective or social-content judgment, assign to Multimodal language, reasoning, and structured knowledge systems.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["uncertainty-aware audiovisual fusion via calibrated and ordinal latent distribution constraints","multi-scale cooperative transformer for unaligned multimodal video sentiment analysis","multimodal aspect-category sentiment analysis with cross-modal graph-based fine-grained fusion"]},{"direction_index":2,"label":"D2","name":"Cross-modal representation and retrieval","description":"Learning reusable aligned representations or shared embedding spaces across modalities for downstream transfer, semantic matching, grounding, search, and retrieval.","inclusion":"Include contrastive, masked, generative, cross-lingual, or self-supervised multimodal pretraining when the output is a general representation; image-text, video-text, audio-text, product, person, composed-query, referring-expression, or fine-grained retrieval and grounding when embedding alignment or matching is primary.","exclusion":"If the main objective is instruction-following with a generative language backbone, assign to Multimodal language, reasoning, and structured knowledge systems; if the main objective is non-text media synthesis, assign to Multimodal media generation.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["masked encoder-decoder pre-training unifying vision and language for text recognition","unified cross-lingual cross-modal pre-training via cross-view mutual information alignment","generative vision-language pretraining via bidirectional masked prediction on shared transformer"]},{"direction_index":3,"label":"D3","name":"Multimodal language, reasoning, and structured knowledge systems","description":"Systems whose primary objective is to use multimodal evidence for language-centered outputs, assistant behavior, explicit reasoning, dialogue, document understanding, QA, captioning, or structured knowledge inference.","inclusion":"Include multimodal instruction-following and large language/foundation models, multimodal chain-of-thought, visual chat, multimodal QA and dialogue, captioning, summarization, document information extraction, visually assisted language tasks, and multimodal knowledge-graph construction, alignment, completion, or reasoning.","exclusion":"If the main output is synthesized image, audio, video, 3D, or other non-text media, assign to Multimodal media generation; if the work is only learning a retrieval embedding without language generation or explicit reasoning, assign to Cross-modal representation and retrieval.","nearest_boundary":"Cross-modal representation and retrieval","representative_topics":["zero-shot multimodal instruction tuning with cross-modal transfer and instruction sensitivity evaluation","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","zero-shot multimodal large language model training from scratch on interleaved text-image corpora"]},{"direction_index":4,"label":"D4","name":"Multimodal media generation","description":"Generative modeling programs whose primary objective is to synthesize, translate, complete, edit, or jointly model non-text media or multimodal outputs across modalities.","inclusion":"Include multimodal VAEs, GANs, diffusion, autoregressive or latent generative models for image, audio, speech, video, 3D, motion, multimodal explanation media, cross-modal synthesis, audio-to-image, text-to-image with multimodal controls, and media completion or editing.","exclusion":"If generation is primarily a language answer, caption, summary, or document output, assign to Multimodal language, reasoning, and structured knowledge systems; if instruction-following alignment is primary, assign there even when generation tools are used.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["multimodal electropalatography and audio fusion for speech generation and enhancement","survey of variational autoencoder-based multimodal deep generative models for cross-modal generation","cross-modal 3d shape generation and manipulation via shared latent spaces for sketches and color images"]},{"direction_index":5,"label":"D5","name":"Sensorimotor and domain perception systems","description":"Multimodal learning for physical-world perception, sensing, tracking, action, planning, and domain-tailored decision systems outside clinical diagnosis.","inclusion":"Include autonomous-driving sensor fusion, RGB-depth/thermal/infrared/multispectral/event visual perception, audio-visual scene or event understanding, embodied agents, robotics, manipulation, multimodal planning, geospatial or disaster sensing, industrial perception, recommendation or tracking systems when the primary objective is a domain-tailored perception/action/decision capability.","exclusion":"If the primary objective is biomedical diagnosis or prognosis, assign to Biomedical multimodal prediction; if the work proposes a general fusion principle independent of the domain objective, assign to Fusion, evaluation, and trustworthy mechanisms.","nearest_boundary":"Fusion, evaluation, and trustworthy mechanisms","representative_topics":["multi-sensor bev perception: key training and sensor integration factors beyond lifting strategies","deformable cross-modal feature aggregation for efficient multi-modal 3d object detection","cross-modal visible-thermal person re-identification via local shortest path alignment and global feature enhancement"]},{"direction_index":6,"label":"D6","name":"Biomedical multimodal prediction","description":"Integrating biomedical, clinical, imaging, omics, pathology, EHR, neuroimaging, report, biomarker, or physiological modalities for health-state prediction and clinical decision support.","inclusion":"Include multimodal diagnosis, prognosis, survival modeling, risk stratification, disease severity prediction, medical decision support, medical image-report fusion, imaging-tabular or imaging-omics fusion, and biomedical reviews when clinical prediction or decision support is the primary objective.","exclusion":"Medical image registration or segmentation without clinical prediction belongs to Sensorimotor and domain perception systems; general medical representation pretraining without a specific clinical endpoint belongs to Cross-modal representation and retrieval.","nearest_boundary":"Sensorimotor and domain perception systems","representative_topics":["multimodal stacking of chest x-ray and clinical data for covid-19 mortality risk prediction","cross-modal attention fusion for multimodal alzheimer's disease and mild cognitive impairment diagnosis","multimodal survival modeling for radiation encephalopathy prediction using weighted concordance index loss"]},{"direction_index":7,"label":"D7","name":"Fusion, evaluation, and trustworthy mechanisms","description":"Research whose primary objective is to understand, design, benchmark, interpret, secure, or make reliable the general mechanisms by which modalities interact and are evaluated.","inclusion":"Include general fusion architectures, attention or graph fusion mechanisms, modality interaction theory, uncertainty-aware or imbalance-aware fusion, adversarial robustness and defenses, poisoning/backdoor studies, privacy-preserving or federated multimodal learning, compression or efficiency analysis, interpretability, surveys, diagnostic benchmarks, and standardized toolkits when mechanism or assessment is primary.","exclusion":"If a fusion module is mainly a contribution to a sustained task or domain such as affect, driving, medicine, retrieval, generation, or language QA, assign to that objective-specific direction; if missing-modality recovery is central, assign to Robust and constrained multimodal transfer.","nearest_boundary":"Robust and constrained multimodal transfer","representative_topics":["defending multimodal fusion models against single-source adversarial perturbations via cross-modal inconsistency detection","uncertainty-aware multimodal fusion via cross-modal random network prediction for feature density estimation","progressive fusion with cross-stage residual connections for multimodal representation refinement"]}],"split":"dev"}
{"episode_id":"F118_T2026-01","field_index":118,"field":"multimodal learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Robust and constrained multimodal transfer","description":"Methods whose primary objective is to make multimodal learning work under modality absence, limited labels or pairing, domain shift, frozen-model adaptation, or teacher-student transfer constraints.","inclusion":"Include papers centered on missing or corrupted modalities, modality hallucination or imputation, incomplete multi-view learning, low-resource or source-free multimodal transfer, prompt/adaptor-style lightweight adaptation, and cross-modal distillation or transfer when the constraint itself is the main scientific objective.","exclusion":"If the main objective is a general reusable pretraining or retrieval embedding rather than constrained transfer, assign to Cross-modal representation and retrieval; if the main objective is safety evaluation or adversarial/federated deployment, assign to Fusion, evaluation, and trustworthy mechanisms.","nearest_boundary":"Cross-modal representation and retrieval","representative_topics":["multimodal inertial sensor fusion with hallucination handling via privileged information learning","efficient multimodal transformer with dual-level feature restoration for robust sentiment analysis under missing modalities","generative adversarial clustering for incomplete image-text with modality-specific subspace learning"]},{"direction_index":1,"label":"D1","name":"Human affect and social content understanding","description":"Multimodal systems whose primary objective is to infer human affective states, subjective attitudes, or socially consequential meanings from multimodal human communication and social media.","inclusion":"Include sentiment, emotion, stress, pain, humor or affective-state recognition, as well as hate, misogyny, ideology, misinformation, fact verification, persuasion, memes, sarcasm when treated as social-pragmatic content, and other social-media integrity tasks.","exclusion":"If the target is a general caption, answer, document extraction, or assistant response rather than affective or social-content judgment, assign to Multimodal language, reasoning, and structured knowledge systems.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["uncertainty-aware audiovisual fusion via calibrated and ordinal latent distribution constraints","multi-scale cooperative transformer for unaligned multimodal video sentiment analysis","multimodal aspect-category sentiment analysis with cross-modal graph-based fine-grained fusion"]},{"direction_index":2,"label":"D2","name":"Cross-modal representation and retrieval","description":"Learning reusable aligned representations or shared embedding spaces across modalities for downstream transfer, semantic matching, grounding, search, and retrieval.","inclusion":"Include contrastive, masked, generative, cross-lingual, or self-supervised multimodal pretraining when the output is a general representation; image-text, video-text, audio-text, product, person, composed-query, referring-expression, or fine-grained retrieval and grounding when embedding alignment or matching is primary.","exclusion":"If the main objective is instruction-following with a generative language backbone, assign to Multimodal language, reasoning, and structured knowledge systems; if the main objective is non-text media synthesis, assign to Multimodal media generation.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["masked encoder-decoder pre-training unifying vision and language for text recognition","unified cross-lingual cross-modal pre-training via cross-view mutual information alignment","generative vision-language pretraining via bidirectional masked prediction on shared transformer"]},{"direction_index":3,"label":"D3","name":"Multimodal language, reasoning, and structured knowledge systems","description":"Systems whose primary objective is to use multimodal evidence for language-centered outputs, assistant behavior, explicit reasoning, dialogue, document understanding, QA, captioning, or structured knowledge inference.","inclusion":"Include multimodal instruction-following and large language/foundation models, multimodal chain-of-thought, visual chat, multimodal QA and dialogue, captioning, summarization, document information extraction, visually assisted language tasks, and multimodal knowledge-graph construction, alignment, completion, or reasoning.","exclusion":"If the main output is synthesized image, audio, video, 3D, or other non-text media, assign to Multimodal media generation; if the work is only learning a retrieval embedding without language generation or explicit reasoning, assign to Cross-modal representation and retrieval.","nearest_boundary":"Cross-modal representation and retrieval","representative_topics":["zero-shot multimodal instruction tuning with cross-modal transfer and instruction sensitivity evaluation","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","zero-shot multimodal large language model training from scratch on interleaved text-image corpora"]},{"direction_index":4,"label":"D4","name":"Multimodal media generation","description":"Generative modeling programs whose primary objective is to synthesize, translate, complete, edit, or jointly model non-text media or multimodal outputs across modalities.","inclusion":"Include multimodal VAEs, GANs, diffusion, autoregressive or latent generative models for image, audio, speech, video, 3D, motion, multimodal explanation media, cross-modal synthesis, audio-to-image, text-to-image with multimodal controls, and media completion or editing.","exclusion":"If generation is primarily a language answer, caption, summary, or document output, assign to Multimodal language, reasoning, and structured knowledge systems; if instruction-following alignment is primary, assign there even when generation tools are used.","nearest_boundary":"Multimodal language, reasoning, and structured knowledge systems","representative_topics":["multimodal electropalatography and audio fusion for speech generation and enhancement","survey of variational autoencoder-based multimodal deep generative models for cross-modal generation","cross-modal 3d shape generation and manipulation via shared latent spaces for sketches and color images"]},{"direction_index":5,"label":"D5","name":"Sensorimotor and domain perception systems","description":"Multimodal learning for physical-world perception, sensing, tracking, action, planning, and domain-tailored decision systems outside clinical diagnosis.","inclusion":"Include autonomous-driving sensor fusion, RGB-depth/thermal/infrared/multispectral/event visual perception, audio-visual scene or event understanding, embodied agents, robotics, manipulation, multimodal planning, geospatial or disaster sensing, industrial perception, recommendation or tracking systems when the primary objective is a domain-tailored perception/action/decision capability.","exclusion":"If the primary objective is biomedical diagnosis or prognosis, assign to Biomedical multimodal prediction; if the work proposes a general fusion principle independent of the domain objective, assign to Fusion, evaluation, and trustworthy mechanisms.","nearest_boundary":"Fusion, evaluation, and trustworthy mechanisms","representative_topics":["multi-sensor bev perception: key training and sensor integration factors beyond lifting strategies","deformable cross-modal feature aggregation for efficient multi-modal 3d object detection","cross-modal visible-thermal person re-identification via local shortest path alignment and global feature enhancement"]},{"direction_index":6,"label":"D6","name":"Biomedical multimodal prediction","description":"Integrating biomedical, clinical, imaging, omics, pathology, EHR, neuroimaging, report, biomarker, or physiological modalities for health-state prediction and clinical decision support.","inclusion":"Include multimodal diagnosis, prognosis, survival modeling, risk stratification, disease severity prediction, medical decision support, medical image-report fusion, imaging-tabular or imaging-omics fusion, and biomedical reviews when clinical prediction or decision support is the primary objective.","exclusion":"Medical image registration or segmentation without clinical prediction belongs to Sensorimotor and domain perception systems; general medical representation pretraining without a specific clinical endpoint belongs to Cross-modal representation and retrieval.","nearest_boundary":"Sensorimotor and domain perception systems","representative_topics":["multimodal stacking of chest x-ray and clinical data for covid-19 mortality risk prediction","cross-modal attention fusion for multimodal alzheimer's disease and mild cognitive impairment diagnosis","multimodal survival modeling for radiation encephalopathy prediction using weighted concordance index loss"]},{"direction_index":7,"label":"D7","name":"Fusion, evaluation, and trustworthy mechanisms","description":"Research whose primary objective is to understand, design, benchmark, interpret, secure, or make reliable the general mechanisms by which modalities interact and are evaluated.","inclusion":"Include general fusion architectures, attention or graph fusion mechanisms, modality interaction theory, uncertainty-aware or imbalance-aware fusion, adversarial robustness and defenses, poisoning/backdoor studies, privacy-preserving or federated multimodal learning, compression or efficiency analysis, interpretability, surveys, diagnostic benchmarks, and standardized toolkits when mechanism or assessment is primary.","exclusion":"If a fusion module is mainly a contribution to a sustained task or domain such as affect, driving, medicine, retrieval, generation, or language QA, assign to that objective-specific direction; if missing-modality recovery is central, assign to Robust and constrained multimodal transfer.","nearest_boundary":"Robust and constrained multimodal transfer","representative_topics":["defending multimodal fusion models against single-source adversarial perturbations via cross-modal inconsistency detection","uncertainty-aware multimodal fusion via cross-modal random network prediction for feature density estimation","progressive fusion with cross-stage residual connections for multimodal representation refinement"]}],"split":"dev"}
{"episode_id":"F123_T2024-01","field_index":123,"field":"active learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Acquisition criteria and label-efficiency mechanisms","description":"Develop broadly reusable query-selection mechanisms that improve sample efficiency through uncertainty, information, diversity, representativeness, hybrid semi/self-supervised signals, or learned acquisition rules.","inclusion":"Include atoms whose primary objective is to design or improve general acquisition functions for static pools or batch selection, including Bayesian/information-theoretic scores, core-set/diversity/coverage criteria, uncertainty-density hybrids, SSL/contrastive/pseudo-label integration, and learned/meta/RL query policies when the learned policy optimizes the acquisition rule rather than controlling an environment.","exclusion":"Put work in the nearest alternative when the dominant objective is formal guarantees, explicit source-target shift/rare-class robustness, sequential/environmental control, modality-specific annotation structure, or socio-technical annotation/evaluation.","nearest_boundary":"D1 Theoretical guarantees and complexity","representative_topics":["inconsistency-based fusion of semi-supervised and active learning via adversarial sample selection","density-aware core-set selection for diverse informative sample acquisition in active learning","sum-product network uncertainty estimation for deep active learning acquisition functions"]},{"direction_index":1,"label":"D1","name":"Theoretical guarantees and complexity","description":"Establish mathematical foundations for active learning, including query complexity, label complexity, convergence, minimax rates, optimality, and provable guarantees for adaptive querying.","inclusion":"Include atoms whose main objective is theorem-driven analysis of active learning algorithms or models, regardless of whether the setting is classification, regression, geometry, nonparametrics, MDP exploration, noise, or teaching-style hypothesis pruning.","exclusion":"Use D0 when theory mainly motivates a practical acquisition heuristic without the formal guarantee being primary; use D5 when symbolic structures or automata are the central learned objects rather than the theory program itself.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["minimax rates for indirect active learning with control variable manipulation constraints","active exploration via markov chain policy design with convergent measurement allocation","black-box teaching hypothesis with tighter slack for convergent hypothesis-pruning active learning"]},{"direction_index":2,"label":"D2","name":"Sequential, control, and distributed querying","description":"Design active learning systems where query decisions unfold over time, are embedded in policies or physical/interactive control loops, or must operate under continual, federated, decentralized, privacy, or forgetting constraints.","inclusion":"Include stream-based and online selective sampling, non-stationary data streams, concept drift, active perception and sensing, active exploration, sequential feature acquisition, embodied or robotic data collection, active imitation/reward learning, federated active learning, and active continual learning when temporal, control, distribution, privacy, or retention constraints define the research objective.","exclusion":"Use D0 for static-pool learned acquisition rules without temporal/environmental constraints; use D6 for scientific or engineering surrogate campaigns where experiments are selected but no policy, stream, embodiment, federation, or continual-learning constraint dominates.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["online active regression with efficient label query budget under ell_p loss","stream-based active learning query strategies for regression using optimal experimental design thresholds","information-theoretic reinforcement learning for active landmark localization and volumetric exploration"]},{"direction_index":3,"label":"D3","name":"Shift-, bias-, and rare-class-resilient learning","description":"Select queries to make models robust to distribution mismatch, sampling bias, imbalance, anomalies, open-set or rare-class discovery, fairness gaps, and underrepresented groups or slices.","inclusion":"Include atoms whose primary objective is active domain adaptation, source-target transfer, domain or label-distribution shift, selective prediction under shift, sampling-bias mitigation, anomaly or rare-slice discovery, long-tail/imbalanced acquisition, fairness-sensitive querying, or robustness to out-of-distribution data.","exclusion":"Use D0 when imbalance or OOD terms are only auxiliary to a general acquisition rule; use D4 when dense visual or medical annotation structure is the dominant objective; use D7 when the bias/noise source is primarily annotator behavior or annotation workflow.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["active data discovery of unknown rare classes and slices using submodular information measures","mitigating sampling bias in risk-based active learning via semi-supervised gaussian mixture models","active learning for algorithmic bias mitigation under limited sensitive attribute annotations"]},{"direction_index":4,"label":"D4","name":"Dense and medical visual annotation","description":"Develop active learning for visual tasks where annotation units, costs, and uncertainty are shaped by dense prediction or clinical/biomedical imaging constraints.","inclusion":"Include active learning for object detection, semantic/instance segmentation, point clouds, LiDAR, aerial/document detection, region/superpixel/frame/box/mask queries, and medical or biomedical image classification/segmentation when the visual annotation structure or clinical imaging context drives the research objective.","exclusion":"Use D3 when the main objective is domain adaptation, fairness, or imbalance independent of visual annotation structure; use D7 when the contribution is primarily a benchmark rather than a new visual active-learning method.","nearest_boundary":"D3 Shift-, bias-, and rare-class-resilient learning","representative_topics":["confidence estimation for active learning annotation selection in document object detection","semi-supervised active learning for multi-label chest x-ray classification using evidential uncertainty","cold start active learning for 3d segmentation via proxy tasks and semi-supervised fine-tuning"]},{"direction_index":5,"label":"D5","name":"Structured modalities and symbolic queries","description":"Adapt active learning to data or query spaces whose structure is central: graphs, relational data, language/sequence representations, retrieval or preference data, pairwise constraints, formal specifications, automata, and other symbolic or combinatorial targets.","inclusion":"Include graph node/edge selection, GNN-oriented acquisition, Laplacian/spectral graph querying, NLP/text/transformer and speech active learning when modality-specific units drive the method, retrieval and ordinal/triplet preference annotation, correlation clustering with pairwise queries, and active learning of formal specifications or automata.","exclusion":"Use D0 for general pool-based acquisition methods merely evaluated on graphs or text; use D1 for papers whose primary contribution is abstract query-complexity theory rather than a structured modality or symbolic target.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["diversity-aware active label sampling for graph neural network training efficiency","multi-task active learning query selection for pre-trained transformer-based nlp models","active learning for few-shot claim verification via pattern exploiting training ensemble"]},{"direction_index":6,"label":"D6","name":"Scientific and engineering experiment design","description":"Use active learning to reduce expensive measurements, simulations, interventions, or experiments in physical sciences, engineering systems, materials, chemistry, causal experimentation, environmental sensing, and other scientific surrogate-modeling campaigns.","inclusion":"Include atoms where the primary objective is adaptive experiment design, surrogate modeling, chemical or materials exploration, molecular simulation, causal intervention selection, power/energy-system proxy training, thermodynamic or reliability-property prediction, environmental data quality assessment, or physical-engineering evaluation reduction.","exclusion":"Use D0 when the contribution is mainly a generic Bayesian or information-theoretic acquisition method; use D2 when the central issue is an embodied/control policy or online/federated/continual querying rather than a scientific surrogate campaign.","nearest_boundary":"D2 Sequential, control, and distributed querying","representative_topics":["active learning for surge distance estimation in multi-pump systems with minimal data","active learning for opf proxy training via bucketized sampling and adaptive acquisition","active learning for thermodynamic property prediction using marginalized graph kernel and molecular simulation"]},{"direction_index":7,"label":"D7","name":"Annotation operations and evaluation methodology","description":"Improve the practical human, organizational, privacy, annotation-quality, and evaluation infrastructure around active learning, including annotator behavior, noisy oracles, deployed systems, benchmarks, surveys, and empirical comparisons.","inclusion":"Include noisy/reliable/multiple annotator modeling, re-labeling and correction, expert deferral, human-in-the-loop annotation platforms, explanation- or workflow-assisted labeling, privacy-constrained annotation systems, active-learning-as-a-service, standardized benchmarks, surveys, empirical reproducibility studies, and comparative evaluation protocols.","exclusion":"Use the task or acquisition direction when a benchmark or annotation interface is only incidental to a new method; use D2 when federation or privacy is primarily a decentralized/continual learning algorithm rather than annotation operation or human-feedback management.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["mlops-based active learning service with automatic strategy selection and parallel pipeline execution","sample-efficient active learning of complementary predictors for expert deferral systems","benchmarking active learning algorithms for long-tailed object detection with standardized evaluation"]}],"split":"dev"}
{"episode_id":"F123_T2024-07","field_index":123,"field":"active learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Acquisition criteria and label-efficiency mechanisms","description":"Develop broadly reusable query-selection mechanisms that improve sample efficiency through uncertainty, information, diversity, representativeness, hybrid semi/self-supervised signals, or learned acquisition rules.","inclusion":"Include atoms whose primary objective is to design or improve general acquisition functions for static pools or batch selection, including Bayesian/information-theoretic scores, core-set/diversity/coverage criteria, uncertainty-density hybrids, SSL/contrastive/pseudo-label integration, and learned/meta/RL query policies when the learned policy optimizes the acquisition rule rather than controlling an environment.","exclusion":"Put work in the nearest alternative when the dominant objective is formal guarantees, explicit source-target shift/rare-class robustness, sequential/environmental control, modality-specific annotation structure, or socio-technical annotation/evaluation.","nearest_boundary":"D1 Theoretical guarantees and complexity","representative_topics":["inconsistency-based fusion of semi-supervised and active learning via adversarial sample selection","density-aware core-set selection for diverse informative sample acquisition in active learning","sum-product network uncertainty estimation for deep active learning acquisition functions"]},{"direction_index":1,"label":"D1","name":"Theoretical guarantees and complexity","description":"Establish mathematical foundations for active learning, including query complexity, label complexity, convergence, minimax rates, optimality, and provable guarantees for adaptive querying.","inclusion":"Include atoms whose main objective is theorem-driven analysis of active learning algorithms or models, regardless of whether the setting is classification, regression, geometry, nonparametrics, MDP exploration, noise, or teaching-style hypothesis pruning.","exclusion":"Use D0 when theory mainly motivates a practical acquisition heuristic without the formal guarantee being primary; use D5 when symbolic structures or automata are the central learned objects rather than the theory program itself.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["minimax rates for indirect active learning with control variable manipulation constraints","active exploration via markov chain policy design with convergent measurement allocation","black-box teaching hypothesis with tighter slack for convergent hypothesis-pruning active learning"]},{"direction_index":2,"label":"D2","name":"Sequential, control, and distributed querying","description":"Design active learning systems where query decisions unfold over time, are embedded in policies or physical/interactive control loops, or must operate under continual, federated, decentralized, privacy, or forgetting constraints.","inclusion":"Include stream-based and online selective sampling, non-stationary data streams, concept drift, active perception and sensing, active exploration, sequential feature acquisition, embodied or robotic data collection, active imitation/reward learning, federated active learning, and active continual learning when temporal, control, distribution, privacy, or retention constraints define the research objective.","exclusion":"Use D0 for static-pool learned acquisition rules without temporal/environmental constraints; use D6 for scientific or engineering surrogate campaigns where experiments are selected but no policy, stream, embodiment, federation, or continual-learning constraint dominates.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["online active regression with efficient label query budget under ell_p loss","stream-based active learning query strategies for regression using optimal experimental design thresholds","information-theoretic reinforcement learning for active landmark localization and volumetric exploration"]},{"direction_index":3,"label":"D3","name":"Shift-, bias-, and rare-class-resilient learning","description":"Select queries to make models robust to distribution mismatch, sampling bias, imbalance, anomalies, open-set or rare-class discovery, fairness gaps, and underrepresented groups or slices.","inclusion":"Include atoms whose primary objective is active domain adaptation, source-target transfer, domain or label-distribution shift, selective prediction under shift, sampling-bias mitigation, anomaly or rare-slice discovery, long-tail/imbalanced acquisition, fairness-sensitive querying, or robustness to out-of-distribution data.","exclusion":"Use D0 when imbalance or OOD terms are only auxiliary to a general acquisition rule; use D4 when dense visual or medical annotation structure is the dominant objective; use D7 when the bias/noise source is primarily annotator behavior or annotation workflow.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["active data discovery of unknown rare classes and slices using submodular information measures","mitigating sampling bias in risk-based active learning via semi-supervised gaussian mixture models","active learning for algorithmic bias mitigation under limited sensitive attribute annotations"]},{"direction_index":4,"label":"D4","name":"Dense and medical visual annotation","description":"Develop active learning for visual tasks where annotation units, costs, and uncertainty are shaped by dense prediction or clinical/biomedical imaging constraints.","inclusion":"Include active learning for object detection, semantic/instance segmentation, point clouds, LiDAR, aerial/document detection, region/superpixel/frame/box/mask queries, and medical or biomedical image classification/segmentation when the visual annotation structure or clinical imaging context drives the research objective.","exclusion":"Use D3 when the main objective is domain adaptation, fairness, or imbalance independent of visual annotation structure; use D7 when the contribution is primarily a benchmark rather than a new visual active-learning method.","nearest_boundary":"D3 Shift-, bias-, and rare-class-resilient learning","representative_topics":["confidence estimation for active learning annotation selection in document object detection","semi-supervised active learning for multi-label chest x-ray classification using evidential uncertainty","cold start active learning for 3d segmentation via proxy tasks and semi-supervised fine-tuning"]},{"direction_index":5,"label":"D5","name":"Structured modalities and symbolic queries","description":"Adapt active learning to data or query spaces whose structure is central: graphs, relational data, language/sequence representations, retrieval or preference data, pairwise constraints, formal specifications, automata, and other symbolic or combinatorial targets.","inclusion":"Include graph node/edge selection, GNN-oriented acquisition, Laplacian/spectral graph querying, NLP/text/transformer and speech active learning when modality-specific units drive the method, retrieval and ordinal/triplet preference annotation, correlation clustering with pairwise queries, and active learning of formal specifications or automata.","exclusion":"Use D0 for general pool-based acquisition methods merely evaluated on graphs or text; use D1 for papers whose primary contribution is abstract query-complexity theory rather than a structured modality or symbolic target.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["diversity-aware active label sampling for graph neural network training efficiency","multi-task active learning query selection for pre-trained transformer-based nlp models","active learning for few-shot claim verification via pattern exploiting training ensemble"]},{"direction_index":6,"label":"D6","name":"Scientific and engineering experiment design","description":"Use active learning to reduce expensive measurements, simulations, interventions, or experiments in physical sciences, engineering systems, materials, chemistry, causal experimentation, environmental sensing, and other scientific surrogate-modeling campaigns.","inclusion":"Include atoms where the primary objective is adaptive experiment design, surrogate modeling, chemical or materials exploration, molecular simulation, causal intervention selection, power/energy-system proxy training, thermodynamic or reliability-property prediction, environmental data quality assessment, or physical-engineering evaluation reduction.","exclusion":"Use D0 when the contribution is mainly a generic Bayesian or information-theoretic acquisition method; use D2 when the central issue is an embodied/control policy or online/federated/continual querying rather than a scientific surrogate campaign.","nearest_boundary":"D2 Sequential, control, and distributed querying","representative_topics":["active learning for surge distance estimation in multi-pump systems with minimal data","active learning for opf proxy training via bucketized sampling and adaptive acquisition","active learning for thermodynamic property prediction using marginalized graph kernel and molecular simulation"]},{"direction_index":7,"label":"D7","name":"Annotation operations and evaluation methodology","description":"Improve the practical human, organizational, privacy, annotation-quality, and evaluation infrastructure around active learning, including annotator behavior, noisy oracles, deployed systems, benchmarks, surveys, and empirical comparisons.","inclusion":"Include noisy/reliable/multiple annotator modeling, re-labeling and correction, expert deferral, human-in-the-loop annotation platforms, explanation- or workflow-assisted labeling, privacy-constrained annotation systems, active-learning-as-a-service, standardized benchmarks, surveys, empirical reproducibility studies, and comparative evaluation protocols.","exclusion":"Use the task or acquisition direction when a benchmark or annotation interface is only incidental to a new method; use D2 when federation or privacy is primarily a decentralized/continual learning algorithm rather than annotation operation or human-feedback management.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["mlops-based active learning service with automatic strategy selection and parallel pipeline execution","sample-efficient active learning of complementary predictors for expert deferral systems","benchmarking active learning algorithms for long-tailed object detection with standardized evaluation"]}],"split":"dev"}
{"episode_id":"F123_T2025-01","field_index":123,"field":"active learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Acquisition criteria and label-efficiency mechanisms","description":"Develop broadly reusable query-selection mechanisms that improve sample efficiency through uncertainty, information, diversity, representativeness, hybrid semi/self-supervised signals, or learned acquisition rules.","inclusion":"Include atoms whose primary objective is to design or improve general acquisition functions for static pools or batch selection, including Bayesian/information-theoretic scores, core-set/diversity/coverage criteria, uncertainty-density hybrids, SSL/contrastive/pseudo-label integration, and learned/meta/RL query policies when the learned policy optimizes the acquisition rule rather than controlling an environment.","exclusion":"Put work in the nearest alternative when the dominant objective is formal guarantees, explicit source-target shift/rare-class robustness, sequential/environmental control, modality-specific annotation structure, or socio-technical annotation/evaluation.","nearest_boundary":"D1 Theoretical guarantees and complexity","representative_topics":["inconsistency-based fusion of semi-supervised and active learning via adversarial sample selection","density-aware core-set selection for diverse informative sample acquisition in active learning","sum-product network uncertainty estimation for deep active learning acquisition functions"]},{"direction_index":1,"label":"D1","name":"Theoretical guarantees and complexity","description":"Establish mathematical foundations for active learning, including query complexity, label complexity, convergence, minimax rates, optimality, and provable guarantees for adaptive querying.","inclusion":"Include atoms whose main objective is theorem-driven analysis of active learning algorithms or models, regardless of whether the setting is classification, regression, geometry, nonparametrics, MDP exploration, noise, or teaching-style hypothesis pruning.","exclusion":"Use D0 when theory mainly motivates a practical acquisition heuristic without the formal guarantee being primary; use D5 when symbolic structures or automata are the central learned objects rather than the theory program itself.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["minimax rates for indirect active learning with control variable manipulation constraints","active exploration via markov chain policy design with convergent measurement allocation","black-box teaching hypothesis with tighter slack for convergent hypothesis-pruning active learning"]},{"direction_index":2,"label":"D2","name":"Sequential, control, and distributed querying","description":"Design active learning systems where query decisions unfold over time, are embedded in policies or physical/interactive control loops, or must operate under continual, federated, decentralized, privacy, or forgetting constraints.","inclusion":"Include stream-based and online selective sampling, non-stationary data streams, concept drift, active perception and sensing, active exploration, sequential feature acquisition, embodied or robotic data collection, active imitation/reward learning, federated active learning, and active continual learning when temporal, control, distribution, privacy, or retention constraints define the research objective.","exclusion":"Use D0 for static-pool learned acquisition rules without temporal/environmental constraints; use D6 for scientific or engineering surrogate campaigns where experiments are selected but no policy, stream, embodiment, federation, or continual-learning constraint dominates.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["online active regression with efficient label query budget under ell_p loss","stream-based active learning query strategies for regression using optimal experimental design thresholds","information-theoretic reinforcement learning for active landmark localization and volumetric exploration"]},{"direction_index":3,"label":"D3","name":"Shift-, bias-, and rare-class-resilient learning","description":"Select queries to make models robust to distribution mismatch, sampling bias, imbalance, anomalies, open-set or rare-class discovery, fairness gaps, and underrepresented groups or slices.","inclusion":"Include atoms whose primary objective is active domain adaptation, source-target transfer, domain or label-distribution shift, selective prediction under shift, sampling-bias mitigation, anomaly or rare-slice discovery, long-tail/imbalanced acquisition, fairness-sensitive querying, or robustness to out-of-distribution data.","exclusion":"Use D0 when imbalance or OOD terms are only auxiliary to a general acquisition rule; use D4 when dense visual or medical annotation structure is the dominant objective; use D7 when the bias/noise source is primarily annotator behavior or annotation workflow.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["active data discovery of unknown rare classes and slices using submodular information measures","mitigating sampling bias in risk-based active learning via semi-supervised gaussian mixture models","active learning for algorithmic bias mitigation under limited sensitive attribute annotations"]},{"direction_index":4,"label":"D4","name":"Dense and medical visual annotation","description":"Develop active learning for visual tasks where annotation units, costs, and uncertainty are shaped by dense prediction or clinical/biomedical imaging constraints.","inclusion":"Include active learning for object detection, semantic/instance segmentation, point clouds, LiDAR, aerial/document detection, region/superpixel/frame/box/mask queries, and medical or biomedical image classification/segmentation when the visual annotation structure or clinical imaging context drives the research objective.","exclusion":"Use D3 when the main objective is domain adaptation, fairness, or imbalance independent of visual annotation structure; use D7 when the contribution is primarily a benchmark rather than a new visual active-learning method.","nearest_boundary":"D3 Shift-, bias-, and rare-class-resilient learning","representative_topics":["confidence estimation for active learning annotation selection in document object detection","semi-supervised active learning for multi-label chest x-ray classification using evidential uncertainty","cold start active learning for 3d segmentation via proxy tasks and semi-supervised fine-tuning"]},{"direction_index":5,"label":"D5","name":"Structured modalities and symbolic queries","description":"Adapt active learning to data or query spaces whose structure is central: graphs, relational data, language/sequence representations, retrieval or preference data, pairwise constraints, formal specifications, automata, and other symbolic or combinatorial targets.","inclusion":"Include graph node/edge selection, GNN-oriented acquisition, Laplacian/spectral graph querying, NLP/text/transformer and speech active learning when modality-specific units drive the method, retrieval and ordinal/triplet preference annotation, correlation clustering with pairwise queries, and active learning of formal specifications or automata.","exclusion":"Use D0 for general pool-based acquisition methods merely evaluated on graphs or text; use D1 for papers whose primary contribution is abstract query-complexity theory rather than a structured modality or symbolic target.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["diversity-aware active label sampling for graph neural network training efficiency","multi-task active learning query selection for pre-trained transformer-based nlp models","active learning for few-shot claim verification via pattern exploiting training ensemble"]},{"direction_index":6,"label":"D6","name":"Scientific and engineering experiment design","description":"Use active learning to reduce expensive measurements, simulations, interventions, or experiments in physical sciences, engineering systems, materials, chemistry, causal experimentation, environmental sensing, and other scientific surrogate-modeling campaigns.","inclusion":"Include atoms where the primary objective is adaptive experiment design, surrogate modeling, chemical or materials exploration, molecular simulation, causal intervention selection, power/energy-system proxy training, thermodynamic or reliability-property prediction, environmental data quality assessment, or physical-engineering evaluation reduction.","exclusion":"Use D0 when the contribution is mainly a generic Bayesian or information-theoretic acquisition method; use D2 when the central issue is an embodied/control policy or online/federated/continual querying rather than a scientific surrogate campaign.","nearest_boundary":"D2 Sequential, control, and distributed querying","representative_topics":["active learning for surge distance estimation in multi-pump systems with minimal data","active learning for opf proxy training via bucketized sampling and adaptive acquisition","active learning for thermodynamic property prediction using marginalized graph kernel and molecular simulation"]},{"direction_index":7,"label":"D7","name":"Annotation operations and evaluation methodology","description":"Improve the practical human, organizational, privacy, annotation-quality, and evaluation infrastructure around active learning, including annotator behavior, noisy oracles, deployed systems, benchmarks, surveys, and empirical comparisons.","inclusion":"Include noisy/reliable/multiple annotator modeling, re-labeling and correction, expert deferral, human-in-the-loop annotation platforms, explanation- or workflow-assisted labeling, privacy-constrained annotation systems, active-learning-as-a-service, standardized benchmarks, surveys, empirical reproducibility studies, and comparative evaluation protocols.","exclusion":"Use the task or acquisition direction when a benchmark or annotation interface is only incidental to a new method; use D2 when federation or privacy is primarily a decentralized/continual learning algorithm rather than annotation operation or human-feedback management.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["mlops-based active learning service with automatic strategy selection and parallel pipeline execution","sample-efficient active learning of complementary predictors for expert deferral systems","benchmarking active learning algorithms for long-tailed object detection with standardized evaluation"]}],"split":"dev"}
{"episode_id":"F123_T2025-07","field_index":123,"field":"active learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Acquisition criteria and label-efficiency mechanisms","description":"Develop broadly reusable query-selection mechanisms that improve sample efficiency through uncertainty, information, diversity, representativeness, hybrid semi/self-supervised signals, or learned acquisition rules.","inclusion":"Include atoms whose primary objective is to design or improve general acquisition functions for static pools or batch selection, including Bayesian/information-theoretic scores, core-set/diversity/coverage criteria, uncertainty-density hybrids, SSL/contrastive/pseudo-label integration, and learned/meta/RL query policies when the learned policy optimizes the acquisition rule rather than controlling an environment.","exclusion":"Put work in the nearest alternative when the dominant objective is formal guarantees, explicit source-target shift/rare-class robustness, sequential/environmental control, modality-specific annotation structure, or socio-technical annotation/evaluation.","nearest_boundary":"D1 Theoretical guarantees and complexity","representative_topics":["inconsistency-based fusion of semi-supervised and active learning via adversarial sample selection","density-aware core-set selection for diverse informative sample acquisition in active learning","sum-product network uncertainty estimation for deep active learning acquisition functions"]},{"direction_index":1,"label":"D1","name":"Theoretical guarantees and complexity","description":"Establish mathematical foundations for active learning, including query complexity, label complexity, convergence, minimax rates, optimality, and provable guarantees for adaptive querying.","inclusion":"Include atoms whose main objective is theorem-driven analysis of active learning algorithms or models, regardless of whether the setting is classification, regression, geometry, nonparametrics, MDP exploration, noise, or teaching-style hypothesis pruning.","exclusion":"Use D0 when theory mainly motivates a practical acquisition heuristic without the formal guarantee being primary; use D5 when symbolic structures or automata are the central learned objects rather than the theory program itself.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["minimax rates for indirect active learning with control variable manipulation constraints","active exploration via markov chain policy design with convergent measurement allocation","black-box teaching hypothesis with tighter slack for convergent hypothesis-pruning active learning"]},{"direction_index":2,"label":"D2","name":"Sequential, control, and distributed querying","description":"Design active learning systems where query decisions unfold over time, are embedded in policies or physical/interactive control loops, or must operate under continual, federated, decentralized, privacy, or forgetting constraints.","inclusion":"Include stream-based and online selective sampling, non-stationary data streams, concept drift, active perception and sensing, active exploration, sequential feature acquisition, embodied or robotic data collection, active imitation/reward learning, federated active learning, and active continual learning when temporal, control, distribution, privacy, or retention constraints define the research objective.","exclusion":"Use D0 for static-pool learned acquisition rules without temporal/environmental constraints; use D6 for scientific or engineering surrogate campaigns where experiments are selected but no policy, stream, embodiment, federation, or continual-learning constraint dominates.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["online active regression with efficient label query budget under ell_p loss","stream-based active learning query strategies for regression using optimal experimental design thresholds","information-theoretic reinforcement learning for active landmark localization and volumetric exploration"]},{"direction_index":3,"label":"D3","name":"Shift-, bias-, and rare-class-resilient learning","description":"Select queries to make models robust to distribution mismatch, sampling bias, imbalance, anomalies, open-set or rare-class discovery, fairness gaps, and underrepresented groups or slices.","inclusion":"Include atoms whose primary objective is active domain adaptation, source-target transfer, domain or label-distribution shift, selective prediction under shift, sampling-bias mitigation, anomaly or rare-slice discovery, long-tail/imbalanced acquisition, fairness-sensitive querying, or robustness to out-of-distribution data.","exclusion":"Use D0 when imbalance or OOD terms are only auxiliary to a general acquisition rule; use D4 when dense visual or medical annotation structure is the dominant objective; use D7 when the bias/noise source is primarily annotator behavior or annotation workflow.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["active data discovery of unknown rare classes and slices using submodular information measures","mitigating sampling bias in risk-based active learning via semi-supervised gaussian mixture models","active learning for algorithmic bias mitigation under limited sensitive attribute annotations"]},{"direction_index":4,"label":"D4","name":"Dense and medical visual annotation","description":"Develop active learning for visual tasks where annotation units, costs, and uncertainty are shaped by dense prediction or clinical/biomedical imaging constraints.","inclusion":"Include active learning for object detection, semantic/instance segmentation, point clouds, LiDAR, aerial/document detection, region/superpixel/frame/box/mask queries, and medical or biomedical image classification/segmentation when the visual annotation structure or clinical imaging context drives the research objective.","exclusion":"Use D3 when the main objective is domain adaptation, fairness, or imbalance independent of visual annotation structure; use D7 when the contribution is primarily a benchmark rather than a new visual active-learning method.","nearest_boundary":"D3 Shift-, bias-, and rare-class-resilient learning","representative_topics":["confidence estimation for active learning annotation selection in document object detection","semi-supervised active learning for multi-label chest x-ray classification using evidential uncertainty","cold start active learning for 3d segmentation via proxy tasks and semi-supervised fine-tuning"]},{"direction_index":5,"label":"D5","name":"Structured modalities and symbolic queries","description":"Adapt active learning to data or query spaces whose structure is central: graphs, relational data, language/sequence representations, retrieval or preference data, pairwise constraints, formal specifications, automata, and other symbolic or combinatorial targets.","inclusion":"Include graph node/edge selection, GNN-oriented acquisition, Laplacian/spectral graph querying, NLP/text/transformer and speech active learning when modality-specific units drive the method, retrieval and ordinal/triplet preference annotation, correlation clustering with pairwise queries, and active learning of formal specifications or automata.","exclusion":"Use D0 for general pool-based acquisition methods merely evaluated on graphs or text; use D1 for papers whose primary contribution is abstract query-complexity theory rather than a structured modality or symbolic target.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["diversity-aware active label sampling for graph neural network training efficiency","multi-task active learning query selection for pre-trained transformer-based nlp models","active learning for few-shot claim verification via pattern exploiting training ensemble"]},{"direction_index":6,"label":"D6","name":"Scientific and engineering experiment design","description":"Use active learning to reduce expensive measurements, simulations, interventions, or experiments in physical sciences, engineering systems, materials, chemistry, causal experimentation, environmental sensing, and other scientific surrogate-modeling campaigns.","inclusion":"Include atoms where the primary objective is adaptive experiment design, surrogate modeling, chemical or materials exploration, molecular simulation, causal intervention selection, power/energy-system proxy training, thermodynamic or reliability-property prediction, environmental data quality assessment, or physical-engineering evaluation reduction.","exclusion":"Use D0 when the contribution is mainly a generic Bayesian or information-theoretic acquisition method; use D2 when the central issue is an embodied/control policy or online/federated/continual querying rather than a scientific surrogate campaign.","nearest_boundary":"D2 Sequential, control, and distributed querying","representative_topics":["active learning for surge distance estimation in multi-pump systems with minimal data","active learning for opf proxy training via bucketized sampling and adaptive acquisition","active learning for thermodynamic property prediction using marginalized graph kernel and molecular simulation"]},{"direction_index":7,"label":"D7","name":"Annotation operations and evaluation methodology","description":"Improve the practical human, organizational, privacy, annotation-quality, and evaluation infrastructure around active learning, including annotator behavior, noisy oracles, deployed systems, benchmarks, surveys, and empirical comparisons.","inclusion":"Include noisy/reliable/multiple annotator modeling, re-labeling and correction, expert deferral, human-in-the-loop annotation platforms, explanation- or workflow-assisted labeling, privacy-constrained annotation systems, active-learning-as-a-service, standardized benchmarks, surveys, empirical reproducibility studies, and comparative evaluation protocols.","exclusion":"Use the task or acquisition direction when a benchmark or annotation interface is only incidental to a new method; use D2 when federation or privacy is primarily a decentralized/continual learning algorithm rather than annotation operation or human-feedback management.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["mlops-based active learning service with automatic strategy selection and parallel pipeline execution","sample-efficient active learning of complementary predictors for expert deferral systems","benchmarking active learning algorithms for long-tailed object detection with standardized evaluation"]}],"split":"dev"}
{"episode_id":"F123_T2026-01","field_index":123,"field":"active learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Acquisition criteria and label-efficiency mechanisms","description":"Develop broadly reusable query-selection mechanisms that improve sample efficiency through uncertainty, information, diversity, representativeness, hybrid semi/self-supervised signals, or learned acquisition rules.","inclusion":"Include atoms whose primary objective is to design or improve general acquisition functions for static pools or batch selection, including Bayesian/information-theoretic scores, core-set/diversity/coverage criteria, uncertainty-density hybrids, SSL/contrastive/pseudo-label integration, and learned/meta/RL query policies when the learned policy optimizes the acquisition rule rather than controlling an environment.","exclusion":"Put work in the nearest alternative when the dominant objective is formal guarantees, explicit source-target shift/rare-class robustness, sequential/environmental control, modality-specific annotation structure, or socio-technical annotation/evaluation.","nearest_boundary":"D1 Theoretical guarantees and complexity","representative_topics":["inconsistency-based fusion of semi-supervised and active learning via adversarial sample selection","density-aware core-set selection for diverse informative sample acquisition in active learning","sum-product network uncertainty estimation for deep active learning acquisition functions"]},{"direction_index":1,"label":"D1","name":"Theoretical guarantees and complexity","description":"Establish mathematical foundations for active learning, including query complexity, label complexity, convergence, minimax rates, optimality, and provable guarantees for adaptive querying.","inclusion":"Include atoms whose main objective is theorem-driven analysis of active learning algorithms or models, regardless of whether the setting is classification, regression, geometry, nonparametrics, MDP exploration, noise, or teaching-style hypothesis pruning.","exclusion":"Use D0 when theory mainly motivates a practical acquisition heuristic without the formal guarantee being primary; use D5 when symbolic structures or automata are the central learned objects rather than the theory program itself.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["minimax rates for indirect active learning with control variable manipulation constraints","active exploration via markov chain policy design with convergent measurement allocation","black-box teaching hypothesis with tighter slack for convergent hypothesis-pruning active learning"]},{"direction_index":2,"label":"D2","name":"Sequential, control, and distributed querying","description":"Design active learning systems where query decisions unfold over time, are embedded in policies or physical/interactive control loops, or must operate under continual, federated, decentralized, privacy, or forgetting constraints.","inclusion":"Include stream-based and online selective sampling, non-stationary data streams, concept drift, active perception and sensing, active exploration, sequential feature acquisition, embodied or robotic data collection, active imitation/reward learning, federated active learning, and active continual learning when temporal, control, distribution, privacy, or retention constraints define the research objective.","exclusion":"Use D0 for static-pool learned acquisition rules without temporal/environmental constraints; use D6 for scientific or engineering surrogate campaigns where experiments are selected but no policy, stream, embodiment, federation, or continual-learning constraint dominates.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["online active regression with efficient label query budget under ell_p loss","stream-based active learning query strategies for regression using optimal experimental design thresholds","information-theoretic reinforcement learning for active landmark localization and volumetric exploration"]},{"direction_index":3,"label":"D3","name":"Shift-, bias-, and rare-class-resilient learning","description":"Select queries to make models robust to distribution mismatch, sampling bias, imbalance, anomalies, open-set or rare-class discovery, fairness gaps, and underrepresented groups or slices.","inclusion":"Include atoms whose primary objective is active domain adaptation, source-target transfer, domain or label-distribution shift, selective prediction under shift, sampling-bias mitigation, anomaly or rare-slice discovery, long-tail/imbalanced acquisition, fairness-sensitive querying, or robustness to out-of-distribution data.","exclusion":"Use D0 when imbalance or OOD terms are only auxiliary to a general acquisition rule; use D4 when dense visual or medical annotation structure is the dominant objective; use D7 when the bias/noise source is primarily annotator behavior or annotation workflow.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["active data discovery of unknown rare classes and slices using submodular information measures","mitigating sampling bias in risk-based active learning via semi-supervised gaussian mixture models","active learning for algorithmic bias mitigation under limited sensitive attribute annotations"]},{"direction_index":4,"label":"D4","name":"Dense and medical visual annotation","description":"Develop active learning for visual tasks where annotation units, costs, and uncertainty are shaped by dense prediction or clinical/biomedical imaging constraints.","inclusion":"Include active learning for object detection, semantic/instance segmentation, point clouds, LiDAR, aerial/document detection, region/superpixel/frame/box/mask queries, and medical or biomedical image classification/segmentation when the visual annotation structure or clinical imaging context drives the research objective.","exclusion":"Use D3 when the main objective is domain adaptation, fairness, or imbalance independent of visual annotation structure; use D7 when the contribution is primarily a benchmark rather than a new visual active-learning method.","nearest_boundary":"D3 Shift-, bias-, and rare-class-resilient learning","representative_topics":["confidence estimation for active learning annotation selection in document object detection","semi-supervised active learning for multi-label chest x-ray classification using evidential uncertainty","cold start active learning for 3d segmentation via proxy tasks and semi-supervised fine-tuning"]},{"direction_index":5,"label":"D5","name":"Structured modalities and symbolic queries","description":"Adapt active learning to data or query spaces whose structure is central: graphs, relational data, language/sequence representations, retrieval or preference data, pairwise constraints, formal specifications, automata, and other symbolic or combinatorial targets.","inclusion":"Include graph node/edge selection, GNN-oriented acquisition, Laplacian/spectral graph querying, NLP/text/transformer and speech active learning when modality-specific units drive the method, retrieval and ordinal/triplet preference annotation, correlation clustering with pairwise queries, and active learning of formal specifications or automata.","exclusion":"Use D0 for general pool-based acquisition methods merely evaluated on graphs or text; use D1 for papers whose primary contribution is abstract query-complexity theory rather than a structured modality or symbolic target.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["diversity-aware active label sampling for graph neural network training efficiency","multi-task active learning query selection for pre-trained transformer-based nlp models","active learning for few-shot claim verification via pattern exploiting training ensemble"]},{"direction_index":6,"label":"D6","name":"Scientific and engineering experiment design","description":"Use active learning to reduce expensive measurements, simulations, interventions, or experiments in physical sciences, engineering systems, materials, chemistry, causal experimentation, environmental sensing, and other scientific surrogate-modeling campaigns.","inclusion":"Include atoms where the primary objective is adaptive experiment design, surrogate modeling, chemical or materials exploration, molecular simulation, causal intervention selection, power/energy-system proxy training, thermodynamic or reliability-property prediction, environmental data quality assessment, or physical-engineering evaluation reduction.","exclusion":"Use D0 when the contribution is mainly a generic Bayesian or information-theoretic acquisition method; use D2 when the central issue is an embodied/control policy or online/federated/continual querying rather than a scientific surrogate campaign.","nearest_boundary":"D2 Sequential, control, and distributed querying","representative_topics":["active learning for surge distance estimation in multi-pump systems with minimal data","active learning for opf proxy training via bucketized sampling and adaptive acquisition","active learning for thermodynamic property prediction using marginalized graph kernel and molecular simulation"]},{"direction_index":7,"label":"D7","name":"Annotation operations and evaluation methodology","description":"Improve the practical human, organizational, privacy, annotation-quality, and evaluation infrastructure around active learning, including annotator behavior, noisy oracles, deployed systems, benchmarks, surveys, and empirical comparisons.","inclusion":"Include noisy/reliable/multiple annotator modeling, re-labeling and correction, expert deferral, human-in-the-loop annotation platforms, explanation- or workflow-assisted labeling, privacy-constrained annotation systems, active-learning-as-a-service, standardized benchmarks, surveys, empirical reproducibility studies, and comparative evaluation protocols.","exclusion":"Use the task or acquisition direction when a benchmark or annotation interface is only incidental to a new method; use D2 when federation or privacy is primarily a decentralized/continual learning algorithm rather than annotation operation or human-feedback management.","nearest_boundary":"D0 Acquisition criteria and label-efficiency mechanisms","representative_topics":["mlops-based active learning service with automatic strategy selection and parallel pipeline execution","sample-efficient active learning of complementary predictors for expert deferral systems","benchmarking active learning algorithms for long-tailed object detection with standardized evaluation"]}],"split":"dev"}
{"episode_id":"F124_T2024-01","field_index":124,"field":"optimization theory","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Stochastic and private first-order guarantees","description":"Convergence, finite-sample, generalization, stability, robustness, and privacy guarantees for stochastic first-order optimization procedures.","inclusion":"Include atoms whose primary objective is proving optimization-error, statistical-generalization, stability, high-probability, robustness-to-noise, or differential-privacy guarantees for SGD, stochastic approximation, DP-SGD, stochastic convex/nonconvex/minimax optimization, or close first-order variants.","exclusion":"If the main objective is optimizer-induced neural solution geometry or neural training regimes, use D1; if adaptivity, preconditioning, or learned optimizer mechanisms are primary, use D2; if communication/federation is primary, use D4.","nearest_boundary":"D1 Neural optimization dynamics and implicit bias","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","algorithmic stability bounds for heavy-tailed stochastic gradient descent generalization","sgd-dependent generalization error bounds via parameter norm dynamics and rademacher complexity"]},{"direction_index":1,"label":"D1","name":"Neural optimization dynamics and implicit bias","description":"Understanding how gradient-based optimization behaves in neural or overparameterized models, including implicit regularization, representation/solution selection, sharpness, flatness, stability regimes, edge-of-stability, and mean-field or architecture-specific training dynamics.","inclusion":"Include atoms whose primary objective is explaining optimizer-induced neural-network trajectories, selected solutions, margins/ranks/sparsity, landscape geometry, sharpness/flatness, phase regimes, overparameterization effects, NTK/mean-field limits, or architecture-specific convergence phenomena.","exclusion":"If the main target is a general SGD convergence or generalization bound not organized around neural training dynamics, use D0; if the contribution is generic nonconvex stationarity or saddle escape rather than neural-specific dynamics, use D3.","nearest_boundary":"D0 Stochastic and private first-order guarantees","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","pde-based theoretical framework for edge of stability and numerical instability in deep network training","gradient descent convergence beyond edge of stability via two-step update and sharpness minimisation analysis"]},{"direction_index":2,"label":"D2","name":"Convex, constrained, and adaptive algorithms","description":"Design and analysis of deterministic or mechanism-focused optimization algorithms for convex, constrained, composite, projection-free, accelerated, proximal, adaptive, momentum, preconditioned, and learned-optimizer settings.","inclusion":"Include atoms whose primary objective is improving algorithmic mechanisms, rates, oracle complexity, conditioning, acceleration, projection avoidance, proximal/cutting-plane/Frank-Wolfe behavior, adaptive learning-rate rules, Adam/AdaGrad-style methods, curvature approximations, or learned optimizer designs for standard optimization problems.","exclusion":"If explicit Hessian/Newton machinery or algebraic relaxations are tied to nonconvex or discrete structure, use D3; if stochastic noise, privacy, or statistical stability is the organizing goal rather than the algorithmic mechanism, use D0.","nearest_boundary":"D3 Nonconvex, second-order, and discrete structure","representative_topics":["accelerated proximal gradient methods for convex optimization with locally lipschitz continuous gradient","relaxed error criterion for efficient catalyst-based convex optimization with logarithmic complexity reduction","lipschitz-free convergence analysis of adam with hyperparameters close to one"]},{"direction_index":3,"label":"D3","name":"Nonconvex, second-order, and discrete structure","description":"Algorithms, complexity guarantees, relaxations, approximation methods, and landscape analyses for nonconvex optimization, second-order stationarity, combinatorial/submodular/discrete structure, and algebraic or semidefinite relaxations.","inclusion":"Include atoms whose primary objective is proving nonconvex convergence or stationarity, escaping saddles, certifying global structure, exploiting matrix-factorization or constrained nonconvex landscapes, developing Newton/trust-region/cubic/quasi-Newton methods, or obtaining approximation/relaxation guarantees for submodular, combinatorial, polynomial, QCQP, SDP, or related structured problems.","exclusion":"If the nonconvex behavior is primarily neural training dynamics or implicit bias, use D1; if the problem is a simultaneous game, minimax problem, or variational inequality, use D5; if the main issue is zeroth-order feedback, use D6.","nearest_boundary":"D1 Neural optimization dynamics and implicit bias","representative_topics":["preconditioned gradient descent for overparameterized nonconvex burer-monteiro factorization with global optimality certification","second-order stationary point methods for nonconvex conic optimization via newton-conjugate gradient barrier","parabolic relaxation of non-convex qcqp via boundary containment and sdp equivalence"]},{"direction_index":4,"label":"D4","name":"Distributed and federated optimization","description":"Optimization theory and algorithms for multi-agent, decentralized, federated, asynchronous, compressed, heterogeneous, or communication-constrained computation.","inclusion":"Include atoms whose primary objective is handling communication efficiency, consensus, compression/quantization, client sampling, delays, local updates, data heterogeneity, gradient tracking, decentralized topologies, or federated aggregation as the central optimization constraint.","exclusion":"If distributed computation is incidental to a bilevel, minimax, manifold, or privacy objective, use D5, D7, or D0 respectively according to the primary scientific objective.","nearest_boundary":"D5 Nested, minimax, and equilibrium optimization","representative_topics":["asynchronous decentralized optimization with network-level variance reduction and asymmetric communication","federated proximal optimization convergence theory with dissimilarity-invariant non-smooth guarantees","asynchronous federated optimization with stochastic aggregation for heterogeneous client delays"]},{"direction_index":5,"label":"D5","name":"Nested, minimax, and equilibrium optimization","description":"Optimization with explicitly nested objectives or simultaneous adversarial/equilibrium structure, including bilevel, multilevel compositional, minimax, saddle-point, game-theoretic, monotone-inclusion, and variational-inequality programs.","inclusion":"Include atoms whose primary objective is algorithm design or convergence theory for lower-level solution maps, hypergradients, compositional expectations, coupled/nested objectives, gradient descent-ascent, extragradient methods, last-iterate behavior, Nash or mixed equilibria, saddle-point systems, monotone inclusions, or variational inequalities.","exclusion":"If the objective is ordinary stochastic optimization without nesting or equilibrium structure, use D0; if the main constraint is federation/communication rather than the game or nesting itself, use D4.","nearest_boundary":"D4 Distributed and federated optimization","representative_topics":["stochastic single-loop algorithms for multi-block min-max bilevel optimization","conditional gradient methods for bilevel optimization with convex lower-level problems and optimal iteration complexity","clipped stochastic extragradient methods for variational inequalities with heavy-tailed noise and unbounded domains"]},{"direction_index":6,"label":"D6","name":"Sequential, black-box, and zeroth-order optimization","description":"Query-efficient optimization under limited, sequential, delayed, noisy, bandit, preference, comparison, function-value, or surrogate-based feedback.","inclusion":"Include atoms whose primary objective is global Bayesian or Lipschitz black-box search, acquisition/surrogate design, online convex regret minimization, bandit or preference feedback, convex chasing, derivative-free local/stationary optimization, zeroth-order gradient estimation, query reuse, safe black-box optimization, or related sequential feedback efficiency.","exclusion":"If exact-gradient deterministic convex mechanism design is primary, use D2; if the problem is a distributed system constraint, use D4; if function-value feedback is used mainly inside a nonconvex structural guarantee, use D3.","nearest_boundary":"D2 Convex, constrained, and adaptive algorithms","representative_topics":["hybrid mcts and gaussian process surrogates for mixed-variable bayesian optimization","minimax optimal global optimization via query rules for multivariate lipschitz functions","structured stochastic zeroth-order descent with orthogonal direction sampling and polyak-łojasiewicz convergence"]},{"direction_index":7,"label":"D7","name":"Geometric and distributional-domain optimization","description":"Optimization over non-Euclidean, manifold, measure, probability-distribution, transport, sampling-dynamics, policy-gradient, and other intrinsically structured domains where the domain geometry or distributional dynamics drive the theory.","inclusion":"Include atoms whose primary objective is adapting optimization to Riemannian or geodesic geometry, measure spaces, Wasserstein/optimal-transport structure, particle or variational transport methods, Langevin/Stein/sampling dynamics, natural-gradient or policy-gradient geometry, distributional optimization, or solution maps tied to structured domains.","exclusion":"If optimal transport or particles are used mainly for game equilibrium computation, use D5; if the main goal is standard convex oracle complexity rather than domain geometry, use D2; if decentralization is the organizing constraint, use D4.","nearest_boundary":"D5 Nested, minimax, and equilibrium optimization","representative_topics":["stochastic approximation convergence on riemannian manifolds via fermi coordinate dynamical systems","mirror descent convergence rates for convex optimization over measure spaces with non-smooth functionals","constrained optimal transport via saddle point methods with soft density and momentum constraints"]}],"split":"dev"}
{"episode_id":"F124_T2024-07","field_index":124,"field":"optimization theory","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Stochastic and private first-order guarantees","description":"Convergence, finite-sample, generalization, stability, robustness, and privacy guarantees for stochastic first-order optimization procedures.","inclusion":"Include atoms whose primary objective is proving optimization-error, statistical-generalization, stability, high-probability, robustness-to-noise, or differential-privacy guarantees for SGD, stochastic approximation, DP-SGD, stochastic convex/nonconvex/minimax optimization, or close first-order variants.","exclusion":"If the main objective is optimizer-induced neural solution geometry or neural training regimes, use D1; if adaptivity, preconditioning, or learned optimizer mechanisms are primary, use D2; if communication/federation is primary, use D4.","nearest_boundary":"D1 Neural optimization dynamics and implicit bias","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","algorithmic stability bounds for heavy-tailed stochastic gradient descent generalization","sgd-dependent generalization error bounds via parameter norm dynamics and rademacher complexity"]},{"direction_index":1,"label":"D1","name":"Neural optimization dynamics and implicit bias","description":"Understanding how gradient-based optimization behaves in neural or overparameterized models, including implicit regularization, representation/solution selection, sharpness, flatness, stability regimes, edge-of-stability, and mean-field or architecture-specific training dynamics.","inclusion":"Include atoms whose primary objective is explaining optimizer-induced neural-network trajectories, selected solutions, margins/ranks/sparsity, landscape geometry, sharpness/flatness, phase regimes, overparameterization effects, NTK/mean-field limits, or architecture-specific convergence phenomena.","exclusion":"If the main target is a general SGD convergence or generalization bound not organized around neural training dynamics, use D0; if the contribution is generic nonconvex stationarity or saddle escape rather than neural-specific dynamics, use D3.","nearest_boundary":"D0 Stochastic and private first-order guarantees","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","pde-based theoretical framework for edge of stability and numerical instability in deep network training","gradient descent convergence beyond edge of stability via two-step update and sharpness minimisation analysis"]},{"direction_index":2,"label":"D2","name":"Convex, constrained, and adaptive algorithms","description":"Design and analysis of deterministic or mechanism-focused optimization algorithms for convex, constrained, composite, projection-free, accelerated, proximal, adaptive, momentum, preconditioned, and learned-optimizer settings.","inclusion":"Include atoms whose primary objective is improving algorithmic mechanisms, rates, oracle complexity, conditioning, acceleration, projection avoidance, proximal/cutting-plane/Frank-Wolfe behavior, adaptive learning-rate rules, Adam/AdaGrad-style methods, curvature approximations, or learned optimizer designs for standard optimization problems.","exclusion":"If explicit Hessian/Newton machinery or algebraic relaxations are tied to nonconvex or discrete structure, use D3; if stochastic noise, privacy, or statistical stability is the organizing goal rather than the algorithmic mechanism, use D0.","nearest_boundary":"D3 Nonconvex, second-order, and discrete structure","representative_topics":["accelerated proximal gradient methods for convex optimization with locally lipschitz continuous gradient","relaxed error criterion for efficient catalyst-based convex optimization with logarithmic complexity reduction","lipschitz-free convergence analysis of adam with hyperparameters close to one"]},{"direction_index":3,"label":"D3","name":"Nonconvex, second-order, and discrete structure","description":"Algorithms, complexity guarantees, relaxations, approximation methods, and landscape analyses for nonconvex optimization, second-order stationarity, combinatorial/submodular/discrete structure, and algebraic or semidefinite relaxations.","inclusion":"Include atoms whose primary objective is proving nonconvex convergence or stationarity, escaping saddles, certifying global structure, exploiting matrix-factorization or constrained nonconvex landscapes, developing Newton/trust-region/cubic/quasi-Newton methods, or obtaining approximation/relaxation guarantees for submodular, combinatorial, polynomial, QCQP, SDP, or related structured problems.","exclusion":"If the nonconvex behavior is primarily neural training dynamics or implicit bias, use D1; if the problem is a simultaneous game, minimax problem, or variational inequality, use D5; if the main issue is zeroth-order feedback, use D6.","nearest_boundary":"D1 Neural optimization dynamics and implicit bias","representative_topics":["preconditioned gradient descent for overparameterized nonconvex burer-monteiro factorization with global optimality certification","second-order stationary point methods for nonconvex conic optimization via newton-conjugate gradient barrier","parabolic relaxation of non-convex qcqp via boundary containment and sdp equivalence"]},{"direction_index":4,"label":"D4","name":"Distributed and federated optimization","description":"Optimization theory and algorithms for multi-agent, decentralized, federated, asynchronous, compressed, heterogeneous, or communication-constrained computation.","inclusion":"Include atoms whose primary objective is handling communication efficiency, consensus, compression/quantization, client sampling, delays, local updates, data heterogeneity, gradient tracking, decentralized topologies, or federated aggregation as the central optimization constraint.","exclusion":"If distributed computation is incidental to a bilevel, minimax, manifold, or privacy objective, use D5, D7, or D0 respectively according to the primary scientific objective.","nearest_boundary":"D5 Nested, minimax, and equilibrium optimization","representative_topics":["asynchronous decentralized optimization with network-level variance reduction and asymmetric communication","federated proximal optimization convergence theory with dissimilarity-invariant non-smooth guarantees","asynchronous federated optimization with stochastic aggregation for heterogeneous client delays"]},{"direction_index":5,"label":"D5","name":"Nested, minimax, and equilibrium optimization","description":"Optimization with explicitly nested objectives or simultaneous adversarial/equilibrium structure, including bilevel, multilevel compositional, minimax, saddle-point, game-theoretic, monotone-inclusion, and variational-inequality programs.","inclusion":"Include atoms whose primary objective is algorithm design or convergence theory for lower-level solution maps, hypergradients, compositional expectations, coupled/nested objectives, gradient descent-ascent, extragradient methods, last-iterate behavior, Nash or mixed equilibria, saddle-point systems, monotone inclusions, or variational inequalities.","exclusion":"If the objective is ordinary stochastic optimization without nesting or equilibrium structure, use D0; if the main constraint is federation/communication rather than the game or nesting itself, use D4.","nearest_boundary":"D4 Distributed and federated optimization","representative_topics":["stochastic single-loop algorithms for multi-block min-max bilevel optimization","conditional gradient methods for bilevel optimization with convex lower-level problems and optimal iteration complexity","clipped stochastic extragradient methods for variational inequalities with heavy-tailed noise and unbounded domains"]},{"direction_index":6,"label":"D6","name":"Sequential, black-box, and zeroth-order optimization","description":"Query-efficient optimization under limited, sequential, delayed, noisy, bandit, preference, comparison, function-value, or surrogate-based feedback.","inclusion":"Include atoms whose primary objective is global Bayesian or Lipschitz black-box search, acquisition/surrogate design, online convex regret minimization, bandit or preference feedback, convex chasing, derivative-free local/stationary optimization, zeroth-order gradient estimation, query reuse, safe black-box optimization, or related sequential feedback efficiency.","exclusion":"If exact-gradient deterministic convex mechanism design is primary, use D2; if the problem is a distributed system constraint, use D4; if function-value feedback is used mainly inside a nonconvex structural guarantee, use D3.","nearest_boundary":"D2 Convex, constrained, and adaptive algorithms","representative_topics":["hybrid mcts and gaussian process surrogates for mixed-variable bayesian optimization","minimax optimal global optimization via query rules for multivariate lipschitz functions","structured stochastic zeroth-order descent with orthogonal direction sampling and polyak-łojasiewicz convergence"]},{"direction_index":7,"label":"D7","name":"Geometric and distributional-domain optimization","description":"Optimization over non-Euclidean, manifold, measure, probability-distribution, transport, sampling-dynamics, policy-gradient, and other intrinsically structured domains where the domain geometry or distributional dynamics drive the theory.","inclusion":"Include atoms whose primary objective is adapting optimization to Riemannian or geodesic geometry, measure spaces, Wasserstein/optimal-transport structure, particle or variational transport methods, Langevin/Stein/sampling dynamics, natural-gradient or policy-gradient geometry, distributional optimization, or solution maps tied to structured domains.","exclusion":"If optimal transport or particles are used mainly for game equilibrium computation, use D5; if the main goal is standard convex oracle complexity rather than domain geometry, use D2; if decentralization is the organizing constraint, use D4.","nearest_boundary":"D5 Nested, minimax, and equilibrium optimization","representative_topics":["stochastic approximation convergence on riemannian manifolds via fermi coordinate dynamical systems","mirror descent convergence rates for convex optimization over measure spaces with non-smooth functionals","constrained optimal transport via saddle point methods with soft density and momentum constraints"]}],"split":"dev"}
{"episode_id":"F124_T2025-01","field_index":124,"field":"optimization theory","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Stochastic and private first-order guarantees","description":"Convergence, finite-sample, generalization, stability, robustness, and privacy guarantees for stochastic first-order optimization procedures.","inclusion":"Include atoms whose primary objective is proving optimization-error, statistical-generalization, stability, high-probability, robustness-to-noise, or differential-privacy guarantees for SGD, stochastic approximation, DP-SGD, stochastic convex/nonconvex/minimax optimization, or close first-order variants.","exclusion":"If the main objective is optimizer-induced neural solution geometry or neural training regimes, use D1; if adaptivity, preconditioning, or learned optimizer mechanisms are primary, use D2; if communication/federation is primary, use D4.","nearest_boundary":"D1 Neural optimization dynamics and implicit bias","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","algorithmic stability bounds for heavy-tailed stochastic gradient descent generalization","sgd-dependent generalization error bounds via parameter norm dynamics and rademacher complexity"]},{"direction_index":1,"label":"D1","name":"Neural optimization dynamics and implicit bias","description":"Understanding how gradient-based optimization behaves in neural or overparameterized models, including implicit regularization, representation/solution selection, sharpness, flatness, stability regimes, edge-of-stability, and mean-field or architecture-specific training dynamics.","inclusion":"Include atoms whose primary objective is explaining optimizer-induced neural-network trajectories, selected solutions, margins/ranks/sparsity, landscape geometry, sharpness/flatness, phase regimes, overparameterization effects, NTK/mean-field limits, or architecture-specific convergence phenomena.","exclusion":"If the main target is a general SGD convergence or generalization bound not organized around neural training dynamics, use D0; if the contribution is generic nonconvex stationarity or saddle escape rather than neural-specific dynamics, use D3.","nearest_boundary":"D0 Stochastic and private first-order guarantees","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","pde-based theoretical framework for edge of stability and numerical instability in deep network training","gradient descent convergence beyond edge of stability via two-step update and sharpness minimisation analysis"]},{"direction_index":2,"label":"D2","name":"Convex, constrained, and adaptive algorithms","description":"Design and analysis of deterministic or mechanism-focused optimization algorithms for convex, constrained, composite, projection-free, accelerated, proximal, adaptive, momentum, preconditioned, and learned-optimizer settings.","inclusion":"Include atoms whose primary objective is improving algorithmic mechanisms, rates, oracle complexity, conditioning, acceleration, projection avoidance, proximal/cutting-plane/Frank-Wolfe behavior, adaptive learning-rate rules, Adam/AdaGrad-style methods, curvature approximations, or learned optimizer designs for standard optimization problems.","exclusion":"If explicit Hessian/Newton machinery or algebraic relaxations are tied to nonconvex or discrete structure, use D3; if stochastic noise, privacy, or statistical stability is the organizing goal rather than the algorithmic mechanism, use D0.","nearest_boundary":"D3 Nonconvex, second-order, and discrete structure","representative_topics":["accelerated proximal gradient methods for convex optimization with locally lipschitz continuous gradient","relaxed error criterion for efficient catalyst-based convex optimization with logarithmic complexity reduction","lipschitz-free convergence analysis of adam with hyperparameters close to one"]},{"direction_index":3,"label":"D3","name":"Nonconvex, second-order, and discrete structure","description":"Algorithms, complexity guarantees, relaxations, approximation methods, and landscape analyses for nonconvex optimization, second-order stationarity, combinatorial/submodular/discrete structure, and algebraic or semidefinite relaxations.","inclusion":"Include atoms whose primary objective is proving nonconvex convergence or stationarity, escaping saddles, certifying global structure, exploiting matrix-factorization or constrained nonconvex landscapes, developing Newton/trust-region/cubic/quasi-Newton methods, or obtaining approximation/relaxation guarantees for submodular, combinatorial, polynomial, QCQP, SDP, or related structured problems.","exclusion":"If the nonconvex behavior is primarily neural training dynamics or implicit bias, use D1; if the problem is a simultaneous game, minimax problem, or variational inequality, use D5; if the main issue is zeroth-order feedback, use D6.","nearest_boundary":"D1 Neural optimization dynamics and implicit bias","representative_topics":["preconditioned gradient descent for overparameterized nonconvex burer-monteiro factorization with global optimality certification","second-order stationary point methods for nonconvex conic optimization via newton-conjugate gradient barrier","parabolic relaxation of non-convex qcqp via boundary containment and sdp equivalence"]},{"direction_index":4,"label":"D4","name":"Distributed and federated optimization","description":"Optimization theory and algorithms for multi-agent, decentralized, federated, asynchronous, compressed, heterogeneous, or communication-constrained computation.","inclusion":"Include atoms whose primary objective is handling communication efficiency, consensus, compression/quantization, client sampling, delays, local updates, data heterogeneity, gradient tracking, decentralized topologies, or federated aggregation as the central optimization constraint.","exclusion":"If distributed computation is incidental to a bilevel, minimax, manifold, or privacy objective, use D5, D7, or D0 respectively according to the primary scientific objective.","nearest_boundary":"D5 Nested, minimax, and equilibrium optimization","representative_topics":["asynchronous decentralized optimization with network-level variance reduction and asymmetric communication","federated proximal optimization convergence theory with dissimilarity-invariant non-smooth guarantees","asynchronous federated optimization with stochastic aggregation for heterogeneous client delays"]},{"direction_index":5,"label":"D5","name":"Nested, minimax, and equilibrium optimization","description":"Optimization with explicitly nested objectives or simultaneous adversarial/equilibrium structure, including bilevel, multilevel compositional, minimax, saddle-point, game-theoretic, monotone-inclusion, and variational-inequality programs.","inclusion":"Include atoms whose primary objective is algorithm design or convergence theory for lower-level solution maps, hypergradients, compositional expectations, coupled/nested objectives, gradient descent-ascent, extragradient methods, last-iterate behavior, Nash or mixed equilibria, saddle-point systems, monotone inclusions, or variational inequalities.","exclusion":"If the objective is ordinary stochastic optimization without nesting or equilibrium structure, use D0; if the main constraint is federation/communication rather than the game or nesting itself, use D4.","nearest_boundary":"D4 Distributed and federated optimization","representative_topics":["stochastic single-loop algorithms for multi-block min-max bilevel optimization","conditional gradient methods for bilevel optimization with convex lower-level problems and optimal iteration complexity","clipped stochastic extragradient methods for variational inequalities with heavy-tailed noise and unbounded domains"]},{"direction_index":6,"label":"D6","name":"Sequential, black-box, and zeroth-order optimization","description":"Query-efficient optimization under limited, sequential, delayed, noisy, bandit, preference, comparison, function-value, or surrogate-based feedback.","inclusion":"Include atoms whose primary objective is global Bayesian or Lipschitz black-box search, acquisition/surrogate design, online convex regret minimization, bandit or preference feedback, convex chasing, derivative-free local/stationary optimization, zeroth-order gradient estimation, query reuse, safe black-box optimization, or related sequential feedback efficiency.","exclusion":"If exact-gradient deterministic convex mechanism design is primary, use D2; if the problem is a distributed system constraint, use D4; if function-value feedback is used mainly inside a nonconvex structural guarantee, use D3.","nearest_boundary":"D2 Convex, constrained, and adaptive algorithms","representative_topics":["hybrid mcts and gaussian process surrogates for mixed-variable bayesian optimization","minimax optimal global optimization via query rules for multivariate lipschitz functions","structured stochastic zeroth-order descent with orthogonal direction sampling and polyak-łojasiewicz convergence"]},{"direction_index":7,"label":"D7","name":"Geometric and distributional-domain optimization","description":"Optimization over non-Euclidean, manifold, measure, probability-distribution, transport, sampling-dynamics, policy-gradient, and other intrinsically structured domains where the domain geometry or distributional dynamics drive the theory.","inclusion":"Include atoms whose primary objective is adapting optimization to Riemannian or geodesic geometry, measure spaces, Wasserstein/optimal-transport structure, particle or variational transport methods, Langevin/Stein/sampling dynamics, natural-gradient or policy-gradient geometry, distributional optimization, or solution maps tied to structured domains.","exclusion":"If optimal transport or particles are used mainly for game equilibrium computation, use D5; if the main goal is standard convex oracle complexity rather than domain geometry, use D2; if decentralization is the organizing constraint, use D4.","nearest_boundary":"D5 Nested, minimax, and equilibrium optimization","representative_topics":["stochastic approximation convergence on riemannian manifolds via fermi coordinate dynamical systems","mirror descent convergence rates for convex optimization over measure spaces with non-smooth functionals","constrained optimal transport via saddle point methods with soft density and momentum constraints"]}],"split":"dev"}
{"episode_id":"F124_T2025-07","field_index":124,"field":"optimization theory","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Stochastic and private first-order guarantees","description":"Convergence, finite-sample, generalization, stability, robustness, and privacy guarantees for stochastic first-order optimization procedures.","inclusion":"Include atoms whose primary objective is proving optimization-error, statistical-generalization, stability, high-probability, robustness-to-noise, or differential-privacy guarantees for SGD, stochastic approximation, DP-SGD, stochastic convex/nonconvex/minimax optimization, or close first-order variants.","exclusion":"If the main objective is optimizer-induced neural solution geometry or neural training regimes, use D1; if adaptivity, preconditioning, or learned optimizer mechanisms are primary, use D2; if communication/federation is primary, use D4.","nearest_boundary":"D1 Neural optimization dynamics and implicit bias","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","algorithmic stability bounds for heavy-tailed stochastic gradient descent generalization","sgd-dependent generalization error bounds via parameter norm dynamics and rademacher complexity"]},{"direction_index":1,"label":"D1","name":"Neural optimization dynamics and implicit bias","description":"Understanding how gradient-based optimization behaves in neural or overparameterized models, including implicit regularization, representation/solution selection, sharpness, flatness, stability regimes, edge-of-stability, and mean-field or architecture-specific training dynamics.","inclusion":"Include atoms whose primary objective is explaining optimizer-induced neural-network trajectories, selected solutions, margins/ranks/sparsity, landscape geometry, sharpness/flatness, phase regimes, overparameterization effects, NTK/mean-field limits, or architecture-specific convergence phenomena.","exclusion":"If the main target is a general SGD convergence or generalization bound not organized around neural training dynamics, use D0; if the contribution is generic nonconvex stationarity or saddle escape rather than neural-specific dynamics, use D3.","nearest_boundary":"D0 Stochastic and private first-order guarantees","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","pde-based theoretical framework for edge of stability and numerical instability in deep network training","gradient descent convergence beyond edge of stability via two-step update and sharpness minimisation analysis"]},{"direction_index":2,"label":"D2","name":"Convex, constrained, and adaptive algorithms","description":"Design and analysis of deterministic or mechanism-focused optimization algorithms for convex, constrained, composite, projection-free, accelerated, proximal, adaptive, momentum, preconditioned, and learned-optimizer settings.","inclusion":"Include atoms whose primary objective is improving algorithmic mechanisms, rates, oracle complexity, conditioning, acceleration, projection avoidance, proximal/cutting-plane/Frank-Wolfe behavior, adaptive learning-rate rules, Adam/AdaGrad-style methods, curvature approximations, or learned optimizer designs for standard optimization problems.","exclusion":"If explicit Hessian/Newton machinery or algebraic relaxations are tied to nonconvex or discrete structure, use D3; if stochastic noise, privacy, or statistical stability is the organizing goal rather than the algorithmic mechanism, use D0.","nearest_boundary":"D3 Nonconvex, second-order, and discrete structure","representative_topics":["accelerated proximal gradient methods for convex optimization with locally lipschitz continuous gradient","relaxed error criterion for efficient catalyst-based convex optimization with logarithmic complexity reduction","lipschitz-free convergence analysis of adam with hyperparameters close to one"]},{"direction_index":3,"label":"D3","name":"Nonconvex, second-order, and discrete structure","description":"Algorithms, complexity guarantees, relaxations, approximation methods, and landscape analyses for nonconvex optimization, second-order stationarity, combinatorial/submodular/discrete structure, and algebraic or semidefinite relaxations.","inclusion":"Include atoms whose primary objective is proving nonconvex convergence or stationarity, escaping saddles, certifying global structure, exploiting matrix-factorization or constrained nonconvex landscapes, developing Newton/trust-region/cubic/quasi-Newton methods, or obtaining approximation/relaxation guarantees for submodular, combinatorial, polynomial, QCQP, SDP, or related structured problems.","exclusion":"If the nonconvex behavior is primarily neural training dynamics or implicit bias, use D1; if the problem is a simultaneous game, minimax problem, or variational inequality, use D5; if the main issue is zeroth-order feedback, use D6.","nearest_boundary":"D1 Neural optimization dynamics and implicit bias","representative_topics":["preconditioned gradient descent for overparameterized nonconvex burer-monteiro factorization with global optimality certification","second-order stationary point methods for nonconvex conic optimization via newton-conjugate gradient barrier","parabolic relaxation of non-convex qcqp via boundary containment and sdp equivalence"]},{"direction_index":4,"label":"D4","name":"Distributed and federated optimization","description":"Optimization theory and algorithms for multi-agent, decentralized, federated, asynchronous, compressed, heterogeneous, or communication-constrained computation.","inclusion":"Include atoms whose primary objective is handling communication efficiency, consensus, compression/quantization, client sampling, delays, local updates, data heterogeneity, gradient tracking, decentralized topologies, or federated aggregation as the central optimization constraint.","exclusion":"If distributed computation is incidental to a bilevel, minimax, manifold, or privacy objective, use D5, D7, or D0 respectively according to the primary scientific objective.","nearest_boundary":"D5 Nested, minimax, and equilibrium optimization","representative_topics":["asynchronous decentralized optimization with network-level variance reduction and asymmetric communication","federated proximal optimization convergence theory with dissimilarity-invariant non-smooth guarantees","asynchronous federated optimization with stochastic aggregation for heterogeneous client delays"]},{"direction_index":5,"label":"D5","name":"Nested, minimax, and equilibrium optimization","description":"Optimization with explicitly nested objectives or simultaneous adversarial/equilibrium structure, including bilevel, multilevel compositional, minimax, saddle-point, game-theoretic, monotone-inclusion, and variational-inequality programs.","inclusion":"Include atoms whose primary objective is algorithm design or convergence theory for lower-level solution maps, hypergradients, compositional expectations, coupled/nested objectives, gradient descent-ascent, extragradient methods, last-iterate behavior, Nash or mixed equilibria, saddle-point systems, monotone inclusions, or variational inequalities.","exclusion":"If the objective is ordinary stochastic optimization without nesting or equilibrium structure, use D0; if the main constraint is federation/communication rather than the game or nesting itself, use D4.","nearest_boundary":"D4 Distributed and federated optimization","representative_topics":["stochastic single-loop algorithms for multi-block min-max bilevel optimization","conditional gradient methods for bilevel optimization with convex lower-level problems and optimal iteration complexity","clipped stochastic extragradient methods for variational inequalities with heavy-tailed noise and unbounded domains"]},{"direction_index":6,"label":"D6","name":"Sequential, black-box, and zeroth-order optimization","description":"Query-efficient optimization under limited, sequential, delayed, noisy, bandit, preference, comparison, function-value, or surrogate-based feedback.","inclusion":"Include atoms whose primary objective is global Bayesian or Lipschitz black-box search, acquisition/surrogate design, online convex regret minimization, bandit or preference feedback, convex chasing, derivative-free local/stationary optimization, zeroth-order gradient estimation, query reuse, safe black-box optimization, or related sequential feedback efficiency.","exclusion":"If exact-gradient deterministic convex mechanism design is primary, use D2; if the problem is a distributed system constraint, use D4; if function-value feedback is used mainly inside a nonconvex structural guarantee, use D3.","nearest_boundary":"D2 Convex, constrained, and adaptive algorithms","representative_topics":["hybrid mcts and gaussian process surrogates for mixed-variable bayesian optimization","minimax optimal global optimization via query rules for multivariate lipschitz functions","structured stochastic zeroth-order descent with orthogonal direction sampling and polyak-łojasiewicz convergence"]},{"direction_index":7,"label":"D7","name":"Geometric and distributional-domain optimization","description":"Optimization over non-Euclidean, manifold, measure, probability-distribution, transport, sampling-dynamics, policy-gradient, and other intrinsically structured domains where the domain geometry or distributional dynamics drive the theory.","inclusion":"Include atoms whose primary objective is adapting optimization to Riemannian or geodesic geometry, measure spaces, Wasserstein/optimal-transport structure, particle or variational transport methods, Langevin/Stein/sampling dynamics, natural-gradient or policy-gradient geometry, distributional optimization, or solution maps tied to structured domains.","exclusion":"If optimal transport or particles are used mainly for game equilibrium computation, use D5; if the main goal is standard convex oracle complexity rather than domain geometry, use D2; if decentralization is the organizing constraint, use D4.","nearest_boundary":"D5 Nested, minimax, and equilibrium optimization","representative_topics":["stochastic approximation convergence on riemannian manifolds via fermi coordinate dynamical systems","mirror descent convergence rates for convex optimization over measure spaces with non-smooth functionals","constrained optimal transport via saddle point methods with soft density and momentum constraints"]}],"split":"dev"}
{"episode_id":"F124_T2026-01","field_index":124,"field":"optimization theory","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Stochastic and private first-order guarantees","description":"Convergence, finite-sample, generalization, stability, robustness, and privacy guarantees for stochastic first-order optimization procedures.","inclusion":"Include atoms whose primary objective is proving optimization-error, statistical-generalization, stability, high-probability, robustness-to-noise, or differential-privacy guarantees for SGD, stochastic approximation, DP-SGD, stochastic convex/nonconvex/minimax optimization, or close first-order variants.","exclusion":"If the main objective is optimizer-induced neural solution geometry or neural training regimes, use D1; if adaptivity, preconditioning, or learned optimizer mechanisms are primary, use D2; if communication/federation is primary, use D4.","nearest_boundary":"D1 Neural optimization dynamics and implicit bias","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","algorithmic stability bounds for heavy-tailed stochastic gradient descent generalization","sgd-dependent generalization error bounds via parameter norm dynamics and rademacher complexity"]},{"direction_index":1,"label":"D1","name":"Neural optimization dynamics and implicit bias","description":"Understanding how gradient-based optimization behaves in neural or overparameterized models, including implicit regularization, representation/solution selection, sharpness, flatness, stability regimes, edge-of-stability, and mean-field or architecture-specific training dynamics.","inclusion":"Include atoms whose primary objective is explaining optimizer-induced neural-network trajectories, selected solutions, margins/ranks/sparsity, landscape geometry, sharpness/flatness, phase regimes, overparameterization effects, NTK/mean-field limits, or architecture-specific convergence phenomena.","exclusion":"If the main target is a general SGD convergence or generalization bound not organized around neural training dynamics, use D0; if the contribution is generic nonconvex stationarity or saddle escape rather than neural-specific dynamics, use D3.","nearest_boundary":"D0 Stochastic and private first-order guarantees","representative_topics":["gradient flow dynamics of shallow relu networks with implicit bias and saddle dynamics","pde-based theoretical framework for edge of stability and numerical instability in deep network training","gradient descent convergence beyond edge of stability via two-step update and sharpness minimisation analysis"]},{"direction_index":2,"label":"D2","name":"Convex, constrained, and adaptive algorithms","description":"Design and analysis of deterministic or mechanism-focused optimization algorithms for convex, constrained, composite, projection-free, accelerated, proximal, adaptive, momentum, preconditioned, and learned-optimizer settings.","inclusion":"Include atoms whose primary objective is improving algorithmic mechanisms, rates, oracle complexity, conditioning, acceleration, projection avoidance, proximal/cutting-plane/Frank-Wolfe behavior, adaptive learning-rate rules, Adam/AdaGrad-style methods, curvature approximations, or learned optimizer designs for standard optimization problems.","exclusion":"If explicit Hessian/Newton machinery or algebraic relaxations are tied to nonconvex or discrete structure, use D3; if stochastic noise, privacy, or statistical stability is the organizing goal rather than the algorithmic mechanism, use D0.","nearest_boundary":"D3 Nonconvex, second-order, and discrete structure","representative_topics":["accelerated proximal gradient methods for convex optimization with locally lipschitz continuous gradient","relaxed error criterion for efficient catalyst-based convex optimization with logarithmic complexity reduction","lipschitz-free convergence analysis of adam with hyperparameters close to one"]},{"direction_index":3,"label":"D3","name":"Nonconvex, second-order, and discrete structure","description":"Algorithms, complexity guarantees, relaxations, approximation methods, and landscape analyses for nonconvex optimization, second-order stationarity, combinatorial/submodular/discrete structure, and algebraic or semidefinite relaxations.","inclusion":"Include atoms whose primary objective is proving nonconvex convergence or stationarity, escaping saddles, certifying global structure, exploiting matrix-factorization or constrained nonconvex landscapes, developing Newton/trust-region/cubic/quasi-Newton methods, or obtaining approximation/relaxation guarantees for submodular, combinatorial, polynomial, QCQP, SDP, or related structured problems.","exclusion":"If the nonconvex behavior is primarily neural training dynamics or implicit bias, use D1; if the problem is a simultaneous game, minimax problem, or variational inequality, use D5; if the main issue is zeroth-order feedback, use D6.","nearest_boundary":"D1 Neural optimization dynamics and implicit bias","representative_topics":["preconditioned gradient descent for overparameterized nonconvex burer-monteiro factorization with global optimality certification","second-order stationary point methods for nonconvex conic optimization via newton-conjugate gradient barrier","parabolic relaxation of non-convex qcqp via boundary containment and sdp equivalence"]},{"direction_index":4,"label":"D4","name":"Distributed and federated optimization","description":"Optimization theory and algorithms for multi-agent, decentralized, federated, asynchronous, compressed, heterogeneous, or communication-constrained computation.","inclusion":"Include atoms whose primary objective is handling communication efficiency, consensus, compression/quantization, client sampling, delays, local updates, data heterogeneity, gradient tracking, decentralized topologies, or federated aggregation as the central optimization constraint.","exclusion":"If distributed computation is incidental to a bilevel, minimax, manifold, or privacy objective, use D5, D7, or D0 respectively according to the primary scientific objective.","nearest_boundary":"D5 Nested, minimax, and equilibrium optimization","representative_topics":["asynchronous decentralized optimization with network-level variance reduction and asymmetric communication","federated proximal optimization convergence theory with dissimilarity-invariant non-smooth guarantees","asynchronous federated optimization with stochastic aggregation for heterogeneous client delays"]},{"direction_index":5,"label":"D5","name":"Nested, minimax, and equilibrium optimization","description":"Optimization with explicitly nested objectives or simultaneous adversarial/equilibrium structure, including bilevel, multilevel compositional, minimax, saddle-point, game-theoretic, monotone-inclusion, and variational-inequality programs.","inclusion":"Include atoms whose primary objective is algorithm design or convergence theory for lower-level solution maps, hypergradients, compositional expectations, coupled/nested objectives, gradient descent-ascent, extragradient methods, last-iterate behavior, Nash or mixed equilibria, saddle-point systems, monotone inclusions, or variational inequalities.","exclusion":"If the objective is ordinary stochastic optimization without nesting or equilibrium structure, use D0; if the main constraint is federation/communication rather than the game or nesting itself, use D4.","nearest_boundary":"D4 Distributed and federated optimization","representative_topics":["stochastic single-loop algorithms for multi-block min-max bilevel optimization","conditional gradient methods for bilevel optimization with convex lower-level problems and optimal iteration complexity","clipped stochastic extragradient methods for variational inequalities with heavy-tailed noise and unbounded domains"]},{"direction_index":6,"label":"D6","name":"Sequential, black-box, and zeroth-order optimization","description":"Query-efficient optimization under limited, sequential, delayed, noisy, bandit, preference, comparison, function-value, or surrogate-based feedback.","inclusion":"Include atoms whose primary objective is global Bayesian or Lipschitz black-box search, acquisition/surrogate design, online convex regret minimization, bandit or preference feedback, convex chasing, derivative-free local/stationary optimization, zeroth-order gradient estimation, query reuse, safe black-box optimization, or related sequential feedback efficiency.","exclusion":"If exact-gradient deterministic convex mechanism design is primary, use D2; if the problem is a distributed system constraint, use D4; if function-value feedback is used mainly inside a nonconvex structural guarantee, use D3.","nearest_boundary":"D2 Convex, constrained, and adaptive algorithms","representative_topics":["hybrid mcts and gaussian process surrogates for mixed-variable bayesian optimization","minimax optimal global optimization via query rules for multivariate lipschitz functions","structured stochastic zeroth-order descent with orthogonal direction sampling and polyak-łojasiewicz convergence"]},{"direction_index":7,"label":"D7","name":"Geometric and distributional-domain optimization","description":"Optimization over non-Euclidean, manifold, measure, probability-distribution, transport, sampling-dynamics, policy-gradient, and other intrinsically structured domains where the domain geometry or distributional dynamics drive the theory.","inclusion":"Include atoms whose primary objective is adapting optimization to Riemannian or geodesic geometry, measure spaces, Wasserstein/optimal-transport structure, particle or variational transport methods, Langevin/Stein/sampling dynamics, natural-gradient or policy-gradient geometry, distributional optimization, or solution maps tied to structured domains.","exclusion":"If optimal transport or particles are used mainly for game equilibrium computation, use D5; if the main goal is standard convex oracle complexity rather than domain geometry, use D2; if decentralization is the organizing constraint, use D4.","nearest_boundary":"D5 Nested, minimax, and equilibrium optimization","representative_topics":["stochastic approximation convergence on riemannian manifolds via fermi coordinate dynamical systems","mirror descent convergence rates for convex optimization over measure spaces with non-smooth functionals","constrained optimal transport via saddle point methods with soft density and momentum constraints"]}],"split":"dev"}
{"episode_id":"F125_T2024-01","field_index":125,"field":"time series anomaly detection","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structural dependency and causal anomaly modeling","description":"Detect or explain time-series anomalies by modeling relationships among variables, entities, subsequences, or system states, including graph, causal, invariant, and functional-dependency structure.","inclusion":"Include work whose primary objective is to learn, exploit, or analyze explicit inter-variable, temporal, graph, causal, invariant, intrinsic, exogenous, or root-cause-oriented structure for anomaly detection.","exclusion":"Assign to reconstruction/generative normality modeling when graph or dependency structure is only an architectural accessory to reconstruction, likelihood, or density scoring; assign to robustness/transfer when causal language is mainly used for distribution-shift robustness.","nearest_boundary":"D1","representative_topics":["intrinsic anomaly detection via adversarial and representation learning for multi-variate time series","causal modular decomposition for multivariate time series anomaly detection and root cause identification","unsupervised graph-based subsequence anomaly detection in long time series without domain knowledge"]},{"direction_index":1,"label":"D1","name":"Reconstruction and generative normality modeling","description":"Learn normal temporal behavior with reconstruction, restoration, density, likelihood, residual, or generative mechanisms and detect anomalies as deviations from that learned normality.","inclusion":"Include autoencoders, recurrent or convolutional reconstruction models, VAEs, GANs, diffusion or score models, normalizing flows, denoising/restoration networks, baseline extraction, forecasting-residual, density, or likelihood-based anomaly scoring when this is the main detector objective.","exclusion":"Assign to self-supervised representation learning when the main contribution is contrastive, degradation-based, or discriminative representation formation rather than reconstruction or generative scoring; assign to structural dependency modeling when explicit graph or causal structure is the main objective.","nearest_boundary":"D2","representative_topics":["semi-supervised point anomaly detection via recurrent analysis of raw gradient distributions","unsupervised recurrent anomaly detection in hpc systems without labeled anomaly data","lstm autoencoder anomaly detection enhanced with denoising architecture for unsupervised time-series"]},{"direction_index":2,"label":"D2","name":"Self-supervised anomaly representation learning","description":"Develop anomaly detectors by learning anomaly-sensitive temporal representations without relying primarily on labeled anomalies or reconstruction error.","inclusion":"Include contrastive learning, one-class representation objectives, synthetic degradation pretext tasks, multi-view learning, memory-augmented representation deviation, prototype or transformer-style self-supervised modeling when representation formation is the central anomaly-detection objective.","exclusion":"Assign to reconstruction/generative normality modeling when anomaly evidence is principally reconstruction error, restoration quality, likelihood, density, or generated-sample discrepancy; assign to robust/label-efficient adaptation when scarce labels or transfer are the main objective.","nearest_boundary":"D1","representative_topics":["self-supervised time series anomaly detection via contrastive one-class sequence representation learning","self-supervised representation learning for unsupervised anomaly detection","multi-scale contrastive memory for unified one-class and unbalanced anomaly detection"]},{"direction_index":3,"label":"D3","name":"Streaming adaptation and timely detection","description":"Maintain or improve anomaly detection in real-time, streaming, non-stationary, drifted, irregular, or sequential sensing settings where timing, delay, or online adaptation is central.","inclusion":"Include online or incremental learning, adaptive model pooling, dynamic thresholds, concept-drift handling, open-ended streams, real-time process limits, precursor-of-anomaly detection, irregular-time early warning, and sequential controlled sensing with delay or measurement-cost objectives.","exclusion":"Assign to robust/label-efficient adaptation when the central problem is contaminated training data, cross-domain transfer, adversarial perturbation, or label scarcity rather than continuous stream adaptation or detection timing.","nearest_boundary":"D4","representative_topics":["adaptive model pooling for online deep anomaly detection under concept drift in evolving streams","unsupervised real-time multivariate anomaly detection for concept-drifting data streams","online inverse cumulative distribution anomaly detection with dynamic process limits for real-time infrastructure"]},{"direction_index":4,"label":"D4","name":"Robust, transferred, and label-efficient detection","description":"Improve anomaly detection when supervision, source domains, or distributions are unreliable, scarce, shifted, contaminated, imbalanced, transferred, or adversarially manipulated.","inclusion":"Include contaminated-normal training, label noise, robust losses, weak or continuous labels, active or interactive labeling, anomaly-only or semi-supervised losses, class-imbalance handling, few-shot detection, domain adaptation, dataset fusion, transfer learning, test-time/domain shift, and adversarial vulnerability or defense.","exclusion":"Assign to streaming adaptation when the main challenge is continuous online concept drift; assign to self-supervised representation learning when labels and domain transfer are absent and the main objective is unsupervised representation formation.","nearest_boundary":"D3","representative_topics":["supervised regression-based anomaly detection using continuous labels and augmentation for low-resource settings","calibrated one-class classification for contamination-tolerant unsupervised time series anomaly detection","robust learning of deep time series anomaly detectors under contaminated training data"]},{"direction_index":5,"label":"D5","name":"Evaluation, interpretability, and model governance","description":"Study how time-series anomaly detectors should be measured, compared, explained, selected, configured, or organized for transparent and reliable use.","inclusion":"Include evaluation metrics, protocol critique, benchmark construction, comparative studies, surveys, taxonomies, baseline assessments, explanation methods, interpretable anomaly rationales, prototype explanations, automated model selection, and neural architecture search when these are the primary objective.","exclusion":"Assign to a detector-development direction when evaluation or interpretability is only secondary validation for a new detection mechanism; assign to domain monitoring when the main objective is a specific operational deployment rather than general measurement, explanation, or model governance.","nearest_boundary":"D7","representative_topics":["interpretable parameter-free evaluation metrics for time series anomaly detection via ground truth affiliation","survey and comparative evaluation of unsupervised time series anomaly detection methods","extensive evaluation protocol for unsupervised time-series anomaly detection methods"]},{"direction_index":6,"label":"D6","name":"Statistical decomposition and lightweight temporal scoring","description":"Detect anomalies through statistical temporal structure, decomposition, subspace, spectral, profile, kernel, probabilistic, or other lightweight non-deep scoring mechanisms.","inclusion":"Include time-frequency analysis, singular spectrum analysis, robust PCA, matrix profiles, subspace differences, seasonal or trend decomposition, probabilistic innovation scoring, kernel novelty detection, gradient-distribution scoring, global-local statistics, and ensemble or feature-isolation scoring when these mechanisms are the core detector objective.","exclusion":"Assign to reconstruction/generative normality modeling for deep autoencoder, GAN, diffusion, or likelihood-generative architectures; assign to domain monitoring when the main contribution is a domain-engineered operational system rather than a general statistical scoring mechanism.","nearest_boundary":"D1","representative_topics":["online robust pca for anomaly detection and imputation in seasonal time series","time-frequency decomposition architecture for univariate and multivariate time series anomaly detection","online nonparametric novelty detection via causal innovations representation and minimax bayes risk"]},{"direction_index":7,"label":"D7","name":"Domain-tailored operational monitoring","description":"Design anomaly-detection systems whose main research objective is adapting detection to a particular operational domain's signals, constraints, maintenance needs, or deployment environment.","inclusion":"Include domain-organized monitoring systems for healthcare, vehicles, power grids, industrial equipment, aviation, IoT, household equipment, distributed systems, HPC, finance, cyber-physical systems, or sensor operations when domain requirements organize the contribution.","exclusion":"Assign to a method-centered direction when the main contribution is a generally applicable graph, reconstruction, representation, statistical, adaptation, or supervised-learning method merely demonstrated on domain data; assign to evaluation/governance when the domain data are used mainly for benchmarking or comparison.","nearest_boundary":"D5","representative_topics":["trajectory-based outlier detection for wearable vital sign monitoring in covid-19 patients","survey of anomaly detection algorithms for real-time cryptocurrency fraud detection","unsupervised recurrent anomaly detection in hpc systems without labeled anomaly data"]}],"split":"dev"}
{"episode_id":"F125_T2024-07","field_index":125,"field":"time series anomaly detection","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structural dependency and causal anomaly modeling","description":"Detect or explain time-series anomalies by modeling relationships among variables, entities, subsequences, or system states, including graph, causal, invariant, and functional-dependency structure.","inclusion":"Include work whose primary objective is to learn, exploit, or analyze explicit inter-variable, temporal, graph, causal, invariant, intrinsic, exogenous, or root-cause-oriented structure for anomaly detection.","exclusion":"Assign to reconstruction/generative normality modeling when graph or dependency structure is only an architectural accessory to reconstruction, likelihood, or density scoring; assign to robustness/transfer when causal language is mainly used for distribution-shift robustness.","nearest_boundary":"D1","representative_topics":["intrinsic anomaly detection via adversarial and representation learning for multi-variate time series","causal modular decomposition for multivariate time series anomaly detection and root cause identification","unsupervised graph-based subsequence anomaly detection in long time series without domain knowledge"]},{"direction_index":1,"label":"D1","name":"Reconstruction and generative normality modeling","description":"Learn normal temporal behavior with reconstruction, restoration, density, likelihood, residual, or generative mechanisms and detect anomalies as deviations from that learned normality.","inclusion":"Include autoencoders, recurrent or convolutional reconstruction models, VAEs, GANs, diffusion or score models, normalizing flows, denoising/restoration networks, baseline extraction, forecasting-residual, density, or likelihood-based anomaly scoring when this is the main detector objective.","exclusion":"Assign to self-supervised representation learning when the main contribution is contrastive, degradation-based, or discriminative representation formation rather than reconstruction or generative scoring; assign to structural dependency modeling when explicit graph or causal structure is the main objective.","nearest_boundary":"D2","representative_topics":["semi-supervised point anomaly detection via recurrent analysis of raw gradient distributions","unsupervised recurrent anomaly detection in hpc systems without labeled anomaly data","lstm autoencoder anomaly detection enhanced with denoising architecture for unsupervised time-series"]},{"direction_index":2,"label":"D2","name":"Self-supervised anomaly representation learning","description":"Develop anomaly detectors by learning anomaly-sensitive temporal representations without relying primarily on labeled anomalies or reconstruction error.","inclusion":"Include contrastive learning, one-class representation objectives, synthetic degradation pretext tasks, multi-view learning, memory-augmented representation deviation, prototype or transformer-style self-supervised modeling when representation formation is the central anomaly-detection objective.","exclusion":"Assign to reconstruction/generative normality modeling when anomaly evidence is principally reconstruction error, restoration quality, likelihood, density, or generated-sample discrepancy; assign to robust/label-efficient adaptation when scarce labels or transfer are the main objective.","nearest_boundary":"D1","representative_topics":["self-supervised time series anomaly detection via contrastive one-class sequence representation learning","self-supervised representation learning for unsupervised anomaly detection","multi-scale contrastive memory for unified one-class and unbalanced anomaly detection"]},{"direction_index":3,"label":"D3","name":"Streaming adaptation and timely detection","description":"Maintain or improve anomaly detection in real-time, streaming, non-stationary, drifted, irregular, or sequential sensing settings where timing, delay, or online adaptation is central.","inclusion":"Include online or incremental learning, adaptive model pooling, dynamic thresholds, concept-drift handling, open-ended streams, real-time process limits, precursor-of-anomaly detection, irregular-time early warning, and sequential controlled sensing with delay or measurement-cost objectives.","exclusion":"Assign to robust/label-efficient adaptation when the central problem is contaminated training data, cross-domain transfer, adversarial perturbation, or label scarcity rather than continuous stream adaptation or detection timing.","nearest_boundary":"D4","representative_topics":["adaptive model pooling for online deep anomaly detection under concept drift in evolving streams","unsupervised real-time multivariate anomaly detection for concept-drifting data streams","online inverse cumulative distribution anomaly detection with dynamic process limits for real-time infrastructure"]},{"direction_index":4,"label":"D4","name":"Robust, transferred, and label-efficient detection","description":"Improve anomaly detection when supervision, source domains, or distributions are unreliable, scarce, shifted, contaminated, imbalanced, transferred, or adversarially manipulated.","inclusion":"Include contaminated-normal training, label noise, robust losses, weak or continuous labels, active or interactive labeling, anomaly-only or semi-supervised losses, class-imbalance handling, few-shot detection, domain adaptation, dataset fusion, transfer learning, test-time/domain shift, and adversarial vulnerability or defense.","exclusion":"Assign to streaming adaptation when the main challenge is continuous online concept drift; assign to self-supervised representation learning when labels and domain transfer are absent and the main objective is unsupervised representation formation.","nearest_boundary":"D3","representative_topics":["supervised regression-based anomaly detection using continuous labels and augmentation for low-resource settings","calibrated one-class classification for contamination-tolerant unsupervised time series anomaly detection","robust learning of deep time series anomaly detectors under contaminated training data"]},{"direction_index":5,"label":"D5","name":"Evaluation, interpretability, and model governance","description":"Study how time-series anomaly detectors should be measured, compared, explained, selected, configured, or organized for transparent and reliable use.","inclusion":"Include evaluation metrics, protocol critique, benchmark construction, comparative studies, surveys, taxonomies, baseline assessments, explanation methods, interpretable anomaly rationales, prototype explanations, automated model selection, and neural architecture search when these are the primary objective.","exclusion":"Assign to a detector-development direction when evaluation or interpretability is only secondary validation for a new detection mechanism; assign to domain monitoring when the main objective is a specific operational deployment rather than general measurement, explanation, or model governance.","nearest_boundary":"D7","representative_topics":["interpretable parameter-free evaluation metrics for time series anomaly detection via ground truth affiliation","survey and comparative evaluation of unsupervised time series anomaly detection methods","extensive evaluation protocol for unsupervised time-series anomaly detection methods"]},{"direction_index":6,"label":"D6","name":"Statistical decomposition and lightweight temporal scoring","description":"Detect anomalies through statistical temporal structure, decomposition, subspace, spectral, profile, kernel, probabilistic, or other lightweight non-deep scoring mechanisms.","inclusion":"Include time-frequency analysis, singular spectrum analysis, robust PCA, matrix profiles, subspace differences, seasonal or trend decomposition, probabilistic innovation scoring, kernel novelty detection, gradient-distribution scoring, global-local statistics, and ensemble or feature-isolation scoring when these mechanisms are the core detector objective.","exclusion":"Assign to reconstruction/generative normality modeling for deep autoencoder, GAN, diffusion, or likelihood-generative architectures; assign to domain monitoring when the main contribution is a domain-engineered operational system rather than a general statistical scoring mechanism.","nearest_boundary":"D1","representative_topics":["online robust pca for anomaly detection and imputation in seasonal time series","time-frequency decomposition architecture for univariate and multivariate time series anomaly detection","online nonparametric novelty detection via causal innovations representation and minimax bayes risk"]},{"direction_index":7,"label":"D7","name":"Domain-tailored operational monitoring","description":"Design anomaly-detection systems whose main research objective is adapting detection to a particular operational domain's signals, constraints, maintenance needs, or deployment environment.","inclusion":"Include domain-organized monitoring systems for healthcare, vehicles, power grids, industrial equipment, aviation, IoT, household equipment, distributed systems, HPC, finance, cyber-physical systems, or sensor operations when domain requirements organize the contribution.","exclusion":"Assign to a method-centered direction when the main contribution is a generally applicable graph, reconstruction, representation, statistical, adaptation, or supervised-learning method merely demonstrated on domain data; assign to evaluation/governance when the domain data are used mainly for benchmarking or comparison.","nearest_boundary":"D5","representative_topics":["trajectory-based outlier detection for wearable vital sign monitoring in covid-19 patients","survey of anomaly detection algorithms for real-time cryptocurrency fraud detection","unsupervised recurrent anomaly detection in hpc systems without labeled anomaly data"]}],"split":"dev"}
{"episode_id":"F125_T2025-01","field_index":125,"field":"time series anomaly detection","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structural dependency and causal anomaly modeling","description":"Detect or explain time-series anomalies by modeling relationships among variables, entities, subsequences, or system states, including graph, causal, invariant, and functional-dependency structure.","inclusion":"Include work whose primary objective is to learn, exploit, or analyze explicit inter-variable, temporal, graph, causal, invariant, intrinsic, exogenous, or root-cause-oriented structure for anomaly detection.","exclusion":"Assign to reconstruction/generative normality modeling when graph or dependency structure is only an architectural accessory to reconstruction, likelihood, or density scoring; assign to robustness/transfer when causal language is mainly used for distribution-shift robustness.","nearest_boundary":"D1","representative_topics":["intrinsic anomaly detection via adversarial and representation learning for multi-variate time series","causal modular decomposition for multivariate time series anomaly detection and root cause identification","unsupervised graph-based subsequence anomaly detection in long time series without domain knowledge"]},{"direction_index":1,"label":"D1","name":"Reconstruction and generative normality modeling","description":"Learn normal temporal behavior with reconstruction, restoration, density, likelihood, residual, or generative mechanisms and detect anomalies as deviations from that learned normality.","inclusion":"Include autoencoders, recurrent or convolutional reconstruction models, VAEs, GANs, diffusion or score models, normalizing flows, denoising/restoration networks, baseline extraction, forecasting-residual, density, or likelihood-based anomaly scoring when this is the main detector objective.","exclusion":"Assign to self-supervised representation learning when the main contribution is contrastive, degradation-based, or discriminative representation formation rather than reconstruction or generative scoring; assign to structural dependency modeling when explicit graph or causal structure is the main objective.","nearest_boundary":"D2","representative_topics":["semi-supervised point anomaly detection via recurrent analysis of raw gradient distributions","unsupervised recurrent anomaly detection in hpc systems without labeled anomaly data","lstm autoencoder anomaly detection enhanced with denoising architecture for unsupervised time-series"]},{"direction_index":2,"label":"D2","name":"Self-supervised anomaly representation learning","description":"Develop anomaly detectors by learning anomaly-sensitive temporal representations without relying primarily on labeled anomalies or reconstruction error.","inclusion":"Include contrastive learning, one-class representation objectives, synthetic degradation pretext tasks, multi-view learning, memory-augmented representation deviation, prototype or transformer-style self-supervised modeling when representation formation is the central anomaly-detection objective.","exclusion":"Assign to reconstruction/generative normality modeling when anomaly evidence is principally reconstruction error, restoration quality, likelihood, density, or generated-sample discrepancy; assign to robust/label-efficient adaptation when scarce labels or transfer are the main objective.","nearest_boundary":"D1","representative_topics":["self-supervised time series anomaly detection via contrastive one-class sequence representation learning","self-supervised representation learning for unsupervised anomaly detection","multi-scale contrastive memory for unified one-class and unbalanced anomaly detection"]},{"direction_index":3,"label":"D3","name":"Streaming adaptation and timely detection","description":"Maintain or improve anomaly detection in real-time, streaming, non-stationary, drifted, irregular, or sequential sensing settings where timing, delay, or online adaptation is central.","inclusion":"Include online or incremental learning, adaptive model pooling, dynamic thresholds, concept-drift handling, open-ended streams, real-time process limits, precursor-of-anomaly detection, irregular-time early warning, and sequential controlled sensing with delay or measurement-cost objectives.","exclusion":"Assign to robust/label-efficient adaptation when the central problem is contaminated training data, cross-domain transfer, adversarial perturbation, or label scarcity rather than continuous stream adaptation or detection timing.","nearest_boundary":"D4","representative_topics":["adaptive model pooling for online deep anomaly detection under concept drift in evolving streams","unsupervised real-time multivariate anomaly detection for concept-drifting data streams","online inverse cumulative distribution anomaly detection with dynamic process limits for real-time infrastructure"]},{"direction_index":4,"label":"D4","name":"Robust, transferred, and label-efficient detection","description":"Improve anomaly detection when supervision, source domains, or distributions are unreliable, scarce, shifted, contaminated, imbalanced, transferred, or adversarially manipulated.","inclusion":"Include contaminated-normal training, label noise, robust losses, weak or continuous labels, active or interactive labeling, anomaly-only or semi-supervised losses, class-imbalance handling, few-shot detection, domain adaptation, dataset fusion, transfer learning, test-time/domain shift, and adversarial vulnerability or defense.","exclusion":"Assign to streaming adaptation when the main challenge is continuous online concept drift; assign to self-supervised representation learning when labels and domain transfer are absent and the main objective is unsupervised representation formation.","nearest_boundary":"D3","representative_topics":["supervised regression-based anomaly detection using continuous labels and augmentation for low-resource settings","calibrated one-class classification for contamination-tolerant unsupervised time series anomaly detection","robust learning of deep time series anomaly detectors under contaminated training data"]},{"direction_index":5,"label":"D5","name":"Evaluation, interpretability, and model governance","description":"Study how time-series anomaly detectors should be measured, compared, explained, selected, configured, or organized for transparent and reliable use.","inclusion":"Include evaluation metrics, protocol critique, benchmark construction, comparative studies, surveys, taxonomies, baseline assessments, explanation methods, interpretable anomaly rationales, prototype explanations, automated model selection, and neural architecture search when these are the primary objective.","exclusion":"Assign to a detector-development direction when evaluation or interpretability is only secondary validation for a new detection mechanism; assign to domain monitoring when the main objective is a specific operational deployment rather than general measurement, explanation, or model governance.","nearest_boundary":"D7","representative_topics":["interpretable parameter-free evaluation metrics for time series anomaly detection via ground truth affiliation","survey and comparative evaluation of unsupervised time series anomaly detection methods","extensive evaluation protocol for unsupervised time-series anomaly detection methods"]},{"direction_index":6,"label":"D6","name":"Statistical decomposition and lightweight temporal scoring","description":"Detect anomalies through statistical temporal structure, decomposition, subspace, spectral, profile, kernel, probabilistic, or other lightweight non-deep scoring mechanisms.","inclusion":"Include time-frequency analysis, singular spectrum analysis, robust PCA, matrix profiles, subspace differences, seasonal or trend decomposition, probabilistic innovation scoring, kernel novelty detection, gradient-distribution scoring, global-local statistics, and ensemble or feature-isolation scoring when these mechanisms are the core detector objective.","exclusion":"Assign to reconstruction/generative normality modeling for deep autoencoder, GAN, diffusion, or likelihood-generative architectures; assign to domain monitoring when the main contribution is a domain-engineered operational system rather than a general statistical scoring mechanism.","nearest_boundary":"D1","representative_topics":["online robust pca for anomaly detection and imputation in seasonal time series","time-frequency decomposition architecture for univariate and multivariate time series anomaly detection","online nonparametric novelty detection via causal innovations representation and minimax bayes risk"]},{"direction_index":7,"label":"D7","name":"Domain-tailored operational monitoring","description":"Design anomaly-detection systems whose main research objective is adapting detection to a particular operational domain's signals, constraints, maintenance needs, or deployment environment.","inclusion":"Include domain-organized monitoring systems for healthcare, vehicles, power grids, industrial equipment, aviation, IoT, household equipment, distributed systems, HPC, finance, cyber-physical systems, or sensor operations when domain requirements organize the contribution.","exclusion":"Assign to a method-centered direction when the main contribution is a generally applicable graph, reconstruction, representation, statistical, adaptation, or supervised-learning method merely demonstrated on domain data; assign to evaluation/governance when the domain data are used mainly for benchmarking or comparison.","nearest_boundary":"D5","representative_topics":["trajectory-based outlier detection for wearable vital sign monitoring in covid-19 patients","survey of anomaly detection algorithms for real-time cryptocurrency fraud detection","unsupervised recurrent anomaly detection in hpc systems without labeled anomaly data"]}],"split":"dev"}
{"episode_id":"F125_T2025-07","field_index":125,"field":"time series anomaly detection","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structural dependency and causal anomaly modeling","description":"Detect or explain time-series anomalies by modeling relationships among variables, entities, subsequences, or system states, including graph, causal, invariant, and functional-dependency structure.","inclusion":"Include work whose primary objective is to learn, exploit, or analyze explicit inter-variable, temporal, graph, causal, invariant, intrinsic, exogenous, or root-cause-oriented structure for anomaly detection.","exclusion":"Assign to reconstruction/generative normality modeling when graph or dependency structure is only an architectural accessory to reconstruction, likelihood, or density scoring; assign to robustness/transfer when causal language is mainly used for distribution-shift robustness.","nearest_boundary":"D1","representative_topics":["intrinsic anomaly detection via adversarial and representation learning for multi-variate time series","causal modular decomposition for multivariate time series anomaly detection and root cause identification","unsupervised graph-based subsequence anomaly detection in long time series without domain knowledge"]},{"direction_index":1,"label":"D1","name":"Reconstruction and generative normality modeling","description":"Learn normal temporal behavior with reconstruction, restoration, density, likelihood, residual, or generative mechanisms and detect anomalies as deviations from that learned normality.","inclusion":"Include autoencoders, recurrent or convolutional reconstruction models, VAEs, GANs, diffusion or score models, normalizing flows, denoising/restoration networks, baseline extraction, forecasting-residual, density, or likelihood-based anomaly scoring when this is the main detector objective.","exclusion":"Assign to self-supervised representation learning when the main contribution is contrastive, degradation-based, or discriminative representation formation rather than reconstruction or generative scoring; assign to structural dependency modeling when explicit graph or causal structure is the main objective.","nearest_boundary":"D2","representative_topics":["semi-supervised point anomaly detection via recurrent analysis of raw gradient distributions","unsupervised recurrent anomaly detection in hpc systems without labeled anomaly data","lstm autoencoder anomaly detection enhanced with denoising architecture for unsupervised time-series"]},{"direction_index":2,"label":"D2","name":"Self-supervised anomaly representation learning","description":"Develop anomaly detectors by learning anomaly-sensitive temporal representations without relying primarily on labeled anomalies or reconstruction error.","inclusion":"Include contrastive learning, one-class representation objectives, synthetic degradation pretext tasks, multi-view learning, memory-augmented representation deviation, prototype or transformer-style self-supervised modeling when representation formation is the central anomaly-detection objective.","exclusion":"Assign to reconstruction/generative normality modeling when anomaly evidence is principally reconstruction error, restoration quality, likelihood, density, or generated-sample discrepancy; assign to robust/label-efficient adaptation when scarce labels or transfer are the main objective.","nearest_boundary":"D1","representative_topics":["self-supervised time series anomaly detection via contrastive one-class sequence representation learning","self-supervised representation learning for unsupervised anomaly detection","multi-scale contrastive memory for unified one-class and unbalanced anomaly detection"]},{"direction_index":3,"label":"D3","name":"Streaming adaptation and timely detection","description":"Maintain or improve anomaly detection in real-time, streaming, non-stationary, drifted, irregular, or sequential sensing settings where timing, delay, or online adaptation is central.","inclusion":"Include online or incremental learning, adaptive model pooling, dynamic thresholds, concept-drift handling, open-ended streams, real-time process limits, precursor-of-anomaly detection, irregular-time early warning, and sequential controlled sensing with delay or measurement-cost objectives.","exclusion":"Assign to robust/label-efficient adaptation when the central problem is contaminated training data, cross-domain transfer, adversarial perturbation, or label scarcity rather than continuous stream adaptation or detection timing.","nearest_boundary":"D4","representative_topics":["adaptive model pooling for online deep anomaly detection under concept drift in evolving streams","unsupervised real-time multivariate anomaly detection for concept-drifting data streams","online inverse cumulative distribution anomaly detection with dynamic process limits for real-time infrastructure"]},{"direction_index":4,"label":"D4","name":"Robust, transferred, and label-efficient detection","description":"Improve anomaly detection when supervision, source domains, or distributions are unreliable, scarce, shifted, contaminated, imbalanced, transferred, or adversarially manipulated.","inclusion":"Include contaminated-normal training, label noise, robust losses, weak or continuous labels, active or interactive labeling, anomaly-only or semi-supervised losses, class-imbalance handling, few-shot detection, domain adaptation, dataset fusion, transfer learning, test-time/domain shift, and adversarial vulnerability or defense.","exclusion":"Assign to streaming adaptation when the main challenge is continuous online concept drift; assign to self-supervised representation learning when labels and domain transfer are absent and the main objective is unsupervised representation formation.","nearest_boundary":"D3","representative_topics":["supervised regression-based anomaly detection using continuous labels and augmentation for low-resource settings","calibrated one-class classification for contamination-tolerant unsupervised time series anomaly detection","robust learning of deep time series anomaly detectors under contaminated training data"]},{"direction_index":5,"label":"D5","name":"Evaluation, interpretability, and model governance","description":"Study how time-series anomaly detectors should be measured, compared, explained, selected, configured, or organized for transparent and reliable use.","inclusion":"Include evaluation metrics, protocol critique, benchmark construction, comparative studies, surveys, taxonomies, baseline assessments, explanation methods, interpretable anomaly rationales, prototype explanations, automated model selection, and neural architecture search when these are the primary objective.","exclusion":"Assign to a detector-development direction when evaluation or interpretability is only secondary validation for a new detection mechanism; assign to domain monitoring when the main objective is a specific operational deployment rather than general measurement, explanation, or model governance.","nearest_boundary":"D7","representative_topics":["interpretable parameter-free evaluation metrics for time series anomaly detection via ground truth affiliation","survey and comparative evaluation of unsupervised time series anomaly detection methods","extensive evaluation protocol for unsupervised time-series anomaly detection methods"]},{"direction_index":6,"label":"D6","name":"Statistical decomposition and lightweight temporal scoring","description":"Detect anomalies through statistical temporal structure, decomposition, subspace, spectral, profile, kernel, probabilistic, or other lightweight non-deep scoring mechanisms.","inclusion":"Include time-frequency analysis, singular spectrum analysis, robust PCA, matrix profiles, subspace differences, seasonal or trend decomposition, probabilistic innovation scoring, kernel novelty detection, gradient-distribution scoring, global-local statistics, and ensemble or feature-isolation scoring when these mechanisms are the core detector objective.","exclusion":"Assign to reconstruction/generative normality modeling for deep autoencoder, GAN, diffusion, or likelihood-generative architectures; assign to domain monitoring when the main contribution is a domain-engineered operational system rather than a general statistical scoring mechanism.","nearest_boundary":"D1","representative_topics":["online robust pca for anomaly detection and imputation in seasonal time series","time-frequency decomposition architecture for univariate and multivariate time series anomaly detection","online nonparametric novelty detection via causal innovations representation and minimax bayes risk"]},{"direction_index":7,"label":"D7","name":"Domain-tailored operational monitoring","description":"Design anomaly-detection systems whose main research objective is adapting detection to a particular operational domain's signals, constraints, maintenance needs, or deployment environment.","inclusion":"Include domain-organized monitoring systems for healthcare, vehicles, power grids, industrial equipment, aviation, IoT, household equipment, distributed systems, HPC, finance, cyber-physical systems, or sensor operations when domain requirements organize the contribution.","exclusion":"Assign to a method-centered direction when the main contribution is a generally applicable graph, reconstruction, representation, statistical, adaptation, or supervised-learning method merely demonstrated on domain data; assign to evaluation/governance when the domain data are used mainly for benchmarking or comparison.","nearest_boundary":"D5","representative_topics":["trajectory-based outlier detection for wearable vital sign monitoring in covid-19 patients","survey of anomaly detection algorithms for real-time cryptocurrency fraud detection","unsupervised recurrent anomaly detection in hpc systems without labeled anomaly data"]}],"split":"dev"}
{"episode_id":"F125_T2026-01","field_index":125,"field":"time series anomaly detection","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structural dependency and causal anomaly modeling","description":"Detect or explain time-series anomalies by modeling relationships among variables, entities, subsequences, or system states, including graph, causal, invariant, and functional-dependency structure.","inclusion":"Include work whose primary objective is to learn, exploit, or analyze explicit inter-variable, temporal, graph, causal, invariant, intrinsic, exogenous, or root-cause-oriented structure for anomaly detection.","exclusion":"Assign to reconstruction/generative normality modeling when graph or dependency structure is only an architectural accessory to reconstruction, likelihood, or density scoring; assign to robustness/transfer when causal language is mainly used for distribution-shift robustness.","nearest_boundary":"D1","representative_topics":["intrinsic anomaly detection via adversarial and representation learning for multi-variate time series","causal modular decomposition for multivariate time series anomaly detection and root cause identification","unsupervised graph-based subsequence anomaly detection in long time series without domain knowledge"]},{"direction_index":1,"label":"D1","name":"Reconstruction and generative normality modeling","description":"Learn normal temporal behavior with reconstruction, restoration, density, likelihood, residual, or generative mechanisms and detect anomalies as deviations from that learned normality.","inclusion":"Include autoencoders, recurrent or convolutional reconstruction models, VAEs, GANs, diffusion or score models, normalizing flows, denoising/restoration networks, baseline extraction, forecasting-residual, density, or likelihood-based anomaly scoring when this is the main detector objective.","exclusion":"Assign to self-supervised representation learning when the main contribution is contrastive, degradation-based, or discriminative representation formation rather than reconstruction or generative scoring; assign to structural dependency modeling when explicit graph or causal structure is the main objective.","nearest_boundary":"D2","representative_topics":["semi-supervised point anomaly detection via recurrent analysis of raw gradient distributions","unsupervised recurrent anomaly detection in hpc systems without labeled anomaly data","lstm autoencoder anomaly detection enhanced with denoising architecture for unsupervised time-series"]},{"direction_index":2,"label":"D2","name":"Self-supervised anomaly representation learning","description":"Develop anomaly detectors by learning anomaly-sensitive temporal representations without relying primarily on labeled anomalies or reconstruction error.","inclusion":"Include contrastive learning, one-class representation objectives, synthetic degradation pretext tasks, multi-view learning, memory-augmented representation deviation, prototype or transformer-style self-supervised modeling when representation formation is the central anomaly-detection objective.","exclusion":"Assign to reconstruction/generative normality modeling when anomaly evidence is principally reconstruction error, restoration quality, likelihood, density, or generated-sample discrepancy; assign to robust/label-efficient adaptation when scarce labels or transfer are the main objective.","nearest_boundary":"D1","representative_topics":["self-supervised time series anomaly detection via contrastive one-class sequence representation learning","self-supervised representation learning for unsupervised anomaly detection","multi-scale contrastive memory for unified one-class and unbalanced anomaly detection"]},{"direction_index":3,"label":"D3","name":"Streaming adaptation and timely detection","description":"Maintain or improve anomaly detection in real-time, streaming, non-stationary, drifted, irregular, or sequential sensing settings where timing, delay, or online adaptation is central.","inclusion":"Include online or incremental learning, adaptive model pooling, dynamic thresholds, concept-drift handling, open-ended streams, real-time process limits, precursor-of-anomaly detection, irregular-time early warning, and sequential controlled sensing with delay or measurement-cost objectives.","exclusion":"Assign to robust/label-efficient adaptation when the central problem is contaminated training data, cross-domain transfer, adversarial perturbation, or label scarcity rather than continuous stream adaptation or detection timing.","nearest_boundary":"D4","representative_topics":["adaptive model pooling for online deep anomaly detection under concept drift in evolving streams","unsupervised real-time multivariate anomaly detection for concept-drifting data streams","online inverse cumulative distribution anomaly detection with dynamic process limits for real-time infrastructure"]},{"direction_index":4,"label":"D4","name":"Robust, transferred, and label-efficient detection","description":"Improve anomaly detection when supervision, source domains, or distributions are unreliable, scarce, shifted, contaminated, imbalanced, transferred, or adversarially manipulated.","inclusion":"Include contaminated-normal training, label noise, robust losses, weak or continuous labels, active or interactive labeling, anomaly-only or semi-supervised losses, class-imbalance handling, few-shot detection, domain adaptation, dataset fusion, transfer learning, test-time/domain shift, and adversarial vulnerability or defense.","exclusion":"Assign to streaming adaptation when the main challenge is continuous online concept drift; assign to self-supervised representation learning when labels and domain transfer are absent and the main objective is unsupervised representation formation.","nearest_boundary":"D3","representative_topics":["supervised regression-based anomaly detection using continuous labels and augmentation for low-resource settings","calibrated one-class classification for contamination-tolerant unsupervised time series anomaly detection","robust learning of deep time series anomaly detectors under contaminated training data"]},{"direction_index":5,"label":"D5","name":"Evaluation, interpretability, and model governance","description":"Study how time-series anomaly detectors should be measured, compared, explained, selected, configured, or organized for transparent and reliable use.","inclusion":"Include evaluation metrics, protocol critique, benchmark construction, comparative studies, surveys, taxonomies, baseline assessments, explanation methods, interpretable anomaly rationales, prototype explanations, automated model selection, and neural architecture search when these are the primary objective.","exclusion":"Assign to a detector-development direction when evaluation or interpretability is only secondary validation for a new detection mechanism; assign to domain monitoring when the main objective is a specific operational deployment rather than general measurement, explanation, or model governance.","nearest_boundary":"D7","representative_topics":["interpretable parameter-free evaluation metrics for time series anomaly detection via ground truth affiliation","survey and comparative evaluation of unsupervised time series anomaly detection methods","extensive evaluation protocol for unsupervised time-series anomaly detection methods"]},{"direction_index":6,"label":"D6","name":"Statistical decomposition and lightweight temporal scoring","description":"Detect anomalies through statistical temporal structure, decomposition, subspace, spectral, profile, kernel, probabilistic, or other lightweight non-deep scoring mechanisms.","inclusion":"Include time-frequency analysis, singular spectrum analysis, robust PCA, matrix profiles, subspace differences, seasonal or trend decomposition, probabilistic innovation scoring, kernel novelty detection, gradient-distribution scoring, global-local statistics, and ensemble or feature-isolation scoring when these mechanisms are the core detector objective.","exclusion":"Assign to reconstruction/generative normality modeling for deep autoencoder, GAN, diffusion, or likelihood-generative architectures; assign to domain monitoring when the main contribution is a domain-engineered operational system rather than a general statistical scoring mechanism.","nearest_boundary":"D1","representative_topics":["online robust pca for anomaly detection and imputation in seasonal time series","time-frequency decomposition architecture for univariate and multivariate time series anomaly detection","online nonparametric novelty detection via causal innovations representation and minimax bayes risk"]},{"direction_index":7,"label":"D7","name":"Domain-tailored operational monitoring","description":"Design anomaly-detection systems whose main research objective is adapting detection to a particular operational domain's signals, constraints, maintenance needs, or deployment environment.","inclusion":"Include domain-organized monitoring systems for healthcare, vehicles, power grids, industrial equipment, aviation, IoT, household equipment, distributed systems, HPC, finance, cyber-physical systems, or sensor operations when domain requirements organize the contribution.","exclusion":"Assign to a method-centered direction when the main contribution is a generally applicable graph, reconstruction, representation, statistical, adaptation, or supervised-learning method merely demonstrated on domain data; assign to evaluation/governance when the domain data are used mainly for benchmarking or comparison.","nearest_boundary":"D5","representative_topics":["trajectory-based outlier detection for wearable vital sign monitoring in covid-19 patients","survey of anomaly detection algorithms for real-time cryptocurrency fraud detection","unsupervised recurrent anomaly detection in hpc systems without labeled anomaly data"]}],"split":"dev"}
{"episode_id":"F128_T2024-01","field_index":128,"field":"safe reinforcement learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Certificate-based safe control","description":"Synthesize, learn, or embed formal control certificates that enforce safety, stability, feasibility, reachability, or invariance for reinforcement-learning controllers.","inclusion":"Include atoms whose central objective is RL control with CBFs, CLFs, Lyapunov or barrier certificates, sum-of-squares certificates, neural or differentiable certificate layers, stochastic or adaptive barrier functions, or invariant-set conditions.","exclusion":"Put work in runtime assurance when the main mechanism is an external action monitor or shield not centered on certificate synthesis; put it in safe exploration when the main objective is safe-set discovery rather than certificate-based control.","nearest_boundary":"Runtime shielding and assurance layers","representative_topics":["sum-of-squares programming with control barrier functions for certified safe reinforcement learning","lyapunov-stabilized online reinforcement learning without pre-training for safety-critical control","sample-efficient safe episodic learning for nonlinear control via stochastic control barrier functions"]},{"direction_index":1,"label":"D1","name":"Runtime shielding and assurance","description":"Design online safety layers that monitor, filter, replace, project, or override actions from an RL policy during training or deployment to prevent unsafe behavior.","inclusion":"Include atoms whose primary objective is an external runtime safety mechanism such as confidence filters, reachability filters, action governors, automata shields, weakest-precondition shields, recovery policies, runtime assurance modules, or analyses of their effect on learning.","exclusion":"Put CBF/CLF-centered filters in certificate-based safe control; put model-predictive or belief-space planning methods in constrained policy and planning optimization when planning itself is the main contribution.","nearest_boundary":"Certificate-based safe control","representative_topics":["control-theoretic safety filters using probabilistic dynamics and confidence-based policy adjustment","ablation study of runtime assurance methods in safe reinforcement learning training","robust action governor for safe reinforcement learning under non-convex constraints and parametric uncertainty"]},{"direction_index":2,"label":"D2","name":"Constrained policy and planning optimization","description":"Optimize or plan safe policies under explicit cost, constraint, belief, or feasibility requirements using CMDP methods, constrained updates, model-based planning, or safety-embedding learning architectures.","inclusion":"Include atoms whose objective is solving safety-constrained policy learning or planning problems through Lagrangian, primal-dual, projection, proximal, trust-region, safety-critic, constrained actor-critic, constrained POMDP, model-predictive, world-model, value-decomposition, or related architectural mechanisms.","exclusion":"Put scalar tail-risk or utility objectives in risk-sensitive safe RL; put external per-action intervention layers in runtime shielding; put fixed-dataset-only methods in offline safe RL when offline data are the central issue.","nearest_boundary":"Risk-sensitive safe RL objectives","representative_topics":["safe policy optimization via constrained update projection and surrogate-based performance bounds","model-based safe deep reinforcement learning via constrained proximal policy optimization with online dynamics learning","safe policy improvement in constrained markov decision processes with formal requirement compliance"]},{"direction_index":3,"label":"D3","name":"Risk-sensitive safe RL","description":"Develop algorithms and theory for safety by optimizing risk, tail, distributional, or utility-based criteria rather than expected return alone.","inclusion":"Include atoms centered on CVaR, iterated CVaR, entropic or exponential utility, EVaR, spectral or dynamic risk measures, optimized certainty equivalents, Gini deviation, mean-semivariance, distributional return objectives, or regret and policy-gradient analysis for such criteria.","exclusion":"Put expected cumulative cost constraints with budgets in constrained policy and planning optimization; put minimax or ambiguity-set robustness against model uncertainty or attacks in robust safe RL.","nearest_boundary":"Robust safe RL under uncertainty","representative_topics":["episodic iterated cvar reinforcement learning with regret minimization and worst-path guarantees","risk-sensitive safe reinforcement learning via conditional value-at-risk constraints under joint disturbances","mean-semivariance policy optimization via perturbation analysis and policy gradient methods"]},{"direction_index":4,"label":"D4","name":"Robust safe RL under uncertainty","description":"Learn policies that remain safe, reliable, or performant under model uncertainty, distribution shift, adversarial perturbations, underspecified environments, attacks, or ambiguity sets.","inclusion":"Include atoms whose primary objective is robustness via minimax formulations, uncertainty sets, distributional robustness, adversarial training, attack or defense analysis, certified robustness, robust offline-data learning, or separation of epistemic and aleatory uncertainty.","exclusion":"Put known-model risk criteria without adversarial or ambiguity-set structure in risk-sensitive safe RL; put CBF-based disturbance handling in certificate-based safe control.","nearest_boundary":"Risk-sensitive safe RL objectives","representative_topics":["uncertainty set regularization for robust policy generalization in continuous control reinforcement learning","feasible adversarial robust reinforcement learning via zero-sum game over environment parameters","minimax game-theoretic analysis of sampling attacks on meta reinforcement learning adaptation"]},{"direction_index":5,"label":"D5","name":"Safe exploration and safe-set learning","description":"Guarantee or improve safety during online data collection by estimating safe regions, feasible actions, uncertainty bounds, safety feedback models, or exploration policies.","inclusion":"Include atoms centered on safe exploration algorithms, Bayesian or bandit safe learning, reward-free safe exploration, active safety-feedback learning, sample-complexity or regret guarantees for safe exploration, and safe-set estimation or expansion.","exclusion":"Put CBF/CLF-enforced exploration in certificate-based safe control; put fixed-dataset methods that avoid online exploration in offline safe RL; put average-cost CMDP training without per-exploration safety guarantees in constrained policy and planning optimization.","nearest_boundary":"Constrained policy and planning optimization","representative_topics":["safe exploration in reward-free rl via truncated value functions with minimal sample complexity overhead","safe reinforcement learning via constraint manifold tangent space exploration for high-dimensional robotics","provable safe reinforcement learning with binary oracle feedback via active learning meta-algorithm"]},{"direction_index":6,"label":"D6","name":"Offline safe RL and evaluation","description":"Evaluate, improve, or synthesize safe policies from fixed offline data while controlling deployment risk, estimation uncertainty, performance degradation, or constraint violations without unrestricted online interaction.","inclusion":"Include atoms where the fixed-dataset restriction is central, including high-confidence off-policy safety evaluation, safe offline policy improvement, offline safe deployment, offline CMDP solving, constrained decision transformers, sequence-model or generative safe policy learning, and offline-to-online safety initialization.","exclusion":"Put online exploration with safety guarantees in safe exploration; put general constrained policy updates using replay data in constrained policy and planning optimization; put dataset or benchmark construction without a new offline method in empirical evaluation and infrastructure.","nearest_boundary":"Constrained policy and planning optimization","representative_topics":["safe policy improvement via uncertainty-aware bootstrapping with provable safety guarantees","safe offline policy deployment via high-confidence off-policy evaluation lower-bound estimation","risk-sensitive dead-end identification via distributional return estimation in offline reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Knowledge-guided and empirical safe RL","description":"Define, guide, validate, or translate safe RL using formal specifications, external knowledge, demonstrations, inferred constraints, domain safety structure, benchmarks, surveys, software, datasets, or evaluation protocols.","inclusion":"Include atoms whose primary objective is incorporating or inferring safety knowledge, logic, rules, temporal specifications, demonstrations, or domain constraints, or building shared empirical resources such as benchmarks, surveys, datasets, infrastructure, and domain-integration studies for safe RL.","exclusion":"Put formal specifications compiled only into online action blockers in runtime shielding; put already-numeric constraints optimized mainly by dual or policy-gradient methods in constrained policy and planning optimization; put algorithmic papers that merely evaluate on a benchmark in the relevant algorithmic direction.","nearest_boundary":"Runtime shielding and assurance layers","representative_topics":["scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","benchmarking inverse constrained reinforcement learning with variational constraint inference","safe lane-change reinforcement learning using human demonstrations as safety constraints"]}],"split":"dev"}
{"episode_id":"F128_T2024-07","field_index":128,"field":"safe reinforcement learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Certificate-based safe control","description":"Synthesize, learn, or embed formal control certificates that enforce safety, stability, feasibility, reachability, or invariance for reinforcement-learning controllers.","inclusion":"Include atoms whose central objective is RL control with CBFs, CLFs, Lyapunov or barrier certificates, sum-of-squares certificates, neural or differentiable certificate layers, stochastic or adaptive barrier functions, or invariant-set conditions.","exclusion":"Put work in runtime assurance when the main mechanism is an external action monitor or shield not centered on certificate synthesis; put it in safe exploration when the main objective is safe-set discovery rather than certificate-based control.","nearest_boundary":"Runtime shielding and assurance layers","representative_topics":["sum-of-squares programming with control barrier functions for certified safe reinforcement learning","lyapunov-stabilized online reinforcement learning without pre-training for safety-critical control","sample-efficient safe episodic learning for nonlinear control via stochastic control barrier functions"]},{"direction_index":1,"label":"D1","name":"Runtime shielding and assurance","description":"Design online safety layers that monitor, filter, replace, project, or override actions from an RL policy during training or deployment to prevent unsafe behavior.","inclusion":"Include atoms whose primary objective is an external runtime safety mechanism such as confidence filters, reachability filters, action governors, automata shields, weakest-precondition shields, recovery policies, runtime assurance modules, or analyses of their effect on learning.","exclusion":"Put CBF/CLF-centered filters in certificate-based safe control; put model-predictive or belief-space planning methods in constrained policy and planning optimization when planning itself is the main contribution.","nearest_boundary":"Certificate-based safe control","representative_topics":["control-theoretic safety filters using probabilistic dynamics and confidence-based policy adjustment","ablation study of runtime assurance methods in safe reinforcement learning training","robust action governor for safe reinforcement learning under non-convex constraints and parametric uncertainty"]},{"direction_index":2,"label":"D2","name":"Constrained policy and planning optimization","description":"Optimize or plan safe policies under explicit cost, constraint, belief, or feasibility requirements using CMDP methods, constrained updates, model-based planning, or safety-embedding learning architectures.","inclusion":"Include atoms whose objective is solving safety-constrained policy learning or planning problems through Lagrangian, primal-dual, projection, proximal, trust-region, safety-critic, constrained actor-critic, constrained POMDP, model-predictive, world-model, value-decomposition, or related architectural mechanisms.","exclusion":"Put scalar tail-risk or utility objectives in risk-sensitive safe RL; put external per-action intervention layers in runtime shielding; put fixed-dataset-only methods in offline safe RL when offline data are the central issue.","nearest_boundary":"Risk-sensitive safe RL objectives","representative_topics":["safe policy optimization via constrained update projection and surrogate-based performance bounds","model-based safe deep reinforcement learning via constrained proximal policy optimization with online dynamics learning","safe policy improvement in constrained markov decision processes with formal requirement compliance"]},{"direction_index":3,"label":"D3","name":"Risk-sensitive safe RL","description":"Develop algorithms and theory for safety by optimizing risk, tail, distributional, or utility-based criteria rather than expected return alone.","inclusion":"Include atoms centered on CVaR, iterated CVaR, entropic or exponential utility, EVaR, spectral or dynamic risk measures, optimized certainty equivalents, Gini deviation, mean-semivariance, distributional return objectives, or regret and policy-gradient analysis for such criteria.","exclusion":"Put expected cumulative cost constraints with budgets in constrained policy and planning optimization; put minimax or ambiguity-set robustness against model uncertainty or attacks in robust safe RL.","nearest_boundary":"Robust safe RL under uncertainty","representative_topics":["episodic iterated cvar reinforcement learning with regret minimization and worst-path guarantees","risk-sensitive safe reinforcement learning via conditional value-at-risk constraints under joint disturbances","mean-semivariance policy optimization via perturbation analysis and policy gradient methods"]},{"direction_index":4,"label":"D4","name":"Robust safe RL under uncertainty","description":"Learn policies that remain safe, reliable, or performant under model uncertainty, distribution shift, adversarial perturbations, underspecified environments, attacks, or ambiguity sets.","inclusion":"Include atoms whose primary objective is robustness via minimax formulations, uncertainty sets, distributional robustness, adversarial training, attack or defense analysis, certified robustness, robust offline-data learning, or separation of epistemic and aleatory uncertainty.","exclusion":"Put known-model risk criteria without adversarial or ambiguity-set structure in risk-sensitive safe RL; put CBF-based disturbance handling in certificate-based safe control.","nearest_boundary":"Risk-sensitive safe RL objectives","representative_topics":["uncertainty set regularization for robust policy generalization in continuous control reinforcement learning","feasible adversarial robust reinforcement learning via zero-sum game over environment parameters","minimax game-theoretic analysis of sampling attacks on meta reinforcement learning adaptation"]},{"direction_index":5,"label":"D5","name":"Safe exploration and safe-set learning","description":"Guarantee or improve safety during online data collection by estimating safe regions, feasible actions, uncertainty bounds, safety feedback models, or exploration policies.","inclusion":"Include atoms centered on safe exploration algorithms, Bayesian or bandit safe learning, reward-free safe exploration, active safety-feedback learning, sample-complexity or regret guarantees for safe exploration, and safe-set estimation or expansion.","exclusion":"Put CBF/CLF-enforced exploration in certificate-based safe control; put fixed-dataset methods that avoid online exploration in offline safe RL; put average-cost CMDP training without per-exploration safety guarantees in constrained policy and planning optimization.","nearest_boundary":"Constrained policy and planning optimization","representative_topics":["safe exploration in reward-free rl via truncated value functions with minimal sample complexity overhead","safe reinforcement learning via constraint manifold tangent space exploration for high-dimensional robotics","provable safe reinforcement learning with binary oracle feedback via active learning meta-algorithm"]},{"direction_index":6,"label":"D6","name":"Offline safe RL and evaluation","description":"Evaluate, improve, or synthesize safe policies from fixed offline data while controlling deployment risk, estimation uncertainty, performance degradation, or constraint violations without unrestricted online interaction.","inclusion":"Include atoms where the fixed-dataset restriction is central, including high-confidence off-policy safety evaluation, safe offline policy improvement, offline safe deployment, offline CMDP solving, constrained decision transformers, sequence-model or generative safe policy learning, and offline-to-online safety initialization.","exclusion":"Put online exploration with safety guarantees in safe exploration; put general constrained policy updates using replay data in constrained policy and planning optimization; put dataset or benchmark construction without a new offline method in empirical evaluation and infrastructure.","nearest_boundary":"Constrained policy and planning optimization","representative_topics":["safe policy improvement via uncertainty-aware bootstrapping with provable safety guarantees","safe offline policy deployment via high-confidence off-policy evaluation lower-bound estimation","risk-sensitive dead-end identification via distributional return estimation in offline reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Knowledge-guided and empirical safe RL","description":"Define, guide, validate, or translate safe RL using formal specifications, external knowledge, demonstrations, inferred constraints, domain safety structure, benchmarks, surveys, software, datasets, or evaluation protocols.","inclusion":"Include atoms whose primary objective is incorporating or inferring safety knowledge, logic, rules, temporal specifications, demonstrations, or domain constraints, or building shared empirical resources such as benchmarks, surveys, datasets, infrastructure, and domain-integration studies for safe RL.","exclusion":"Put formal specifications compiled only into online action blockers in runtime shielding; put already-numeric constraints optimized mainly by dual or policy-gradient methods in constrained policy and planning optimization; put algorithmic papers that merely evaluate on a benchmark in the relevant algorithmic direction.","nearest_boundary":"Runtime shielding and assurance layers","representative_topics":["scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","benchmarking inverse constrained reinforcement learning with variational constraint inference","safe lane-change reinforcement learning using human demonstrations as safety constraints"]}],"split":"dev"}
{"episode_id":"F128_T2025-01","field_index":128,"field":"safe reinforcement learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Certificate-based safe control","description":"Synthesize, learn, or embed formal control certificates that enforce safety, stability, feasibility, reachability, or invariance for reinforcement-learning controllers.","inclusion":"Include atoms whose central objective is RL control with CBFs, CLFs, Lyapunov or barrier certificates, sum-of-squares certificates, neural or differentiable certificate layers, stochastic or adaptive barrier functions, or invariant-set conditions.","exclusion":"Put work in runtime assurance when the main mechanism is an external action monitor or shield not centered on certificate synthesis; put it in safe exploration when the main objective is safe-set discovery rather than certificate-based control.","nearest_boundary":"Runtime shielding and assurance layers","representative_topics":["sum-of-squares programming with control barrier functions for certified safe reinforcement learning","lyapunov-stabilized online reinforcement learning without pre-training for safety-critical control","sample-efficient safe episodic learning for nonlinear control via stochastic control barrier functions"]},{"direction_index":1,"label":"D1","name":"Runtime shielding and assurance","description":"Design online safety layers that monitor, filter, replace, project, or override actions from an RL policy during training or deployment to prevent unsafe behavior.","inclusion":"Include atoms whose primary objective is an external runtime safety mechanism such as confidence filters, reachability filters, action governors, automata shields, weakest-precondition shields, recovery policies, runtime assurance modules, or analyses of their effect on learning.","exclusion":"Put CBF/CLF-centered filters in certificate-based safe control; put model-predictive or belief-space planning methods in constrained policy and planning optimization when planning itself is the main contribution.","nearest_boundary":"Certificate-based safe control","representative_topics":["control-theoretic safety filters using probabilistic dynamics and confidence-based policy adjustment","ablation study of runtime assurance methods in safe reinforcement learning training","robust action governor for safe reinforcement learning under non-convex constraints and parametric uncertainty"]},{"direction_index":2,"label":"D2","name":"Constrained policy and planning optimization","description":"Optimize or plan safe policies under explicit cost, constraint, belief, or feasibility requirements using CMDP methods, constrained updates, model-based planning, or safety-embedding learning architectures.","inclusion":"Include atoms whose objective is solving safety-constrained policy learning or planning problems through Lagrangian, primal-dual, projection, proximal, trust-region, safety-critic, constrained actor-critic, constrained POMDP, model-predictive, world-model, value-decomposition, or related architectural mechanisms.","exclusion":"Put scalar tail-risk or utility objectives in risk-sensitive safe RL; put external per-action intervention layers in runtime shielding; put fixed-dataset-only methods in offline safe RL when offline data are the central issue.","nearest_boundary":"Risk-sensitive safe RL objectives","representative_topics":["safe policy optimization via constrained update projection and surrogate-based performance bounds","model-based safe deep reinforcement learning via constrained proximal policy optimization with online dynamics learning","safe policy improvement in constrained markov decision processes with formal requirement compliance"]},{"direction_index":3,"label":"D3","name":"Risk-sensitive safe RL","description":"Develop algorithms and theory for safety by optimizing risk, tail, distributional, or utility-based criteria rather than expected return alone.","inclusion":"Include atoms centered on CVaR, iterated CVaR, entropic or exponential utility, EVaR, spectral or dynamic risk measures, optimized certainty equivalents, Gini deviation, mean-semivariance, distributional return objectives, or regret and policy-gradient analysis for such criteria.","exclusion":"Put expected cumulative cost constraints with budgets in constrained policy and planning optimization; put minimax or ambiguity-set robustness against model uncertainty or attacks in robust safe RL.","nearest_boundary":"Robust safe RL under uncertainty","representative_topics":["episodic iterated cvar reinforcement learning with regret minimization and worst-path guarantees","risk-sensitive safe reinforcement learning via conditional value-at-risk constraints under joint disturbances","mean-semivariance policy optimization via perturbation analysis and policy gradient methods"]},{"direction_index":4,"label":"D4","name":"Robust safe RL under uncertainty","description":"Learn policies that remain safe, reliable, or performant under model uncertainty, distribution shift, adversarial perturbations, underspecified environments, attacks, or ambiguity sets.","inclusion":"Include atoms whose primary objective is robustness via minimax formulations, uncertainty sets, distributional robustness, adversarial training, attack or defense analysis, certified robustness, robust offline-data learning, or separation of epistemic and aleatory uncertainty.","exclusion":"Put known-model risk criteria without adversarial or ambiguity-set structure in risk-sensitive safe RL; put CBF-based disturbance handling in certificate-based safe control.","nearest_boundary":"Risk-sensitive safe RL objectives","representative_topics":["uncertainty set regularization for robust policy generalization in continuous control reinforcement learning","feasible adversarial robust reinforcement learning via zero-sum game over environment parameters","minimax game-theoretic analysis of sampling attacks on meta reinforcement learning adaptation"]},{"direction_index":5,"label":"D5","name":"Safe exploration and safe-set learning","description":"Guarantee or improve safety during online data collection by estimating safe regions, feasible actions, uncertainty bounds, safety feedback models, or exploration policies.","inclusion":"Include atoms centered on safe exploration algorithms, Bayesian or bandit safe learning, reward-free safe exploration, active safety-feedback learning, sample-complexity or regret guarantees for safe exploration, and safe-set estimation or expansion.","exclusion":"Put CBF/CLF-enforced exploration in certificate-based safe control; put fixed-dataset methods that avoid online exploration in offline safe RL; put average-cost CMDP training without per-exploration safety guarantees in constrained policy and planning optimization.","nearest_boundary":"Constrained policy and planning optimization","representative_topics":["safe exploration in reward-free rl via truncated value functions with minimal sample complexity overhead","safe reinforcement learning via constraint manifold tangent space exploration for high-dimensional robotics","provable safe reinforcement learning with binary oracle feedback via active learning meta-algorithm"]},{"direction_index":6,"label":"D6","name":"Offline safe RL and evaluation","description":"Evaluate, improve, or synthesize safe policies from fixed offline data while controlling deployment risk, estimation uncertainty, performance degradation, or constraint violations without unrestricted online interaction.","inclusion":"Include atoms where the fixed-dataset restriction is central, including high-confidence off-policy safety evaluation, safe offline policy improvement, offline safe deployment, offline CMDP solving, constrained decision transformers, sequence-model or generative safe policy learning, and offline-to-online safety initialization.","exclusion":"Put online exploration with safety guarantees in safe exploration; put general constrained policy updates using replay data in constrained policy and planning optimization; put dataset or benchmark construction without a new offline method in empirical evaluation and infrastructure.","nearest_boundary":"Constrained policy and planning optimization","representative_topics":["safe policy improvement via uncertainty-aware bootstrapping with provable safety guarantees","safe offline policy deployment via high-confidence off-policy evaluation lower-bound estimation","risk-sensitive dead-end identification via distributional return estimation in offline reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Knowledge-guided and empirical safe RL","description":"Define, guide, validate, or translate safe RL using formal specifications, external knowledge, demonstrations, inferred constraints, domain safety structure, benchmarks, surveys, software, datasets, or evaluation protocols.","inclusion":"Include atoms whose primary objective is incorporating or inferring safety knowledge, logic, rules, temporal specifications, demonstrations, or domain constraints, or building shared empirical resources such as benchmarks, surveys, datasets, infrastructure, and domain-integration studies for safe RL.","exclusion":"Put formal specifications compiled only into online action blockers in runtime shielding; put already-numeric constraints optimized mainly by dual or policy-gradient methods in constrained policy and planning optimization; put algorithmic papers that merely evaluate on a benchmark in the relevant algorithmic direction.","nearest_boundary":"Runtime shielding and assurance layers","representative_topics":["scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","benchmarking inverse constrained reinforcement learning with variational constraint inference","safe lane-change reinforcement learning using human demonstrations as safety constraints"]}],"split":"dev"}
{"episode_id":"F128_T2025-07","field_index":128,"field":"safe reinforcement learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Certificate-based safe control","description":"Synthesize, learn, or embed formal control certificates that enforce safety, stability, feasibility, reachability, or invariance for reinforcement-learning controllers.","inclusion":"Include atoms whose central objective is RL control with CBFs, CLFs, Lyapunov or barrier certificates, sum-of-squares certificates, neural or differentiable certificate layers, stochastic or adaptive barrier functions, or invariant-set conditions.","exclusion":"Put work in runtime assurance when the main mechanism is an external action monitor or shield not centered on certificate synthesis; put it in safe exploration when the main objective is safe-set discovery rather than certificate-based control.","nearest_boundary":"Runtime shielding and assurance layers","representative_topics":["sum-of-squares programming with control barrier functions for certified safe reinforcement learning","lyapunov-stabilized online reinforcement learning without pre-training for safety-critical control","sample-efficient safe episodic learning for nonlinear control via stochastic control barrier functions"]},{"direction_index":1,"label":"D1","name":"Runtime shielding and assurance","description":"Design online safety layers that monitor, filter, replace, project, or override actions from an RL policy during training or deployment to prevent unsafe behavior.","inclusion":"Include atoms whose primary objective is an external runtime safety mechanism such as confidence filters, reachability filters, action governors, automata shields, weakest-precondition shields, recovery policies, runtime assurance modules, or analyses of their effect on learning.","exclusion":"Put CBF/CLF-centered filters in certificate-based safe control; put model-predictive or belief-space planning methods in constrained policy and planning optimization when planning itself is the main contribution.","nearest_boundary":"Certificate-based safe control","representative_topics":["control-theoretic safety filters using probabilistic dynamics and confidence-based policy adjustment","ablation study of runtime assurance methods in safe reinforcement learning training","robust action governor for safe reinforcement learning under non-convex constraints and parametric uncertainty"]},{"direction_index":2,"label":"D2","name":"Constrained policy and planning optimization","description":"Optimize or plan safe policies under explicit cost, constraint, belief, or feasibility requirements using CMDP methods, constrained updates, model-based planning, or safety-embedding learning architectures.","inclusion":"Include atoms whose objective is solving safety-constrained policy learning or planning problems through Lagrangian, primal-dual, projection, proximal, trust-region, safety-critic, constrained actor-critic, constrained POMDP, model-predictive, world-model, value-decomposition, or related architectural mechanisms.","exclusion":"Put scalar tail-risk or utility objectives in risk-sensitive safe RL; put external per-action intervention layers in runtime shielding; put fixed-dataset-only methods in offline safe RL when offline data are the central issue.","nearest_boundary":"Risk-sensitive safe RL objectives","representative_topics":["safe policy optimization via constrained update projection and surrogate-based performance bounds","model-based safe deep reinforcement learning via constrained proximal policy optimization with online dynamics learning","safe policy improvement in constrained markov decision processes with formal requirement compliance"]},{"direction_index":3,"label":"D3","name":"Risk-sensitive safe RL","description":"Develop algorithms and theory for safety by optimizing risk, tail, distributional, or utility-based criteria rather than expected return alone.","inclusion":"Include atoms centered on CVaR, iterated CVaR, entropic or exponential utility, EVaR, spectral or dynamic risk measures, optimized certainty equivalents, Gini deviation, mean-semivariance, distributional return objectives, or regret and policy-gradient analysis for such criteria.","exclusion":"Put expected cumulative cost constraints with budgets in constrained policy and planning optimization; put minimax or ambiguity-set robustness against model uncertainty or attacks in robust safe RL.","nearest_boundary":"Robust safe RL under uncertainty","representative_topics":["episodic iterated cvar reinforcement learning with regret minimization and worst-path guarantees","risk-sensitive safe reinforcement learning via conditional value-at-risk constraints under joint disturbances","mean-semivariance policy optimization via perturbation analysis and policy gradient methods"]},{"direction_index":4,"label":"D4","name":"Robust safe RL under uncertainty","description":"Learn policies that remain safe, reliable, or performant under model uncertainty, distribution shift, adversarial perturbations, underspecified environments, attacks, or ambiguity sets.","inclusion":"Include atoms whose primary objective is robustness via minimax formulations, uncertainty sets, distributional robustness, adversarial training, attack or defense analysis, certified robustness, robust offline-data learning, or separation of epistemic and aleatory uncertainty.","exclusion":"Put known-model risk criteria without adversarial or ambiguity-set structure in risk-sensitive safe RL; put CBF-based disturbance handling in certificate-based safe control.","nearest_boundary":"Risk-sensitive safe RL objectives","representative_topics":["uncertainty set regularization for robust policy generalization in continuous control reinforcement learning","feasible adversarial robust reinforcement learning via zero-sum game over environment parameters","minimax game-theoretic analysis of sampling attacks on meta reinforcement learning adaptation"]},{"direction_index":5,"label":"D5","name":"Safe exploration and safe-set learning","description":"Guarantee or improve safety during online data collection by estimating safe regions, feasible actions, uncertainty bounds, safety feedback models, or exploration policies.","inclusion":"Include atoms centered on safe exploration algorithms, Bayesian or bandit safe learning, reward-free safe exploration, active safety-feedback learning, sample-complexity or regret guarantees for safe exploration, and safe-set estimation or expansion.","exclusion":"Put CBF/CLF-enforced exploration in certificate-based safe control; put fixed-dataset methods that avoid online exploration in offline safe RL; put average-cost CMDP training without per-exploration safety guarantees in constrained policy and planning optimization.","nearest_boundary":"Constrained policy and planning optimization","representative_topics":["safe exploration in reward-free rl via truncated value functions with minimal sample complexity overhead","safe reinforcement learning via constraint manifold tangent space exploration for high-dimensional robotics","provable safe reinforcement learning with binary oracle feedback via active learning meta-algorithm"]},{"direction_index":6,"label":"D6","name":"Offline safe RL and evaluation","description":"Evaluate, improve, or synthesize safe policies from fixed offline data while controlling deployment risk, estimation uncertainty, performance degradation, or constraint violations without unrestricted online interaction.","inclusion":"Include atoms where the fixed-dataset restriction is central, including high-confidence off-policy safety evaluation, safe offline policy improvement, offline safe deployment, offline CMDP solving, constrained decision transformers, sequence-model or generative safe policy learning, and offline-to-online safety initialization.","exclusion":"Put online exploration with safety guarantees in safe exploration; put general constrained policy updates using replay data in constrained policy and planning optimization; put dataset or benchmark construction without a new offline method in empirical evaluation and infrastructure.","nearest_boundary":"Constrained policy and planning optimization","representative_topics":["safe policy improvement via uncertainty-aware bootstrapping with provable safety guarantees","safe offline policy deployment via high-confidence off-policy evaluation lower-bound estimation","risk-sensitive dead-end identification via distributional return estimation in offline reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Knowledge-guided and empirical safe RL","description":"Define, guide, validate, or translate safe RL using formal specifications, external knowledge, demonstrations, inferred constraints, domain safety structure, benchmarks, surveys, software, datasets, or evaluation protocols.","inclusion":"Include atoms whose primary objective is incorporating or inferring safety knowledge, logic, rules, temporal specifications, demonstrations, or domain constraints, or building shared empirical resources such as benchmarks, surveys, datasets, infrastructure, and domain-integration studies for safe RL.","exclusion":"Put formal specifications compiled only into online action blockers in runtime shielding; put already-numeric constraints optimized mainly by dual or policy-gradient methods in constrained policy and planning optimization; put algorithmic papers that merely evaluate on a benchmark in the relevant algorithmic direction.","nearest_boundary":"Runtime shielding and assurance layers","representative_topics":["scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","benchmarking inverse constrained reinforcement learning with variational constraint inference","safe lane-change reinforcement learning using human demonstrations as safety constraints"]}],"split":"dev"}
{"episode_id":"F128_T2026-01","field_index":128,"field":"safe reinforcement learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Certificate-based safe control","description":"Synthesize, learn, or embed formal control certificates that enforce safety, stability, feasibility, reachability, or invariance for reinforcement-learning controllers.","inclusion":"Include atoms whose central objective is RL control with CBFs, CLFs, Lyapunov or barrier certificates, sum-of-squares certificates, neural or differentiable certificate layers, stochastic or adaptive barrier functions, or invariant-set conditions.","exclusion":"Put work in runtime assurance when the main mechanism is an external action monitor or shield not centered on certificate synthesis; put it in safe exploration when the main objective is safe-set discovery rather than certificate-based control.","nearest_boundary":"Runtime shielding and assurance layers","representative_topics":["sum-of-squares programming with control barrier functions for certified safe reinforcement learning","lyapunov-stabilized online reinforcement learning without pre-training for safety-critical control","sample-efficient safe episodic learning for nonlinear control via stochastic control barrier functions"]},{"direction_index":1,"label":"D1","name":"Runtime shielding and assurance","description":"Design online safety layers that monitor, filter, replace, project, or override actions from an RL policy during training or deployment to prevent unsafe behavior.","inclusion":"Include atoms whose primary objective is an external runtime safety mechanism such as confidence filters, reachability filters, action governors, automata shields, weakest-precondition shields, recovery policies, runtime assurance modules, or analyses of their effect on learning.","exclusion":"Put CBF/CLF-centered filters in certificate-based safe control; put model-predictive or belief-space planning methods in constrained policy and planning optimization when planning itself is the main contribution.","nearest_boundary":"Certificate-based safe control","representative_topics":["control-theoretic safety filters using probabilistic dynamics and confidence-based policy adjustment","ablation study of runtime assurance methods in safe reinforcement learning training","robust action governor for safe reinforcement learning under non-convex constraints and parametric uncertainty"]},{"direction_index":2,"label":"D2","name":"Constrained policy and planning optimization","description":"Optimize or plan safe policies under explicit cost, constraint, belief, or feasibility requirements using CMDP methods, constrained updates, model-based planning, or safety-embedding learning architectures.","inclusion":"Include atoms whose objective is solving safety-constrained policy learning or planning problems through Lagrangian, primal-dual, projection, proximal, trust-region, safety-critic, constrained actor-critic, constrained POMDP, model-predictive, world-model, value-decomposition, or related architectural mechanisms.","exclusion":"Put scalar tail-risk or utility objectives in risk-sensitive safe RL; put external per-action intervention layers in runtime shielding; put fixed-dataset-only methods in offline safe RL when offline data are the central issue.","nearest_boundary":"Risk-sensitive safe RL objectives","representative_topics":["safe policy optimization via constrained update projection and surrogate-based performance bounds","model-based safe deep reinforcement learning via constrained proximal policy optimization with online dynamics learning","safe policy improvement in constrained markov decision processes with formal requirement compliance"]},{"direction_index":3,"label":"D3","name":"Risk-sensitive safe RL","description":"Develop algorithms and theory for safety by optimizing risk, tail, distributional, or utility-based criteria rather than expected return alone.","inclusion":"Include atoms centered on CVaR, iterated CVaR, entropic or exponential utility, EVaR, spectral or dynamic risk measures, optimized certainty equivalents, Gini deviation, mean-semivariance, distributional return objectives, or regret and policy-gradient analysis for such criteria.","exclusion":"Put expected cumulative cost constraints with budgets in constrained policy and planning optimization; put minimax or ambiguity-set robustness against model uncertainty or attacks in robust safe RL.","nearest_boundary":"Robust safe RL under uncertainty","representative_topics":["episodic iterated cvar reinforcement learning with regret minimization and worst-path guarantees","risk-sensitive safe reinforcement learning via conditional value-at-risk constraints under joint disturbances","mean-semivariance policy optimization via perturbation analysis and policy gradient methods"]},{"direction_index":4,"label":"D4","name":"Robust safe RL under uncertainty","description":"Learn policies that remain safe, reliable, or performant under model uncertainty, distribution shift, adversarial perturbations, underspecified environments, attacks, or ambiguity sets.","inclusion":"Include atoms whose primary objective is robustness via minimax formulations, uncertainty sets, distributional robustness, adversarial training, attack or defense analysis, certified robustness, robust offline-data learning, or separation of epistemic and aleatory uncertainty.","exclusion":"Put known-model risk criteria without adversarial or ambiguity-set structure in risk-sensitive safe RL; put CBF-based disturbance handling in certificate-based safe control.","nearest_boundary":"Risk-sensitive safe RL objectives","representative_topics":["uncertainty set regularization for robust policy generalization in continuous control reinforcement learning","feasible adversarial robust reinforcement learning via zero-sum game over environment parameters","minimax game-theoretic analysis of sampling attacks on meta reinforcement learning adaptation"]},{"direction_index":5,"label":"D5","name":"Safe exploration and safe-set learning","description":"Guarantee or improve safety during online data collection by estimating safe regions, feasible actions, uncertainty bounds, safety feedback models, or exploration policies.","inclusion":"Include atoms centered on safe exploration algorithms, Bayesian or bandit safe learning, reward-free safe exploration, active safety-feedback learning, sample-complexity or regret guarantees for safe exploration, and safe-set estimation or expansion.","exclusion":"Put CBF/CLF-enforced exploration in certificate-based safe control; put fixed-dataset methods that avoid online exploration in offline safe RL; put average-cost CMDP training without per-exploration safety guarantees in constrained policy and planning optimization.","nearest_boundary":"Constrained policy and planning optimization","representative_topics":["safe exploration in reward-free rl via truncated value functions with minimal sample complexity overhead","safe reinforcement learning via constraint manifold tangent space exploration for high-dimensional robotics","provable safe reinforcement learning with binary oracle feedback via active learning meta-algorithm"]},{"direction_index":6,"label":"D6","name":"Offline safe RL and evaluation","description":"Evaluate, improve, or synthesize safe policies from fixed offline data while controlling deployment risk, estimation uncertainty, performance degradation, or constraint violations without unrestricted online interaction.","inclusion":"Include atoms where the fixed-dataset restriction is central, including high-confidence off-policy safety evaluation, safe offline policy improvement, offline safe deployment, offline CMDP solving, constrained decision transformers, sequence-model or generative safe policy learning, and offline-to-online safety initialization.","exclusion":"Put online exploration with safety guarantees in safe exploration; put general constrained policy updates using replay data in constrained policy and planning optimization; put dataset or benchmark construction without a new offline method in empirical evaluation and infrastructure.","nearest_boundary":"Constrained policy and planning optimization","representative_topics":["safe policy improvement via uncertainty-aware bootstrapping with provable safety guarantees","safe offline policy deployment via high-confidence off-policy evaluation lower-bound estimation","risk-sensitive dead-end identification via distributional return estimation in offline reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Knowledge-guided and empirical safe RL","description":"Define, guide, validate, or translate safe RL using formal specifications, external knowledge, demonstrations, inferred constraints, domain safety structure, benchmarks, surveys, software, datasets, or evaluation protocols.","inclusion":"Include atoms whose primary objective is incorporating or inferring safety knowledge, logic, rules, temporal specifications, demonstrations, or domain constraints, or building shared empirical resources such as benchmarks, surveys, datasets, infrastructure, and domain-integration studies for safe RL.","exclusion":"Put formal specifications compiled only into online action blockers in runtime shielding; put already-numeric constraints optimized mainly by dual or policy-gradient methods in constrained policy and planning optimization; put algorithmic papers that merely evaluate on a benchmark in the relevant algorithmic direction.","nearest_boundary":"Runtime shielding and assurance layers","representative_topics":["scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","benchmarking inverse constrained reinforcement learning with variational constraint inference","safe lane-change reinforcement learning using human demonstrations as safety constraints"]}],"split":"dev"}
{"episode_id":"F129_T2024-01","field_index":129,"field":"object pose estimation","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Novel-object pose generalization","description":"Estimate or track 6D pose for object instances not seen during pose training, including zero-shot, one-shot, few-shot, category-agnostic, no-CAD, template-based, foundation-feature, meta-learning, and render-and-compare generalization settings.","inclusion":"Include atoms whose dominant objective is transferring pose estimation capability to previously unseen rigid object instances with little or no object-specific pose supervision.","exclusion":"If the object class is fixed and the main challenge is intra-category shape variation, assign to category-level pose and shape modeling instead.","nearest_boundary":"Category-level pose and shape modeling","representative_topics":["category-agnostic 6d pose estimation via neural process meta-learning and geometry-aware keypoint decoding","unseen object 6d pose estimation via 3d correspondences and symmetry-robust pose inference","zero-shot zero-cad 6dof tracking of unknown objects using image-only synthetic training"]},{"direction_index":1,"label":"D1","name":"Category-level pose and shape modeling","description":"Estimate pose, scale, size, canonical alignment, shape, or articulated part pose for objects drawn from a known semantic category with intra-class geometric variation.","inclusion":"Include atoms whose dominant objective is category-level 6D/9DoF pose, pose-and-shape recovery, canonical-space prediction, semantic keypoint reasoning, shape-prior alignment, or category-level refinement.","exclusion":"If the main objective is pose generalization to arbitrary unseen instances without a fixed category assumption, assign to novel-object pose generalization instead.","nearest_boundary":"Novel-object pose generalization","representative_topics":["optimal and outlier-robust category-level 3d pose and shape estimation from semantic keypoints","semi-supervised category-level 6d pose estimation in the wild using rendering and silhouette matching","iterative point cloud and shape prior alignment for category-level 9dof object pose refinement"]},{"direction_index":2,"label":"D2","name":"Domain-robust deployment pose","description":"Make pose estimation work under deployment gaps caused by synthetic-to-real transfer, source-to-target domain shift, harsh environments, or nonstandard sensing modalities such as event cameras, X-ray, LiDAR, radar, underwater imagery, spacecraft imagery, and industrial sensing.","inclusion":"Include atoms whose dominant objective is robustness to a deployment domain, sensor physics, environment, or source-target gap, including sim-to-real adaptation, synthetic-only training, test-time adaptation, and modality-specific pose formulations.","exclusion":"If the main contribution is a general RGB/RGB-D estimator architecture without a domain or modality gap as the central problem, assign to deterministic geometric estimator design instead.","nearest_boundary":"Deterministic geometric estimator design","representative_topics":["counterfactual causal inference for robust 6d pose estimation of space objects","sim-to-real 6d pose estimation for robotic grasping via synthetic rgb-d data generation","event-based synthetic-only training for satellite pose estimation under extreme illumination domain gap"]},{"direction_index":3,"label":"D3","name":"Deterministic geometric estimator design","description":"Design deterministic pose estimators based on explicit correspondences, keypoints, point-pair relations, geometric solvers, learned geometric features, transformer or attention fusion, equivariant representations, RGB-D or multi-view feature integration, and geometry-consistent neural architectures.","inclusion":"Include atoms whose dominant objective is improving the core pose estimator through correspondence learning, geometric feature representations, algebraic solvers, equivariant or invariant architecture, transformer-based pose reasoning, or deterministic feature fusion.","exclusion":"If the dominant objective is representing multiple ambiguous pose hypotheses or optimizing pose by rendering model-observation agreement, assign to inverse-graphics and probabilistic pose inference instead.","nearest_boundary":"Inverse-graphics and probabilistic pose inference","representative_topics":["graph-based 2d-to-3d feature mapping for 6dof pose estimation in cluttered scenes","sparse 3d model-guided keypoint selection for accurate single-image 6dof camera localization","two-view rgb 6d pose estimation via implicit 3d feature fusion strategies"]},{"direction_index":4,"label":"D4","name":"Inverse-graphics and probabilistic pose inference","description":"Infer, track, refine, or sample pose by modeling ambiguity or by optimizing agreement between observations and explicit, CAD, neural, rendered, implicit, or probabilistic object or scene models.","inclusion":"Include atoms whose dominant objective is pose distributions, implicit likelihoods, diffusion or score-based inference, ambiguity-aware optimization, probabilistic inverse graphics, differentiable rendering, render-and-compare, NeRF or scene-representation inversion, feature-metric rendering alignment, or iterative model-to-image pose refinement.","exclusion":"If rendering or probability modeling is not central and the method instead solves a deterministic correspondence, feature-regression, transformer, or algebraic estimator problem, assign to deterministic geometric estimator design.","nearest_boundary":"Deterministic geometric estimator design","representative_topics":["multi-view 6dof object pose refinement via differentiable renderer and iterative icp-like optimization","render-and-compare 6d pose estimation using center and curvature representations for textureless objects","6dof object pose tracking via nerf-based relocalization and feature-metric alignment"]},{"direction_index":5,"label":"D5","name":"Annotation-efficient pose learning","description":"Learn pose estimators while reducing or eliminating ground-truth pose labels, dense 3D labels, manual 2D annotations, or full supervision, using self-supervision, weak supervision, minimal annotation, contrastive learning, distillation, reconstruction, cycle consistency, pseudo-labeling, or adversarial signals.","inclusion":"Include atoms whose dominant objective is lowering annotation requirements for pose learning rather than primarily improving a supervised estimator, adapting across domains, or handling unseen objects.","exclusion":"If the main objective is synthetic-to-real transfer with labeled synthetic data, assign to domain-robust deployment pose; if the main objective is arbitrary novel-instance generalization, assign to novel-object pose generalization.","nearest_boundary":"Domain-robust deployment pose","representative_topics":["contrastive knowledge distillation for 3d-augmented image-based category-agnostic object pose estimation","instance-specific 6-dof pose estimation via geometry-assisted dataset generation from minimal annotations","self-supervised category-level 6d pose estimation via canonical shape reconstruction and geometric cycle-consistency"]},{"direction_index":6,"label":"D6","name":"Interaction and articulated-agent pose","description":"Estimate, track, or reconstruct poses in scenes where hands, human bodies, manipulated objects, contact, mutual occlusion, articulated interaction, or multi-person structure are the central target or constraint.","inclusion":"Include atoms whose dominant objective is joint hand-object pose, hand-object reconstruction, in-hand object pose, manipulation-aware pose, articulated object interaction with hands, or human body pose and shape estimation as an articulated pose-estimation target.","exclusion":"If the paper estimates object pose without explicit hand, body, contact, or interaction modeling, assign to the relevant object-only program such as category-level pose, deterministic geometric estimation, or domain-robust deployment.","nearest_boundary":"Category-level pose and shape modeling","representative_topics":["joint hand joint tracking and object sdf reconstruction from depth point cloud sequences","multi-person 3d mesh estimation via inverse kinematics and relation-aware refinement under occlusion","dense mutual attention on mesh graphs for fine-grained hand-object pose interaction modeling"]},{"direction_index":7,"label":"D7","name":"Pose data and evaluation infrastructure","description":"Construct datasets, benchmarks, annotation pipelines, evaluation protocols, bias analyses, challenge reports, surveys, reviews, or field-level measurement resources for pose estimation.","inclusion":"Include atoms whose dominant objective is resource construction, benchmarking, evaluation methodology, annotation infrastructure, comparative analysis, or synthesis of the pose-estimation field rather than proposing a new estimator.","exclusion":"If a dataset is mainly a vehicle for a new estimator, adaptation method, modality-specific algorithm, or deployment pipeline, assign to that scientific program instead.","nearest_boundary":"Domain-robust deployment pose","representative_topics":["6d pose estimation for reflective industrial metal objects via sim-to-real dataset","6d pose estimation dataset for occluded hand-held tools in imitation learning demonstrations","mixed-reality dataset for category-level 6d pose estimation of hand-occluded household containers"]}],"split":"dev"}
{"episode_id":"F129_T2024-07","field_index":129,"field":"object pose estimation","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Novel-object pose generalization","description":"Estimate or track 6D pose for object instances not seen during pose training, including zero-shot, one-shot, few-shot, category-agnostic, no-CAD, template-based, foundation-feature, meta-learning, and render-and-compare generalization settings.","inclusion":"Include atoms whose dominant objective is transferring pose estimation capability to previously unseen rigid object instances with little or no object-specific pose supervision.","exclusion":"If the object class is fixed and the main challenge is intra-category shape variation, assign to category-level pose and shape modeling instead.","nearest_boundary":"Category-level pose and shape modeling","representative_topics":["category-agnostic 6d pose estimation via neural process meta-learning and geometry-aware keypoint decoding","unseen object 6d pose estimation via 3d correspondences and symmetry-robust pose inference","zero-shot zero-cad 6dof tracking of unknown objects using image-only synthetic training"]},{"direction_index":1,"label":"D1","name":"Category-level pose and shape modeling","description":"Estimate pose, scale, size, canonical alignment, shape, or articulated part pose for objects drawn from a known semantic category with intra-class geometric variation.","inclusion":"Include atoms whose dominant objective is category-level 6D/9DoF pose, pose-and-shape recovery, canonical-space prediction, semantic keypoint reasoning, shape-prior alignment, or category-level refinement.","exclusion":"If the main objective is pose generalization to arbitrary unseen instances without a fixed category assumption, assign to novel-object pose generalization instead.","nearest_boundary":"Novel-object pose generalization","representative_topics":["optimal and outlier-robust category-level 3d pose and shape estimation from semantic keypoints","semi-supervised category-level 6d pose estimation in the wild using rendering and silhouette matching","iterative point cloud and shape prior alignment for category-level 9dof object pose refinement"]},{"direction_index":2,"label":"D2","name":"Domain-robust deployment pose","description":"Make pose estimation work under deployment gaps caused by synthetic-to-real transfer, source-to-target domain shift, harsh environments, or nonstandard sensing modalities such as event cameras, X-ray, LiDAR, radar, underwater imagery, spacecraft imagery, and industrial sensing.","inclusion":"Include atoms whose dominant objective is robustness to a deployment domain, sensor physics, environment, or source-target gap, including sim-to-real adaptation, synthetic-only training, test-time adaptation, and modality-specific pose formulations.","exclusion":"If the main contribution is a general RGB/RGB-D estimator architecture without a domain or modality gap as the central problem, assign to deterministic geometric estimator design instead.","nearest_boundary":"Deterministic geometric estimator design","representative_topics":["counterfactual causal inference for robust 6d pose estimation of space objects","sim-to-real 6d pose estimation for robotic grasping via synthetic rgb-d data generation","event-based synthetic-only training for satellite pose estimation under extreme illumination domain gap"]},{"direction_index":3,"label":"D3","name":"Deterministic geometric estimator design","description":"Design deterministic pose estimators based on explicit correspondences, keypoints, point-pair relations, geometric solvers, learned geometric features, transformer or attention fusion, equivariant representations, RGB-D or multi-view feature integration, and geometry-consistent neural architectures.","inclusion":"Include atoms whose dominant objective is improving the core pose estimator through correspondence learning, geometric feature representations, algebraic solvers, equivariant or invariant architecture, transformer-based pose reasoning, or deterministic feature fusion.","exclusion":"If the dominant objective is representing multiple ambiguous pose hypotheses or optimizing pose by rendering model-observation agreement, assign to inverse-graphics and probabilistic pose inference instead.","nearest_boundary":"Inverse-graphics and probabilistic pose inference","representative_topics":["graph-based 2d-to-3d feature mapping for 6dof pose estimation in cluttered scenes","sparse 3d model-guided keypoint selection for accurate single-image 6dof camera localization","two-view rgb 6d pose estimation via implicit 3d feature fusion strategies"]},{"direction_index":4,"label":"D4","name":"Inverse-graphics and probabilistic pose inference","description":"Infer, track, refine, or sample pose by modeling ambiguity or by optimizing agreement between observations and explicit, CAD, neural, rendered, implicit, or probabilistic object or scene models.","inclusion":"Include atoms whose dominant objective is pose distributions, implicit likelihoods, diffusion or score-based inference, ambiguity-aware optimization, probabilistic inverse graphics, differentiable rendering, render-and-compare, NeRF or scene-representation inversion, feature-metric rendering alignment, or iterative model-to-image pose refinement.","exclusion":"If rendering or probability modeling is not central and the method instead solves a deterministic correspondence, feature-regression, transformer, or algebraic estimator problem, assign to deterministic geometric estimator design.","nearest_boundary":"Deterministic geometric estimator design","representative_topics":["multi-view 6dof object pose refinement via differentiable renderer and iterative icp-like optimization","render-and-compare 6d pose estimation using center and curvature representations for textureless objects","6dof object pose tracking via nerf-based relocalization and feature-metric alignment"]},{"direction_index":5,"label":"D5","name":"Annotation-efficient pose learning","description":"Learn pose estimators while reducing or eliminating ground-truth pose labels, dense 3D labels, manual 2D annotations, or full supervision, using self-supervision, weak supervision, minimal annotation, contrastive learning, distillation, reconstruction, cycle consistency, pseudo-labeling, or adversarial signals.","inclusion":"Include atoms whose dominant objective is lowering annotation requirements for pose learning rather than primarily improving a supervised estimator, adapting across domains, or handling unseen objects.","exclusion":"If the main objective is synthetic-to-real transfer with labeled synthetic data, assign to domain-robust deployment pose; if the main objective is arbitrary novel-instance generalization, assign to novel-object pose generalization.","nearest_boundary":"Domain-robust deployment pose","representative_topics":["contrastive knowledge distillation for 3d-augmented image-based category-agnostic object pose estimation","instance-specific 6-dof pose estimation via geometry-assisted dataset generation from minimal annotations","self-supervised category-level 6d pose estimation via canonical shape reconstruction and geometric cycle-consistency"]},{"direction_index":6,"label":"D6","name":"Interaction and articulated-agent pose","description":"Estimate, track, or reconstruct poses in scenes where hands, human bodies, manipulated objects, contact, mutual occlusion, articulated interaction, or multi-person structure are the central target or constraint.","inclusion":"Include atoms whose dominant objective is joint hand-object pose, hand-object reconstruction, in-hand object pose, manipulation-aware pose, articulated object interaction with hands, or human body pose and shape estimation as an articulated pose-estimation target.","exclusion":"If the paper estimates object pose without explicit hand, body, contact, or interaction modeling, assign to the relevant object-only program such as category-level pose, deterministic geometric estimation, or domain-robust deployment.","nearest_boundary":"Category-level pose and shape modeling","representative_topics":["joint hand joint tracking and object sdf reconstruction from depth point cloud sequences","multi-person 3d mesh estimation via inverse kinematics and relation-aware refinement under occlusion","dense mutual attention on mesh graphs for fine-grained hand-object pose interaction modeling"]},{"direction_index":7,"label":"D7","name":"Pose data and evaluation infrastructure","description":"Construct datasets, benchmarks, annotation pipelines, evaluation protocols, bias analyses, challenge reports, surveys, reviews, or field-level measurement resources for pose estimation.","inclusion":"Include atoms whose dominant objective is resource construction, benchmarking, evaluation methodology, annotation infrastructure, comparative analysis, or synthesis of the pose-estimation field rather than proposing a new estimator.","exclusion":"If a dataset is mainly a vehicle for a new estimator, adaptation method, modality-specific algorithm, or deployment pipeline, assign to that scientific program instead.","nearest_boundary":"Domain-robust deployment pose","representative_topics":["6d pose estimation for reflective industrial metal objects via sim-to-real dataset","6d pose estimation dataset for occluded hand-held tools in imitation learning demonstrations","mixed-reality dataset for category-level 6d pose estimation of hand-occluded household containers"]}],"split":"dev"}
{"episode_id":"F129_T2025-01","field_index":129,"field":"object pose estimation","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Novel-object pose generalization","description":"Estimate or track 6D pose for object instances not seen during pose training, including zero-shot, one-shot, few-shot, category-agnostic, no-CAD, template-based, foundation-feature, meta-learning, and render-and-compare generalization settings.","inclusion":"Include atoms whose dominant objective is transferring pose estimation capability to previously unseen rigid object instances with little or no object-specific pose supervision.","exclusion":"If the object class is fixed and the main challenge is intra-category shape variation, assign to category-level pose and shape modeling instead.","nearest_boundary":"Category-level pose and shape modeling","representative_topics":["category-agnostic 6d pose estimation via neural process meta-learning and geometry-aware keypoint decoding","unseen object 6d pose estimation via 3d correspondences and symmetry-robust pose inference","zero-shot zero-cad 6dof tracking of unknown objects using image-only synthetic training"]},{"direction_index":1,"label":"D1","name":"Category-level pose and shape modeling","description":"Estimate pose, scale, size, canonical alignment, shape, or articulated part pose for objects drawn from a known semantic category with intra-class geometric variation.","inclusion":"Include atoms whose dominant objective is category-level 6D/9DoF pose, pose-and-shape recovery, canonical-space prediction, semantic keypoint reasoning, shape-prior alignment, or category-level refinement.","exclusion":"If the main objective is pose generalization to arbitrary unseen instances without a fixed category assumption, assign to novel-object pose generalization instead.","nearest_boundary":"Novel-object pose generalization","representative_topics":["optimal and outlier-robust category-level 3d pose and shape estimation from semantic keypoints","semi-supervised category-level 6d pose estimation in the wild using rendering and silhouette matching","iterative point cloud and shape prior alignment for category-level 9dof object pose refinement"]},{"direction_index":2,"label":"D2","name":"Domain-robust deployment pose","description":"Make pose estimation work under deployment gaps caused by synthetic-to-real transfer, source-to-target domain shift, harsh environments, or nonstandard sensing modalities such as event cameras, X-ray, LiDAR, radar, underwater imagery, spacecraft imagery, and industrial sensing.","inclusion":"Include atoms whose dominant objective is robustness to a deployment domain, sensor physics, environment, or source-target gap, including sim-to-real adaptation, synthetic-only training, test-time adaptation, and modality-specific pose formulations.","exclusion":"If the main contribution is a general RGB/RGB-D estimator architecture without a domain or modality gap as the central problem, assign to deterministic geometric estimator design instead.","nearest_boundary":"Deterministic geometric estimator design","representative_topics":["counterfactual causal inference for robust 6d pose estimation of space objects","sim-to-real 6d pose estimation for robotic grasping via synthetic rgb-d data generation","event-based synthetic-only training for satellite pose estimation under extreme illumination domain gap"]},{"direction_index":3,"label":"D3","name":"Deterministic geometric estimator design","description":"Design deterministic pose estimators based on explicit correspondences, keypoints, point-pair relations, geometric solvers, learned geometric features, transformer or attention fusion, equivariant representations, RGB-D or multi-view feature integration, and geometry-consistent neural architectures.","inclusion":"Include atoms whose dominant objective is improving the core pose estimator through correspondence learning, geometric feature representations, algebraic solvers, equivariant or invariant architecture, transformer-based pose reasoning, or deterministic feature fusion.","exclusion":"If the dominant objective is representing multiple ambiguous pose hypotheses or optimizing pose by rendering model-observation agreement, assign to inverse-graphics and probabilistic pose inference instead.","nearest_boundary":"Inverse-graphics and probabilistic pose inference","representative_topics":["graph-based 2d-to-3d feature mapping for 6dof pose estimation in cluttered scenes","sparse 3d model-guided keypoint selection for accurate single-image 6dof camera localization","two-view rgb 6d pose estimation via implicit 3d feature fusion strategies"]},{"direction_index":4,"label":"D4","name":"Inverse-graphics and probabilistic pose inference","description":"Infer, track, refine, or sample pose by modeling ambiguity or by optimizing agreement between observations and explicit, CAD, neural, rendered, implicit, or probabilistic object or scene models.","inclusion":"Include atoms whose dominant objective is pose distributions, implicit likelihoods, diffusion or score-based inference, ambiguity-aware optimization, probabilistic inverse graphics, differentiable rendering, render-and-compare, NeRF or scene-representation inversion, feature-metric rendering alignment, or iterative model-to-image pose refinement.","exclusion":"If rendering or probability modeling is not central and the method instead solves a deterministic correspondence, feature-regression, transformer, or algebraic estimator problem, assign to deterministic geometric estimator design.","nearest_boundary":"Deterministic geometric estimator design","representative_topics":["multi-view 6dof object pose refinement via differentiable renderer and iterative icp-like optimization","render-and-compare 6d pose estimation using center and curvature representations for textureless objects","6dof object pose tracking via nerf-based relocalization and feature-metric alignment"]},{"direction_index":5,"label":"D5","name":"Annotation-efficient pose learning","description":"Learn pose estimators while reducing or eliminating ground-truth pose labels, dense 3D labels, manual 2D annotations, or full supervision, using self-supervision, weak supervision, minimal annotation, contrastive learning, distillation, reconstruction, cycle consistency, pseudo-labeling, or adversarial signals.","inclusion":"Include atoms whose dominant objective is lowering annotation requirements for pose learning rather than primarily improving a supervised estimator, adapting across domains, or handling unseen objects.","exclusion":"If the main objective is synthetic-to-real transfer with labeled synthetic data, assign to domain-robust deployment pose; if the main objective is arbitrary novel-instance generalization, assign to novel-object pose generalization.","nearest_boundary":"Domain-robust deployment pose","representative_topics":["contrastive knowledge distillation for 3d-augmented image-based category-agnostic object pose estimation","instance-specific 6-dof pose estimation via geometry-assisted dataset generation from minimal annotations","self-supervised category-level 6d pose estimation via canonical shape reconstruction and geometric cycle-consistency"]},{"direction_index":6,"label":"D6","name":"Interaction and articulated-agent pose","description":"Estimate, track, or reconstruct poses in scenes where hands, human bodies, manipulated objects, contact, mutual occlusion, articulated interaction, or multi-person structure are the central target or constraint.","inclusion":"Include atoms whose dominant objective is joint hand-object pose, hand-object reconstruction, in-hand object pose, manipulation-aware pose, articulated object interaction with hands, or human body pose and shape estimation as an articulated pose-estimation target.","exclusion":"If the paper estimates object pose without explicit hand, body, contact, or interaction modeling, assign to the relevant object-only program such as category-level pose, deterministic geometric estimation, or domain-robust deployment.","nearest_boundary":"Category-level pose and shape modeling","representative_topics":["joint hand joint tracking and object sdf reconstruction from depth point cloud sequences","multi-person 3d mesh estimation via inverse kinematics and relation-aware refinement under occlusion","dense mutual attention on mesh graphs for fine-grained hand-object pose interaction modeling"]},{"direction_index":7,"label":"D7","name":"Pose data and evaluation infrastructure","description":"Construct datasets, benchmarks, annotation pipelines, evaluation protocols, bias analyses, challenge reports, surveys, reviews, or field-level measurement resources for pose estimation.","inclusion":"Include atoms whose dominant objective is resource construction, benchmarking, evaluation methodology, annotation infrastructure, comparative analysis, or synthesis of the pose-estimation field rather than proposing a new estimator.","exclusion":"If a dataset is mainly a vehicle for a new estimator, adaptation method, modality-specific algorithm, or deployment pipeline, assign to that scientific program instead.","nearest_boundary":"Domain-robust deployment pose","representative_topics":["6d pose estimation for reflective industrial metal objects via sim-to-real dataset","6d pose estimation dataset for occluded hand-held tools in imitation learning demonstrations","mixed-reality dataset for category-level 6d pose estimation of hand-occluded household containers"]}],"split":"dev"}
{"episode_id":"F129_T2025-07","field_index":129,"field":"object pose estimation","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Novel-object pose generalization","description":"Estimate or track 6D pose for object instances not seen during pose training, including zero-shot, one-shot, few-shot, category-agnostic, no-CAD, template-based, foundation-feature, meta-learning, and render-and-compare generalization settings.","inclusion":"Include atoms whose dominant objective is transferring pose estimation capability to previously unseen rigid object instances with little or no object-specific pose supervision.","exclusion":"If the object class is fixed and the main challenge is intra-category shape variation, assign to category-level pose and shape modeling instead.","nearest_boundary":"Category-level pose and shape modeling","representative_topics":["category-agnostic 6d pose estimation via neural process meta-learning and geometry-aware keypoint decoding","unseen object 6d pose estimation via 3d correspondences and symmetry-robust pose inference","zero-shot zero-cad 6dof tracking of unknown objects using image-only synthetic training"]},{"direction_index":1,"label":"D1","name":"Category-level pose and shape modeling","description":"Estimate pose, scale, size, canonical alignment, shape, or articulated part pose for objects drawn from a known semantic category with intra-class geometric variation.","inclusion":"Include atoms whose dominant objective is category-level 6D/9DoF pose, pose-and-shape recovery, canonical-space prediction, semantic keypoint reasoning, shape-prior alignment, or category-level refinement.","exclusion":"If the main objective is pose generalization to arbitrary unseen instances without a fixed category assumption, assign to novel-object pose generalization instead.","nearest_boundary":"Novel-object pose generalization","representative_topics":["optimal and outlier-robust category-level 3d pose and shape estimation from semantic keypoints","semi-supervised category-level 6d pose estimation in the wild using rendering and silhouette matching","iterative point cloud and shape prior alignment for category-level 9dof object pose refinement"]},{"direction_index":2,"label":"D2","name":"Domain-robust deployment pose","description":"Make pose estimation work under deployment gaps caused by synthetic-to-real transfer, source-to-target domain shift, harsh environments, or nonstandard sensing modalities such as event cameras, X-ray, LiDAR, radar, underwater imagery, spacecraft imagery, and industrial sensing.","inclusion":"Include atoms whose dominant objective is robustness to a deployment domain, sensor physics, environment, or source-target gap, including sim-to-real adaptation, synthetic-only training, test-time adaptation, and modality-specific pose formulations.","exclusion":"If the main contribution is a general RGB/RGB-D estimator architecture without a domain or modality gap as the central problem, assign to deterministic geometric estimator design instead.","nearest_boundary":"Deterministic geometric estimator design","representative_topics":["counterfactual causal inference for robust 6d pose estimation of space objects","sim-to-real 6d pose estimation for robotic grasping via synthetic rgb-d data generation","event-based synthetic-only training for satellite pose estimation under extreme illumination domain gap"]},{"direction_index":3,"label":"D3","name":"Deterministic geometric estimator design","description":"Design deterministic pose estimators based on explicit correspondences, keypoints, point-pair relations, geometric solvers, learned geometric features, transformer or attention fusion, equivariant representations, RGB-D or multi-view feature integration, and geometry-consistent neural architectures.","inclusion":"Include atoms whose dominant objective is improving the core pose estimator through correspondence learning, geometric feature representations, algebraic solvers, equivariant or invariant architecture, transformer-based pose reasoning, or deterministic feature fusion.","exclusion":"If the dominant objective is representing multiple ambiguous pose hypotheses or optimizing pose by rendering model-observation agreement, assign to inverse-graphics and probabilistic pose inference instead.","nearest_boundary":"Inverse-graphics and probabilistic pose inference","representative_topics":["graph-based 2d-to-3d feature mapping for 6dof pose estimation in cluttered scenes","sparse 3d model-guided keypoint selection for accurate single-image 6dof camera localization","two-view rgb 6d pose estimation via implicit 3d feature fusion strategies"]},{"direction_index":4,"label":"D4","name":"Inverse-graphics and probabilistic pose inference","description":"Infer, track, refine, or sample pose by modeling ambiguity or by optimizing agreement between observations and explicit, CAD, neural, rendered, implicit, or probabilistic object or scene models.","inclusion":"Include atoms whose dominant objective is pose distributions, implicit likelihoods, diffusion or score-based inference, ambiguity-aware optimization, probabilistic inverse graphics, differentiable rendering, render-and-compare, NeRF or scene-representation inversion, feature-metric rendering alignment, or iterative model-to-image pose refinement.","exclusion":"If rendering or probability modeling is not central and the method instead solves a deterministic correspondence, feature-regression, transformer, or algebraic estimator problem, assign to deterministic geometric estimator design.","nearest_boundary":"Deterministic geometric estimator design","representative_topics":["multi-view 6dof object pose refinement via differentiable renderer and iterative icp-like optimization","render-and-compare 6d pose estimation using center and curvature representations for textureless objects","6dof object pose tracking via nerf-based relocalization and feature-metric alignment"]},{"direction_index":5,"label":"D5","name":"Annotation-efficient pose learning","description":"Learn pose estimators while reducing or eliminating ground-truth pose labels, dense 3D labels, manual 2D annotations, or full supervision, using self-supervision, weak supervision, minimal annotation, contrastive learning, distillation, reconstruction, cycle consistency, pseudo-labeling, or adversarial signals.","inclusion":"Include atoms whose dominant objective is lowering annotation requirements for pose learning rather than primarily improving a supervised estimator, adapting across domains, or handling unseen objects.","exclusion":"If the main objective is synthetic-to-real transfer with labeled synthetic data, assign to domain-robust deployment pose; if the main objective is arbitrary novel-instance generalization, assign to novel-object pose generalization.","nearest_boundary":"Domain-robust deployment pose","representative_topics":["contrastive knowledge distillation for 3d-augmented image-based category-agnostic object pose estimation","instance-specific 6-dof pose estimation via geometry-assisted dataset generation from minimal annotations","self-supervised category-level 6d pose estimation via canonical shape reconstruction and geometric cycle-consistency"]},{"direction_index":6,"label":"D6","name":"Interaction and articulated-agent pose","description":"Estimate, track, or reconstruct poses in scenes where hands, human bodies, manipulated objects, contact, mutual occlusion, articulated interaction, or multi-person structure are the central target or constraint.","inclusion":"Include atoms whose dominant objective is joint hand-object pose, hand-object reconstruction, in-hand object pose, manipulation-aware pose, articulated object interaction with hands, or human body pose and shape estimation as an articulated pose-estimation target.","exclusion":"If the paper estimates object pose without explicit hand, body, contact, or interaction modeling, assign to the relevant object-only program such as category-level pose, deterministic geometric estimation, or domain-robust deployment.","nearest_boundary":"Category-level pose and shape modeling","representative_topics":["joint hand joint tracking and object sdf reconstruction from depth point cloud sequences","multi-person 3d mesh estimation via inverse kinematics and relation-aware refinement under occlusion","dense mutual attention on mesh graphs for fine-grained hand-object pose interaction modeling"]},{"direction_index":7,"label":"D7","name":"Pose data and evaluation infrastructure","description":"Construct datasets, benchmarks, annotation pipelines, evaluation protocols, bias analyses, challenge reports, surveys, reviews, or field-level measurement resources for pose estimation.","inclusion":"Include atoms whose dominant objective is resource construction, benchmarking, evaluation methodology, annotation infrastructure, comparative analysis, or synthesis of the pose-estimation field rather than proposing a new estimator.","exclusion":"If a dataset is mainly a vehicle for a new estimator, adaptation method, modality-specific algorithm, or deployment pipeline, assign to that scientific program instead.","nearest_boundary":"Domain-robust deployment pose","representative_topics":["6d pose estimation for reflective industrial metal objects via sim-to-real dataset","6d pose estimation dataset for occluded hand-held tools in imitation learning demonstrations","mixed-reality dataset for category-level 6d pose estimation of hand-occluded household containers"]}],"split":"dev"}
{"episode_id":"F129_T2026-01","field_index":129,"field":"object pose estimation","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Novel-object pose generalization","description":"Estimate or track 6D pose for object instances not seen during pose training, including zero-shot, one-shot, few-shot, category-agnostic, no-CAD, template-based, foundation-feature, meta-learning, and render-and-compare generalization settings.","inclusion":"Include atoms whose dominant objective is transferring pose estimation capability to previously unseen rigid object instances with little or no object-specific pose supervision.","exclusion":"If the object class is fixed and the main challenge is intra-category shape variation, assign to category-level pose and shape modeling instead.","nearest_boundary":"Category-level pose and shape modeling","representative_topics":["category-agnostic 6d pose estimation via neural process meta-learning and geometry-aware keypoint decoding","unseen object 6d pose estimation via 3d correspondences and symmetry-robust pose inference","zero-shot zero-cad 6dof tracking of unknown objects using image-only synthetic training"]},{"direction_index":1,"label":"D1","name":"Category-level pose and shape modeling","description":"Estimate pose, scale, size, canonical alignment, shape, or articulated part pose for objects drawn from a known semantic category with intra-class geometric variation.","inclusion":"Include atoms whose dominant objective is category-level 6D/9DoF pose, pose-and-shape recovery, canonical-space prediction, semantic keypoint reasoning, shape-prior alignment, or category-level refinement.","exclusion":"If the main objective is pose generalization to arbitrary unseen instances without a fixed category assumption, assign to novel-object pose generalization instead.","nearest_boundary":"Novel-object pose generalization","representative_topics":["optimal and outlier-robust category-level 3d pose and shape estimation from semantic keypoints","semi-supervised category-level 6d pose estimation in the wild using rendering and silhouette matching","iterative point cloud and shape prior alignment for category-level 9dof object pose refinement"]},{"direction_index":2,"label":"D2","name":"Domain-robust deployment pose","description":"Make pose estimation work under deployment gaps caused by synthetic-to-real transfer, source-to-target domain shift, harsh environments, or nonstandard sensing modalities such as event cameras, X-ray, LiDAR, radar, underwater imagery, spacecraft imagery, and industrial sensing.","inclusion":"Include atoms whose dominant objective is robustness to a deployment domain, sensor physics, environment, or source-target gap, including sim-to-real adaptation, synthetic-only training, test-time adaptation, and modality-specific pose formulations.","exclusion":"If the main contribution is a general RGB/RGB-D estimator architecture without a domain or modality gap as the central problem, assign to deterministic geometric estimator design instead.","nearest_boundary":"Deterministic geometric estimator design","representative_topics":["counterfactual causal inference for robust 6d pose estimation of space objects","sim-to-real 6d pose estimation for robotic grasping via synthetic rgb-d data generation","event-based synthetic-only training for satellite pose estimation under extreme illumination domain gap"]},{"direction_index":3,"label":"D3","name":"Deterministic geometric estimator design","description":"Design deterministic pose estimators based on explicit correspondences, keypoints, point-pair relations, geometric solvers, learned geometric features, transformer or attention fusion, equivariant representations, RGB-D or multi-view feature integration, and geometry-consistent neural architectures.","inclusion":"Include atoms whose dominant objective is improving the core pose estimator through correspondence learning, geometric feature representations, algebraic solvers, equivariant or invariant architecture, transformer-based pose reasoning, or deterministic feature fusion.","exclusion":"If the dominant objective is representing multiple ambiguous pose hypotheses or optimizing pose by rendering model-observation agreement, assign to inverse-graphics and probabilistic pose inference instead.","nearest_boundary":"Inverse-graphics and probabilistic pose inference","representative_topics":["graph-based 2d-to-3d feature mapping for 6dof pose estimation in cluttered scenes","sparse 3d model-guided keypoint selection for accurate single-image 6dof camera localization","two-view rgb 6d pose estimation via implicit 3d feature fusion strategies"]},{"direction_index":4,"label":"D4","name":"Inverse-graphics and probabilistic pose inference","description":"Infer, track, refine, or sample pose by modeling ambiguity or by optimizing agreement between observations and explicit, CAD, neural, rendered, implicit, or probabilistic object or scene models.","inclusion":"Include atoms whose dominant objective is pose distributions, implicit likelihoods, diffusion or score-based inference, ambiguity-aware optimization, probabilistic inverse graphics, differentiable rendering, render-and-compare, NeRF or scene-representation inversion, feature-metric rendering alignment, or iterative model-to-image pose refinement.","exclusion":"If rendering or probability modeling is not central and the method instead solves a deterministic correspondence, feature-regression, transformer, or algebraic estimator problem, assign to deterministic geometric estimator design.","nearest_boundary":"Deterministic geometric estimator design","representative_topics":["multi-view 6dof object pose refinement via differentiable renderer and iterative icp-like optimization","render-and-compare 6d pose estimation using center and curvature representations for textureless objects","6dof object pose tracking via nerf-based relocalization and feature-metric alignment"]},{"direction_index":5,"label":"D5","name":"Annotation-efficient pose learning","description":"Learn pose estimators while reducing or eliminating ground-truth pose labels, dense 3D labels, manual 2D annotations, or full supervision, using self-supervision, weak supervision, minimal annotation, contrastive learning, distillation, reconstruction, cycle consistency, pseudo-labeling, or adversarial signals.","inclusion":"Include atoms whose dominant objective is lowering annotation requirements for pose learning rather than primarily improving a supervised estimator, adapting across domains, or handling unseen objects.","exclusion":"If the main objective is synthetic-to-real transfer with labeled synthetic data, assign to domain-robust deployment pose; if the main objective is arbitrary novel-instance generalization, assign to novel-object pose generalization.","nearest_boundary":"Domain-robust deployment pose","representative_topics":["contrastive knowledge distillation for 3d-augmented image-based category-agnostic object pose estimation","instance-specific 6-dof pose estimation via geometry-assisted dataset generation from minimal annotations","self-supervised category-level 6d pose estimation via canonical shape reconstruction and geometric cycle-consistency"]},{"direction_index":6,"label":"D6","name":"Interaction and articulated-agent pose","description":"Estimate, track, or reconstruct poses in scenes where hands, human bodies, manipulated objects, contact, mutual occlusion, articulated interaction, or multi-person structure are the central target or constraint.","inclusion":"Include atoms whose dominant objective is joint hand-object pose, hand-object reconstruction, in-hand object pose, manipulation-aware pose, articulated object interaction with hands, or human body pose and shape estimation as an articulated pose-estimation target.","exclusion":"If the paper estimates object pose without explicit hand, body, contact, or interaction modeling, assign to the relevant object-only program such as category-level pose, deterministic geometric estimation, or domain-robust deployment.","nearest_boundary":"Category-level pose and shape modeling","representative_topics":["joint hand joint tracking and object sdf reconstruction from depth point cloud sequences","multi-person 3d mesh estimation via inverse kinematics and relation-aware refinement under occlusion","dense mutual attention on mesh graphs for fine-grained hand-object pose interaction modeling"]},{"direction_index":7,"label":"D7","name":"Pose data and evaluation infrastructure","description":"Construct datasets, benchmarks, annotation pipelines, evaluation protocols, bias analyses, challenge reports, surveys, reviews, or field-level measurement resources for pose estimation.","inclusion":"Include atoms whose dominant objective is resource construction, benchmarking, evaluation methodology, annotation infrastructure, comparative analysis, or synthesis of the pose-estimation field rather than proposing a new estimator.","exclusion":"If a dataset is mainly a vehicle for a new estimator, adaptation method, modality-specific algorithm, or deployment pipeline, assign to that scientific program instead.","nearest_boundary":"Domain-robust deployment pose","representative_topics":["6d pose estimation for reflective industrial metal objects via sim-to-real dataset","6d pose estimation dataset for occluded hand-held tools in imitation learning demonstrations","mixed-reality dataset for category-level 6d pose estimation of hand-occluded household containers"]}],"split":"dev"}
{"episode_id":"F130_T2024-01","field_index":130,"field":"multi-agent systems","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cooperative MARL policy learning","description":"Develop algorithms that let cooperative agents learn coordinated policies or values under shared objectives, partial observability, sparse rewards, decentralization, or centralized-training/decentralized-execution constraints.","inclusion":"Include source atoms whose primary objective is cooperative MARL credit assignment, intrinsic reward or reward/value decomposition, CTDE, value or policy factorization, decentralized cooperative actor-critic/Q-learning, parameter sharing, coordination architectures, or scalable cooperative policy optimization.","exclusion":"Work whose main objective is strategic equilibrium against self-interested agents, partner generalization to unseen teammates, explicit communication protocol design, or domain-specific infrastructure optimization belongs to the nearest alternative instead.","nearest_boundary":"Game-theoretic learning and bandits","representative_topics":["relation-aware value function factorization for ad-hoc cooperation in multi-agent reinforcement learning","consensus-based communication-free cooperation in decentralized multi-agent reinforcement learning","policy gradient methods for cooperative marl beyond value decomposition and parameter sharing"]},{"direction_index":1,"label":"D1","name":"Game-theoretic learning and bandits","description":"Compute, learn, or analyze strategic solution concepts and low-feedback multi-agent learning procedures in Markov games, stochastic games, leader-follower games, population games, and bandit settings.","inclusion":"Include source atoms whose primary objective is Nash, correlated, Stackelberg, mean-field, potential-game, no-regret, PSRO, offline game-learning, regret-optimal bandit, best-arm identification, fair bandit, or competitive/cooperative bandit analysis.","exclusion":"Work aimed mainly at changing incentives or norms to induce cooperation, rather than solving or analyzing a strategic learning model, belongs to partner adaptation, norms, and incentives.","nearest_boundary":"Partner adaptation, norms, and incentives","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","decentralized optimistic hyperpolicy mirror descent for no-regret learning in markov games","convergence guarantees and price of anarchy bounds for softmax policy gradient in markov potential games"]},{"direction_index":2,"label":"D2","name":"Partner adaptation, norms, and incentives","description":"Enable agents to coordinate with unfamiliar or changing partners and to induce cooperative, fair, norm-following, or socially aligned behavior in mixed-motive multi-agent settings.","inclusion":"Include source atoms whose primary objective is zero-shot coordination, ad hoc teamwork, partner or opponent modeling, convention inference, teammate diversity, dynamic agent populations, social dilemmas, contracts, fairness, norms, moral rewards, reputation, negotiation, or incentive design.","exclusion":"Fixed-population equilibrium computation or generic cooperative MARL without unseen-partner or incentive-alignment goals belongs to game-theoretic learning and bandits or cooperative MARL policy learning.","nearest_boundary":"Game-theoretic learning and bandits","representative_topics":["belief-based convention adaptation for cooperative ai in self-play markov games","zero-shot coordination in cooperative games via synchronous k-level reasoning and best response","market-based incentive mechanisms for cooperative learning in mixed-motive multi-agent games"]},{"direction_index":3,"label":"D3","name":"Communication and semantic alignment","description":"Design, optimize, robustify, or study the emergence and grounding of inter-agent communication protocols, messages, abstractions, symbols, and shared meanings.","inclusion":"Include source atoms whose primary objective is explicit communication scheduling, bandwidth reduction, sparse or unreliable messaging, expressive protocol design, communication-control coupling, collaborative perception sharing as communication, emergent language, semantic grounding, abstraction learning, referential signaling, or inter-agent explainability.","exclusion":"If communication is only incidental to cooperative policy learning, assign to cooperative MARL policy learning; if adversarial communication is treated primarily as certifiable security or safety, assign to assurance, verification, and resilient control.","nearest_boundary":"Assurance, verification, and resilient control","representative_topics":["universally expressive communication protocols in multi-agent reinforcement learning via gnn augmentation","decentralized attention-based communication for cooperative routing in partially observable queueing networks","cooperative multi-agent reinforcement learning with adaptive communication strategies under unreliable wireless channels"]},{"direction_index":4,"label":"D4","name":"Assurance, verification, and resilient control","description":"Establish, verify, or enforce reliable multi-agent behavior under formal specifications, adversarial perturbations, privacy constraints, attacks, uncertainty, or networked dynamical-system requirements.","inclusion":"Include source atoms whose primary objective is robustness, safety, security, privacy, attack resilience, barrier or specification-guided control, formal logic, rational verification, epistemic or strategy reasoning, decidability, model checking, distributed observers, estimation, filtering, consensus, or output regulation.","exclusion":"If the main contribution is ordinary communication-efficiency or cooperative learning performance without assurance or formal-property focus, assign to communication and semantic alignment or cooperative MARL policy learning.","nearest_boundary":"Communication and semantic alignment","representative_topics":["certifiably robust multi-agent policy learning under adversarial communication manipulation","process calculus for asynchronous epistemic interaction modeling in multi-agent systems","rational verification complexity in multiagent systems under equilibrium and social welfare constraints"]},{"direction_index":5,"label":"D5","name":"Embodied planning, robotics, and forecasting","description":"Plan, control, perceive, or forecast the motion and interactions of embodied agents, robots, vehicles, swarms, or agents moving in shared spatial environments.","inclusion":"Include source atoms whose primary objective is MAPF, MAPD, task-and-path planning, collision-free execution, coverage, flocking, pursuit-evasion, formation control, swarm coordination, multi-robot control, social navigation, collaborative perception, trajectory forecasting, motion prediction, or relation discovery for multi-agent scenes.","exclusion":"Infrastructure service optimization such as traffic, wireless, energy, logistics, or supply-chain management belongs to resource-system optimization and MAS design unless the core objective is embodied collision-free planning/control or motion prediction.","nearest_boundary":"Resource-system optimization and MAS design","representative_topics":["decentralized multi-target pursuit via fuzzy task decomposition and cooperative coevolution in partially observable markov games","dynamic group-aware relational reasoning for multi-agent trajectory prediction","adaptive swarm entrapping via artificial potential field formation control"]},{"direction_index":6,"label":"D6","name":"Resource-system optimization and MAS design","description":"Use multi-agent methods to optimize concrete infrastructure, resource-allocation, scheduling, marketplace, logistics, network, energy, transportation, industrial, or general MAS design problems.","inclusion":"Include source atoms whose primary objective is domain-system optimization, multi-agent scheduling, negotiation frameworks, vehicle routing, air-traffic conflict resolution, VNF placement, supply-chain replanning, cloud marketplaces, black-box or combinatorial optimization, Bayesian optimization, TSP-style decomposition, or classical MAS planning/application frameworks.","exclusion":"If the central objective is reusable cooperative MARL algorithm design independent of the domain, assign to cooperative MARL policy learning; if the objective is spatial collision-free multi-agent execution, assign to embodied planning, robotics, and forecasting.","nearest_boundary":"Cooperative MARL policy learning","representative_topics":["multi-agent neural rewriting for vehicle routing under partial observability and limited cost disclosure","multiagent graph convolutional reinforcement learning for automated tactical conflict resolution in air traffic control","bdi multi-agent negotiation framework for cloud marketplace resource acquisition"]},{"direction_index":7,"label":"D7","name":"Agent societies and evaluation infrastructure","description":"Construct, simulate, orchestrate, benchmark, or evaluate multi-agent systems and agent societies, including LLM-based agent collectives and scalable experimental platforms.","inclusion":"Include source atoms whose primary objective is LLM-agent collaboration, multi-agent debate or role-play, multi-agent QA or orchestration platforms, agent-based social or behavioral simulation, believable generative agents, surveys, benchmarks, datasets, simulators, evaluation protocols, scalable simulation infrastructure, or standardized MARL environments.","exclusion":"If LLM or simulation agents are mainly used to solve a control, planning, communication, or infrastructure optimization problem, assign to that primary problem program rather than to system/evaluation construction.","nearest_boundary":"Resource-system optimization and MAS design","representative_topics":["scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research","distributed local simulation with influence-aware models for scalable multi-agent reinforcement learning","vectorized simulator framework for scalable multi-robot multi-agent reinforcement learning benchmarking"]}],"split":"dev"}
{"episode_id":"F130_T2024-07","field_index":130,"field":"multi-agent systems","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cooperative MARL policy learning","description":"Develop algorithms that let cooperative agents learn coordinated policies or values under shared objectives, partial observability, sparse rewards, decentralization, or centralized-training/decentralized-execution constraints.","inclusion":"Include source atoms whose primary objective is cooperative MARL credit assignment, intrinsic reward or reward/value decomposition, CTDE, value or policy factorization, decentralized cooperative actor-critic/Q-learning, parameter sharing, coordination architectures, or scalable cooperative policy optimization.","exclusion":"Work whose main objective is strategic equilibrium against self-interested agents, partner generalization to unseen teammates, explicit communication protocol design, or domain-specific infrastructure optimization belongs to the nearest alternative instead.","nearest_boundary":"Game-theoretic learning and bandits","representative_topics":["relation-aware value function factorization for ad-hoc cooperation in multi-agent reinforcement learning","consensus-based communication-free cooperation in decentralized multi-agent reinforcement learning","policy gradient methods for cooperative marl beyond value decomposition and parameter sharing"]},{"direction_index":1,"label":"D1","name":"Game-theoretic learning and bandits","description":"Compute, learn, or analyze strategic solution concepts and low-feedback multi-agent learning procedures in Markov games, stochastic games, leader-follower games, population games, and bandit settings.","inclusion":"Include source atoms whose primary objective is Nash, correlated, Stackelberg, mean-field, potential-game, no-regret, PSRO, offline game-learning, regret-optimal bandit, best-arm identification, fair bandit, or competitive/cooperative bandit analysis.","exclusion":"Work aimed mainly at changing incentives or norms to induce cooperation, rather than solving or analyzing a strategic learning model, belongs to partner adaptation, norms, and incentives.","nearest_boundary":"Partner adaptation, norms, and incentives","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","decentralized optimistic hyperpolicy mirror descent for no-regret learning in markov games","convergence guarantees and price of anarchy bounds for softmax policy gradient in markov potential games"]},{"direction_index":2,"label":"D2","name":"Partner adaptation, norms, and incentives","description":"Enable agents to coordinate with unfamiliar or changing partners and to induce cooperative, fair, norm-following, or socially aligned behavior in mixed-motive multi-agent settings.","inclusion":"Include source atoms whose primary objective is zero-shot coordination, ad hoc teamwork, partner or opponent modeling, convention inference, teammate diversity, dynamic agent populations, social dilemmas, contracts, fairness, norms, moral rewards, reputation, negotiation, or incentive design.","exclusion":"Fixed-population equilibrium computation or generic cooperative MARL without unseen-partner or incentive-alignment goals belongs to game-theoretic learning and bandits or cooperative MARL policy learning.","nearest_boundary":"Game-theoretic learning and bandits","representative_topics":["belief-based convention adaptation for cooperative ai in self-play markov games","zero-shot coordination in cooperative games via synchronous k-level reasoning and best response","market-based incentive mechanisms for cooperative learning in mixed-motive multi-agent games"]},{"direction_index":3,"label":"D3","name":"Communication and semantic alignment","description":"Design, optimize, robustify, or study the emergence and grounding of inter-agent communication protocols, messages, abstractions, symbols, and shared meanings.","inclusion":"Include source atoms whose primary objective is explicit communication scheduling, bandwidth reduction, sparse or unreliable messaging, expressive protocol design, communication-control coupling, collaborative perception sharing as communication, emergent language, semantic grounding, abstraction learning, referential signaling, or inter-agent explainability.","exclusion":"If communication is only incidental to cooperative policy learning, assign to cooperative MARL policy learning; if adversarial communication is treated primarily as certifiable security or safety, assign to assurance, verification, and resilient control.","nearest_boundary":"Assurance, verification, and resilient control","representative_topics":["universally expressive communication protocols in multi-agent reinforcement learning via gnn augmentation","decentralized attention-based communication for cooperative routing in partially observable queueing networks","cooperative multi-agent reinforcement learning with adaptive communication strategies under unreliable wireless channels"]},{"direction_index":4,"label":"D4","name":"Assurance, verification, and resilient control","description":"Establish, verify, or enforce reliable multi-agent behavior under formal specifications, adversarial perturbations, privacy constraints, attacks, uncertainty, or networked dynamical-system requirements.","inclusion":"Include source atoms whose primary objective is robustness, safety, security, privacy, attack resilience, barrier or specification-guided control, formal logic, rational verification, epistemic or strategy reasoning, decidability, model checking, distributed observers, estimation, filtering, consensus, or output regulation.","exclusion":"If the main contribution is ordinary communication-efficiency or cooperative learning performance without assurance or formal-property focus, assign to communication and semantic alignment or cooperative MARL policy learning.","nearest_boundary":"Communication and semantic alignment","representative_topics":["certifiably robust multi-agent policy learning under adversarial communication manipulation","process calculus for asynchronous epistemic interaction modeling in multi-agent systems","rational verification complexity in multiagent systems under equilibrium and social welfare constraints"]},{"direction_index":5,"label":"D5","name":"Embodied planning, robotics, and forecasting","description":"Plan, control, perceive, or forecast the motion and interactions of embodied agents, robots, vehicles, swarms, or agents moving in shared spatial environments.","inclusion":"Include source atoms whose primary objective is MAPF, MAPD, task-and-path planning, collision-free execution, coverage, flocking, pursuit-evasion, formation control, swarm coordination, multi-robot control, social navigation, collaborative perception, trajectory forecasting, motion prediction, or relation discovery for multi-agent scenes.","exclusion":"Infrastructure service optimization such as traffic, wireless, energy, logistics, or supply-chain management belongs to resource-system optimization and MAS design unless the core objective is embodied collision-free planning/control or motion prediction.","nearest_boundary":"Resource-system optimization and MAS design","representative_topics":["decentralized multi-target pursuit via fuzzy task decomposition and cooperative coevolution in partially observable markov games","dynamic group-aware relational reasoning for multi-agent trajectory prediction","adaptive swarm entrapping via artificial potential field formation control"]},{"direction_index":6,"label":"D6","name":"Resource-system optimization and MAS design","description":"Use multi-agent methods to optimize concrete infrastructure, resource-allocation, scheduling, marketplace, logistics, network, energy, transportation, industrial, or general MAS design problems.","inclusion":"Include source atoms whose primary objective is domain-system optimization, multi-agent scheduling, negotiation frameworks, vehicle routing, air-traffic conflict resolution, VNF placement, supply-chain replanning, cloud marketplaces, black-box or combinatorial optimization, Bayesian optimization, TSP-style decomposition, or classical MAS planning/application frameworks.","exclusion":"If the central objective is reusable cooperative MARL algorithm design independent of the domain, assign to cooperative MARL policy learning; if the objective is spatial collision-free multi-agent execution, assign to embodied planning, robotics, and forecasting.","nearest_boundary":"Cooperative MARL policy learning","representative_topics":["multi-agent neural rewriting for vehicle routing under partial observability and limited cost disclosure","multiagent graph convolutional reinforcement learning for automated tactical conflict resolution in air traffic control","bdi multi-agent negotiation framework for cloud marketplace resource acquisition"]},{"direction_index":7,"label":"D7","name":"Agent societies and evaluation infrastructure","description":"Construct, simulate, orchestrate, benchmark, or evaluate multi-agent systems and agent societies, including LLM-based agent collectives and scalable experimental platforms.","inclusion":"Include source atoms whose primary objective is LLM-agent collaboration, multi-agent debate or role-play, multi-agent QA or orchestration platforms, agent-based social or behavioral simulation, believable generative agents, surveys, benchmarks, datasets, simulators, evaluation protocols, scalable simulation infrastructure, or standardized MARL environments.","exclusion":"If LLM or simulation agents are mainly used to solve a control, planning, communication, or infrastructure optimization problem, assign to that primary problem program rather than to system/evaluation construction.","nearest_boundary":"Resource-system optimization and MAS design","representative_topics":["scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research","distributed local simulation with influence-aware models for scalable multi-agent reinforcement learning","vectorized simulator framework for scalable multi-robot multi-agent reinforcement learning benchmarking"]}],"split":"dev"}
{"episode_id":"F130_T2025-01","field_index":130,"field":"multi-agent systems","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cooperative MARL policy learning","description":"Develop algorithms that let cooperative agents learn coordinated policies or values under shared objectives, partial observability, sparse rewards, decentralization, or centralized-training/decentralized-execution constraints.","inclusion":"Include source atoms whose primary objective is cooperative MARL credit assignment, intrinsic reward or reward/value decomposition, CTDE, value or policy factorization, decentralized cooperative actor-critic/Q-learning, parameter sharing, coordination architectures, or scalable cooperative policy optimization.","exclusion":"Work whose main objective is strategic equilibrium against self-interested agents, partner generalization to unseen teammates, explicit communication protocol design, or domain-specific infrastructure optimization belongs to the nearest alternative instead.","nearest_boundary":"Game-theoretic learning and bandits","representative_topics":["relation-aware value function factorization for ad-hoc cooperation in multi-agent reinforcement learning","consensus-based communication-free cooperation in decentralized multi-agent reinforcement learning","policy gradient methods for cooperative marl beyond value decomposition and parameter sharing"]},{"direction_index":1,"label":"D1","name":"Game-theoretic learning and bandits","description":"Compute, learn, or analyze strategic solution concepts and low-feedback multi-agent learning procedures in Markov games, stochastic games, leader-follower games, population games, and bandit settings.","inclusion":"Include source atoms whose primary objective is Nash, correlated, Stackelberg, mean-field, potential-game, no-regret, PSRO, offline game-learning, regret-optimal bandit, best-arm identification, fair bandit, or competitive/cooperative bandit analysis.","exclusion":"Work aimed mainly at changing incentives or norms to induce cooperation, rather than solving or analyzing a strategic learning model, belongs to partner adaptation, norms, and incentives.","nearest_boundary":"Partner adaptation, norms, and incentives","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","decentralized optimistic hyperpolicy mirror descent for no-regret learning in markov games","convergence guarantees and price of anarchy bounds for softmax policy gradient in markov potential games"]},{"direction_index":2,"label":"D2","name":"Partner adaptation, norms, and incentives","description":"Enable agents to coordinate with unfamiliar or changing partners and to induce cooperative, fair, norm-following, or socially aligned behavior in mixed-motive multi-agent settings.","inclusion":"Include source atoms whose primary objective is zero-shot coordination, ad hoc teamwork, partner or opponent modeling, convention inference, teammate diversity, dynamic agent populations, social dilemmas, contracts, fairness, norms, moral rewards, reputation, negotiation, or incentive design.","exclusion":"Fixed-population equilibrium computation or generic cooperative MARL without unseen-partner or incentive-alignment goals belongs to game-theoretic learning and bandits or cooperative MARL policy learning.","nearest_boundary":"Game-theoretic learning and bandits","representative_topics":["belief-based convention adaptation for cooperative ai in self-play markov games","zero-shot coordination in cooperative games via synchronous k-level reasoning and best response","market-based incentive mechanisms for cooperative learning in mixed-motive multi-agent games"]},{"direction_index":3,"label":"D3","name":"Communication and semantic alignment","description":"Design, optimize, robustify, or study the emergence and grounding of inter-agent communication protocols, messages, abstractions, symbols, and shared meanings.","inclusion":"Include source atoms whose primary objective is explicit communication scheduling, bandwidth reduction, sparse or unreliable messaging, expressive protocol design, communication-control coupling, collaborative perception sharing as communication, emergent language, semantic grounding, abstraction learning, referential signaling, or inter-agent explainability.","exclusion":"If communication is only incidental to cooperative policy learning, assign to cooperative MARL policy learning; if adversarial communication is treated primarily as certifiable security or safety, assign to assurance, verification, and resilient control.","nearest_boundary":"Assurance, verification, and resilient control","representative_topics":["universally expressive communication protocols in multi-agent reinforcement learning via gnn augmentation","decentralized attention-based communication for cooperative routing in partially observable queueing networks","cooperative multi-agent reinforcement learning with adaptive communication strategies under unreliable wireless channels"]},{"direction_index":4,"label":"D4","name":"Assurance, verification, and resilient control","description":"Establish, verify, or enforce reliable multi-agent behavior under formal specifications, adversarial perturbations, privacy constraints, attacks, uncertainty, or networked dynamical-system requirements.","inclusion":"Include source atoms whose primary objective is robustness, safety, security, privacy, attack resilience, barrier or specification-guided control, formal logic, rational verification, epistemic or strategy reasoning, decidability, model checking, distributed observers, estimation, filtering, consensus, or output regulation.","exclusion":"If the main contribution is ordinary communication-efficiency or cooperative learning performance without assurance or formal-property focus, assign to communication and semantic alignment or cooperative MARL policy learning.","nearest_boundary":"Communication and semantic alignment","representative_topics":["certifiably robust multi-agent policy learning under adversarial communication manipulation","process calculus for asynchronous epistemic interaction modeling in multi-agent systems","rational verification complexity in multiagent systems under equilibrium and social welfare constraints"]},{"direction_index":5,"label":"D5","name":"Embodied planning, robotics, and forecasting","description":"Plan, control, perceive, or forecast the motion and interactions of embodied agents, robots, vehicles, swarms, or agents moving in shared spatial environments.","inclusion":"Include source atoms whose primary objective is MAPF, MAPD, task-and-path planning, collision-free execution, coverage, flocking, pursuit-evasion, formation control, swarm coordination, multi-robot control, social navigation, collaborative perception, trajectory forecasting, motion prediction, or relation discovery for multi-agent scenes.","exclusion":"Infrastructure service optimization such as traffic, wireless, energy, logistics, or supply-chain management belongs to resource-system optimization and MAS design unless the core objective is embodied collision-free planning/control or motion prediction.","nearest_boundary":"Resource-system optimization and MAS design","representative_topics":["decentralized multi-target pursuit via fuzzy task decomposition and cooperative coevolution in partially observable markov games","dynamic group-aware relational reasoning for multi-agent trajectory prediction","adaptive swarm entrapping via artificial potential field formation control"]},{"direction_index":6,"label":"D6","name":"Resource-system optimization and MAS design","description":"Use multi-agent methods to optimize concrete infrastructure, resource-allocation, scheduling, marketplace, logistics, network, energy, transportation, industrial, or general MAS design problems.","inclusion":"Include source atoms whose primary objective is domain-system optimization, multi-agent scheduling, negotiation frameworks, vehicle routing, air-traffic conflict resolution, VNF placement, supply-chain replanning, cloud marketplaces, black-box or combinatorial optimization, Bayesian optimization, TSP-style decomposition, or classical MAS planning/application frameworks.","exclusion":"If the central objective is reusable cooperative MARL algorithm design independent of the domain, assign to cooperative MARL policy learning; if the objective is spatial collision-free multi-agent execution, assign to embodied planning, robotics, and forecasting.","nearest_boundary":"Cooperative MARL policy learning","representative_topics":["multi-agent neural rewriting for vehicle routing under partial observability and limited cost disclosure","multiagent graph convolutional reinforcement learning for automated tactical conflict resolution in air traffic control","bdi multi-agent negotiation framework for cloud marketplace resource acquisition"]},{"direction_index":7,"label":"D7","name":"Agent societies and evaluation infrastructure","description":"Construct, simulate, orchestrate, benchmark, or evaluate multi-agent systems and agent societies, including LLM-based agent collectives and scalable experimental platforms.","inclusion":"Include source atoms whose primary objective is LLM-agent collaboration, multi-agent debate or role-play, multi-agent QA or orchestration platforms, agent-based social or behavioral simulation, believable generative agents, surveys, benchmarks, datasets, simulators, evaluation protocols, scalable simulation infrastructure, or standardized MARL environments.","exclusion":"If LLM or simulation agents are mainly used to solve a control, planning, communication, or infrastructure optimization problem, assign to that primary problem program rather than to system/evaluation construction.","nearest_boundary":"Resource-system optimization and MAS design","representative_topics":["scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research","distributed local simulation with influence-aware models for scalable multi-agent reinforcement learning","vectorized simulator framework for scalable multi-robot multi-agent reinforcement learning benchmarking"]}],"split":"dev"}
{"episode_id":"F130_T2025-07","field_index":130,"field":"multi-agent systems","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cooperative MARL policy learning","description":"Develop algorithms that let cooperative agents learn coordinated policies or values under shared objectives, partial observability, sparse rewards, decentralization, or centralized-training/decentralized-execution constraints.","inclusion":"Include source atoms whose primary objective is cooperative MARL credit assignment, intrinsic reward or reward/value decomposition, CTDE, value or policy factorization, decentralized cooperative actor-critic/Q-learning, parameter sharing, coordination architectures, or scalable cooperative policy optimization.","exclusion":"Work whose main objective is strategic equilibrium against self-interested agents, partner generalization to unseen teammates, explicit communication protocol design, or domain-specific infrastructure optimization belongs to the nearest alternative instead.","nearest_boundary":"Game-theoretic learning and bandits","representative_topics":["relation-aware value function factorization for ad-hoc cooperation in multi-agent reinforcement learning","consensus-based communication-free cooperation in decentralized multi-agent reinforcement learning","policy gradient methods for cooperative marl beyond value decomposition and parameter sharing"]},{"direction_index":1,"label":"D1","name":"Game-theoretic learning and bandits","description":"Compute, learn, or analyze strategic solution concepts and low-feedback multi-agent learning procedures in Markov games, stochastic games, leader-follower games, population games, and bandit settings.","inclusion":"Include source atoms whose primary objective is Nash, correlated, Stackelberg, mean-field, potential-game, no-regret, PSRO, offline game-learning, regret-optimal bandit, best-arm identification, fair bandit, or competitive/cooperative bandit analysis.","exclusion":"Work aimed mainly at changing incentives or norms to induce cooperation, rather than solving or analyzing a strategic learning model, belongs to partner adaptation, norms, and incentives.","nearest_boundary":"Partner adaptation, norms, and incentives","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","decentralized optimistic hyperpolicy mirror descent for no-regret learning in markov games","convergence guarantees and price of anarchy bounds for softmax policy gradient in markov potential games"]},{"direction_index":2,"label":"D2","name":"Partner adaptation, norms, and incentives","description":"Enable agents to coordinate with unfamiliar or changing partners and to induce cooperative, fair, norm-following, or socially aligned behavior in mixed-motive multi-agent settings.","inclusion":"Include source atoms whose primary objective is zero-shot coordination, ad hoc teamwork, partner or opponent modeling, convention inference, teammate diversity, dynamic agent populations, social dilemmas, contracts, fairness, norms, moral rewards, reputation, negotiation, or incentive design.","exclusion":"Fixed-population equilibrium computation or generic cooperative MARL without unseen-partner or incentive-alignment goals belongs to game-theoretic learning and bandits or cooperative MARL policy learning.","nearest_boundary":"Game-theoretic learning and bandits","representative_topics":["belief-based convention adaptation for cooperative ai in self-play markov games","zero-shot coordination in cooperative games via synchronous k-level reasoning and best response","market-based incentive mechanisms for cooperative learning in mixed-motive multi-agent games"]},{"direction_index":3,"label":"D3","name":"Communication and semantic alignment","description":"Design, optimize, robustify, or study the emergence and grounding of inter-agent communication protocols, messages, abstractions, symbols, and shared meanings.","inclusion":"Include source atoms whose primary objective is explicit communication scheduling, bandwidth reduction, sparse or unreliable messaging, expressive protocol design, communication-control coupling, collaborative perception sharing as communication, emergent language, semantic grounding, abstraction learning, referential signaling, or inter-agent explainability.","exclusion":"If communication is only incidental to cooperative policy learning, assign to cooperative MARL policy learning; if adversarial communication is treated primarily as certifiable security or safety, assign to assurance, verification, and resilient control.","nearest_boundary":"Assurance, verification, and resilient control","representative_topics":["universally expressive communication protocols in multi-agent reinforcement learning via gnn augmentation","decentralized attention-based communication for cooperative routing in partially observable queueing networks","cooperative multi-agent reinforcement learning with adaptive communication strategies under unreliable wireless channels"]},{"direction_index":4,"label":"D4","name":"Assurance, verification, and resilient control","description":"Establish, verify, or enforce reliable multi-agent behavior under formal specifications, adversarial perturbations, privacy constraints, attacks, uncertainty, or networked dynamical-system requirements.","inclusion":"Include source atoms whose primary objective is robustness, safety, security, privacy, attack resilience, barrier or specification-guided control, formal logic, rational verification, epistemic or strategy reasoning, decidability, model checking, distributed observers, estimation, filtering, consensus, or output regulation.","exclusion":"If the main contribution is ordinary communication-efficiency or cooperative learning performance without assurance or formal-property focus, assign to communication and semantic alignment or cooperative MARL policy learning.","nearest_boundary":"Communication and semantic alignment","representative_topics":["certifiably robust multi-agent policy learning under adversarial communication manipulation","process calculus for asynchronous epistemic interaction modeling in multi-agent systems","rational verification complexity in multiagent systems under equilibrium and social welfare constraints"]},{"direction_index":5,"label":"D5","name":"Embodied planning, robotics, and forecasting","description":"Plan, control, perceive, or forecast the motion and interactions of embodied agents, robots, vehicles, swarms, or agents moving in shared spatial environments.","inclusion":"Include source atoms whose primary objective is MAPF, MAPD, task-and-path planning, collision-free execution, coverage, flocking, pursuit-evasion, formation control, swarm coordination, multi-robot control, social navigation, collaborative perception, trajectory forecasting, motion prediction, or relation discovery for multi-agent scenes.","exclusion":"Infrastructure service optimization such as traffic, wireless, energy, logistics, or supply-chain management belongs to resource-system optimization and MAS design unless the core objective is embodied collision-free planning/control or motion prediction.","nearest_boundary":"Resource-system optimization and MAS design","representative_topics":["decentralized multi-target pursuit via fuzzy task decomposition and cooperative coevolution in partially observable markov games","dynamic group-aware relational reasoning for multi-agent trajectory prediction","adaptive swarm entrapping via artificial potential field formation control"]},{"direction_index":6,"label":"D6","name":"Resource-system optimization and MAS design","description":"Use multi-agent methods to optimize concrete infrastructure, resource-allocation, scheduling, marketplace, logistics, network, energy, transportation, industrial, or general MAS design problems.","inclusion":"Include source atoms whose primary objective is domain-system optimization, multi-agent scheduling, negotiation frameworks, vehicle routing, air-traffic conflict resolution, VNF placement, supply-chain replanning, cloud marketplaces, black-box or combinatorial optimization, Bayesian optimization, TSP-style decomposition, or classical MAS planning/application frameworks.","exclusion":"If the central objective is reusable cooperative MARL algorithm design independent of the domain, assign to cooperative MARL policy learning; if the objective is spatial collision-free multi-agent execution, assign to embodied planning, robotics, and forecasting.","nearest_boundary":"Cooperative MARL policy learning","representative_topics":["multi-agent neural rewriting for vehicle routing under partial observability and limited cost disclosure","multiagent graph convolutional reinforcement learning for automated tactical conflict resolution in air traffic control","bdi multi-agent negotiation framework for cloud marketplace resource acquisition"]},{"direction_index":7,"label":"D7","name":"Agent societies and evaluation infrastructure","description":"Construct, simulate, orchestrate, benchmark, or evaluate multi-agent systems and agent societies, including LLM-based agent collectives and scalable experimental platforms.","inclusion":"Include source atoms whose primary objective is LLM-agent collaboration, multi-agent debate or role-play, multi-agent QA or orchestration platforms, agent-based social or behavioral simulation, believable generative agents, surveys, benchmarks, datasets, simulators, evaluation protocols, scalable simulation infrastructure, or standardized MARL environments.","exclusion":"If LLM or simulation agents are mainly used to solve a control, planning, communication, or infrastructure optimization problem, assign to that primary problem program rather than to system/evaluation construction.","nearest_boundary":"Resource-system optimization and MAS design","representative_topics":["scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research","distributed local simulation with influence-aware models for scalable multi-agent reinforcement learning","vectorized simulator framework for scalable multi-robot multi-agent reinforcement learning benchmarking"]}],"split":"dev"}
{"episode_id":"F130_T2026-01","field_index":130,"field":"multi-agent systems","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Cooperative MARL policy learning","description":"Develop algorithms that let cooperative agents learn coordinated policies or values under shared objectives, partial observability, sparse rewards, decentralization, or centralized-training/decentralized-execution constraints.","inclusion":"Include source atoms whose primary objective is cooperative MARL credit assignment, intrinsic reward or reward/value decomposition, CTDE, value or policy factorization, decentralized cooperative actor-critic/Q-learning, parameter sharing, coordination architectures, or scalable cooperative policy optimization.","exclusion":"Work whose main objective is strategic equilibrium against self-interested agents, partner generalization to unseen teammates, explicit communication protocol design, or domain-specific infrastructure optimization belongs to the nearest alternative instead.","nearest_boundary":"Game-theoretic learning and bandits","representative_topics":["relation-aware value function factorization for ad-hoc cooperation in multi-agent reinforcement learning","consensus-based communication-free cooperation in decentralized multi-agent reinforcement learning","policy gradient methods for cooperative marl beyond value decomposition and parameter sharing"]},{"direction_index":1,"label":"D1","name":"Game-theoretic learning and bandits","description":"Compute, learn, or analyze strategic solution concepts and low-feedback multi-agent learning procedures in Markov games, stochastic games, leader-follower games, population games, and bandit settings.","inclusion":"Include source atoms whose primary objective is Nash, correlated, Stackelberg, mean-field, potential-game, no-regret, PSRO, offline game-learning, regret-optimal bandit, best-arm identification, fair bandit, or competitive/cooperative bandit analysis.","exclusion":"Work aimed mainly at changing incentives or norms to induce cooperation, rather than solving or analyzing a strategic learning model, belongs to partner adaptation, norms, and incentives.","nearest_boundary":"Partner adaptation, norms, and incentives","representative_topics":["offline multi-agent markov games via strategy-wise bonus and convexity-based sample complexity","decentralized optimistic hyperpolicy mirror descent for no-regret learning in markov games","convergence guarantees and price of anarchy bounds for softmax policy gradient in markov potential games"]},{"direction_index":2,"label":"D2","name":"Partner adaptation, norms, and incentives","description":"Enable agents to coordinate with unfamiliar or changing partners and to induce cooperative, fair, norm-following, or socially aligned behavior in mixed-motive multi-agent settings.","inclusion":"Include source atoms whose primary objective is zero-shot coordination, ad hoc teamwork, partner or opponent modeling, convention inference, teammate diversity, dynamic agent populations, social dilemmas, contracts, fairness, norms, moral rewards, reputation, negotiation, or incentive design.","exclusion":"Fixed-population equilibrium computation or generic cooperative MARL without unseen-partner or incentive-alignment goals belongs to game-theoretic learning and bandits or cooperative MARL policy learning.","nearest_boundary":"Game-theoretic learning and bandits","representative_topics":["belief-based convention adaptation for cooperative ai in self-play markov games","zero-shot coordination in cooperative games via synchronous k-level reasoning and best response","market-based incentive mechanisms for cooperative learning in mixed-motive multi-agent games"]},{"direction_index":3,"label":"D3","name":"Communication and semantic alignment","description":"Design, optimize, robustify, or study the emergence and grounding of inter-agent communication protocols, messages, abstractions, symbols, and shared meanings.","inclusion":"Include source atoms whose primary objective is explicit communication scheduling, bandwidth reduction, sparse or unreliable messaging, expressive protocol design, communication-control coupling, collaborative perception sharing as communication, emergent language, semantic grounding, abstraction learning, referential signaling, or inter-agent explainability.","exclusion":"If communication is only incidental to cooperative policy learning, assign to cooperative MARL policy learning; if adversarial communication is treated primarily as certifiable security or safety, assign to assurance, verification, and resilient control.","nearest_boundary":"Assurance, verification, and resilient control","representative_topics":["universally expressive communication protocols in multi-agent reinforcement learning via gnn augmentation","decentralized attention-based communication for cooperative routing in partially observable queueing networks","cooperative multi-agent reinforcement learning with adaptive communication strategies under unreliable wireless channels"]},{"direction_index":4,"label":"D4","name":"Assurance, verification, and resilient control","description":"Establish, verify, or enforce reliable multi-agent behavior under formal specifications, adversarial perturbations, privacy constraints, attacks, uncertainty, or networked dynamical-system requirements.","inclusion":"Include source atoms whose primary objective is robustness, safety, security, privacy, attack resilience, barrier or specification-guided control, formal logic, rational verification, epistemic or strategy reasoning, decidability, model checking, distributed observers, estimation, filtering, consensus, or output regulation.","exclusion":"If the main contribution is ordinary communication-efficiency or cooperative learning performance without assurance or formal-property focus, assign to communication and semantic alignment or cooperative MARL policy learning.","nearest_boundary":"Communication and semantic alignment","representative_topics":["certifiably robust multi-agent policy learning under adversarial communication manipulation","process calculus for asynchronous epistemic interaction modeling in multi-agent systems","rational verification complexity in multiagent systems under equilibrium and social welfare constraints"]},{"direction_index":5,"label":"D5","name":"Embodied planning, robotics, and forecasting","description":"Plan, control, perceive, or forecast the motion and interactions of embodied agents, robots, vehicles, swarms, or agents moving in shared spatial environments.","inclusion":"Include source atoms whose primary objective is MAPF, MAPD, task-and-path planning, collision-free execution, coverage, flocking, pursuit-evasion, formation control, swarm coordination, multi-robot control, social navigation, collaborative perception, trajectory forecasting, motion prediction, or relation discovery for multi-agent scenes.","exclusion":"Infrastructure service optimization such as traffic, wireless, energy, logistics, or supply-chain management belongs to resource-system optimization and MAS design unless the core objective is embodied collision-free planning/control or motion prediction.","nearest_boundary":"Resource-system optimization and MAS design","representative_topics":["decentralized multi-target pursuit via fuzzy task decomposition and cooperative coevolution in partially observable markov games","dynamic group-aware relational reasoning for multi-agent trajectory prediction","adaptive swarm entrapping via artificial potential field formation control"]},{"direction_index":6,"label":"D6","name":"Resource-system optimization and MAS design","description":"Use multi-agent methods to optimize concrete infrastructure, resource-allocation, scheduling, marketplace, logistics, network, energy, transportation, industrial, or general MAS design problems.","inclusion":"Include source atoms whose primary objective is domain-system optimization, multi-agent scheduling, negotiation frameworks, vehicle routing, air-traffic conflict resolution, VNF placement, supply-chain replanning, cloud marketplaces, black-box or combinatorial optimization, Bayesian optimization, TSP-style decomposition, or classical MAS planning/application frameworks.","exclusion":"If the central objective is reusable cooperative MARL algorithm design independent of the domain, assign to cooperative MARL policy learning; if the objective is spatial collision-free multi-agent execution, assign to embodied planning, robotics, and forecasting.","nearest_boundary":"Cooperative MARL policy learning","representative_topics":["multi-agent neural rewriting for vehicle routing under partial observability and limited cost disclosure","multiagent graph convolutional reinforcement learning for automated tactical conflict resolution in air traffic control","bdi multi-agent negotiation framework for cloud marketplace resource acquisition"]},{"direction_index":7,"label":"D7","name":"Agent societies and evaluation infrastructure","description":"Construct, simulate, orchestrate, benchmark, or evaluate multi-agent systems and agent societies, including LLM-based agent collectives and scalable experimental platforms.","inclusion":"Include source atoms whose primary objective is LLM-agent collaboration, multi-agent debate or role-play, multi-agent QA or orchestration platforms, agent-based social or behavioral simulation, believable generative agents, surveys, benchmarks, datasets, simulators, evaluation protocols, scalable simulation infrastructure, or standardized MARL environments.","exclusion":"If LLM or simulation agents are mainly used to solve a control, planning, communication, or infrastructure optimization problem, assign to that primary problem program rather than to system/evaluation construction.","nearest_boundary":"Resource-system optimization and MAS design","representative_topics":["scalable multi-agent driving benchmark for partial-observable coordination and theory of mind research","distributed local simulation with influence-aware models for scalable multi-agent reinforcement learning","vectorized simulator framework for scalable multi-robot multi-agent reinforcement learning benchmarking"]}],"split":"dev"}
{"episode_id":"F136_T2024-01","field_index":136,"field":"knowledge graph completion","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Inductive and low-resource completion","description":"Develop KG completion methods that generalize beyond the training graph or operate with very limited support, including unseen entities, unseen relations, emerging graph components, disconnected subgraphs, and few-shot link or relation prediction.","inclusion":"Assign atoms here when the central objective is out-of-training-graph generalization, emerging-component reasoning, fully inductive triple prediction, or few-shot completion for sparse entities or relations.","exclusion":"Fixed-entity/relation transductive link prediction belongs to core static model learning; temporally annotated few-shot forecasting belongs to temporal completion when temporal dynamics dominate; formal logical-query generalization belongs to KG reasoning and query answering when query-operator semantics dominate.","nearest_boundary":"Core static model learning","representative_topics":["unifying factorisation-based and message-passing models for inductive knowledge graph completion","inductive knowledge graph reasoning for continuously emerging multi-batch entities via walk-based aggregation","hierarchical relational learning for few-shot knowledge graph completion via multi-level meta relation representation"]},{"direction_index":1,"label":"D1","name":"Core static model learning","description":"Improve standard static KG completion by designing representations, neural architectures, scoring functions, geometric or algebraic embedding spaces, and training objectives for fixed binary subject-predicate-object graphs.","inclusion":"Assign atoms here when the primary objective is better transductive link or relation prediction through embeddings, GNNs, transformers, convolutions, attention, factorization, geometric/algebraic spaces, regularization, contrastive/self-supervised learning, negative-sampling improvements, denoising, or bias-aware training, without a specialized temporal, inductive, multimodal, logical, or hyper-relational regime being central.","exclusion":"Unseen-entity or few-shot generalization belongs to inductive and low-resource completion; PLM/LLM or non-textual evidence fusion belongs to external-evidence completion; explicit rule/query semantics belong to KG reasoning and query answering.","nearest_boundary":"Inductive and low-resource completion","representative_topics":["ultrahyperbolic geometry embeddings for heterogeneous knowledge graph topological structures","spatio-functional knowledge graph embedding for joint composition and hierarchy inference patterns","attention-based entity representation learning with semantic path context for knowledge graph relation prediction"]},{"direction_index":2,"label":"D2","name":"Temporal completion and forecasting","description":"Predict missing, current, or future facts in temporally annotated KGs by modeling timestamps, intervals, historical events, time-varying structure, temporal causality, and event sequences.","inclusion":"Assign atoms here when the central scientific objective is temporal KG interpolation, extrapolation, future event prediction, temporal representation learning, temporal rule use, or temporal multi-hop reasoning.","exclusion":"Static snapshot link prediction belongs to core static model learning; n-ary or qualifier modeling belongs to hyper-relational completion when the non-binary fact structure is central; unseen-entity temporal work belongs here only when temporal dynamics dominate over inductive generalization.","nearest_boundary":"Hyper-relational completion","representative_topics":["autoregressive temporal knowledge graph reasoning via curvature-variable hyperbolic space embeddings","meta-learning for few-shot temporal knowledge graph reasoning over emerging entities","temporal knowledge graph future prediction using historical structure matching and background knowledge"]},{"direction_index":3,"label":"D3","name":"Hyper-relational completion","description":"Complete and reason over KGs whose facts extend beyond ordinary binary triples, including qualified facts, n-ary relations, RDF-star or nested triples, bi-level fact structures, and fact-level numeric literals.","inclusion":"Assign atoms here when the primary objective is link prediction, entity typing, or representation learning for hyper-relational, qualified, n-ary, nested-triple, or numeric-qualified facts, and modeling the extra fact components is the main challenge.","exclusion":"Ordinary timestamped triples belong to temporal completion when time dynamics are central; standard binary link prediction with entity attributes belongs to core static model learning or external-evidence completion depending on the evidence source.","nearest_boundary":"Temporal completion and forecasting","representative_topics":["dual-view hyper-relational knowledge graph embedding for link prediction and entity typing","hyper-relational knowledge graph completion via bidirectional base triple and qualifier representation learning","bi-level knowledge graph embedding for triplet prediction and conditional link prediction"]},{"direction_index":4,"label":"D4","name":"External-evidence completion","description":"Complete KGs by integrating structured triples with pretrained language models, large language models, textual descriptions, retrieval and reading signals, images, web evidence, biological or molecular modalities, video, or other heterogeneous external inputs.","inclusion":"Assign atoms here when the primary evaluated objective is missing-link, relation-label, fact, or KB completion and the main contribution is fusing, aligning, retrieving, prompting, or reading external textual, language-model, multimodal, or heterogeneous evidence.","exclusion":"Natural-language question answering or fact verification over KGs belongs to KG reasoning and query answering; KG construction from text belongs to KG resource creation and assessment; purely structural models with incidental descriptions belong to core static model learning.","nearest_boundary":"KG reasoning and query answering","representative_topics":["language model-based knowledge graph embeddings for long-tail entity representation via contrastive learning","sparse knowledge graph completion via variational em fusion of text and structure knowledge","hypercomplex space integration of knowledge graph and language model embeddings for link prediction"]},{"direction_index":5,"label":"D5","name":"KG reasoning and query answering","description":"Answer complex queries or perform explicit reasoning over KGs using logical operators, symbolic rules, constraints, neuro-symbolic methods, KG-grounded retrieval, LLM reasoning pipelines, or natural-language QA systems.","inclusion":"Assign atoms here when the primary objective is complex logical query answering, rule-guided KGC, differentiable rule learning, logical constraint integration, explainable rule aggregation, natural-language KGQA, KG-grounded fact verification, or multi-step LLM reasoning over KG paths or subgraphs.","exclusion":"Single-triple link prediction without explicit logical, rule, or QA semantics belongs to core static model learning; PLM/LLM use for direct link prediction belongs to external-evidence completion; KG construction or extraction pipelines belong to KG resource creation and assessment.","nearest_boundary":"External-evidence completion","representative_topics":["transformer-based masked pre-training for complex logical query reasoning over knowledge graphs","knowledge representation learning for logical query answering over multi-view knowledge graphs","inductive logical query answering in knowledge graphs using graph neural network representations"]},{"direction_index":6,"label":"D6","name":"Entity-centric inference","description":"Infer entity-level properties or cross-graph entity correspondences, including entity alignment, multilingual or heterogeneous KG matching, continual alignment under graph growth, and entity type prediction.","inclusion":"Assign atoms here when the primary task is identifying equivalent entities across KGs or assigning semantic type labels to entities using KG structure, descriptions, auxiliary relations, multilingual evidence, or heterogeneous graph features.","exclusion":"Triple-level missing-link prediction belongs to the relevant completion program; KG merging as a construction pipeline belongs to KG resource creation and assessment; multimodal link prediction belongs to external-evidence completion unless entity matching is the main target.","nearest_boundary":"Core static model learning","representative_topics":["continual entity alignment for growing knowledge graphs via inductive neighbor-based representation learning","entity type prediction in knowledge graphs via graph walk embeddings and textual descriptions","scalable knowledge graph entity alignment via subgraph partitioning, reconstruction, and embedding merging"]},{"direction_index":7,"label":"D7","name":"KG resource creation and assessment","description":"Create, expand, manage, benchmark, evaluate, analyze, or survey KGs, KG schemas, datasets, toolkits, and completion systems as research infrastructure rather than proposing a primary completion model.","inclusion":"Assign atoms here when the main contribution is KG construction, schema expansion or evolution, human-in-the-loop graph growth, KG generation from text, bootstrapping, schema collection, benchmark design, evaluation protocols, leaderboard tooling, metric critique, dataset analysis, theoretical or empirical model analysis, libraries, or systematic surveys.","exclusion":"A paper proposing a new completion model with standard evaluation belongs to that model’s primary objective; direct link prediction over a fixed graph belongs to core static model learning or a specialized completion regime; KG-grounded QA systems belong to KG reasoning and query answering.","nearest_boundary":"Core static model learning","representative_topics":["theoretical analysis of negative sampling hyperparameters for knowledge graph embedding scoring functions","scoring function discovery and evaluation bias in knowledge graph link prediction","explainable interactive leaderboard for fine-grained knowledge graph completion evaluation"]}],"split":"dev"}
{"episode_id":"F136_T2024-07","field_index":136,"field":"knowledge graph completion","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Inductive and low-resource completion","description":"Develop KG completion methods that generalize beyond the training graph or operate with very limited support, including unseen entities, unseen relations, emerging graph components, disconnected subgraphs, and few-shot link or relation prediction.","inclusion":"Assign atoms here when the central objective is out-of-training-graph generalization, emerging-component reasoning, fully inductive triple prediction, or few-shot completion for sparse entities or relations.","exclusion":"Fixed-entity/relation transductive link prediction belongs to core static model learning; temporally annotated few-shot forecasting belongs to temporal completion when temporal dynamics dominate; formal logical-query generalization belongs to KG reasoning and query answering when query-operator semantics dominate.","nearest_boundary":"Core static model learning","representative_topics":["unifying factorisation-based and message-passing models for inductive knowledge graph completion","inductive knowledge graph reasoning for continuously emerging multi-batch entities via walk-based aggregation","hierarchical relational learning for few-shot knowledge graph completion via multi-level meta relation representation"]},{"direction_index":1,"label":"D1","name":"Core static model learning","description":"Improve standard static KG completion by designing representations, neural architectures, scoring functions, geometric or algebraic embedding spaces, and training objectives for fixed binary subject-predicate-object graphs.","inclusion":"Assign atoms here when the primary objective is better transductive link or relation prediction through embeddings, GNNs, transformers, convolutions, attention, factorization, geometric/algebraic spaces, regularization, contrastive/self-supervised learning, negative-sampling improvements, denoising, or bias-aware training, without a specialized temporal, inductive, multimodal, logical, or hyper-relational regime being central.","exclusion":"Unseen-entity or few-shot generalization belongs to inductive and low-resource completion; PLM/LLM or non-textual evidence fusion belongs to external-evidence completion; explicit rule/query semantics belong to KG reasoning and query answering.","nearest_boundary":"Inductive and low-resource completion","representative_topics":["ultrahyperbolic geometry embeddings for heterogeneous knowledge graph topological structures","spatio-functional knowledge graph embedding for joint composition and hierarchy inference patterns","attention-based entity representation learning with semantic path context for knowledge graph relation prediction"]},{"direction_index":2,"label":"D2","name":"Temporal completion and forecasting","description":"Predict missing, current, or future facts in temporally annotated KGs by modeling timestamps, intervals, historical events, time-varying structure, temporal causality, and event sequences.","inclusion":"Assign atoms here when the central scientific objective is temporal KG interpolation, extrapolation, future event prediction, temporal representation learning, temporal rule use, or temporal multi-hop reasoning.","exclusion":"Static snapshot link prediction belongs to core static model learning; n-ary or qualifier modeling belongs to hyper-relational completion when the non-binary fact structure is central; unseen-entity temporal work belongs here only when temporal dynamics dominate over inductive generalization.","nearest_boundary":"Hyper-relational completion","representative_topics":["autoregressive temporal knowledge graph reasoning via curvature-variable hyperbolic space embeddings","meta-learning for few-shot temporal knowledge graph reasoning over emerging entities","temporal knowledge graph future prediction using historical structure matching and background knowledge"]},{"direction_index":3,"label":"D3","name":"Hyper-relational completion","description":"Complete and reason over KGs whose facts extend beyond ordinary binary triples, including qualified facts, n-ary relations, RDF-star or nested triples, bi-level fact structures, and fact-level numeric literals.","inclusion":"Assign atoms here when the primary objective is link prediction, entity typing, or representation learning for hyper-relational, qualified, n-ary, nested-triple, or numeric-qualified facts, and modeling the extra fact components is the main challenge.","exclusion":"Ordinary timestamped triples belong to temporal completion when time dynamics are central; standard binary link prediction with entity attributes belongs to core static model learning or external-evidence completion depending on the evidence source.","nearest_boundary":"Temporal completion and forecasting","representative_topics":["dual-view hyper-relational knowledge graph embedding for link prediction and entity typing","hyper-relational knowledge graph completion via bidirectional base triple and qualifier representation learning","bi-level knowledge graph embedding for triplet prediction and conditional link prediction"]},{"direction_index":4,"label":"D4","name":"External-evidence completion","description":"Complete KGs by integrating structured triples with pretrained language models, large language models, textual descriptions, retrieval and reading signals, images, web evidence, biological or molecular modalities, video, or other heterogeneous external inputs.","inclusion":"Assign atoms here when the primary evaluated objective is missing-link, relation-label, fact, or KB completion and the main contribution is fusing, aligning, retrieving, prompting, or reading external textual, language-model, multimodal, or heterogeneous evidence.","exclusion":"Natural-language question answering or fact verification over KGs belongs to KG reasoning and query answering; KG construction from text belongs to KG resource creation and assessment; purely structural models with incidental descriptions belong to core static model learning.","nearest_boundary":"KG reasoning and query answering","representative_topics":["language model-based knowledge graph embeddings for long-tail entity representation via contrastive learning","sparse knowledge graph completion via variational em fusion of text and structure knowledge","hypercomplex space integration of knowledge graph and language model embeddings for link prediction"]},{"direction_index":5,"label":"D5","name":"KG reasoning and query answering","description":"Answer complex queries or perform explicit reasoning over KGs using logical operators, symbolic rules, constraints, neuro-symbolic methods, KG-grounded retrieval, LLM reasoning pipelines, or natural-language QA systems.","inclusion":"Assign atoms here when the primary objective is complex logical query answering, rule-guided KGC, differentiable rule learning, logical constraint integration, explainable rule aggregation, natural-language KGQA, KG-grounded fact verification, or multi-step LLM reasoning over KG paths or subgraphs.","exclusion":"Single-triple link prediction without explicit logical, rule, or QA semantics belongs to core static model learning; PLM/LLM use for direct link prediction belongs to external-evidence completion; KG construction or extraction pipelines belong to KG resource creation and assessment.","nearest_boundary":"External-evidence completion","representative_topics":["transformer-based masked pre-training for complex logical query reasoning over knowledge graphs","knowledge representation learning for logical query answering over multi-view knowledge graphs","inductive logical query answering in knowledge graphs using graph neural network representations"]},{"direction_index":6,"label":"D6","name":"Entity-centric inference","description":"Infer entity-level properties or cross-graph entity correspondences, including entity alignment, multilingual or heterogeneous KG matching, continual alignment under graph growth, and entity type prediction.","inclusion":"Assign atoms here when the primary task is identifying equivalent entities across KGs or assigning semantic type labels to entities using KG structure, descriptions, auxiliary relations, multilingual evidence, or heterogeneous graph features.","exclusion":"Triple-level missing-link prediction belongs to the relevant completion program; KG merging as a construction pipeline belongs to KG resource creation and assessment; multimodal link prediction belongs to external-evidence completion unless entity matching is the main target.","nearest_boundary":"Core static model learning","representative_topics":["continual entity alignment for growing knowledge graphs via inductive neighbor-based representation learning","entity type prediction in knowledge graphs via graph walk embeddings and textual descriptions","scalable knowledge graph entity alignment via subgraph partitioning, reconstruction, and embedding merging"]},{"direction_index":7,"label":"D7","name":"KG resource creation and assessment","description":"Create, expand, manage, benchmark, evaluate, analyze, or survey KGs, KG schemas, datasets, toolkits, and completion systems as research infrastructure rather than proposing a primary completion model.","inclusion":"Assign atoms here when the main contribution is KG construction, schema expansion or evolution, human-in-the-loop graph growth, KG generation from text, bootstrapping, schema collection, benchmark design, evaluation protocols, leaderboard tooling, metric critique, dataset analysis, theoretical or empirical model analysis, libraries, or systematic surveys.","exclusion":"A paper proposing a new completion model with standard evaluation belongs to that model’s primary objective; direct link prediction over a fixed graph belongs to core static model learning or a specialized completion regime; KG-grounded QA systems belong to KG reasoning and query answering.","nearest_boundary":"Core static model learning","representative_topics":["theoretical analysis of negative sampling hyperparameters for knowledge graph embedding scoring functions","scoring function discovery and evaluation bias in knowledge graph link prediction","explainable interactive leaderboard for fine-grained knowledge graph completion evaluation"]}],"split":"dev"}
{"episode_id":"F136_T2025-01","field_index":136,"field":"knowledge graph completion","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Inductive and low-resource completion","description":"Develop KG completion methods that generalize beyond the training graph or operate with very limited support, including unseen entities, unseen relations, emerging graph components, disconnected subgraphs, and few-shot link or relation prediction.","inclusion":"Assign atoms here when the central objective is out-of-training-graph generalization, emerging-component reasoning, fully inductive triple prediction, or few-shot completion for sparse entities or relations.","exclusion":"Fixed-entity/relation transductive link prediction belongs to core static model learning; temporally annotated few-shot forecasting belongs to temporal completion when temporal dynamics dominate; formal logical-query generalization belongs to KG reasoning and query answering when query-operator semantics dominate.","nearest_boundary":"Core static model learning","representative_topics":["unifying factorisation-based and message-passing models for inductive knowledge graph completion","inductive knowledge graph reasoning for continuously emerging multi-batch entities via walk-based aggregation","hierarchical relational learning for few-shot knowledge graph completion via multi-level meta relation representation"]},{"direction_index":1,"label":"D1","name":"Core static model learning","description":"Improve standard static KG completion by designing representations, neural architectures, scoring functions, geometric or algebraic embedding spaces, and training objectives for fixed binary subject-predicate-object graphs.","inclusion":"Assign atoms here when the primary objective is better transductive link or relation prediction through embeddings, GNNs, transformers, convolutions, attention, factorization, geometric/algebraic spaces, regularization, contrastive/self-supervised learning, negative-sampling improvements, denoising, or bias-aware training, without a specialized temporal, inductive, multimodal, logical, or hyper-relational regime being central.","exclusion":"Unseen-entity or few-shot generalization belongs to inductive and low-resource completion; PLM/LLM or non-textual evidence fusion belongs to external-evidence completion; explicit rule/query semantics belong to KG reasoning and query answering.","nearest_boundary":"Inductive and low-resource completion","representative_topics":["ultrahyperbolic geometry embeddings for heterogeneous knowledge graph topological structures","spatio-functional knowledge graph embedding for joint composition and hierarchy inference patterns","attention-based entity representation learning with semantic path context for knowledge graph relation prediction"]},{"direction_index":2,"label":"D2","name":"Temporal completion and forecasting","description":"Predict missing, current, or future facts in temporally annotated KGs by modeling timestamps, intervals, historical events, time-varying structure, temporal causality, and event sequences.","inclusion":"Assign atoms here when the central scientific objective is temporal KG interpolation, extrapolation, future event prediction, temporal representation learning, temporal rule use, or temporal multi-hop reasoning.","exclusion":"Static snapshot link prediction belongs to core static model learning; n-ary or qualifier modeling belongs to hyper-relational completion when the non-binary fact structure is central; unseen-entity temporal work belongs here only when temporal dynamics dominate over inductive generalization.","nearest_boundary":"Hyper-relational completion","representative_topics":["autoregressive temporal knowledge graph reasoning via curvature-variable hyperbolic space embeddings","meta-learning for few-shot temporal knowledge graph reasoning over emerging entities","temporal knowledge graph future prediction using historical structure matching and background knowledge"]},{"direction_index":3,"label":"D3","name":"Hyper-relational completion","description":"Complete and reason over KGs whose facts extend beyond ordinary binary triples, including qualified facts, n-ary relations, RDF-star or nested triples, bi-level fact structures, and fact-level numeric literals.","inclusion":"Assign atoms here when the primary objective is link prediction, entity typing, or representation learning for hyper-relational, qualified, n-ary, nested-triple, or numeric-qualified facts, and modeling the extra fact components is the main challenge.","exclusion":"Ordinary timestamped triples belong to temporal completion when time dynamics are central; standard binary link prediction with entity attributes belongs to core static model learning or external-evidence completion depending on the evidence source.","nearest_boundary":"Temporal completion and forecasting","representative_topics":["dual-view hyper-relational knowledge graph embedding for link prediction and entity typing","hyper-relational knowledge graph completion via bidirectional base triple and qualifier representation learning","bi-level knowledge graph embedding for triplet prediction and conditional link prediction"]},{"direction_index":4,"label":"D4","name":"External-evidence completion","description":"Complete KGs by integrating structured triples with pretrained language models, large language models, textual descriptions, retrieval and reading signals, images, web evidence, biological or molecular modalities, video, or other heterogeneous external inputs.","inclusion":"Assign atoms here when the primary evaluated objective is missing-link, relation-label, fact, or KB completion and the main contribution is fusing, aligning, retrieving, prompting, or reading external textual, language-model, multimodal, or heterogeneous evidence.","exclusion":"Natural-language question answering or fact verification over KGs belongs to KG reasoning and query answering; KG construction from text belongs to KG resource creation and assessment; purely structural models with incidental descriptions belong to core static model learning.","nearest_boundary":"KG reasoning and query answering","representative_topics":["language model-based knowledge graph embeddings for long-tail entity representation via contrastive learning","sparse knowledge graph completion via variational em fusion of text and structure knowledge","hypercomplex space integration of knowledge graph and language model embeddings for link prediction"]},{"direction_index":5,"label":"D5","name":"KG reasoning and query answering","description":"Answer complex queries or perform explicit reasoning over KGs using logical operators, symbolic rules, constraints, neuro-symbolic methods, KG-grounded retrieval, LLM reasoning pipelines, or natural-language QA systems.","inclusion":"Assign atoms here when the primary objective is complex logical query answering, rule-guided KGC, differentiable rule learning, logical constraint integration, explainable rule aggregation, natural-language KGQA, KG-grounded fact verification, or multi-step LLM reasoning over KG paths or subgraphs.","exclusion":"Single-triple link prediction without explicit logical, rule, or QA semantics belongs to core static model learning; PLM/LLM use for direct link prediction belongs to external-evidence completion; KG construction or extraction pipelines belong to KG resource creation and assessment.","nearest_boundary":"External-evidence completion","representative_topics":["transformer-based masked pre-training for complex logical query reasoning over knowledge graphs","knowledge representation learning for logical query answering over multi-view knowledge graphs","inductive logical query answering in knowledge graphs using graph neural network representations"]},{"direction_index":6,"label":"D6","name":"Entity-centric inference","description":"Infer entity-level properties or cross-graph entity correspondences, including entity alignment, multilingual or heterogeneous KG matching, continual alignment under graph growth, and entity type prediction.","inclusion":"Assign atoms here when the primary task is identifying equivalent entities across KGs or assigning semantic type labels to entities using KG structure, descriptions, auxiliary relations, multilingual evidence, or heterogeneous graph features.","exclusion":"Triple-level missing-link prediction belongs to the relevant completion program; KG merging as a construction pipeline belongs to KG resource creation and assessment; multimodal link prediction belongs to external-evidence completion unless entity matching is the main target.","nearest_boundary":"Core static model learning","representative_topics":["continual entity alignment for growing knowledge graphs via inductive neighbor-based representation learning","entity type prediction in knowledge graphs via graph walk embeddings and textual descriptions","scalable knowledge graph entity alignment via subgraph partitioning, reconstruction, and embedding merging"]},{"direction_index":7,"label":"D7","name":"KG resource creation and assessment","description":"Create, expand, manage, benchmark, evaluate, analyze, or survey KGs, KG schemas, datasets, toolkits, and completion systems as research infrastructure rather than proposing a primary completion model.","inclusion":"Assign atoms here when the main contribution is KG construction, schema expansion or evolution, human-in-the-loop graph growth, KG generation from text, bootstrapping, schema collection, benchmark design, evaluation protocols, leaderboard tooling, metric critique, dataset analysis, theoretical or empirical model analysis, libraries, or systematic surveys.","exclusion":"A paper proposing a new completion model with standard evaluation belongs to that model’s primary objective; direct link prediction over a fixed graph belongs to core static model learning or a specialized completion regime; KG-grounded QA systems belong to KG reasoning and query answering.","nearest_boundary":"Core static model learning","representative_topics":["theoretical analysis of negative sampling hyperparameters for knowledge graph embedding scoring functions","scoring function discovery and evaluation bias in knowledge graph link prediction","explainable interactive leaderboard for fine-grained knowledge graph completion evaluation"]}],"split":"dev"}
{"episode_id":"F136_T2025-07","field_index":136,"field":"knowledge graph completion","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Inductive and low-resource completion","description":"Develop KG completion methods that generalize beyond the training graph or operate with very limited support, including unseen entities, unseen relations, emerging graph components, disconnected subgraphs, and few-shot link or relation prediction.","inclusion":"Assign atoms here when the central objective is out-of-training-graph generalization, emerging-component reasoning, fully inductive triple prediction, or few-shot completion for sparse entities or relations.","exclusion":"Fixed-entity/relation transductive link prediction belongs to core static model learning; temporally annotated few-shot forecasting belongs to temporal completion when temporal dynamics dominate; formal logical-query generalization belongs to KG reasoning and query answering when query-operator semantics dominate.","nearest_boundary":"Core static model learning","representative_topics":["unifying factorisation-based and message-passing models for inductive knowledge graph completion","inductive knowledge graph reasoning for continuously emerging multi-batch entities via walk-based aggregation","hierarchical relational learning for few-shot knowledge graph completion via multi-level meta relation representation"]},{"direction_index":1,"label":"D1","name":"Core static model learning","description":"Improve standard static KG completion by designing representations, neural architectures, scoring functions, geometric or algebraic embedding spaces, and training objectives for fixed binary subject-predicate-object graphs.","inclusion":"Assign atoms here when the primary objective is better transductive link or relation prediction through embeddings, GNNs, transformers, convolutions, attention, factorization, geometric/algebraic spaces, regularization, contrastive/self-supervised learning, negative-sampling improvements, denoising, or bias-aware training, without a specialized temporal, inductive, multimodal, logical, or hyper-relational regime being central.","exclusion":"Unseen-entity or few-shot generalization belongs to inductive and low-resource completion; PLM/LLM or non-textual evidence fusion belongs to external-evidence completion; explicit rule/query semantics belong to KG reasoning and query answering.","nearest_boundary":"Inductive and low-resource completion","representative_topics":["ultrahyperbolic geometry embeddings for heterogeneous knowledge graph topological structures","spatio-functional knowledge graph embedding for joint composition and hierarchy inference patterns","attention-based entity representation learning with semantic path context for knowledge graph relation prediction"]},{"direction_index":2,"label":"D2","name":"Temporal completion and forecasting","description":"Predict missing, current, or future facts in temporally annotated KGs by modeling timestamps, intervals, historical events, time-varying structure, temporal causality, and event sequences.","inclusion":"Assign atoms here when the central scientific objective is temporal KG interpolation, extrapolation, future event prediction, temporal representation learning, temporal rule use, or temporal multi-hop reasoning.","exclusion":"Static snapshot link prediction belongs to core static model learning; n-ary or qualifier modeling belongs to hyper-relational completion when the non-binary fact structure is central; unseen-entity temporal work belongs here only when temporal dynamics dominate over inductive generalization.","nearest_boundary":"Hyper-relational completion","representative_topics":["autoregressive temporal knowledge graph reasoning via curvature-variable hyperbolic space embeddings","meta-learning for few-shot temporal knowledge graph reasoning over emerging entities","temporal knowledge graph future prediction using historical structure matching and background knowledge"]},{"direction_index":3,"label":"D3","name":"Hyper-relational completion","description":"Complete and reason over KGs whose facts extend beyond ordinary binary triples, including qualified facts, n-ary relations, RDF-star or nested triples, bi-level fact structures, and fact-level numeric literals.","inclusion":"Assign atoms here when the primary objective is link prediction, entity typing, or representation learning for hyper-relational, qualified, n-ary, nested-triple, or numeric-qualified facts, and modeling the extra fact components is the main challenge.","exclusion":"Ordinary timestamped triples belong to temporal completion when time dynamics are central; standard binary link prediction with entity attributes belongs to core static model learning or external-evidence completion depending on the evidence source.","nearest_boundary":"Temporal completion and forecasting","representative_topics":["dual-view hyper-relational knowledge graph embedding for link prediction and entity typing","hyper-relational knowledge graph completion via bidirectional base triple and qualifier representation learning","bi-level knowledge graph embedding for triplet prediction and conditional link prediction"]},{"direction_index":4,"label":"D4","name":"External-evidence completion","description":"Complete KGs by integrating structured triples with pretrained language models, large language models, textual descriptions, retrieval and reading signals, images, web evidence, biological or molecular modalities, video, or other heterogeneous external inputs.","inclusion":"Assign atoms here when the primary evaluated objective is missing-link, relation-label, fact, or KB completion and the main contribution is fusing, aligning, retrieving, prompting, or reading external textual, language-model, multimodal, or heterogeneous evidence.","exclusion":"Natural-language question answering or fact verification over KGs belongs to KG reasoning and query answering; KG construction from text belongs to KG resource creation and assessment; purely structural models with incidental descriptions belong to core static model learning.","nearest_boundary":"KG reasoning and query answering","representative_topics":["language model-based knowledge graph embeddings for long-tail entity representation via contrastive learning","sparse knowledge graph completion via variational em fusion of text and structure knowledge","hypercomplex space integration of knowledge graph and language model embeddings for link prediction"]},{"direction_index":5,"label":"D5","name":"KG reasoning and query answering","description":"Answer complex queries or perform explicit reasoning over KGs using logical operators, symbolic rules, constraints, neuro-symbolic methods, KG-grounded retrieval, LLM reasoning pipelines, or natural-language QA systems.","inclusion":"Assign atoms here when the primary objective is complex logical query answering, rule-guided KGC, differentiable rule learning, logical constraint integration, explainable rule aggregation, natural-language KGQA, KG-grounded fact verification, or multi-step LLM reasoning over KG paths or subgraphs.","exclusion":"Single-triple link prediction without explicit logical, rule, or QA semantics belongs to core static model learning; PLM/LLM use for direct link prediction belongs to external-evidence completion; KG construction or extraction pipelines belong to KG resource creation and assessment.","nearest_boundary":"External-evidence completion","representative_topics":["transformer-based masked pre-training for complex logical query reasoning over knowledge graphs","knowledge representation learning for logical query answering over multi-view knowledge graphs","inductive logical query answering in knowledge graphs using graph neural network representations"]},{"direction_index":6,"label":"D6","name":"Entity-centric inference","description":"Infer entity-level properties or cross-graph entity correspondences, including entity alignment, multilingual or heterogeneous KG matching, continual alignment under graph growth, and entity type prediction.","inclusion":"Assign atoms here when the primary task is identifying equivalent entities across KGs or assigning semantic type labels to entities using KG structure, descriptions, auxiliary relations, multilingual evidence, or heterogeneous graph features.","exclusion":"Triple-level missing-link prediction belongs to the relevant completion program; KG merging as a construction pipeline belongs to KG resource creation and assessment; multimodal link prediction belongs to external-evidence completion unless entity matching is the main target.","nearest_boundary":"Core static model learning","representative_topics":["continual entity alignment for growing knowledge graphs via inductive neighbor-based representation learning","entity type prediction in knowledge graphs via graph walk embeddings and textual descriptions","scalable knowledge graph entity alignment via subgraph partitioning, reconstruction, and embedding merging"]},{"direction_index":7,"label":"D7","name":"KG resource creation and assessment","description":"Create, expand, manage, benchmark, evaluate, analyze, or survey KGs, KG schemas, datasets, toolkits, and completion systems as research infrastructure rather than proposing a primary completion model.","inclusion":"Assign atoms here when the main contribution is KG construction, schema expansion or evolution, human-in-the-loop graph growth, KG generation from text, bootstrapping, schema collection, benchmark design, evaluation protocols, leaderboard tooling, metric critique, dataset analysis, theoretical or empirical model analysis, libraries, or systematic surveys.","exclusion":"A paper proposing a new completion model with standard evaluation belongs to that model’s primary objective; direct link prediction over a fixed graph belongs to core static model learning or a specialized completion regime; KG-grounded QA systems belong to KG reasoning and query answering.","nearest_boundary":"Core static model learning","representative_topics":["theoretical analysis of negative sampling hyperparameters for knowledge graph embedding scoring functions","scoring function discovery and evaluation bias in knowledge graph link prediction","explainable interactive leaderboard for fine-grained knowledge graph completion evaluation"]}],"split":"dev"}
{"episode_id":"F136_T2026-01","field_index":136,"field":"knowledge graph completion","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Inductive and low-resource completion","description":"Develop KG completion methods that generalize beyond the training graph or operate with very limited support, including unseen entities, unseen relations, emerging graph components, disconnected subgraphs, and few-shot link or relation prediction.","inclusion":"Assign atoms here when the central objective is out-of-training-graph generalization, emerging-component reasoning, fully inductive triple prediction, or few-shot completion for sparse entities or relations.","exclusion":"Fixed-entity/relation transductive link prediction belongs to core static model learning; temporally annotated few-shot forecasting belongs to temporal completion when temporal dynamics dominate; formal logical-query generalization belongs to KG reasoning and query answering when query-operator semantics dominate.","nearest_boundary":"Core static model learning","representative_topics":["unifying factorisation-based and message-passing models for inductive knowledge graph completion","inductive knowledge graph reasoning for continuously emerging multi-batch entities via walk-based aggregation","hierarchical relational learning for few-shot knowledge graph completion via multi-level meta relation representation"]},{"direction_index":1,"label":"D1","name":"Core static model learning","description":"Improve standard static KG completion by designing representations, neural architectures, scoring functions, geometric or algebraic embedding spaces, and training objectives for fixed binary subject-predicate-object graphs.","inclusion":"Assign atoms here when the primary objective is better transductive link or relation prediction through embeddings, GNNs, transformers, convolutions, attention, factorization, geometric/algebraic spaces, regularization, contrastive/self-supervised learning, negative-sampling improvements, denoising, or bias-aware training, without a specialized temporal, inductive, multimodal, logical, or hyper-relational regime being central.","exclusion":"Unseen-entity or few-shot generalization belongs to inductive and low-resource completion; PLM/LLM or non-textual evidence fusion belongs to external-evidence completion; explicit rule/query semantics belong to KG reasoning and query answering.","nearest_boundary":"Inductive and low-resource completion","representative_topics":["ultrahyperbolic geometry embeddings for heterogeneous knowledge graph topological structures","spatio-functional knowledge graph embedding for joint composition and hierarchy inference patterns","attention-based entity representation learning with semantic path context for knowledge graph relation prediction"]},{"direction_index":2,"label":"D2","name":"Temporal completion and forecasting","description":"Predict missing, current, or future facts in temporally annotated KGs by modeling timestamps, intervals, historical events, time-varying structure, temporal causality, and event sequences.","inclusion":"Assign atoms here when the central scientific objective is temporal KG interpolation, extrapolation, future event prediction, temporal representation learning, temporal rule use, or temporal multi-hop reasoning.","exclusion":"Static snapshot link prediction belongs to core static model learning; n-ary or qualifier modeling belongs to hyper-relational completion when the non-binary fact structure is central; unseen-entity temporal work belongs here only when temporal dynamics dominate over inductive generalization.","nearest_boundary":"Hyper-relational completion","representative_topics":["autoregressive temporal knowledge graph reasoning via curvature-variable hyperbolic space embeddings","meta-learning for few-shot temporal knowledge graph reasoning over emerging entities","temporal knowledge graph future prediction using historical structure matching and background knowledge"]},{"direction_index":3,"label":"D3","name":"Hyper-relational completion","description":"Complete and reason over KGs whose facts extend beyond ordinary binary triples, including qualified facts, n-ary relations, RDF-star or nested triples, bi-level fact structures, and fact-level numeric literals.","inclusion":"Assign atoms here when the primary objective is link prediction, entity typing, or representation learning for hyper-relational, qualified, n-ary, nested-triple, or numeric-qualified facts, and modeling the extra fact components is the main challenge.","exclusion":"Ordinary timestamped triples belong to temporal completion when time dynamics are central; standard binary link prediction with entity attributes belongs to core static model learning or external-evidence completion depending on the evidence source.","nearest_boundary":"Temporal completion and forecasting","representative_topics":["dual-view hyper-relational knowledge graph embedding for link prediction and entity typing","hyper-relational knowledge graph completion via bidirectional base triple and qualifier representation learning","bi-level knowledge graph embedding for triplet prediction and conditional link prediction"]},{"direction_index":4,"label":"D4","name":"External-evidence completion","description":"Complete KGs by integrating structured triples with pretrained language models, large language models, textual descriptions, retrieval and reading signals, images, web evidence, biological or molecular modalities, video, or other heterogeneous external inputs.","inclusion":"Assign atoms here when the primary evaluated objective is missing-link, relation-label, fact, or KB completion and the main contribution is fusing, aligning, retrieving, prompting, or reading external textual, language-model, multimodal, or heterogeneous evidence.","exclusion":"Natural-language question answering or fact verification over KGs belongs to KG reasoning and query answering; KG construction from text belongs to KG resource creation and assessment; purely structural models with incidental descriptions belong to core static model learning.","nearest_boundary":"KG reasoning and query answering","representative_topics":["language model-based knowledge graph embeddings for long-tail entity representation via contrastive learning","sparse knowledge graph completion via variational em fusion of text and structure knowledge","hypercomplex space integration of knowledge graph and language model embeddings for link prediction"]},{"direction_index":5,"label":"D5","name":"KG reasoning and query answering","description":"Answer complex queries or perform explicit reasoning over KGs using logical operators, symbolic rules, constraints, neuro-symbolic methods, KG-grounded retrieval, LLM reasoning pipelines, or natural-language QA systems.","inclusion":"Assign atoms here when the primary objective is complex logical query answering, rule-guided KGC, differentiable rule learning, logical constraint integration, explainable rule aggregation, natural-language KGQA, KG-grounded fact verification, or multi-step LLM reasoning over KG paths or subgraphs.","exclusion":"Single-triple link prediction without explicit logical, rule, or QA semantics belongs to core static model learning; PLM/LLM use for direct link prediction belongs to external-evidence completion; KG construction or extraction pipelines belong to KG resource creation and assessment.","nearest_boundary":"External-evidence completion","representative_topics":["transformer-based masked pre-training for complex logical query reasoning over knowledge graphs","knowledge representation learning for logical query answering over multi-view knowledge graphs","inductive logical query answering in knowledge graphs using graph neural network representations"]},{"direction_index":6,"label":"D6","name":"Entity-centric inference","description":"Infer entity-level properties or cross-graph entity correspondences, including entity alignment, multilingual or heterogeneous KG matching, continual alignment under graph growth, and entity type prediction.","inclusion":"Assign atoms here when the primary task is identifying equivalent entities across KGs or assigning semantic type labels to entities using KG structure, descriptions, auxiliary relations, multilingual evidence, or heterogeneous graph features.","exclusion":"Triple-level missing-link prediction belongs to the relevant completion program; KG merging as a construction pipeline belongs to KG resource creation and assessment; multimodal link prediction belongs to external-evidence completion unless entity matching is the main target.","nearest_boundary":"Core static model learning","representative_topics":["continual entity alignment for growing knowledge graphs via inductive neighbor-based representation learning","entity type prediction in knowledge graphs via graph walk embeddings and textual descriptions","scalable knowledge graph entity alignment via subgraph partitioning, reconstruction, and embedding merging"]},{"direction_index":7,"label":"D7","name":"KG resource creation and assessment","description":"Create, expand, manage, benchmark, evaluate, analyze, or survey KGs, KG schemas, datasets, toolkits, and completion systems as research infrastructure rather than proposing a primary completion model.","inclusion":"Assign atoms here when the main contribution is KG construction, schema expansion or evolution, human-in-the-loop graph growth, KG generation from text, bootstrapping, schema collection, benchmark design, evaluation protocols, leaderboard tooling, metric critique, dataset analysis, theoretical or empirical model analysis, libraries, or systematic surveys.","exclusion":"A paper proposing a new completion model with standard evaluation belongs to that model’s primary objective; direct link prediction over a fixed graph belongs to core static model learning or a specialized completion regime; KG-grounded QA systems belong to KG reasoning and query answering.","nearest_boundary":"Core static model learning","representative_topics":["theoretical analysis of negative sampling hyperparameters for knowledge graph embedding scoring functions","scoring function discovery and evaluation bias in knowledge graph link prediction","explainable interactive leaderboard for fine-grained knowledge graph completion evaluation"]}],"split":"dev"}
{"episode_id":"F147_T2024-01","field_index":147,"field":"video anomaly detection","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Predictive normality modeling","description":"Learning normal spatiotemporal video, appearance, motion, object, or latent dynamics and detecting anomalies as failures of reconstruction, prediction, likelihood, or consistency with learned normal patterns.","inclusion":"Include methods whose primary detection signal is reconstruction error, future-frame or motion prediction error, feature/latent deviation, generative likelihood, or normal-pattern consistency learned from normal or mostly normal video.","exclusion":"Methods centered on explicit memory/prototype retrieval belong to memory-based normality modeling; methods centered on pose or skeleton representations belong to pose-centric behavior modeling.","nearest_boundary":"Memory-based normality modeling","representative_topics":["multi-contextual transformer prediction for spatio-temporal video anomaly detection","prediction-based video anomaly detection via bidirectional lstm with enhanced spatio-temporal memory exchange","appearance-motion decoupled normality modeling for video anomaly detection via adversarial two-stream autoencoder"]},{"direction_index":1,"label":"D1","name":"Pose-centric behavior modeling","description":"Detecting abnormal human behavior through skeletons, pose graphs, joint trajectories, pedestrian paths, kinematics, crowd motion consistency, or other person-centered motion abstractions.","inclusion":"Include work whose primary representation and anomaly objective operate on skeleton sequences, body keypoints, pose-derived attributes, person trajectories, or multi-person kinematic relations.","exclusion":"Trajectory or pose features used only as auxiliary targets inside a broader pixel, appearance, flow, or object-normality model belong to predictive normality modeling.","nearest_boundary":"Predictive normality modeling","representative_topics":["skeletal trajectory-based human-related crime classification in surveillance video","trajectory prediction error-based unsupervised pedestrian anomaly detection in video","bi-directional trajectory prediction with pose constraints for pedestrian anomaly detection"]},{"direction_index":2,"label":"D2","name":"Weak-label temporal localization","description":"Localizing anomalous snippets, segments, or frames in untrimmed videos from coarse video-level supervision using temporal aggregation, MIL, attention, uncertainty, pseudo-label refinement, or weakly supervised semantic guidance.","inclusion":"Include methods whose primary scientific objective is converting video-level anomaly labels into reliable temporal anomaly scores through MIL, temporal attention, transformers, recurrent aggregation, open-set/uncertainty modeling, clustering guidance, or weak-label pseudo-label refinement.","exclusion":"Fully unsupervised pseudo-label or synthetic-anomaly training without manual video-level labels belongs to label-free pseudo-supervised training; domain-specific or multimodal systems whose main objective is deployment adaptation belong to context-specialized recognition.","nearest_boundary":"Label-free pseudo-supervised training","representative_topics":["weakly supervised surveillance video anomaly detection using transformer and dilated convolution temporal features","weakly supervised video anomaly detection via discriminative temporal dynamics and locality-aware attention","weakly supervised open-world video anomaly detection via evidential learning and normalizing flow pseudo-labeling"]},{"direction_index":3,"label":"D3","name":"Self-supervised representation learning","description":"Learning anomaly-sensitive video representations through self-supervised pretext, proxy, contrastive, masked, teacher-student, or multi-task objectives rather than direct anomaly labels or explicit synthetic anomaly negatives.","inclusion":"Include methods whose main contribution is a self-supervised task or representation objective such as jigsaw solving, temporal-order learning, masked modeling, contrastive agreement, distillation discrepancy, or multi-task pretext learning for anomaly scoring.","exclusion":"Methods whose anomaly score is primarily pixel/frame/flow reconstruction or prediction error belong to predictive normality modeling; methods that create explicit pseudo anomalies or pseudo abnormal labels belong to label-free pseudo-supervised training.","nearest_boundary":"Predictive normality modeling","representative_topics":["self-supervised autoencoder training for visual anomaly detection via submanifold regularization","self-supervised multi-task learning with vision transformer backbone for video anomaly detection","self-supervised video anomaly detection via decoupled spatio-temporal jigsaw puzzle classification"]},{"direction_index":4,"label":"D4","name":"Memory-based normality modeling","description":"Detecting anomalies by explicitly storing, clustering, retrieving, contrasting, or comparing events against learned memories, prototypes, local-global normality stores, or compact normal-pattern repositories.","inclusion":"Include methods whose primary novelty is a memory bank, prototype set, adaptive clusterer, retrieval stream, local-global normality repository, or contrastive memory mechanism used to distinguish normal from abnormal video patterns.","exclusion":"Reconstruction or prediction methods without explicit memory/prototype retrieval as the central mechanism belong to predictive normality modeling; memory used only inside a weak-label MIL localization framework belongs to weak-label temporal localization.","nearest_boundary":"Predictive normality modeling","representative_topics":["dynamic local aggregation and adaptive clustering for memory-efficient video anomaly detection","two-stream video anomaly detection via spatiotemporal context recovery and knowledge retrieval","unsupervised video anomaly detection via local-global normality representation fusion"]},{"direction_index":5,"label":"D5","name":"Label-free pseudo-supervised training","description":"Training anomaly detectors without manual anomaly labels by synthesizing pseudo anomalies, deriving pseudo labels, self-training on unlabeled video, modeling relative normalcy, or using generative one-class objectives with pseudo negative supervision.","inclusion":"Include methods whose primary objective is bootstrapping discriminative anomaly supervision from normal-only, unlabeled, or target-free data through synthetic anomalies, pseudo abnormal samples, dynamicity-guided self-training, coarse-to-fine pseudo labels, relative normalcy, or diffusion/generative unsupervised modeling.","exclusion":"Pseudo-label refinement that starts from video-level anomaly labels belongs to weak-label temporal localization; self-supervised pretext learning without pseudo abnormal labels or synthetic negatives belongs to self-supervised representation learning.","nearest_boundary":"Self-supervised representation learning","representative_topics":["unsupervised video anomaly detection using dynamicity-guided pseudo-label refinement","zero-shot cross-domain video anomaly detection via relative normalcy learning and pseudo-abnormal synthesis","masked autoencoder pseudo anomaly synthesis for unsupervised video anomaly detection"]},{"direction_index":6,"label":"D6","name":"Context-specialized recognition","description":"Adapting video anomaly detection to specific operational contexts, anomaly semantics, event families, camera regimes, or heterogeneous modalities such as UAV, driving, violence, mass gatherings, traffic audio-visual cues, promptable categories, or vision-language alignment.","inclusion":"Include work whose dominant objective is recognizing anomalies under a concrete deployment setting, event taxonomy, sensor/viewpoint constraint, semantic prior, language-conditioned representation, audio-visual fusion, or other domain-specific formulation.","exclusion":"General-purpose MIL, reconstruction, memory, or self-supervised methods merely evaluated on a domain dataset belong to their corresponding modeling program; resources or surveys for a domain belong to resources and evaluation frameworks.","nearest_boundary":"Weak-label temporal localization","representative_topics":["hybrid cnn-random forest spatio-temporal abnormal behavior detection and recognition in large-scale crowds","video vision transformer for violence detection in surveillance systems","transformer-based anomaly detection in aerial uav video via tubelet sequence modeling"]},{"direction_index":7,"label":"D7","name":"Resources and evaluation frameworks","description":"Constructing, organizing, benchmarking, surveying, or reframing video anomaly detection through datasets, taxonomies, systematic comparisons, unified frameworks, long-sequence formulations, or new evaluation paradigms.","inclusion":"Include papers whose primary contribution is a dataset, benchmark, survey, taxonomy, comparative framework, task-agnostic architecture, long-sequence evaluation formulation, or systematic redefinition of the anomaly detection problem.","exclusion":"Papers primarily introducing a new detector belong to the detector program matching their main scientific objective, even if they also provide experiments, datasets, or domain analysis.","nearest_boundary":"Context-specialized recognition","representative_topics":["multi-view skeleton-based video anomaly detection with person identity and pose annotations","survey of generalized video anomaly event detection across supervision regimes","semi-supervised video anomaly detection and anticipation with scene-dependent anomaly support"]}],"split":"dev"}
{"episode_id":"F147_T2024-07","field_index":147,"field":"video anomaly detection","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Predictive normality modeling","description":"Learning normal spatiotemporal video, appearance, motion, object, or latent dynamics and detecting anomalies as failures of reconstruction, prediction, likelihood, or consistency with learned normal patterns.","inclusion":"Include methods whose primary detection signal is reconstruction error, future-frame or motion prediction error, feature/latent deviation, generative likelihood, or normal-pattern consistency learned from normal or mostly normal video.","exclusion":"Methods centered on explicit memory/prototype retrieval belong to memory-based normality modeling; methods centered on pose or skeleton representations belong to pose-centric behavior modeling.","nearest_boundary":"Memory-based normality modeling","representative_topics":["multi-contextual transformer prediction for spatio-temporal video anomaly detection","prediction-based video anomaly detection via bidirectional lstm with enhanced spatio-temporal memory exchange","appearance-motion decoupled normality modeling for video anomaly detection via adversarial two-stream autoencoder"]},{"direction_index":1,"label":"D1","name":"Pose-centric behavior modeling","description":"Detecting abnormal human behavior through skeletons, pose graphs, joint trajectories, pedestrian paths, kinematics, crowd motion consistency, or other person-centered motion abstractions.","inclusion":"Include work whose primary representation and anomaly objective operate on skeleton sequences, body keypoints, pose-derived attributes, person trajectories, or multi-person kinematic relations.","exclusion":"Trajectory or pose features used only as auxiliary targets inside a broader pixel, appearance, flow, or object-normality model belong to predictive normality modeling.","nearest_boundary":"Predictive normality modeling","representative_topics":["skeletal trajectory-based human-related crime classification in surveillance video","trajectory prediction error-based unsupervised pedestrian anomaly detection in video","bi-directional trajectory prediction with pose constraints for pedestrian anomaly detection"]},{"direction_index":2,"label":"D2","name":"Weak-label temporal localization","description":"Localizing anomalous snippets, segments, or frames in untrimmed videos from coarse video-level supervision using temporal aggregation, MIL, attention, uncertainty, pseudo-label refinement, or weakly supervised semantic guidance.","inclusion":"Include methods whose primary scientific objective is converting video-level anomaly labels into reliable temporal anomaly scores through MIL, temporal attention, transformers, recurrent aggregation, open-set/uncertainty modeling, clustering guidance, or weak-label pseudo-label refinement.","exclusion":"Fully unsupervised pseudo-label or synthetic-anomaly training without manual video-level labels belongs to label-free pseudo-supervised training; domain-specific or multimodal systems whose main objective is deployment adaptation belong to context-specialized recognition.","nearest_boundary":"Label-free pseudo-supervised training","representative_topics":["weakly supervised surveillance video anomaly detection using transformer and dilated convolution temporal features","weakly supervised video anomaly detection via discriminative temporal dynamics and locality-aware attention","weakly supervised open-world video anomaly detection via evidential learning and normalizing flow pseudo-labeling"]},{"direction_index":3,"label":"D3","name":"Self-supervised representation learning","description":"Learning anomaly-sensitive video representations through self-supervised pretext, proxy, contrastive, masked, teacher-student, or multi-task objectives rather than direct anomaly labels or explicit synthetic anomaly negatives.","inclusion":"Include methods whose main contribution is a self-supervised task or representation objective such as jigsaw solving, temporal-order learning, masked modeling, contrastive agreement, distillation discrepancy, or multi-task pretext learning for anomaly scoring.","exclusion":"Methods whose anomaly score is primarily pixel/frame/flow reconstruction or prediction error belong to predictive normality modeling; methods that create explicit pseudo anomalies or pseudo abnormal labels belong to label-free pseudo-supervised training.","nearest_boundary":"Predictive normality modeling","representative_topics":["self-supervised autoencoder training for visual anomaly detection via submanifold regularization","self-supervised multi-task learning with vision transformer backbone for video anomaly detection","self-supervised video anomaly detection via decoupled spatio-temporal jigsaw puzzle classification"]},{"direction_index":4,"label":"D4","name":"Memory-based normality modeling","description":"Detecting anomalies by explicitly storing, clustering, retrieving, contrasting, or comparing events against learned memories, prototypes, local-global normality stores, or compact normal-pattern repositories.","inclusion":"Include methods whose primary novelty is a memory bank, prototype set, adaptive clusterer, retrieval stream, local-global normality repository, or contrastive memory mechanism used to distinguish normal from abnormal video patterns.","exclusion":"Reconstruction or prediction methods without explicit memory/prototype retrieval as the central mechanism belong to predictive normality modeling; memory used only inside a weak-label MIL localization framework belongs to weak-label temporal localization.","nearest_boundary":"Predictive normality modeling","representative_topics":["dynamic local aggregation and adaptive clustering for memory-efficient video anomaly detection","two-stream video anomaly detection via spatiotemporal context recovery and knowledge retrieval","unsupervised video anomaly detection via local-global normality representation fusion"]},{"direction_index":5,"label":"D5","name":"Label-free pseudo-supervised training","description":"Training anomaly detectors without manual anomaly labels by synthesizing pseudo anomalies, deriving pseudo labels, self-training on unlabeled video, modeling relative normalcy, or using generative one-class objectives with pseudo negative supervision.","inclusion":"Include methods whose primary objective is bootstrapping discriminative anomaly supervision from normal-only, unlabeled, or target-free data through synthetic anomalies, pseudo abnormal samples, dynamicity-guided self-training, coarse-to-fine pseudo labels, relative normalcy, or diffusion/generative unsupervised modeling.","exclusion":"Pseudo-label refinement that starts from video-level anomaly labels belongs to weak-label temporal localization; self-supervised pretext learning without pseudo abnormal labels or synthetic negatives belongs to self-supervised representation learning.","nearest_boundary":"Self-supervised representation learning","representative_topics":["unsupervised video anomaly detection using dynamicity-guided pseudo-label refinement","zero-shot cross-domain video anomaly detection via relative normalcy learning and pseudo-abnormal synthesis","masked autoencoder pseudo anomaly synthesis for unsupervised video anomaly detection"]},{"direction_index":6,"label":"D6","name":"Context-specialized recognition","description":"Adapting video anomaly detection to specific operational contexts, anomaly semantics, event families, camera regimes, or heterogeneous modalities such as UAV, driving, violence, mass gatherings, traffic audio-visual cues, promptable categories, or vision-language alignment.","inclusion":"Include work whose dominant objective is recognizing anomalies under a concrete deployment setting, event taxonomy, sensor/viewpoint constraint, semantic prior, language-conditioned representation, audio-visual fusion, or other domain-specific formulation.","exclusion":"General-purpose MIL, reconstruction, memory, or self-supervised methods merely evaluated on a domain dataset belong to their corresponding modeling program; resources or surveys for a domain belong to resources and evaluation frameworks.","nearest_boundary":"Weak-label temporal localization","representative_topics":["hybrid cnn-random forest spatio-temporal abnormal behavior detection and recognition in large-scale crowds","video vision transformer for violence detection in surveillance systems","transformer-based anomaly detection in aerial uav video via tubelet sequence modeling"]},{"direction_index":7,"label":"D7","name":"Resources and evaluation frameworks","description":"Constructing, organizing, benchmarking, surveying, or reframing video anomaly detection through datasets, taxonomies, systematic comparisons, unified frameworks, long-sequence formulations, or new evaluation paradigms.","inclusion":"Include papers whose primary contribution is a dataset, benchmark, survey, taxonomy, comparative framework, task-agnostic architecture, long-sequence evaluation formulation, or systematic redefinition of the anomaly detection problem.","exclusion":"Papers primarily introducing a new detector belong to the detector program matching their main scientific objective, even if they also provide experiments, datasets, or domain analysis.","nearest_boundary":"Context-specialized recognition","representative_topics":["multi-view skeleton-based video anomaly detection with person identity and pose annotations","survey of generalized video anomaly event detection across supervision regimes","semi-supervised video anomaly detection and anticipation with scene-dependent anomaly support"]}],"split":"dev"}
{"episode_id":"F147_T2025-01","field_index":147,"field":"video anomaly detection","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Predictive normality modeling","description":"Learning normal spatiotemporal video, appearance, motion, object, or latent dynamics and detecting anomalies as failures of reconstruction, prediction, likelihood, or consistency with learned normal patterns.","inclusion":"Include methods whose primary detection signal is reconstruction error, future-frame or motion prediction error, feature/latent deviation, generative likelihood, or normal-pattern consistency learned from normal or mostly normal video.","exclusion":"Methods centered on explicit memory/prototype retrieval belong to memory-based normality modeling; methods centered on pose or skeleton representations belong to pose-centric behavior modeling.","nearest_boundary":"Memory-based normality modeling","representative_topics":["multi-contextual transformer prediction for spatio-temporal video anomaly detection","prediction-based video anomaly detection via bidirectional lstm with enhanced spatio-temporal memory exchange","appearance-motion decoupled normality modeling for video anomaly detection via adversarial two-stream autoencoder"]},{"direction_index":1,"label":"D1","name":"Pose-centric behavior modeling","description":"Detecting abnormal human behavior through skeletons, pose graphs, joint trajectories, pedestrian paths, kinematics, crowd motion consistency, or other person-centered motion abstractions.","inclusion":"Include work whose primary representation and anomaly objective operate on skeleton sequences, body keypoints, pose-derived attributes, person trajectories, or multi-person kinematic relations.","exclusion":"Trajectory or pose features used only as auxiliary targets inside a broader pixel, appearance, flow, or object-normality model belong to predictive normality modeling.","nearest_boundary":"Predictive normality modeling","representative_topics":["skeletal trajectory-based human-related crime classification in surveillance video","trajectory prediction error-based unsupervised pedestrian anomaly detection in video","bi-directional trajectory prediction with pose constraints for pedestrian anomaly detection"]},{"direction_index":2,"label":"D2","name":"Weak-label temporal localization","description":"Localizing anomalous snippets, segments, or frames in untrimmed videos from coarse video-level supervision using temporal aggregation, MIL, attention, uncertainty, pseudo-label refinement, or weakly supervised semantic guidance.","inclusion":"Include methods whose primary scientific objective is converting video-level anomaly labels into reliable temporal anomaly scores through MIL, temporal attention, transformers, recurrent aggregation, open-set/uncertainty modeling, clustering guidance, or weak-label pseudo-label refinement.","exclusion":"Fully unsupervised pseudo-label or synthetic-anomaly training without manual video-level labels belongs to label-free pseudo-supervised training; domain-specific or multimodal systems whose main objective is deployment adaptation belong to context-specialized recognition.","nearest_boundary":"Label-free pseudo-supervised training","representative_topics":["weakly supervised surveillance video anomaly detection using transformer and dilated convolution temporal features","weakly supervised video anomaly detection via discriminative temporal dynamics and locality-aware attention","weakly supervised open-world video anomaly detection via evidential learning and normalizing flow pseudo-labeling"]},{"direction_index":3,"label":"D3","name":"Self-supervised representation learning","description":"Learning anomaly-sensitive video representations through self-supervised pretext, proxy, contrastive, masked, teacher-student, or multi-task objectives rather than direct anomaly labels or explicit synthetic anomaly negatives.","inclusion":"Include methods whose main contribution is a self-supervised task or representation objective such as jigsaw solving, temporal-order learning, masked modeling, contrastive agreement, distillation discrepancy, or multi-task pretext learning for anomaly scoring.","exclusion":"Methods whose anomaly score is primarily pixel/frame/flow reconstruction or prediction error belong to predictive normality modeling; methods that create explicit pseudo anomalies or pseudo abnormal labels belong to label-free pseudo-supervised training.","nearest_boundary":"Predictive normality modeling","representative_topics":["self-supervised autoencoder training for visual anomaly detection via submanifold regularization","self-supervised multi-task learning with vision transformer backbone for video anomaly detection","self-supervised video anomaly detection via decoupled spatio-temporal jigsaw puzzle classification"]},{"direction_index":4,"label":"D4","name":"Memory-based normality modeling","description":"Detecting anomalies by explicitly storing, clustering, retrieving, contrasting, or comparing events against learned memories, prototypes, local-global normality stores, or compact normal-pattern repositories.","inclusion":"Include methods whose primary novelty is a memory bank, prototype set, adaptive clusterer, retrieval stream, local-global normality repository, or contrastive memory mechanism used to distinguish normal from abnormal video patterns.","exclusion":"Reconstruction or prediction methods without explicit memory/prototype retrieval as the central mechanism belong to predictive normality modeling; memory used only inside a weak-label MIL localization framework belongs to weak-label temporal localization.","nearest_boundary":"Predictive normality modeling","representative_topics":["dynamic local aggregation and adaptive clustering for memory-efficient video anomaly detection","two-stream video anomaly detection via spatiotemporal context recovery and knowledge retrieval","unsupervised video anomaly detection via local-global normality representation fusion"]},{"direction_index":5,"label":"D5","name":"Label-free pseudo-supervised training","description":"Training anomaly detectors without manual anomaly labels by synthesizing pseudo anomalies, deriving pseudo labels, self-training on unlabeled video, modeling relative normalcy, or using generative one-class objectives with pseudo negative supervision.","inclusion":"Include methods whose primary objective is bootstrapping discriminative anomaly supervision from normal-only, unlabeled, or target-free data through synthetic anomalies, pseudo abnormal samples, dynamicity-guided self-training, coarse-to-fine pseudo labels, relative normalcy, or diffusion/generative unsupervised modeling.","exclusion":"Pseudo-label refinement that starts from video-level anomaly labels belongs to weak-label temporal localization; self-supervised pretext learning without pseudo abnormal labels or synthetic negatives belongs to self-supervised representation learning.","nearest_boundary":"Self-supervised representation learning","representative_topics":["unsupervised video anomaly detection using dynamicity-guided pseudo-label refinement","zero-shot cross-domain video anomaly detection via relative normalcy learning and pseudo-abnormal synthesis","masked autoencoder pseudo anomaly synthesis for unsupervised video anomaly detection"]},{"direction_index":6,"label":"D6","name":"Context-specialized recognition","description":"Adapting video anomaly detection to specific operational contexts, anomaly semantics, event families, camera regimes, or heterogeneous modalities such as UAV, driving, violence, mass gatherings, traffic audio-visual cues, promptable categories, or vision-language alignment.","inclusion":"Include work whose dominant objective is recognizing anomalies under a concrete deployment setting, event taxonomy, sensor/viewpoint constraint, semantic prior, language-conditioned representation, audio-visual fusion, or other domain-specific formulation.","exclusion":"General-purpose MIL, reconstruction, memory, or self-supervised methods merely evaluated on a domain dataset belong to their corresponding modeling program; resources or surveys for a domain belong to resources and evaluation frameworks.","nearest_boundary":"Weak-label temporal localization","representative_topics":["hybrid cnn-random forest spatio-temporal abnormal behavior detection and recognition in large-scale crowds","video vision transformer for violence detection in surveillance systems","transformer-based anomaly detection in aerial uav video via tubelet sequence modeling"]},{"direction_index":7,"label":"D7","name":"Resources and evaluation frameworks","description":"Constructing, organizing, benchmarking, surveying, or reframing video anomaly detection through datasets, taxonomies, systematic comparisons, unified frameworks, long-sequence formulations, or new evaluation paradigms.","inclusion":"Include papers whose primary contribution is a dataset, benchmark, survey, taxonomy, comparative framework, task-agnostic architecture, long-sequence evaluation formulation, or systematic redefinition of the anomaly detection problem.","exclusion":"Papers primarily introducing a new detector belong to the detector program matching their main scientific objective, even if they also provide experiments, datasets, or domain analysis.","nearest_boundary":"Context-specialized recognition","representative_topics":["multi-view skeleton-based video anomaly detection with person identity and pose annotations","survey of generalized video anomaly event detection across supervision regimes","semi-supervised video anomaly detection and anticipation with scene-dependent anomaly support"]}],"split":"dev"}
{"episode_id":"F147_T2025-07","field_index":147,"field":"video anomaly detection","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Predictive normality modeling","description":"Learning normal spatiotemporal video, appearance, motion, object, or latent dynamics and detecting anomalies as failures of reconstruction, prediction, likelihood, or consistency with learned normal patterns.","inclusion":"Include methods whose primary detection signal is reconstruction error, future-frame or motion prediction error, feature/latent deviation, generative likelihood, or normal-pattern consistency learned from normal or mostly normal video.","exclusion":"Methods centered on explicit memory/prototype retrieval belong to memory-based normality modeling; methods centered on pose or skeleton representations belong to pose-centric behavior modeling.","nearest_boundary":"Memory-based normality modeling","representative_topics":["multi-contextual transformer prediction for spatio-temporal video anomaly detection","prediction-based video anomaly detection via bidirectional lstm with enhanced spatio-temporal memory exchange","appearance-motion decoupled normality modeling for video anomaly detection via adversarial two-stream autoencoder"]},{"direction_index":1,"label":"D1","name":"Pose-centric behavior modeling","description":"Detecting abnormal human behavior through skeletons, pose graphs, joint trajectories, pedestrian paths, kinematics, crowd motion consistency, or other person-centered motion abstractions.","inclusion":"Include work whose primary representation and anomaly objective operate on skeleton sequences, body keypoints, pose-derived attributes, person trajectories, or multi-person kinematic relations.","exclusion":"Trajectory or pose features used only as auxiliary targets inside a broader pixel, appearance, flow, or object-normality model belong to predictive normality modeling.","nearest_boundary":"Predictive normality modeling","representative_topics":["skeletal trajectory-based human-related crime classification in surveillance video","trajectory prediction error-based unsupervised pedestrian anomaly detection in video","bi-directional trajectory prediction with pose constraints for pedestrian anomaly detection"]},{"direction_index":2,"label":"D2","name":"Weak-label temporal localization","description":"Localizing anomalous snippets, segments, or frames in untrimmed videos from coarse video-level supervision using temporal aggregation, MIL, attention, uncertainty, pseudo-label refinement, or weakly supervised semantic guidance.","inclusion":"Include methods whose primary scientific objective is converting video-level anomaly labels into reliable temporal anomaly scores through MIL, temporal attention, transformers, recurrent aggregation, open-set/uncertainty modeling, clustering guidance, or weak-label pseudo-label refinement.","exclusion":"Fully unsupervised pseudo-label or synthetic-anomaly training without manual video-level labels belongs to label-free pseudo-supervised training; domain-specific or multimodal systems whose main objective is deployment adaptation belong to context-specialized recognition.","nearest_boundary":"Label-free pseudo-supervised training","representative_topics":["weakly supervised surveillance video anomaly detection using transformer and dilated convolution temporal features","weakly supervised video anomaly detection via discriminative temporal dynamics and locality-aware attention","weakly supervised open-world video anomaly detection via evidential learning and normalizing flow pseudo-labeling"]},{"direction_index":3,"label":"D3","name":"Self-supervised representation learning","description":"Learning anomaly-sensitive video representations through self-supervised pretext, proxy, contrastive, masked, teacher-student, or multi-task objectives rather than direct anomaly labels or explicit synthetic anomaly negatives.","inclusion":"Include methods whose main contribution is a self-supervised task or representation objective such as jigsaw solving, temporal-order learning, masked modeling, contrastive agreement, distillation discrepancy, or multi-task pretext learning for anomaly scoring.","exclusion":"Methods whose anomaly score is primarily pixel/frame/flow reconstruction or prediction error belong to predictive normality modeling; methods that create explicit pseudo anomalies or pseudo abnormal labels belong to label-free pseudo-supervised training.","nearest_boundary":"Predictive normality modeling","representative_topics":["self-supervised autoencoder training for visual anomaly detection via submanifold regularization","self-supervised multi-task learning with vision transformer backbone for video anomaly detection","self-supervised video anomaly detection via decoupled spatio-temporal jigsaw puzzle classification"]},{"direction_index":4,"label":"D4","name":"Memory-based normality modeling","description":"Detecting anomalies by explicitly storing, clustering, retrieving, contrasting, or comparing events against learned memories, prototypes, local-global normality stores, or compact normal-pattern repositories.","inclusion":"Include methods whose primary novelty is a memory bank, prototype set, adaptive clusterer, retrieval stream, local-global normality repository, or contrastive memory mechanism used to distinguish normal from abnormal video patterns.","exclusion":"Reconstruction or prediction methods without explicit memory/prototype retrieval as the central mechanism belong to predictive normality modeling; memory used only inside a weak-label MIL localization framework belongs to weak-label temporal localization.","nearest_boundary":"Predictive normality modeling","representative_topics":["dynamic local aggregation and adaptive clustering for memory-efficient video anomaly detection","two-stream video anomaly detection via spatiotemporal context recovery and knowledge retrieval","unsupervised video anomaly detection via local-global normality representation fusion"]},{"direction_index":5,"label":"D5","name":"Label-free pseudo-supervised training","description":"Training anomaly detectors without manual anomaly labels by synthesizing pseudo anomalies, deriving pseudo labels, self-training on unlabeled video, modeling relative normalcy, or using generative one-class objectives with pseudo negative supervision.","inclusion":"Include methods whose primary objective is bootstrapping discriminative anomaly supervision from normal-only, unlabeled, or target-free data through synthetic anomalies, pseudo abnormal samples, dynamicity-guided self-training, coarse-to-fine pseudo labels, relative normalcy, or diffusion/generative unsupervised modeling.","exclusion":"Pseudo-label refinement that starts from video-level anomaly labels belongs to weak-label temporal localization; self-supervised pretext learning without pseudo abnormal labels or synthetic negatives belongs to self-supervised representation learning.","nearest_boundary":"Self-supervised representation learning","representative_topics":["unsupervised video anomaly detection using dynamicity-guided pseudo-label refinement","zero-shot cross-domain video anomaly detection via relative normalcy learning and pseudo-abnormal synthesis","masked autoencoder pseudo anomaly synthesis for unsupervised video anomaly detection"]},{"direction_index":6,"label":"D6","name":"Context-specialized recognition","description":"Adapting video anomaly detection to specific operational contexts, anomaly semantics, event families, camera regimes, or heterogeneous modalities such as UAV, driving, violence, mass gatherings, traffic audio-visual cues, promptable categories, or vision-language alignment.","inclusion":"Include work whose dominant objective is recognizing anomalies under a concrete deployment setting, event taxonomy, sensor/viewpoint constraint, semantic prior, language-conditioned representation, audio-visual fusion, or other domain-specific formulation.","exclusion":"General-purpose MIL, reconstruction, memory, or self-supervised methods merely evaluated on a domain dataset belong to their corresponding modeling program; resources or surveys for a domain belong to resources and evaluation frameworks.","nearest_boundary":"Weak-label temporal localization","representative_topics":["hybrid cnn-random forest spatio-temporal abnormal behavior detection and recognition in large-scale crowds","video vision transformer for violence detection in surveillance systems","transformer-based anomaly detection in aerial uav video via tubelet sequence modeling"]},{"direction_index":7,"label":"D7","name":"Resources and evaluation frameworks","description":"Constructing, organizing, benchmarking, surveying, or reframing video anomaly detection through datasets, taxonomies, systematic comparisons, unified frameworks, long-sequence formulations, or new evaluation paradigms.","inclusion":"Include papers whose primary contribution is a dataset, benchmark, survey, taxonomy, comparative framework, task-agnostic architecture, long-sequence evaluation formulation, or systematic redefinition of the anomaly detection problem.","exclusion":"Papers primarily introducing a new detector belong to the detector program matching their main scientific objective, even if they also provide experiments, datasets, or domain analysis.","nearest_boundary":"Context-specialized recognition","representative_topics":["multi-view skeleton-based video anomaly detection with person identity and pose annotations","survey of generalized video anomaly event detection across supervision regimes","semi-supervised video anomaly detection and anticipation with scene-dependent anomaly support"]}],"split":"dev"}
{"episode_id":"F147_T2026-01","field_index":147,"field":"video anomaly detection","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Predictive normality modeling","description":"Learning normal spatiotemporal video, appearance, motion, object, or latent dynamics and detecting anomalies as failures of reconstruction, prediction, likelihood, or consistency with learned normal patterns.","inclusion":"Include methods whose primary detection signal is reconstruction error, future-frame or motion prediction error, feature/latent deviation, generative likelihood, or normal-pattern consistency learned from normal or mostly normal video.","exclusion":"Methods centered on explicit memory/prototype retrieval belong to memory-based normality modeling; methods centered on pose or skeleton representations belong to pose-centric behavior modeling.","nearest_boundary":"Memory-based normality modeling","representative_topics":["multi-contextual transformer prediction for spatio-temporal video anomaly detection","prediction-based video anomaly detection via bidirectional lstm with enhanced spatio-temporal memory exchange","appearance-motion decoupled normality modeling for video anomaly detection via adversarial two-stream autoencoder"]},{"direction_index":1,"label":"D1","name":"Pose-centric behavior modeling","description":"Detecting abnormal human behavior through skeletons, pose graphs, joint trajectories, pedestrian paths, kinematics, crowd motion consistency, or other person-centered motion abstractions.","inclusion":"Include work whose primary representation and anomaly objective operate on skeleton sequences, body keypoints, pose-derived attributes, person trajectories, or multi-person kinematic relations.","exclusion":"Trajectory or pose features used only as auxiliary targets inside a broader pixel, appearance, flow, or object-normality model belong to predictive normality modeling.","nearest_boundary":"Predictive normality modeling","representative_topics":["skeletal trajectory-based human-related crime classification in surveillance video","trajectory prediction error-based unsupervised pedestrian anomaly detection in video","bi-directional trajectory prediction with pose constraints for pedestrian anomaly detection"]},{"direction_index":2,"label":"D2","name":"Weak-label temporal localization","description":"Localizing anomalous snippets, segments, or frames in untrimmed videos from coarse video-level supervision using temporal aggregation, MIL, attention, uncertainty, pseudo-label refinement, or weakly supervised semantic guidance.","inclusion":"Include methods whose primary scientific objective is converting video-level anomaly labels into reliable temporal anomaly scores through MIL, temporal attention, transformers, recurrent aggregation, open-set/uncertainty modeling, clustering guidance, or weak-label pseudo-label refinement.","exclusion":"Fully unsupervised pseudo-label or synthetic-anomaly training without manual video-level labels belongs to label-free pseudo-supervised training; domain-specific or multimodal systems whose main objective is deployment adaptation belong to context-specialized recognition.","nearest_boundary":"Label-free pseudo-supervised training","representative_topics":["weakly supervised surveillance video anomaly detection using transformer and dilated convolution temporal features","weakly supervised video anomaly detection via discriminative temporal dynamics and locality-aware attention","weakly supervised open-world video anomaly detection via evidential learning and normalizing flow pseudo-labeling"]},{"direction_index":3,"label":"D3","name":"Self-supervised representation learning","description":"Learning anomaly-sensitive video representations through self-supervised pretext, proxy, contrastive, masked, teacher-student, or multi-task objectives rather than direct anomaly labels or explicit synthetic anomaly negatives.","inclusion":"Include methods whose main contribution is a self-supervised task or representation objective such as jigsaw solving, temporal-order learning, masked modeling, contrastive agreement, distillation discrepancy, or multi-task pretext learning for anomaly scoring.","exclusion":"Methods whose anomaly score is primarily pixel/frame/flow reconstruction or prediction error belong to predictive normality modeling; methods that create explicit pseudo anomalies or pseudo abnormal labels belong to label-free pseudo-supervised training.","nearest_boundary":"Predictive normality modeling","representative_topics":["self-supervised autoencoder training for visual anomaly detection via submanifold regularization","self-supervised multi-task learning with vision transformer backbone for video anomaly detection","self-supervised video anomaly detection via decoupled spatio-temporal jigsaw puzzle classification"]},{"direction_index":4,"label":"D4","name":"Memory-based normality modeling","description":"Detecting anomalies by explicitly storing, clustering, retrieving, contrasting, or comparing events against learned memories, prototypes, local-global normality stores, or compact normal-pattern repositories.","inclusion":"Include methods whose primary novelty is a memory bank, prototype set, adaptive clusterer, retrieval stream, local-global normality repository, or contrastive memory mechanism used to distinguish normal from abnormal video patterns.","exclusion":"Reconstruction or prediction methods without explicit memory/prototype retrieval as the central mechanism belong to predictive normality modeling; memory used only inside a weak-label MIL localization framework belongs to weak-label temporal localization.","nearest_boundary":"Predictive normality modeling","representative_topics":["dynamic local aggregation and adaptive clustering for memory-efficient video anomaly detection","two-stream video anomaly detection via spatiotemporal context recovery and knowledge retrieval","unsupervised video anomaly detection via local-global normality representation fusion"]},{"direction_index":5,"label":"D5","name":"Label-free pseudo-supervised training","description":"Training anomaly detectors without manual anomaly labels by synthesizing pseudo anomalies, deriving pseudo labels, self-training on unlabeled video, modeling relative normalcy, or using generative one-class objectives with pseudo negative supervision.","inclusion":"Include methods whose primary objective is bootstrapping discriminative anomaly supervision from normal-only, unlabeled, or target-free data through synthetic anomalies, pseudo abnormal samples, dynamicity-guided self-training, coarse-to-fine pseudo labels, relative normalcy, or diffusion/generative unsupervised modeling.","exclusion":"Pseudo-label refinement that starts from video-level anomaly labels belongs to weak-label temporal localization; self-supervised pretext learning without pseudo abnormal labels or synthetic negatives belongs to self-supervised representation learning.","nearest_boundary":"Self-supervised representation learning","representative_topics":["unsupervised video anomaly detection using dynamicity-guided pseudo-label refinement","zero-shot cross-domain video anomaly detection via relative normalcy learning and pseudo-abnormal synthesis","masked autoencoder pseudo anomaly synthesis for unsupervised video anomaly detection"]},{"direction_index":6,"label":"D6","name":"Context-specialized recognition","description":"Adapting video anomaly detection to specific operational contexts, anomaly semantics, event families, camera regimes, or heterogeneous modalities such as UAV, driving, violence, mass gatherings, traffic audio-visual cues, promptable categories, or vision-language alignment.","inclusion":"Include work whose dominant objective is recognizing anomalies under a concrete deployment setting, event taxonomy, sensor/viewpoint constraint, semantic prior, language-conditioned representation, audio-visual fusion, or other domain-specific formulation.","exclusion":"General-purpose MIL, reconstruction, memory, or self-supervised methods merely evaluated on a domain dataset belong to their corresponding modeling program; resources or surveys for a domain belong to resources and evaluation frameworks.","nearest_boundary":"Weak-label temporal localization","representative_topics":["hybrid cnn-random forest spatio-temporal abnormal behavior detection and recognition in large-scale crowds","video vision transformer for violence detection in surveillance systems","transformer-based anomaly detection in aerial uav video via tubelet sequence modeling"]},{"direction_index":7,"label":"D7","name":"Resources and evaluation frameworks","description":"Constructing, organizing, benchmarking, surveying, or reframing video anomaly detection through datasets, taxonomies, systematic comparisons, unified frameworks, long-sequence formulations, or new evaluation paradigms.","inclusion":"Include papers whose primary contribution is a dataset, benchmark, survey, taxonomy, comparative framework, task-agnostic architecture, long-sequence evaluation formulation, or systematic redefinition of the anomaly detection problem.","exclusion":"Papers primarily introducing a new detector belong to the detector program matching their main scientific objective, even if they also provide experiments, datasets, or domain analysis.","nearest_boundary":"Context-specialized recognition","representative_topics":["multi-view skeleton-based video anomaly detection with person identity and pose annotations","survey of generalized video anomaly event detection across supervision regimes","semi-supervised video anomaly detection and anticipation with scene-dependent anomaly support"]}],"split":"dev"}
{"episode_id":"F149_T2024-01","field_index":149,"field":"model-based reinforcement learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Offline conservative MBRL","description":"Reliable policy improvement from fixed logged datasets using learned or implicit dynamics while controlling batch-distribution shift, unsupported rollouts, and model exploitation.","inclusion":"Include work whose main objective is offline model-based policy learning under limited support, using pessimism, conservatism, coverage guarantees, distributional robustness tied to dataset support, confidence-filtered imagination, or causal offline world models.","exclusion":"If uncertainty is used mainly for online exploration, safety, or calibrated planning rather than fixed-dataset support, place it in reliability under uncertainty; if the main contribution is a generic model architecture, place it in world-model representation and architecture.","nearest_boundary":"D6","representative_topics":["model generation with provable coverability for offline dynamics-aware policy learning","causal world models for offline reinforcement learning generalization","confidence-aware bidirectional model-based imagination for offline reinforcement learning generalization"]},{"direction_index":1,"label":"D1","name":"Transfer, meta, and continual adaptation","description":"Rapidly adapting model-based RL systems across tasks, users, environments, or time-varying deployments through meta-learning, task representations, continual model maintenance, and lifelong adaptation.","inclusion":"Include work whose main objective is task-level transfer, few-shot adaptation, meta-RL, personalized or robust meta-learning, context inference, continual world-model adaptation, change handling, selective replay, or lifelong/non-stationary model reuse.","exclusion":"If the main objective is invariant or physically structured transition-model discovery rather than adaptation over task distributions or time, place it in structured dynamics; if the main issue is offline dataset support, place it in offline conservative MBRL.","nearest_boundary":"D7","representative_topics":["fine-tuning versus meta-rl on novel tasks in vision-based reinforcement learning benchmarks","action translation for sample-efficient meta-rl on sparse-reward heterogeneous tasks","context-based meta reinforcement learning via successor feature reward decomposition for data-efficient multi-task adaptation"]},{"direction_index":2,"label":"D2","name":"Theory and sample-efficiency foundations","description":"Formal characterization of model-based RL efficiency through regret, PAC, sample-complexity, representation-complexity, model-selection, exploration-scheduling, and structural guarantee frameworks.","inclusion":"Include work whose primary contribution is theoretical analysis or principled sample-efficiency framework design, including posterior-sampling guarantees, model selection, structural MDP regret, generalized eluder/PAC frameworks, confidence scheduling, and sample-complexity bounds.","exclusion":"If the formal result supports a concrete offline, planning, uncertainty, robust, or representation algorithm whose scientific objective dominates, place it in that mechanism-specific direction.","nearest_boundary":"D3","representative_topics":["optimistic posterior sampling for model-based rl with hellinger-distance regret analysis","model-based reinforcement learning for adaptive online influence maximization with regret guarantees","adaptive model selection in reinforcement learning over nested function approximation classes"]},{"direction_index":3,"label":"D3","name":"Model-based policy improvement","description":"Converting learned models, predictive simulators, or model-derived objectives into better actions and policies through planning, rollout control, value expansion, MPC, imitation, preference optimization, or policy-optimization loops.","inclusion":"Include work whose main objective is how model information is used for decision making or policy improvement, including adaptive rollout length, decision-time or background planning, goal/subgoal planning when aimed at control, visual foresight, decentralized model-based optimization, model-based imitation, preference-based policy optimization, and broad surveys or cross-cutting methods centered on policy-learning pipelines.","exclusion":"If the main contribution is the world-model representation itself, place it in world-model representation and architecture; if goal or subgoal machinery is primarily for reward discovery, place it in exploration and sparse-reward discovery; if deployment constraints dominate, place it in structured dynamics and domain models.","nearest_boundary":"D5","representative_topics":["adaptive rollout length optimization in model-based rl via meta-level deep reinforcement learning","value-based decision-time versus background planning in model-based reinforcement learning and transfer learning","model-based entropy-regularized imitation learning with joint policy and model adversarial training"]},{"direction_index":4,"label":"D4","name":"Exploration and sparse-reward discovery","description":"Using world models, planning, intrinsic motivation, reachability, memory, or reward shaping to discover rewards, cover environments, or guide behavior when external rewards are sparse, delayed, absent, or hard to reach.","inclusion":"Include work whose main objective is exploration itself, including intrinsic reward learning, curiosity, reward-free or sparse-reward coverage, goal-space exploration, hierarchical subgoal generation for discovery, memory-guided exploration, self-imitation with intrinsic motivation, and information-directed exploration.","exclusion":"If uncertainty is mainly used to calibrate risk, safety, or robustness rather than to discover new behavior, place it in reliability under uncertainty; if the method is primarily a generic planning/control algorithm, place it in model-based policy improvement.","nearest_boundary":"D6","representative_topics":["goal-space background planning with subgoal-conditioned models for model-based reinforcement learning","pixel-based hierarchical planning via latent world model goal selection and action decomposition","deterministic exploration-exploitation scheduling for joint mdp learning and policy optimization"]},{"direction_index":5,"label":"D5","name":"World-model representation and architecture","description":"Designing latent states, perceptual encoders, model architectures, compression criteria, and training objectives that make learned world models compact, predictive, decision-relevant, and useful for downstream control.","inclusion":"Include work whose main objective is the model or representation rather than the downstream planner, including visual/object/entity-centric representations, masked or sparse latent world models, value-equivalent models, rate-distortion compression, graph or memory world models, policy-adapted dynamics fitting, demonstration-guided world-model learning, and multi-agent latent world models.","exclusion":"If the key claim is causal, invariant, physics-informed, simulator-structured, or domain-structured transition modeling, place it in structured dynamics and domain models; if the main contribution is the planner or rollout policy, place it in model-based policy improvement.","nearest_boundary":"D7","representative_topics":["rate-distortion-based approximate value equivalence for tractable model-based reinforcement learning","value-equivalent model compression via rate-distortion for capacity-constrained model-based reinforcement learning","entity-based reinforcement learning using slot attention and graph neural networks for structured observations"]},{"direction_index":6,"label":"D6","name":"Reliability under uncertainty and risk","description":"Making model-based or model-informed decision making reliable under epistemic uncertainty, aleatoric uncertainty, adversarial perturbations, risk-sensitive criteria, safety constraints, or robust MDP uncertainty sets.","inclusion":"Include work whose main objective is uncertainty quantification, uncertainty-aware planning, safety certificates, Bayesian or ensemble-style probabilistic control, value-function uncertainty, explicit robust MDP optimization, adversarial robustness, distributional or worst-case objectives, or risk-sensitive criteria such as exponential utility.","exclusion":"If uncertainty is specifically a pessimistic mechanism for fixed offline data support, place it in offline conservative MBRL; if the focus is formal sample-complexity theory rather than a reliability objective, place it in theory and sample-efficiency foundations.","nearest_boundary":"D0","representative_topics":["bayesian coreset compression for model-based reinforcement learning via kernelized stein discrepancy","safe model-based reinforcement learning via uncertainty-aware distributional reachability certificates","variational recurrent kalman networks for principled aleatoric and epistemic uncertainty in model-based reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Structured dynamics and domain models","description":"Learning or exploiting transition models with causal, invariant, physical, simulator, or domain-specific structure to improve generalization, sample efficiency, and control in structured environments.","inclusion":"Include work whose main objective is structured model learning or structured deployment, including causal and intervention-invariant dynamics, modular world models, parsimonious dynamics, physics-informed or Hamiltonian/ODE/SINDy models, differentiable simulators or renderers, surrogate domain simulators, and domain-specific model-based RL systems where environment structure is central.","exclusion":"If the work targets task-level meta-adaptation or continual change rather than the structure of the transition model, place it in transfer, meta, and continual adaptation; if the architecture is generic and not justified by causal, physical, simulator, or domain structure, place it in world-model representation and architecture.","nearest_boundary":"D1","representative_topics":["unsupervised dynamics generalization via relational intervention for model-based reinforcement learning","variational causal world models for modular adaptation across intervened environments","causal dynamics learning for task-independent state abstraction in model-based reinforcement learning"]}],"split":"dev"}
{"episode_id":"F149_T2024-07","field_index":149,"field":"model-based reinforcement learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Offline conservative MBRL","description":"Reliable policy improvement from fixed logged datasets using learned or implicit dynamics while controlling batch-distribution shift, unsupported rollouts, and model exploitation.","inclusion":"Include work whose main objective is offline model-based policy learning under limited support, using pessimism, conservatism, coverage guarantees, distributional robustness tied to dataset support, confidence-filtered imagination, or causal offline world models.","exclusion":"If uncertainty is used mainly for online exploration, safety, or calibrated planning rather than fixed-dataset support, place it in reliability under uncertainty; if the main contribution is a generic model architecture, place it in world-model representation and architecture.","nearest_boundary":"D6","representative_topics":["model generation with provable coverability for offline dynamics-aware policy learning","causal world models for offline reinforcement learning generalization","confidence-aware bidirectional model-based imagination for offline reinforcement learning generalization"]},{"direction_index":1,"label":"D1","name":"Transfer, meta, and continual adaptation","description":"Rapidly adapting model-based RL systems across tasks, users, environments, or time-varying deployments through meta-learning, task representations, continual model maintenance, and lifelong adaptation.","inclusion":"Include work whose main objective is task-level transfer, few-shot adaptation, meta-RL, personalized or robust meta-learning, context inference, continual world-model adaptation, change handling, selective replay, or lifelong/non-stationary model reuse.","exclusion":"If the main objective is invariant or physically structured transition-model discovery rather than adaptation over task distributions or time, place it in structured dynamics; if the main issue is offline dataset support, place it in offline conservative MBRL.","nearest_boundary":"D7","representative_topics":["fine-tuning versus meta-rl on novel tasks in vision-based reinforcement learning benchmarks","action translation for sample-efficient meta-rl on sparse-reward heterogeneous tasks","context-based meta reinforcement learning via successor feature reward decomposition for data-efficient multi-task adaptation"]},{"direction_index":2,"label":"D2","name":"Theory and sample-efficiency foundations","description":"Formal characterization of model-based RL efficiency through regret, PAC, sample-complexity, representation-complexity, model-selection, exploration-scheduling, and structural guarantee frameworks.","inclusion":"Include work whose primary contribution is theoretical analysis or principled sample-efficiency framework design, including posterior-sampling guarantees, model selection, structural MDP regret, generalized eluder/PAC frameworks, confidence scheduling, and sample-complexity bounds.","exclusion":"If the formal result supports a concrete offline, planning, uncertainty, robust, or representation algorithm whose scientific objective dominates, place it in that mechanism-specific direction.","nearest_boundary":"D3","representative_topics":["optimistic posterior sampling for model-based rl with hellinger-distance regret analysis","model-based reinforcement learning for adaptive online influence maximization with regret guarantees","adaptive model selection in reinforcement learning over nested function approximation classes"]},{"direction_index":3,"label":"D3","name":"Model-based policy improvement","description":"Converting learned models, predictive simulators, or model-derived objectives into better actions and policies through planning, rollout control, value expansion, MPC, imitation, preference optimization, or policy-optimization loops.","inclusion":"Include work whose main objective is how model information is used for decision making or policy improvement, including adaptive rollout length, decision-time or background planning, goal/subgoal planning when aimed at control, visual foresight, decentralized model-based optimization, model-based imitation, preference-based policy optimization, and broad surveys or cross-cutting methods centered on policy-learning pipelines.","exclusion":"If the main contribution is the world-model representation itself, place it in world-model representation and architecture; if goal or subgoal machinery is primarily for reward discovery, place it in exploration and sparse-reward discovery; if deployment constraints dominate, place it in structured dynamics and domain models.","nearest_boundary":"D5","representative_topics":["adaptive rollout length optimization in model-based rl via meta-level deep reinforcement learning","value-based decision-time versus background planning in model-based reinforcement learning and transfer learning","model-based entropy-regularized imitation learning with joint policy and model adversarial training"]},{"direction_index":4,"label":"D4","name":"Exploration and sparse-reward discovery","description":"Using world models, planning, intrinsic motivation, reachability, memory, or reward shaping to discover rewards, cover environments, or guide behavior when external rewards are sparse, delayed, absent, or hard to reach.","inclusion":"Include work whose main objective is exploration itself, including intrinsic reward learning, curiosity, reward-free or sparse-reward coverage, goal-space exploration, hierarchical subgoal generation for discovery, memory-guided exploration, self-imitation with intrinsic motivation, and information-directed exploration.","exclusion":"If uncertainty is mainly used to calibrate risk, safety, or robustness rather than to discover new behavior, place it in reliability under uncertainty; if the method is primarily a generic planning/control algorithm, place it in model-based policy improvement.","nearest_boundary":"D6","representative_topics":["goal-space background planning with subgoal-conditioned models for model-based reinforcement learning","pixel-based hierarchical planning via latent world model goal selection and action decomposition","deterministic exploration-exploitation scheduling for joint mdp learning and policy optimization"]},{"direction_index":5,"label":"D5","name":"World-model representation and architecture","description":"Designing latent states, perceptual encoders, model architectures, compression criteria, and training objectives that make learned world models compact, predictive, decision-relevant, and useful for downstream control.","inclusion":"Include work whose main objective is the model or representation rather than the downstream planner, including visual/object/entity-centric representations, masked or sparse latent world models, value-equivalent models, rate-distortion compression, graph or memory world models, policy-adapted dynamics fitting, demonstration-guided world-model learning, and multi-agent latent world models.","exclusion":"If the key claim is causal, invariant, physics-informed, simulator-structured, or domain-structured transition modeling, place it in structured dynamics and domain models; if the main contribution is the planner or rollout policy, place it in model-based policy improvement.","nearest_boundary":"D7","representative_topics":["rate-distortion-based approximate value equivalence for tractable model-based reinforcement learning","value-equivalent model compression via rate-distortion for capacity-constrained model-based reinforcement learning","entity-based reinforcement learning using slot attention and graph neural networks for structured observations"]},{"direction_index":6,"label":"D6","name":"Reliability under uncertainty and risk","description":"Making model-based or model-informed decision making reliable under epistemic uncertainty, aleatoric uncertainty, adversarial perturbations, risk-sensitive criteria, safety constraints, or robust MDP uncertainty sets.","inclusion":"Include work whose main objective is uncertainty quantification, uncertainty-aware planning, safety certificates, Bayesian or ensemble-style probabilistic control, value-function uncertainty, explicit robust MDP optimization, adversarial robustness, distributional or worst-case objectives, or risk-sensitive criteria such as exponential utility.","exclusion":"If uncertainty is specifically a pessimistic mechanism for fixed offline data support, place it in offline conservative MBRL; if the focus is formal sample-complexity theory rather than a reliability objective, place it in theory and sample-efficiency foundations.","nearest_boundary":"D0","representative_topics":["bayesian coreset compression for model-based reinforcement learning via kernelized stein discrepancy","safe model-based reinforcement learning via uncertainty-aware distributional reachability certificates","variational recurrent kalman networks for principled aleatoric and epistemic uncertainty in model-based reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Structured dynamics and domain models","description":"Learning or exploiting transition models with causal, invariant, physical, simulator, or domain-specific structure to improve generalization, sample efficiency, and control in structured environments.","inclusion":"Include work whose main objective is structured model learning or structured deployment, including causal and intervention-invariant dynamics, modular world models, parsimonious dynamics, physics-informed or Hamiltonian/ODE/SINDy models, differentiable simulators or renderers, surrogate domain simulators, and domain-specific model-based RL systems where environment structure is central.","exclusion":"If the work targets task-level meta-adaptation or continual change rather than the structure of the transition model, place it in transfer, meta, and continual adaptation; if the architecture is generic and not justified by causal, physical, simulator, or domain structure, place it in world-model representation and architecture.","nearest_boundary":"D1","representative_topics":["unsupervised dynamics generalization via relational intervention for model-based reinforcement learning","variational causal world models for modular adaptation across intervened environments","causal dynamics learning for task-independent state abstraction in model-based reinforcement learning"]}],"split":"dev"}
{"episode_id":"F149_T2025-01","field_index":149,"field":"model-based reinforcement learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Offline conservative MBRL","description":"Reliable policy improvement from fixed logged datasets using learned or implicit dynamics while controlling batch-distribution shift, unsupported rollouts, and model exploitation.","inclusion":"Include work whose main objective is offline model-based policy learning under limited support, using pessimism, conservatism, coverage guarantees, distributional robustness tied to dataset support, confidence-filtered imagination, or causal offline world models.","exclusion":"If uncertainty is used mainly for online exploration, safety, or calibrated planning rather than fixed-dataset support, place it in reliability under uncertainty; if the main contribution is a generic model architecture, place it in world-model representation and architecture.","nearest_boundary":"D6","representative_topics":["model generation with provable coverability for offline dynamics-aware policy learning","causal world models for offline reinforcement learning generalization","confidence-aware bidirectional model-based imagination for offline reinforcement learning generalization"]},{"direction_index":1,"label":"D1","name":"Transfer, meta, and continual adaptation","description":"Rapidly adapting model-based RL systems across tasks, users, environments, or time-varying deployments through meta-learning, task representations, continual model maintenance, and lifelong adaptation.","inclusion":"Include work whose main objective is task-level transfer, few-shot adaptation, meta-RL, personalized or robust meta-learning, context inference, continual world-model adaptation, change handling, selective replay, or lifelong/non-stationary model reuse.","exclusion":"If the main objective is invariant or physically structured transition-model discovery rather than adaptation over task distributions or time, place it in structured dynamics; if the main issue is offline dataset support, place it in offline conservative MBRL.","nearest_boundary":"D7","representative_topics":["fine-tuning versus meta-rl on novel tasks in vision-based reinforcement learning benchmarks","action translation for sample-efficient meta-rl on sparse-reward heterogeneous tasks","context-based meta reinforcement learning via successor feature reward decomposition for data-efficient multi-task adaptation"]},{"direction_index":2,"label":"D2","name":"Theory and sample-efficiency foundations","description":"Formal characterization of model-based RL efficiency through regret, PAC, sample-complexity, representation-complexity, model-selection, exploration-scheduling, and structural guarantee frameworks.","inclusion":"Include work whose primary contribution is theoretical analysis or principled sample-efficiency framework design, including posterior-sampling guarantees, model selection, structural MDP regret, generalized eluder/PAC frameworks, confidence scheduling, and sample-complexity bounds.","exclusion":"If the formal result supports a concrete offline, planning, uncertainty, robust, or representation algorithm whose scientific objective dominates, place it in that mechanism-specific direction.","nearest_boundary":"D3","representative_topics":["optimistic posterior sampling for model-based rl with hellinger-distance regret analysis","model-based reinforcement learning for adaptive online influence maximization with regret guarantees","adaptive model selection in reinforcement learning over nested function approximation classes"]},{"direction_index":3,"label":"D3","name":"Model-based policy improvement","description":"Converting learned models, predictive simulators, or model-derived objectives into better actions and policies through planning, rollout control, value expansion, MPC, imitation, preference optimization, or policy-optimization loops.","inclusion":"Include work whose main objective is how model information is used for decision making or policy improvement, including adaptive rollout length, decision-time or background planning, goal/subgoal planning when aimed at control, visual foresight, decentralized model-based optimization, model-based imitation, preference-based policy optimization, and broad surveys or cross-cutting methods centered on policy-learning pipelines.","exclusion":"If the main contribution is the world-model representation itself, place it in world-model representation and architecture; if goal or subgoal machinery is primarily for reward discovery, place it in exploration and sparse-reward discovery; if deployment constraints dominate, place it in structured dynamics and domain models.","nearest_boundary":"D5","representative_topics":["adaptive rollout length optimization in model-based rl via meta-level deep reinforcement learning","value-based decision-time versus background planning in model-based reinforcement learning and transfer learning","model-based entropy-regularized imitation learning with joint policy and model adversarial training"]},{"direction_index":4,"label":"D4","name":"Exploration and sparse-reward discovery","description":"Using world models, planning, intrinsic motivation, reachability, memory, or reward shaping to discover rewards, cover environments, or guide behavior when external rewards are sparse, delayed, absent, or hard to reach.","inclusion":"Include work whose main objective is exploration itself, including intrinsic reward learning, curiosity, reward-free or sparse-reward coverage, goal-space exploration, hierarchical subgoal generation for discovery, memory-guided exploration, self-imitation with intrinsic motivation, and information-directed exploration.","exclusion":"If uncertainty is mainly used to calibrate risk, safety, or robustness rather than to discover new behavior, place it in reliability under uncertainty; if the method is primarily a generic planning/control algorithm, place it in model-based policy improvement.","nearest_boundary":"D6","representative_topics":["goal-space background planning with subgoal-conditioned models for model-based reinforcement learning","pixel-based hierarchical planning via latent world model goal selection and action decomposition","deterministic exploration-exploitation scheduling for joint mdp learning and policy optimization"]},{"direction_index":5,"label":"D5","name":"World-model representation and architecture","description":"Designing latent states, perceptual encoders, model architectures, compression criteria, and training objectives that make learned world models compact, predictive, decision-relevant, and useful for downstream control.","inclusion":"Include work whose main objective is the model or representation rather than the downstream planner, including visual/object/entity-centric representations, masked or sparse latent world models, value-equivalent models, rate-distortion compression, graph or memory world models, policy-adapted dynamics fitting, demonstration-guided world-model learning, and multi-agent latent world models.","exclusion":"If the key claim is causal, invariant, physics-informed, simulator-structured, or domain-structured transition modeling, place it in structured dynamics and domain models; if the main contribution is the planner or rollout policy, place it in model-based policy improvement.","nearest_boundary":"D7","representative_topics":["rate-distortion-based approximate value equivalence for tractable model-based reinforcement learning","value-equivalent model compression via rate-distortion for capacity-constrained model-based reinforcement learning","entity-based reinforcement learning using slot attention and graph neural networks for structured observations"]},{"direction_index":6,"label":"D6","name":"Reliability under uncertainty and risk","description":"Making model-based or model-informed decision making reliable under epistemic uncertainty, aleatoric uncertainty, adversarial perturbations, risk-sensitive criteria, safety constraints, or robust MDP uncertainty sets.","inclusion":"Include work whose main objective is uncertainty quantification, uncertainty-aware planning, safety certificates, Bayesian or ensemble-style probabilistic control, value-function uncertainty, explicit robust MDP optimization, adversarial robustness, distributional or worst-case objectives, or risk-sensitive criteria such as exponential utility.","exclusion":"If uncertainty is specifically a pessimistic mechanism for fixed offline data support, place it in offline conservative MBRL; if the focus is formal sample-complexity theory rather than a reliability objective, place it in theory and sample-efficiency foundations.","nearest_boundary":"D0","representative_topics":["bayesian coreset compression for model-based reinforcement learning via kernelized stein discrepancy","safe model-based reinforcement learning via uncertainty-aware distributional reachability certificates","variational recurrent kalman networks for principled aleatoric and epistemic uncertainty in model-based reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Structured dynamics and domain models","description":"Learning or exploiting transition models with causal, invariant, physical, simulator, or domain-specific structure to improve generalization, sample efficiency, and control in structured environments.","inclusion":"Include work whose main objective is structured model learning or structured deployment, including causal and intervention-invariant dynamics, modular world models, parsimonious dynamics, physics-informed or Hamiltonian/ODE/SINDy models, differentiable simulators or renderers, surrogate domain simulators, and domain-specific model-based RL systems where environment structure is central.","exclusion":"If the work targets task-level meta-adaptation or continual change rather than the structure of the transition model, place it in transfer, meta, and continual adaptation; if the architecture is generic and not justified by causal, physical, simulator, or domain structure, place it in world-model representation and architecture.","nearest_boundary":"D1","representative_topics":["unsupervised dynamics generalization via relational intervention for model-based reinforcement learning","variational causal world models for modular adaptation across intervened environments","causal dynamics learning for task-independent state abstraction in model-based reinforcement learning"]}],"split":"dev"}
{"episode_id":"F149_T2025-07","field_index":149,"field":"model-based reinforcement learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Offline conservative MBRL","description":"Reliable policy improvement from fixed logged datasets using learned or implicit dynamics while controlling batch-distribution shift, unsupported rollouts, and model exploitation.","inclusion":"Include work whose main objective is offline model-based policy learning under limited support, using pessimism, conservatism, coverage guarantees, distributional robustness tied to dataset support, confidence-filtered imagination, or causal offline world models.","exclusion":"If uncertainty is used mainly for online exploration, safety, or calibrated planning rather than fixed-dataset support, place it in reliability under uncertainty; if the main contribution is a generic model architecture, place it in world-model representation and architecture.","nearest_boundary":"D6","representative_topics":["model generation with provable coverability for offline dynamics-aware policy learning","causal world models for offline reinforcement learning generalization","confidence-aware bidirectional model-based imagination for offline reinforcement learning generalization"]},{"direction_index":1,"label":"D1","name":"Transfer, meta, and continual adaptation","description":"Rapidly adapting model-based RL systems across tasks, users, environments, or time-varying deployments through meta-learning, task representations, continual model maintenance, and lifelong adaptation.","inclusion":"Include work whose main objective is task-level transfer, few-shot adaptation, meta-RL, personalized or robust meta-learning, context inference, continual world-model adaptation, change handling, selective replay, or lifelong/non-stationary model reuse.","exclusion":"If the main objective is invariant or physically structured transition-model discovery rather than adaptation over task distributions or time, place it in structured dynamics; if the main issue is offline dataset support, place it in offline conservative MBRL.","nearest_boundary":"D7","representative_topics":["fine-tuning versus meta-rl on novel tasks in vision-based reinforcement learning benchmarks","action translation for sample-efficient meta-rl on sparse-reward heterogeneous tasks","context-based meta reinforcement learning via successor feature reward decomposition for data-efficient multi-task adaptation"]},{"direction_index":2,"label":"D2","name":"Theory and sample-efficiency foundations","description":"Formal characterization of model-based RL efficiency through regret, PAC, sample-complexity, representation-complexity, model-selection, exploration-scheduling, and structural guarantee frameworks.","inclusion":"Include work whose primary contribution is theoretical analysis or principled sample-efficiency framework design, including posterior-sampling guarantees, model selection, structural MDP regret, generalized eluder/PAC frameworks, confidence scheduling, and sample-complexity bounds.","exclusion":"If the formal result supports a concrete offline, planning, uncertainty, robust, or representation algorithm whose scientific objective dominates, place it in that mechanism-specific direction.","nearest_boundary":"D3","representative_topics":["optimistic posterior sampling for model-based rl with hellinger-distance regret analysis","model-based reinforcement learning for adaptive online influence maximization with regret guarantees","adaptive model selection in reinforcement learning over nested function approximation classes"]},{"direction_index":3,"label":"D3","name":"Model-based policy improvement","description":"Converting learned models, predictive simulators, or model-derived objectives into better actions and policies through planning, rollout control, value expansion, MPC, imitation, preference optimization, or policy-optimization loops.","inclusion":"Include work whose main objective is how model information is used for decision making or policy improvement, including adaptive rollout length, decision-time or background planning, goal/subgoal planning when aimed at control, visual foresight, decentralized model-based optimization, model-based imitation, preference-based policy optimization, and broad surveys or cross-cutting methods centered on policy-learning pipelines.","exclusion":"If the main contribution is the world-model representation itself, place it in world-model representation and architecture; if goal or subgoal machinery is primarily for reward discovery, place it in exploration and sparse-reward discovery; if deployment constraints dominate, place it in structured dynamics and domain models.","nearest_boundary":"D5","representative_topics":["adaptive rollout length optimization in model-based rl via meta-level deep reinforcement learning","value-based decision-time versus background planning in model-based reinforcement learning and transfer learning","model-based entropy-regularized imitation learning with joint policy and model adversarial training"]},{"direction_index":4,"label":"D4","name":"Exploration and sparse-reward discovery","description":"Using world models, planning, intrinsic motivation, reachability, memory, or reward shaping to discover rewards, cover environments, or guide behavior when external rewards are sparse, delayed, absent, or hard to reach.","inclusion":"Include work whose main objective is exploration itself, including intrinsic reward learning, curiosity, reward-free or sparse-reward coverage, goal-space exploration, hierarchical subgoal generation for discovery, memory-guided exploration, self-imitation with intrinsic motivation, and information-directed exploration.","exclusion":"If uncertainty is mainly used to calibrate risk, safety, or robustness rather than to discover new behavior, place it in reliability under uncertainty; if the method is primarily a generic planning/control algorithm, place it in model-based policy improvement.","nearest_boundary":"D6","representative_topics":["goal-space background planning with subgoal-conditioned models for model-based reinforcement learning","pixel-based hierarchical planning via latent world model goal selection and action decomposition","deterministic exploration-exploitation scheduling for joint mdp learning and policy optimization"]},{"direction_index":5,"label":"D5","name":"World-model representation and architecture","description":"Designing latent states, perceptual encoders, model architectures, compression criteria, and training objectives that make learned world models compact, predictive, decision-relevant, and useful for downstream control.","inclusion":"Include work whose main objective is the model or representation rather than the downstream planner, including visual/object/entity-centric representations, masked or sparse latent world models, value-equivalent models, rate-distortion compression, graph or memory world models, policy-adapted dynamics fitting, demonstration-guided world-model learning, and multi-agent latent world models.","exclusion":"If the key claim is causal, invariant, physics-informed, simulator-structured, or domain-structured transition modeling, place it in structured dynamics and domain models; if the main contribution is the planner or rollout policy, place it in model-based policy improvement.","nearest_boundary":"D7","representative_topics":["rate-distortion-based approximate value equivalence for tractable model-based reinforcement learning","value-equivalent model compression via rate-distortion for capacity-constrained model-based reinforcement learning","entity-based reinforcement learning using slot attention and graph neural networks for structured observations"]},{"direction_index":6,"label":"D6","name":"Reliability under uncertainty and risk","description":"Making model-based or model-informed decision making reliable under epistemic uncertainty, aleatoric uncertainty, adversarial perturbations, risk-sensitive criteria, safety constraints, or robust MDP uncertainty sets.","inclusion":"Include work whose main objective is uncertainty quantification, uncertainty-aware planning, safety certificates, Bayesian or ensemble-style probabilistic control, value-function uncertainty, explicit robust MDP optimization, adversarial robustness, distributional or worst-case objectives, or risk-sensitive criteria such as exponential utility.","exclusion":"If uncertainty is specifically a pessimistic mechanism for fixed offline data support, place it in offline conservative MBRL; if the focus is formal sample-complexity theory rather than a reliability objective, place it in theory and sample-efficiency foundations.","nearest_boundary":"D0","representative_topics":["bayesian coreset compression for model-based reinforcement learning via kernelized stein discrepancy","safe model-based reinforcement learning via uncertainty-aware distributional reachability certificates","variational recurrent kalman networks for principled aleatoric and epistemic uncertainty in model-based reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Structured dynamics and domain models","description":"Learning or exploiting transition models with causal, invariant, physical, simulator, or domain-specific structure to improve generalization, sample efficiency, and control in structured environments.","inclusion":"Include work whose main objective is structured model learning or structured deployment, including causal and intervention-invariant dynamics, modular world models, parsimonious dynamics, physics-informed or Hamiltonian/ODE/SINDy models, differentiable simulators or renderers, surrogate domain simulators, and domain-specific model-based RL systems where environment structure is central.","exclusion":"If the work targets task-level meta-adaptation or continual change rather than the structure of the transition model, place it in transfer, meta, and continual adaptation; if the architecture is generic and not justified by causal, physical, simulator, or domain structure, place it in world-model representation and architecture.","nearest_boundary":"D1","representative_topics":["unsupervised dynamics generalization via relational intervention for model-based reinforcement learning","variational causal world models for modular adaptation across intervened environments","causal dynamics learning for task-independent state abstraction in model-based reinforcement learning"]}],"split":"dev"}
{"episode_id":"F149_T2026-01","field_index":149,"field":"model-based reinforcement learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Offline conservative MBRL","description":"Reliable policy improvement from fixed logged datasets using learned or implicit dynamics while controlling batch-distribution shift, unsupported rollouts, and model exploitation.","inclusion":"Include work whose main objective is offline model-based policy learning under limited support, using pessimism, conservatism, coverage guarantees, distributional robustness tied to dataset support, confidence-filtered imagination, or causal offline world models.","exclusion":"If uncertainty is used mainly for online exploration, safety, or calibrated planning rather than fixed-dataset support, place it in reliability under uncertainty; if the main contribution is a generic model architecture, place it in world-model representation and architecture.","nearest_boundary":"D6","representative_topics":["model generation with provable coverability for offline dynamics-aware policy learning","causal world models for offline reinforcement learning generalization","confidence-aware bidirectional model-based imagination for offline reinforcement learning generalization"]},{"direction_index":1,"label":"D1","name":"Transfer, meta, and continual adaptation","description":"Rapidly adapting model-based RL systems across tasks, users, environments, or time-varying deployments through meta-learning, task representations, continual model maintenance, and lifelong adaptation.","inclusion":"Include work whose main objective is task-level transfer, few-shot adaptation, meta-RL, personalized or robust meta-learning, context inference, continual world-model adaptation, change handling, selective replay, or lifelong/non-stationary model reuse.","exclusion":"If the main objective is invariant or physically structured transition-model discovery rather than adaptation over task distributions or time, place it in structured dynamics; if the main issue is offline dataset support, place it in offline conservative MBRL.","nearest_boundary":"D7","representative_topics":["fine-tuning versus meta-rl on novel tasks in vision-based reinforcement learning benchmarks","action translation for sample-efficient meta-rl on sparse-reward heterogeneous tasks","context-based meta reinforcement learning via successor feature reward decomposition for data-efficient multi-task adaptation"]},{"direction_index":2,"label":"D2","name":"Theory and sample-efficiency foundations","description":"Formal characterization of model-based RL efficiency through regret, PAC, sample-complexity, representation-complexity, model-selection, exploration-scheduling, and structural guarantee frameworks.","inclusion":"Include work whose primary contribution is theoretical analysis or principled sample-efficiency framework design, including posterior-sampling guarantees, model selection, structural MDP regret, generalized eluder/PAC frameworks, confidence scheduling, and sample-complexity bounds.","exclusion":"If the formal result supports a concrete offline, planning, uncertainty, robust, or representation algorithm whose scientific objective dominates, place it in that mechanism-specific direction.","nearest_boundary":"D3","representative_topics":["optimistic posterior sampling for model-based rl with hellinger-distance regret analysis","model-based reinforcement learning for adaptive online influence maximization with regret guarantees","adaptive model selection in reinforcement learning over nested function approximation classes"]},{"direction_index":3,"label":"D3","name":"Model-based policy improvement","description":"Converting learned models, predictive simulators, or model-derived objectives into better actions and policies through planning, rollout control, value expansion, MPC, imitation, preference optimization, or policy-optimization loops.","inclusion":"Include work whose main objective is how model information is used for decision making or policy improvement, including adaptive rollout length, decision-time or background planning, goal/subgoal planning when aimed at control, visual foresight, decentralized model-based optimization, model-based imitation, preference-based policy optimization, and broad surveys or cross-cutting methods centered on policy-learning pipelines.","exclusion":"If the main contribution is the world-model representation itself, place it in world-model representation and architecture; if goal or subgoal machinery is primarily for reward discovery, place it in exploration and sparse-reward discovery; if deployment constraints dominate, place it in structured dynamics and domain models.","nearest_boundary":"D5","representative_topics":["adaptive rollout length optimization in model-based rl via meta-level deep reinforcement learning","value-based decision-time versus background planning in model-based reinforcement learning and transfer learning","model-based entropy-regularized imitation learning with joint policy and model adversarial training"]},{"direction_index":4,"label":"D4","name":"Exploration and sparse-reward discovery","description":"Using world models, planning, intrinsic motivation, reachability, memory, or reward shaping to discover rewards, cover environments, or guide behavior when external rewards are sparse, delayed, absent, or hard to reach.","inclusion":"Include work whose main objective is exploration itself, including intrinsic reward learning, curiosity, reward-free or sparse-reward coverage, goal-space exploration, hierarchical subgoal generation for discovery, memory-guided exploration, self-imitation with intrinsic motivation, and information-directed exploration.","exclusion":"If uncertainty is mainly used to calibrate risk, safety, or robustness rather than to discover new behavior, place it in reliability under uncertainty; if the method is primarily a generic planning/control algorithm, place it in model-based policy improvement.","nearest_boundary":"D6","representative_topics":["goal-space background planning with subgoal-conditioned models for model-based reinforcement learning","pixel-based hierarchical planning via latent world model goal selection and action decomposition","deterministic exploration-exploitation scheduling for joint mdp learning and policy optimization"]},{"direction_index":5,"label":"D5","name":"World-model representation and architecture","description":"Designing latent states, perceptual encoders, model architectures, compression criteria, and training objectives that make learned world models compact, predictive, decision-relevant, and useful for downstream control.","inclusion":"Include work whose main objective is the model or representation rather than the downstream planner, including visual/object/entity-centric representations, masked or sparse latent world models, value-equivalent models, rate-distortion compression, graph or memory world models, policy-adapted dynamics fitting, demonstration-guided world-model learning, and multi-agent latent world models.","exclusion":"If the key claim is causal, invariant, physics-informed, simulator-structured, or domain-structured transition modeling, place it in structured dynamics and domain models; if the main contribution is the planner or rollout policy, place it in model-based policy improvement.","nearest_boundary":"D7","representative_topics":["rate-distortion-based approximate value equivalence for tractable model-based reinforcement learning","value-equivalent model compression via rate-distortion for capacity-constrained model-based reinforcement learning","entity-based reinforcement learning using slot attention and graph neural networks for structured observations"]},{"direction_index":6,"label":"D6","name":"Reliability under uncertainty and risk","description":"Making model-based or model-informed decision making reliable under epistemic uncertainty, aleatoric uncertainty, adversarial perturbations, risk-sensitive criteria, safety constraints, or robust MDP uncertainty sets.","inclusion":"Include work whose main objective is uncertainty quantification, uncertainty-aware planning, safety certificates, Bayesian or ensemble-style probabilistic control, value-function uncertainty, explicit robust MDP optimization, adversarial robustness, distributional or worst-case objectives, or risk-sensitive criteria such as exponential utility.","exclusion":"If uncertainty is specifically a pessimistic mechanism for fixed offline data support, place it in offline conservative MBRL; if the focus is formal sample-complexity theory rather than a reliability objective, place it in theory and sample-efficiency foundations.","nearest_boundary":"D0","representative_topics":["bayesian coreset compression for model-based reinforcement learning via kernelized stein discrepancy","safe model-based reinforcement learning via uncertainty-aware distributional reachability certificates","variational recurrent kalman networks for principled aleatoric and epistemic uncertainty in model-based reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Structured dynamics and domain models","description":"Learning or exploiting transition models with causal, invariant, physical, simulator, or domain-specific structure to improve generalization, sample efficiency, and control in structured environments.","inclusion":"Include work whose main objective is structured model learning or structured deployment, including causal and intervention-invariant dynamics, modular world models, parsimonious dynamics, physics-informed or Hamiltonian/ODE/SINDy models, differentiable simulators or renderers, surrogate domain simulators, and domain-specific model-based RL systems where environment structure is central.","exclusion":"If the work targets task-level meta-adaptation or continual change rather than the structure of the transition model, place it in transfer, meta, and continual adaptation; if the architecture is generic and not justified by causal, physical, simulator, or domain structure, place it in world-model representation and architecture.","nearest_boundary":"D1","representative_topics":["unsupervised dynamics generalization via relational intervention for model-based reinforcement learning","variational causal world models for modular adaptation across intervened environments","causal dynamics learning for task-independent state abstraction in model-based reinforcement learning"]}],"split":"dev"}
{"episode_id":"F150_T2024-01","field_index":150,"field":"human motion generation","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Semantic motion grounding","description":"Generate, retrieve, edit, or evaluate 3D human motion from natural-language descriptions, action labels, open-vocabulary prompts, or compositional semantic specifications.","inclusion":"Assign atoms whose primary objective is grounding semantic intent in plausible human motion, including text-to-motion, action-label-to-motion, reciprocal text-motion modeling, open-vocabulary generalization, temporal action composition, and semantic alignment metrics.","exclusion":"If the central objective is learning a reusable generative backbone independent of semantic grounding, assign to general motion priors; if scenes, objects, music, speech, other humans, or physical control are the main conditioning problem, assign to that direction.","nearest_boundary":"General motion priors","representative_topics":["tokenized autoregressive modeling for reciprocal text and 3d motion generation","diffusion-based text-conditioned motion synthesis and joint-wise editing without fine-tuning","hierarchical autoregressive temporal action composition for text-conditioned 3d human motion synthesis"]},{"direction_index":1,"label":"D1","name":"General motion priors","description":"Learn reusable human-motion distributions, representations, and generative mechanisms for synthesizing, sampling, completing, or weakly conditioning motion across tasks.","inclusion":"Assign atoms whose primary objective is a broadly reusable generative prior or backbone, including discrete motion tokens, language-model-style sequence generation, VAEs, diffusion models, latent diffusion, neural fields, masked modeling, and generic multi-condition architectures.","exclusion":"If the primary research objective is a specific external grounding signal such as text, music, speech, objects, scenes, other humans, future extrapolation, editing, or simulation control, assign to that more specific direction.","nearest_boundary":"Semantic motion grounding","representative_topics":["implicit neural motion fields as continuous generative priors for motion interpolation and editing","recurrent transformer variational autoencoders for multi-action arbitrary-length motion sequence synthesis","unsupervised unconditional motion synthesis with diverse motion manifold and semantic editing"]},{"direction_index":2,"label":"D2","name":"Music choreography synthesis","description":"Generate, regenerate, transfer, or evaluate dance and choreographic motion from music, rhythm, choreography labels, key poses, or dance-specific creative supervision.","inclusion":"Assign atoms whose primary objective is dance or choreography, including audio-conditioned dance, beat synchronization, keyframe-constrained dance, artist-centric choreography, dance style transfer, group choreography, and dance-specific datasets.","exclusion":"Speech-driven communicative gesture belongs to communicative gesture synthesis; generic audio-conditioned motion belongs here only when the output objective is explicitly dance or choreography.","nearest_boundary":"Communicative gesture synthesis","representative_topics":["music-driven dance motion synthesis with controllable key pose constraints via cross-modal transformer","keyframe-based dance motion synthesis using hierarchical rnn and transformer constraint controllers","audio-conditioned dance motion synthesis for complex acrobatic poses via hybrid keypoint annotation"]},{"direction_index":3,"label":"D3","name":"Communicative gesture synthesis","description":"Generate co-speech gestures, sign poses, upper-body motion, hands, face, or holistic speaking-avatar motion from speech, transcripts, notation, or multimodal communicative signals.","inclusion":"Assign atoms whose primary objective is communicative body motion aligned with speech, text, sign notation, or speaker style, including speech-to-gesture, sign-language pose animation, holistic speech-driven avatars, style control, and gesture-generation benchmarks.","exclusion":"Music-driven dance belongs to music choreography synthesis; generic full-body text-to-motion without communicative speech, sign, or avatar alignment belongs to semantic motion grounding.","nearest_boundary":"Music choreography synthesis","representative_topics":["zero-shot multimodal style transfer for gesture animation via adversarial disentanglement of speaker style embeddings","large-scale crowdsourced benchmarking of co-speech gesture generation systems","weakly supervised sign language notation animation into pose sequences via transformer encoders"]},{"direction_index":4,"label":"D4","name":"Embodied environment interaction","description":"Generate motion that is physically or spatially grounded in external environments, objects, contacts, affordances, navigation constraints, or simulated dynamics.","inclusion":"Assign atoms whose primary objective is coherent embodied interaction with scenes, objects, or physical dynamics, including human-object manipulation, task-oriented affordance use, scene-aware locomotion, collision-aware navigation, environment-conditioned motion, physics simulation, reinforcement-learning control, humanoid imitation, and dynamics-feasible animation.","exclusion":"If the central coupling is with another person rather than the environment, assign to multi-human interaction; if physical losses only support a generic kinematic generator, assign to general motion priors; if semantic text alone is the main condition without scene or dynamics grounding, assign to semantic motion grounding.","nearest_boundary":"Multi-human interaction","representative_topics":["language-conditioned 3d human motion generation with scene and object interaction","intent-driven full-body motion synthesis for virtual human-object interactions","physics-simulated motion in-betweening via skeleton optimization and humanoid robot imitation"]},{"direction_index":5,"label":"D5","name":"Temporal forecasting and completion","description":"Predict, continue, interpolate, or fill human motion from observed past poses, keyframes, partial temporal constraints, or incomplete motion sequences.","inclusion":"Assign atoms whose primary objective is temporal extrapolation or completion from motion observations, including stochastic future prediction, multimodal forecasting, long-term continuation, pose interpolation, gap filling, and temporal consistency from observed history.","exclusion":"If the output is generated de novo from text, music, speech, scenes, objects, social prompts, or physical policies rather than observed motion history, assign to that conditioning objective; physics-simulated in-betweening belongs to embodied environment interaction.","nearest_boundary":"Motion editing and transfer","representative_topics":["diverse multimodal human motion prediction via auxiliary space gumbel-softmax sampling","privileged knowledge distillation for human motion prediction via pose-based interpolation","diverse multi-motion 3d human joint prediction using attention-based generative adversarial networks"]},{"direction_index":6,"label":"D6","name":"Motion editing and transfer","description":"Transform, retarget, stylize, complete, steer, or visually realize existing or partially specified human motion under user constraints, source-target mappings, sparse signals, or video-space requirements.","inclusion":"Assign atoms whose primary objective is controllable transformation or realization of motion content, including skeleton or person motion transfer, retargeting, style transfer, latent editing, sparse-tracking full-body completion, directable characters, motion-to-video transfer, human dance video synthesis, and video/2D-consistency-driven motion generation.","exclusion":"If the primary objective is de novo semantic, music, speech, object, scene, social, or physics-conditioned synthesis, assign to that direction; if the core contribution is a general reusable prior rather than editing or transfer, assign to general motion priors.","nearest_boundary":"General motion priors","representative_topics":["implicit neural motion fields as continuous generative priors for motion interpolation and editing","unsupervised unconditional motion synthesis with diverse motion manifold and semantic editing","progressive region-to-whole gan framework for realistic human video motion transfer"]},{"direction_index":7,"label":"D7","name":"Multi-human interaction","description":"Generate coordinated, reactive, or socially coupled motion for two or more humans, emphasizing partner dependence, mutual response, interaction semantics, and multi-agent coherence.","inclusion":"Assign atoms whose primary objective is modeling inter-person motion dependencies, including dyadic reaction synthesis, partner-conditioned generation, two-person interaction diffusion, text-guided interpersonal motion, and highly interactive multi-person prediction.","exclusion":"Group dance belongs to music choreography synthesis when music or choreography is the primary condition; forecasting belongs to temporal forecasting when extrapolating from history is the main objective; object or scene coupling belongs to embodied environment interaction.","nearest_boundary":"Temporal forecasting and completion","representative_topics":["interaction transformer for reaction motion generation using temporal-spatial graph attention","partner-conditioned neural operator generation of dyadic human motion in function spaces","bipartite graph diffusion models for two-person human motion interaction generation"]}],"split":"dev"}
{"episode_id":"F150_T2024-07","field_index":150,"field":"human motion generation","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Semantic motion grounding","description":"Generate, retrieve, edit, or evaluate 3D human motion from natural-language descriptions, action labels, open-vocabulary prompts, or compositional semantic specifications.","inclusion":"Assign atoms whose primary objective is grounding semantic intent in plausible human motion, including text-to-motion, action-label-to-motion, reciprocal text-motion modeling, open-vocabulary generalization, temporal action composition, and semantic alignment metrics.","exclusion":"If the central objective is learning a reusable generative backbone independent of semantic grounding, assign to general motion priors; if scenes, objects, music, speech, other humans, or physical control are the main conditioning problem, assign to that direction.","nearest_boundary":"General motion priors","representative_topics":["tokenized autoregressive modeling for reciprocal text and 3d motion generation","diffusion-based text-conditioned motion synthesis and joint-wise editing without fine-tuning","hierarchical autoregressive temporal action composition for text-conditioned 3d human motion synthesis"]},{"direction_index":1,"label":"D1","name":"General motion priors","description":"Learn reusable human-motion distributions, representations, and generative mechanisms for synthesizing, sampling, completing, or weakly conditioning motion across tasks.","inclusion":"Assign atoms whose primary objective is a broadly reusable generative prior or backbone, including discrete motion tokens, language-model-style sequence generation, VAEs, diffusion models, latent diffusion, neural fields, masked modeling, and generic multi-condition architectures.","exclusion":"If the primary research objective is a specific external grounding signal such as text, music, speech, objects, scenes, other humans, future extrapolation, editing, or simulation control, assign to that more specific direction.","nearest_boundary":"Semantic motion grounding","representative_topics":["implicit neural motion fields as continuous generative priors for motion interpolation and editing","recurrent transformer variational autoencoders for multi-action arbitrary-length motion sequence synthesis","unsupervised unconditional motion synthesis with diverse motion manifold and semantic editing"]},{"direction_index":2,"label":"D2","name":"Music choreography synthesis","description":"Generate, regenerate, transfer, or evaluate dance and choreographic motion from music, rhythm, choreography labels, key poses, or dance-specific creative supervision.","inclusion":"Assign atoms whose primary objective is dance or choreography, including audio-conditioned dance, beat synchronization, keyframe-constrained dance, artist-centric choreography, dance style transfer, group choreography, and dance-specific datasets.","exclusion":"Speech-driven communicative gesture belongs to communicative gesture synthesis; generic audio-conditioned motion belongs here only when the output objective is explicitly dance or choreography.","nearest_boundary":"Communicative gesture synthesis","representative_topics":["music-driven dance motion synthesis with controllable key pose constraints via cross-modal transformer","keyframe-based dance motion synthesis using hierarchical rnn and transformer constraint controllers","audio-conditioned dance motion synthesis for complex acrobatic poses via hybrid keypoint annotation"]},{"direction_index":3,"label":"D3","name":"Communicative gesture synthesis","description":"Generate co-speech gestures, sign poses, upper-body motion, hands, face, or holistic speaking-avatar motion from speech, transcripts, notation, or multimodal communicative signals.","inclusion":"Assign atoms whose primary objective is communicative body motion aligned with speech, text, sign notation, or speaker style, including speech-to-gesture, sign-language pose animation, holistic speech-driven avatars, style control, and gesture-generation benchmarks.","exclusion":"Music-driven dance belongs to music choreography synthesis; generic full-body text-to-motion without communicative speech, sign, or avatar alignment belongs to semantic motion grounding.","nearest_boundary":"Music choreography synthesis","representative_topics":["zero-shot multimodal style transfer for gesture animation via adversarial disentanglement of speaker style embeddings","large-scale crowdsourced benchmarking of co-speech gesture generation systems","weakly supervised sign language notation animation into pose sequences via transformer encoders"]},{"direction_index":4,"label":"D4","name":"Embodied environment interaction","description":"Generate motion that is physically or spatially grounded in external environments, objects, contacts, affordances, navigation constraints, or simulated dynamics.","inclusion":"Assign atoms whose primary objective is coherent embodied interaction with scenes, objects, or physical dynamics, including human-object manipulation, task-oriented affordance use, scene-aware locomotion, collision-aware navigation, environment-conditioned motion, physics simulation, reinforcement-learning control, humanoid imitation, and dynamics-feasible animation.","exclusion":"If the central coupling is with another person rather than the environment, assign to multi-human interaction; if physical losses only support a generic kinematic generator, assign to general motion priors; if semantic text alone is the main condition without scene or dynamics grounding, assign to semantic motion grounding.","nearest_boundary":"Multi-human interaction","representative_topics":["language-conditioned 3d human motion generation with scene and object interaction","intent-driven full-body motion synthesis for virtual human-object interactions","physics-simulated motion in-betweening via skeleton optimization and humanoid robot imitation"]},{"direction_index":5,"label":"D5","name":"Temporal forecasting and completion","description":"Predict, continue, interpolate, or fill human motion from observed past poses, keyframes, partial temporal constraints, or incomplete motion sequences.","inclusion":"Assign atoms whose primary objective is temporal extrapolation or completion from motion observations, including stochastic future prediction, multimodal forecasting, long-term continuation, pose interpolation, gap filling, and temporal consistency from observed history.","exclusion":"If the output is generated de novo from text, music, speech, scenes, objects, social prompts, or physical policies rather than observed motion history, assign to that conditioning objective; physics-simulated in-betweening belongs to embodied environment interaction.","nearest_boundary":"Motion editing and transfer","representative_topics":["diverse multimodal human motion prediction via auxiliary space gumbel-softmax sampling","privileged knowledge distillation for human motion prediction via pose-based interpolation","diverse multi-motion 3d human joint prediction using attention-based generative adversarial networks"]},{"direction_index":6,"label":"D6","name":"Motion editing and transfer","description":"Transform, retarget, stylize, complete, steer, or visually realize existing or partially specified human motion under user constraints, source-target mappings, sparse signals, or video-space requirements.","inclusion":"Assign atoms whose primary objective is controllable transformation or realization of motion content, including skeleton or person motion transfer, retargeting, style transfer, latent editing, sparse-tracking full-body completion, directable characters, motion-to-video transfer, human dance video synthesis, and video/2D-consistency-driven motion generation.","exclusion":"If the primary objective is de novo semantic, music, speech, object, scene, social, or physics-conditioned synthesis, assign to that direction; if the core contribution is a general reusable prior rather than editing or transfer, assign to general motion priors.","nearest_boundary":"General motion priors","representative_topics":["implicit neural motion fields as continuous generative priors for motion interpolation and editing","unsupervised unconditional motion synthesis with diverse motion manifold and semantic editing","progressive region-to-whole gan framework for realistic human video motion transfer"]},{"direction_index":7,"label":"D7","name":"Multi-human interaction","description":"Generate coordinated, reactive, or socially coupled motion for two or more humans, emphasizing partner dependence, mutual response, interaction semantics, and multi-agent coherence.","inclusion":"Assign atoms whose primary objective is modeling inter-person motion dependencies, including dyadic reaction synthesis, partner-conditioned generation, two-person interaction diffusion, text-guided interpersonal motion, and highly interactive multi-person prediction.","exclusion":"Group dance belongs to music choreography synthesis when music or choreography is the primary condition; forecasting belongs to temporal forecasting when extrapolating from history is the main objective; object or scene coupling belongs to embodied environment interaction.","nearest_boundary":"Temporal forecasting and completion","representative_topics":["interaction transformer for reaction motion generation using temporal-spatial graph attention","partner-conditioned neural operator generation of dyadic human motion in function spaces","bipartite graph diffusion models for two-person human motion interaction generation"]}],"split":"dev"}
{"episode_id":"F150_T2025-01","field_index":150,"field":"human motion generation","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Semantic motion grounding","description":"Generate, retrieve, edit, or evaluate 3D human motion from natural-language descriptions, action labels, open-vocabulary prompts, or compositional semantic specifications.","inclusion":"Assign atoms whose primary objective is grounding semantic intent in plausible human motion, including text-to-motion, action-label-to-motion, reciprocal text-motion modeling, open-vocabulary generalization, temporal action composition, and semantic alignment metrics.","exclusion":"If the central objective is learning a reusable generative backbone independent of semantic grounding, assign to general motion priors; if scenes, objects, music, speech, other humans, or physical control are the main conditioning problem, assign to that direction.","nearest_boundary":"General motion priors","representative_topics":["tokenized autoregressive modeling for reciprocal text and 3d motion generation","diffusion-based text-conditioned motion synthesis and joint-wise editing without fine-tuning","hierarchical autoregressive temporal action composition for text-conditioned 3d human motion synthesis"]},{"direction_index":1,"label":"D1","name":"General motion priors","description":"Learn reusable human-motion distributions, representations, and generative mechanisms for synthesizing, sampling, completing, or weakly conditioning motion across tasks.","inclusion":"Assign atoms whose primary objective is a broadly reusable generative prior or backbone, including discrete motion tokens, language-model-style sequence generation, VAEs, diffusion models, latent diffusion, neural fields, masked modeling, and generic multi-condition architectures.","exclusion":"If the primary research objective is a specific external grounding signal such as text, music, speech, objects, scenes, other humans, future extrapolation, editing, or simulation control, assign to that more specific direction.","nearest_boundary":"Semantic motion grounding","representative_topics":["implicit neural motion fields as continuous generative priors for motion interpolation and editing","recurrent transformer variational autoencoders for multi-action arbitrary-length motion sequence synthesis","unsupervised unconditional motion synthesis with diverse motion manifold and semantic editing"]},{"direction_index":2,"label":"D2","name":"Music choreography synthesis","description":"Generate, regenerate, transfer, or evaluate dance and choreographic motion from music, rhythm, choreography labels, key poses, or dance-specific creative supervision.","inclusion":"Assign atoms whose primary objective is dance or choreography, including audio-conditioned dance, beat synchronization, keyframe-constrained dance, artist-centric choreography, dance style transfer, group choreography, and dance-specific datasets.","exclusion":"Speech-driven communicative gesture belongs to communicative gesture synthesis; generic audio-conditioned motion belongs here only when the output objective is explicitly dance or choreography.","nearest_boundary":"Communicative gesture synthesis","representative_topics":["music-driven dance motion synthesis with controllable key pose constraints via cross-modal transformer","keyframe-based dance motion synthesis using hierarchical rnn and transformer constraint controllers","audio-conditioned dance motion synthesis for complex acrobatic poses via hybrid keypoint annotation"]},{"direction_index":3,"label":"D3","name":"Communicative gesture synthesis","description":"Generate co-speech gestures, sign poses, upper-body motion, hands, face, or holistic speaking-avatar motion from speech, transcripts, notation, or multimodal communicative signals.","inclusion":"Assign atoms whose primary objective is communicative body motion aligned with speech, text, sign notation, or speaker style, including speech-to-gesture, sign-language pose animation, holistic speech-driven avatars, style control, and gesture-generation benchmarks.","exclusion":"Music-driven dance belongs to music choreography synthesis; generic full-body text-to-motion without communicative speech, sign, or avatar alignment belongs to semantic motion grounding.","nearest_boundary":"Music choreography synthesis","representative_topics":["zero-shot multimodal style transfer for gesture animation via adversarial disentanglement of speaker style embeddings","large-scale crowdsourced benchmarking of co-speech gesture generation systems","weakly supervised sign language notation animation into pose sequences via transformer encoders"]},{"direction_index":4,"label":"D4","name":"Embodied environment interaction","description":"Generate motion that is physically or spatially grounded in external environments, objects, contacts, affordances, navigation constraints, or simulated dynamics.","inclusion":"Assign atoms whose primary objective is coherent embodied interaction with scenes, objects, or physical dynamics, including human-object manipulation, task-oriented affordance use, scene-aware locomotion, collision-aware navigation, environment-conditioned motion, physics simulation, reinforcement-learning control, humanoid imitation, and dynamics-feasible animation.","exclusion":"If the central coupling is with another person rather than the environment, assign to multi-human interaction; if physical losses only support a generic kinematic generator, assign to general motion priors; if semantic text alone is the main condition without scene or dynamics grounding, assign to semantic motion grounding.","nearest_boundary":"Multi-human interaction","representative_topics":["language-conditioned 3d human motion generation with scene and object interaction","intent-driven full-body motion synthesis for virtual human-object interactions","physics-simulated motion in-betweening via skeleton optimization and humanoid robot imitation"]},{"direction_index":5,"label":"D5","name":"Temporal forecasting and completion","description":"Predict, continue, interpolate, or fill human motion from observed past poses, keyframes, partial temporal constraints, or incomplete motion sequences.","inclusion":"Assign atoms whose primary objective is temporal extrapolation or completion from motion observations, including stochastic future prediction, multimodal forecasting, long-term continuation, pose interpolation, gap filling, and temporal consistency from observed history.","exclusion":"If the output is generated de novo from text, music, speech, scenes, objects, social prompts, or physical policies rather than observed motion history, assign to that conditioning objective; physics-simulated in-betweening belongs to embodied environment interaction.","nearest_boundary":"Motion editing and transfer","representative_topics":["diverse multimodal human motion prediction via auxiliary space gumbel-softmax sampling","privileged knowledge distillation for human motion prediction via pose-based interpolation","diverse multi-motion 3d human joint prediction using attention-based generative adversarial networks"]},{"direction_index":6,"label":"D6","name":"Motion editing and transfer","description":"Transform, retarget, stylize, complete, steer, or visually realize existing or partially specified human motion under user constraints, source-target mappings, sparse signals, or video-space requirements.","inclusion":"Assign atoms whose primary objective is controllable transformation or realization of motion content, including skeleton or person motion transfer, retargeting, style transfer, latent editing, sparse-tracking full-body completion, directable characters, motion-to-video transfer, human dance video synthesis, and video/2D-consistency-driven motion generation.","exclusion":"If the primary objective is de novo semantic, music, speech, object, scene, social, or physics-conditioned synthesis, assign to that direction; if the core contribution is a general reusable prior rather than editing or transfer, assign to general motion priors.","nearest_boundary":"General motion priors","representative_topics":["implicit neural motion fields as continuous generative priors for motion interpolation and editing","unsupervised unconditional motion synthesis with diverse motion manifold and semantic editing","progressive region-to-whole gan framework for realistic human video motion transfer"]},{"direction_index":7,"label":"D7","name":"Multi-human interaction","description":"Generate coordinated, reactive, or socially coupled motion for two or more humans, emphasizing partner dependence, mutual response, interaction semantics, and multi-agent coherence.","inclusion":"Assign atoms whose primary objective is modeling inter-person motion dependencies, including dyadic reaction synthesis, partner-conditioned generation, two-person interaction diffusion, text-guided interpersonal motion, and highly interactive multi-person prediction.","exclusion":"Group dance belongs to music choreography synthesis when music or choreography is the primary condition; forecasting belongs to temporal forecasting when extrapolating from history is the main objective; object or scene coupling belongs to embodied environment interaction.","nearest_boundary":"Temporal forecasting and completion","representative_topics":["interaction transformer for reaction motion generation using temporal-spatial graph attention","partner-conditioned neural operator generation of dyadic human motion in function spaces","bipartite graph diffusion models for two-person human motion interaction generation"]}],"split":"dev"}
{"episode_id":"F150_T2025-07","field_index":150,"field":"human motion generation","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Semantic motion grounding","description":"Generate, retrieve, edit, or evaluate 3D human motion from natural-language descriptions, action labels, open-vocabulary prompts, or compositional semantic specifications.","inclusion":"Assign atoms whose primary objective is grounding semantic intent in plausible human motion, including text-to-motion, action-label-to-motion, reciprocal text-motion modeling, open-vocabulary generalization, temporal action composition, and semantic alignment metrics.","exclusion":"If the central objective is learning a reusable generative backbone independent of semantic grounding, assign to general motion priors; if scenes, objects, music, speech, other humans, or physical control are the main conditioning problem, assign to that direction.","nearest_boundary":"General motion priors","representative_topics":["tokenized autoregressive modeling for reciprocal text and 3d motion generation","diffusion-based text-conditioned motion synthesis and joint-wise editing without fine-tuning","hierarchical autoregressive temporal action composition for text-conditioned 3d human motion synthesis"]},{"direction_index":1,"label":"D1","name":"General motion priors","description":"Learn reusable human-motion distributions, representations, and generative mechanisms for synthesizing, sampling, completing, or weakly conditioning motion across tasks.","inclusion":"Assign atoms whose primary objective is a broadly reusable generative prior or backbone, including discrete motion tokens, language-model-style sequence generation, VAEs, diffusion models, latent diffusion, neural fields, masked modeling, and generic multi-condition architectures.","exclusion":"If the primary research objective is a specific external grounding signal such as text, music, speech, objects, scenes, other humans, future extrapolation, editing, or simulation control, assign to that more specific direction.","nearest_boundary":"Semantic motion grounding","representative_topics":["implicit neural motion fields as continuous generative priors for motion interpolation and editing","recurrent transformer variational autoencoders for multi-action arbitrary-length motion sequence synthesis","unsupervised unconditional motion synthesis with diverse motion manifold and semantic editing"]},{"direction_index":2,"label":"D2","name":"Music choreography synthesis","description":"Generate, regenerate, transfer, or evaluate dance and choreographic motion from music, rhythm, choreography labels, key poses, or dance-specific creative supervision.","inclusion":"Assign atoms whose primary objective is dance or choreography, including audio-conditioned dance, beat synchronization, keyframe-constrained dance, artist-centric choreography, dance style transfer, group choreography, and dance-specific datasets.","exclusion":"Speech-driven communicative gesture belongs to communicative gesture synthesis; generic audio-conditioned motion belongs here only when the output objective is explicitly dance or choreography.","nearest_boundary":"Communicative gesture synthesis","representative_topics":["music-driven dance motion synthesis with controllable key pose constraints via cross-modal transformer","keyframe-based dance motion synthesis using hierarchical rnn and transformer constraint controllers","audio-conditioned dance motion synthesis for complex acrobatic poses via hybrid keypoint annotation"]},{"direction_index":3,"label":"D3","name":"Communicative gesture synthesis","description":"Generate co-speech gestures, sign poses, upper-body motion, hands, face, or holistic speaking-avatar motion from speech, transcripts, notation, or multimodal communicative signals.","inclusion":"Assign atoms whose primary objective is communicative body motion aligned with speech, text, sign notation, or speaker style, including speech-to-gesture, sign-language pose animation, holistic speech-driven avatars, style control, and gesture-generation benchmarks.","exclusion":"Music-driven dance belongs to music choreography synthesis; generic full-body text-to-motion without communicative speech, sign, or avatar alignment belongs to semantic motion grounding.","nearest_boundary":"Music choreography synthesis","representative_topics":["zero-shot multimodal style transfer for gesture animation via adversarial disentanglement of speaker style embeddings","large-scale crowdsourced benchmarking of co-speech gesture generation systems","weakly supervised sign language notation animation into pose sequences via transformer encoders"]},{"direction_index":4,"label":"D4","name":"Embodied environment interaction","description":"Generate motion that is physically or spatially grounded in external environments, objects, contacts, affordances, navigation constraints, or simulated dynamics.","inclusion":"Assign atoms whose primary objective is coherent embodied interaction with scenes, objects, or physical dynamics, including human-object manipulation, task-oriented affordance use, scene-aware locomotion, collision-aware navigation, environment-conditioned motion, physics simulation, reinforcement-learning control, humanoid imitation, and dynamics-feasible animation.","exclusion":"If the central coupling is with another person rather than the environment, assign to multi-human interaction; if physical losses only support a generic kinematic generator, assign to general motion priors; if semantic text alone is the main condition without scene or dynamics grounding, assign to semantic motion grounding.","nearest_boundary":"Multi-human interaction","representative_topics":["language-conditioned 3d human motion generation with scene and object interaction","intent-driven full-body motion synthesis for virtual human-object interactions","physics-simulated motion in-betweening via skeleton optimization and humanoid robot imitation"]},{"direction_index":5,"label":"D5","name":"Temporal forecasting and completion","description":"Predict, continue, interpolate, or fill human motion from observed past poses, keyframes, partial temporal constraints, or incomplete motion sequences.","inclusion":"Assign atoms whose primary objective is temporal extrapolation or completion from motion observations, including stochastic future prediction, multimodal forecasting, long-term continuation, pose interpolation, gap filling, and temporal consistency from observed history.","exclusion":"If the output is generated de novo from text, music, speech, scenes, objects, social prompts, or physical policies rather than observed motion history, assign to that conditioning objective; physics-simulated in-betweening belongs to embodied environment interaction.","nearest_boundary":"Motion editing and transfer","representative_topics":["diverse multimodal human motion prediction via auxiliary space gumbel-softmax sampling","privileged knowledge distillation for human motion prediction via pose-based interpolation","diverse multi-motion 3d human joint prediction using attention-based generative adversarial networks"]},{"direction_index":6,"label":"D6","name":"Motion editing and transfer","description":"Transform, retarget, stylize, complete, steer, or visually realize existing or partially specified human motion under user constraints, source-target mappings, sparse signals, or video-space requirements.","inclusion":"Assign atoms whose primary objective is controllable transformation or realization of motion content, including skeleton or person motion transfer, retargeting, style transfer, latent editing, sparse-tracking full-body completion, directable characters, motion-to-video transfer, human dance video synthesis, and video/2D-consistency-driven motion generation.","exclusion":"If the primary objective is de novo semantic, music, speech, object, scene, social, or physics-conditioned synthesis, assign to that direction; if the core contribution is a general reusable prior rather than editing or transfer, assign to general motion priors.","nearest_boundary":"General motion priors","representative_topics":["implicit neural motion fields as continuous generative priors for motion interpolation and editing","unsupervised unconditional motion synthesis with diverse motion manifold and semantic editing","progressive region-to-whole gan framework for realistic human video motion transfer"]},{"direction_index":7,"label":"D7","name":"Multi-human interaction","description":"Generate coordinated, reactive, or socially coupled motion for two or more humans, emphasizing partner dependence, mutual response, interaction semantics, and multi-agent coherence.","inclusion":"Assign atoms whose primary objective is modeling inter-person motion dependencies, including dyadic reaction synthesis, partner-conditioned generation, two-person interaction diffusion, text-guided interpersonal motion, and highly interactive multi-person prediction.","exclusion":"Group dance belongs to music choreography synthesis when music or choreography is the primary condition; forecasting belongs to temporal forecasting when extrapolating from history is the main objective; object or scene coupling belongs to embodied environment interaction.","nearest_boundary":"Temporal forecasting and completion","representative_topics":["interaction transformer for reaction motion generation using temporal-spatial graph attention","partner-conditioned neural operator generation of dyadic human motion in function spaces","bipartite graph diffusion models for two-person human motion interaction generation"]}],"split":"dev"}
{"episode_id":"F150_T2026-01","field_index":150,"field":"human motion generation","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Semantic motion grounding","description":"Generate, retrieve, edit, or evaluate 3D human motion from natural-language descriptions, action labels, open-vocabulary prompts, or compositional semantic specifications.","inclusion":"Assign atoms whose primary objective is grounding semantic intent in plausible human motion, including text-to-motion, action-label-to-motion, reciprocal text-motion modeling, open-vocabulary generalization, temporal action composition, and semantic alignment metrics.","exclusion":"If the central objective is learning a reusable generative backbone independent of semantic grounding, assign to general motion priors; if scenes, objects, music, speech, other humans, or physical control are the main conditioning problem, assign to that direction.","nearest_boundary":"General motion priors","representative_topics":["tokenized autoregressive modeling for reciprocal text and 3d motion generation","diffusion-based text-conditioned motion synthesis and joint-wise editing without fine-tuning","hierarchical autoregressive temporal action composition for text-conditioned 3d human motion synthesis"]},{"direction_index":1,"label":"D1","name":"General motion priors","description":"Learn reusable human-motion distributions, representations, and generative mechanisms for synthesizing, sampling, completing, or weakly conditioning motion across tasks.","inclusion":"Assign atoms whose primary objective is a broadly reusable generative prior or backbone, including discrete motion tokens, language-model-style sequence generation, VAEs, diffusion models, latent diffusion, neural fields, masked modeling, and generic multi-condition architectures.","exclusion":"If the primary research objective is a specific external grounding signal such as text, music, speech, objects, scenes, other humans, future extrapolation, editing, or simulation control, assign to that more specific direction.","nearest_boundary":"Semantic motion grounding","representative_topics":["implicit neural motion fields as continuous generative priors for motion interpolation and editing","recurrent transformer variational autoencoders for multi-action arbitrary-length motion sequence synthesis","unsupervised unconditional motion synthesis with diverse motion manifold and semantic editing"]},{"direction_index":2,"label":"D2","name":"Music choreography synthesis","description":"Generate, regenerate, transfer, or evaluate dance and choreographic motion from music, rhythm, choreography labels, key poses, or dance-specific creative supervision.","inclusion":"Assign atoms whose primary objective is dance or choreography, including audio-conditioned dance, beat synchronization, keyframe-constrained dance, artist-centric choreography, dance style transfer, group choreography, and dance-specific datasets.","exclusion":"Speech-driven communicative gesture belongs to communicative gesture synthesis; generic audio-conditioned motion belongs here only when the output objective is explicitly dance or choreography.","nearest_boundary":"Communicative gesture synthesis","representative_topics":["music-driven dance motion synthesis with controllable key pose constraints via cross-modal transformer","keyframe-based dance motion synthesis using hierarchical rnn and transformer constraint controllers","audio-conditioned dance motion synthesis for complex acrobatic poses via hybrid keypoint annotation"]},{"direction_index":3,"label":"D3","name":"Communicative gesture synthesis","description":"Generate co-speech gestures, sign poses, upper-body motion, hands, face, or holistic speaking-avatar motion from speech, transcripts, notation, or multimodal communicative signals.","inclusion":"Assign atoms whose primary objective is communicative body motion aligned with speech, text, sign notation, or speaker style, including speech-to-gesture, sign-language pose animation, holistic speech-driven avatars, style control, and gesture-generation benchmarks.","exclusion":"Music-driven dance belongs to music choreography synthesis; generic full-body text-to-motion without communicative speech, sign, or avatar alignment belongs to semantic motion grounding.","nearest_boundary":"Music choreography synthesis","representative_topics":["zero-shot multimodal style transfer for gesture animation via adversarial disentanglement of speaker style embeddings","large-scale crowdsourced benchmarking of co-speech gesture generation systems","weakly supervised sign language notation animation into pose sequences via transformer encoders"]},{"direction_index":4,"label":"D4","name":"Embodied environment interaction","description":"Generate motion that is physically or spatially grounded in external environments, objects, contacts, affordances, navigation constraints, or simulated dynamics.","inclusion":"Assign atoms whose primary objective is coherent embodied interaction with scenes, objects, or physical dynamics, including human-object manipulation, task-oriented affordance use, scene-aware locomotion, collision-aware navigation, environment-conditioned motion, physics simulation, reinforcement-learning control, humanoid imitation, and dynamics-feasible animation.","exclusion":"If the central coupling is with another person rather than the environment, assign to multi-human interaction; if physical losses only support a generic kinematic generator, assign to general motion priors; if semantic text alone is the main condition without scene or dynamics grounding, assign to semantic motion grounding.","nearest_boundary":"Multi-human interaction","representative_topics":["language-conditioned 3d human motion generation with scene and object interaction","intent-driven full-body motion synthesis for virtual human-object interactions","physics-simulated motion in-betweening via skeleton optimization and humanoid robot imitation"]},{"direction_index":5,"label":"D5","name":"Temporal forecasting and completion","description":"Predict, continue, interpolate, or fill human motion from observed past poses, keyframes, partial temporal constraints, or incomplete motion sequences.","inclusion":"Assign atoms whose primary objective is temporal extrapolation or completion from motion observations, including stochastic future prediction, multimodal forecasting, long-term continuation, pose interpolation, gap filling, and temporal consistency from observed history.","exclusion":"If the output is generated de novo from text, music, speech, scenes, objects, social prompts, or physical policies rather than observed motion history, assign to that conditioning objective; physics-simulated in-betweening belongs to embodied environment interaction.","nearest_boundary":"Motion editing and transfer","representative_topics":["diverse multimodal human motion prediction via auxiliary space gumbel-softmax sampling","privileged knowledge distillation for human motion prediction via pose-based interpolation","diverse multi-motion 3d human joint prediction using attention-based generative adversarial networks"]},{"direction_index":6,"label":"D6","name":"Motion editing and transfer","description":"Transform, retarget, stylize, complete, steer, or visually realize existing or partially specified human motion under user constraints, source-target mappings, sparse signals, or video-space requirements.","inclusion":"Assign atoms whose primary objective is controllable transformation or realization of motion content, including skeleton or person motion transfer, retargeting, style transfer, latent editing, sparse-tracking full-body completion, directable characters, motion-to-video transfer, human dance video synthesis, and video/2D-consistency-driven motion generation.","exclusion":"If the primary objective is de novo semantic, music, speech, object, scene, social, or physics-conditioned synthesis, assign to that direction; if the core contribution is a general reusable prior rather than editing or transfer, assign to general motion priors.","nearest_boundary":"General motion priors","representative_topics":["implicit neural motion fields as continuous generative priors for motion interpolation and editing","unsupervised unconditional motion synthesis with diverse motion manifold and semantic editing","progressive region-to-whole gan framework for realistic human video motion transfer"]},{"direction_index":7,"label":"D7","name":"Multi-human interaction","description":"Generate coordinated, reactive, or socially coupled motion for two or more humans, emphasizing partner dependence, mutual response, interaction semantics, and multi-agent coherence.","inclusion":"Assign atoms whose primary objective is modeling inter-person motion dependencies, including dyadic reaction synthesis, partner-conditioned generation, two-person interaction diffusion, text-guided interpersonal motion, and highly interactive multi-person prediction.","exclusion":"Group dance belongs to music choreography synthesis when music or choreography is the primary condition; forecasting belongs to temporal forecasting when extrapolating from history is the main objective; object or scene coupling belongs to embodied environment interaction.","nearest_boundary":"Temporal forecasting and completion","representative_topics":["interaction transformer for reaction motion generation using temporal-spatial graph attention","partner-conditioned neural operator generation of dyadic human motion in function spaces","bipartite graph diffusion models for two-person human motion interaction generation"]}],"split":"dev"}
{"episode_id":"F156_T2024-01","field_index":156,"field":"misinformation detection","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Textual content detection","description":"Detect misinformation primarily from linguistic content, document representations, stylistic cues, narratives, sentiment, or text-classification architectures, including domain-specific textual settings such as health misinformation when the research objective remains content classification.","inclusion":"Assign work whose main objective is veracity, fake-news, disinformation, deception, propaganda, misleading-narrative, false-review, or health-misinformation classification from intrinsic text or content signals without central reliance on external evidence, social propagation, multimodal consistency, language-resource scarcity, or LLM-era generation questions.","exclusion":"Use evidence-grounded verification when explicit claim-evidence retrieval or reasoning is central; use social propagation modeling when diffusion, users, or conversation structure are central; use generalization across domains and languages when transfer or low-resource language coverage is the main objective; use LLM-mediated detection when LLM generation or LLM capability is the object.","nearest_boundary":"Evidence-grounded verification","representative_topics":["fake news classification under class imbalance using pre-trained encoders and negation-based data augmentation","multi-policy deep learning framework for joint statement and article fake news detection","hybrid ensemble of classical and deep learning models for fake news detection"]},{"direction_index":1,"label":"D1","name":"Social propagation modeling","description":"Detect misinformation by modeling how information spreads through users, reports, conversations, temporal cascades, propagation trees, social graphs, heterogeneous networks, or interaction contexts.","inclusion":"Assign work whose primary detection signal is social context, user behavior, feedback reports, graph structure, temporal diffusion, thread structure, or propagation dynamics, including GNN, hypergraph, heterogeneous-graph, streaming-rumor, and online propagation models.","exclusion":"Use textual content detection when social context is absent or only auxiliary; use evidence-grounded verification when graphs represent evidence or knowledge rather than social spread; use multimodal semantic verification when image/video-text reasoning is central.","nearest_boundary":"Textual content detection","representative_topics":["diverse example-based explanations for streaming rumour graph detection","graph neural network fake news detection via social propagation context and transformer content features","user-report-driven classification of reported misinformation content across platforms and regions"]},{"direction_index":2,"label":"D2","name":"Generalization across domains and languages","description":"Improve misinformation detection or verification under distribution shift, temporal drift, emerging topics, scarce labels, dataset bias, cross-domain transfer, multilingual settings, or low-resource languages.","inclusion":"Assign work whose primary objective is robustness or transfer across domains, time periods, topics, modalities, or languages, including domain adaptation, adversarial or contrastive feature alignment, weak/self-supervised learning, temporal knowledgebase construction, language-specific dataset creation, multilingual modeling, cross-lingual transfer, and low-resource language adaptation.","exclusion":"Use textual content detection for single-domain or single-language classification without a transfer, drift, or resource-scarcity objective; use multimodal semantic verification when cross-modal reasoning is the main contribution; use evidence-grounded verification when claim-evidence reasoning dominates and language or domain transfer is secondary.","nearest_boundary":"Textual content detection","representative_topics":["multi-domain fake news detection via memory-guided multi-view domain adaptation","low-resource urdu fake news detection using transformer and traditional feature representations","low-resource urdu fake news detection using transformer-based neural classification"]},{"direction_index":3,"label":"D3","name":"Evidence-grounded verification","description":"Verify claims by retrieving, selecting, reasoning over, distilling, or explaining textual or structured evidence, including interpretable claim-evidence reasoning and evidence-aware fake-news detection.","inclusion":"Assign work whose main objective is grounding veracity judgments in explicit evidence or producing human-interpretable reasoning for verification, including evidence retrieval, evidence selection, multi-hop reasoning, rationale extraction, justification generation, table verification, evidence distillation, claim-evidence entailment, and evidence-aware explainable detection.","exclusion":"Use textual content detection when no explicit evidence comparison is central; use multimodal semantic verification when visual or cross-modal evidence is central; use fact-checking workflow and intervention when the goal is selecting claims to check or producing corrective responses rather than verifying a claim against evidence.","nearest_boundary":"Fact-checking workflow and intervention","representative_topics":["chinese evidence-based fact-checking with joint evidence retrieval and veracity prediction","multi-modal explainable ai for interpretable fake news detection on social media","interpretable fake news classification using bert embeddings and word-token activation mapping"]},{"direction_index":4,"label":"D4","name":"Multimodal semantic verification","description":"Detect or verify misinformation by jointly modeling text with images, video, captions, OCR, charts, or other non-textual signals to assess cross-modal consistency or multimodal veracity.","inclusion":"Assign work whose primary objective is multimodal veracity assessment, image-text consistency, cross-modal entailment, multimodal fusion, visual-textual grounding, multimodal benchmark construction, cross-modal manipulation localization, or interpretable multimodal reasoning for misinformation.","exclusion":"Use synthetic media forensics when the main question is whether visual media are manipulated or generated at the signal/artifact level; use generalization across domains and languages when multimodal data are mainly used to study domain shift; use evidence-grounded verification for primarily textual or structured evidence reasoning.","nearest_boundary":"Synthetic media forensics","representative_topics":["bootstrapping multi-view representations for cross-modal consistency-aware fake news detection","multimodal fusion and ensemble learning for social media rumor veracity classification","parameter-efficient multimodal fact verification with cross-modal multi-type fusion and ensemble methods"]},{"direction_index":5,"label":"D5","name":"Synthetic media forensics","description":"Detect manipulated, generated, or inauthentic visual media by identifying forensic traces, generation artifacts, signal inconsistencies, physiological anomalies, facial dynamics, frequency-domain cues, or robustness failures in images, faces, or videos.","inclusion":"Assign work focused on authenticity detection for deepfakes, synthetic images, face forgeries, manipulated images, or AI-generated visual media using artifact analysis, forensic networks, physiological signals, landmark dynamics, degradation robustness, generator generalization, or visual-forensic prompts.","exclusion":"Use multimodal semantic verification when the central task is semantic truthfulness or image-text claim consistency; use LLM-mediated detection when the central artifact is generated text or LLM reasoning; use textual content detection for non-visual misinformation classification.","nearest_boundary":"Multimodal semantic verification","representative_topics":["human-centered deepfake detection using artifact attention and visual caricature amplification","robustness analysis of deepfake detectors under quality-degrading in-the-wild corruptions","multimodal fake video detection via physiological signal visual representations and graph fusion"]},{"direction_index":6,"label":"D6","name":"LLM-mediated detection","description":"Study misinformation detection problems whose primary object is LLM-generated content, LLM detector behavior, or LLM-based reasoning, prompting, and labeling for misinformation tasks.","inclusion":"Assign work where large language models are either the object of detection or the central methodological tool, including detecting machine-generated fake reviews or news, evaluating LLMs as misinformation detectors, prompt-engineered LLM classification, LLM-based false-statement judgment, detector bias in LLM-era settings, and LLM-assisted conspiracy or narrative identification.","exclusion":"Use textual content detection for ordinary BERT/RoBERTa-style text classifiers where LLM generation or LLM capability is not central; use multimodal semantic verification for image-text veracity unless LLM prompting/capability is the primary question; use synthetic media forensics for visual deepfake detection.","nearest_boundary":"Textual content detection","representative_topics":["fine-tuned language model detection of machine-generated fake restaurant reviews","llm-based prompt engineering for out-of-context image-caption coherence detection","fine-tuned transformer models for ai-generated fake news detection"]},{"direction_index":7,"label":"D7","name":"Fact-checking workflow and intervention","description":"Support the upstream and downstream human-facing fact-checking pipeline by identifying check-worthy claims, locating claim spans, retrieving prior fact checks, rewriting discovery queries, generating or recommending rebuttals, and evaluating corrective interventions.","inclusion":"Assign work whose primary objective is deciding what should be checked, extracting or reformulating candidate claims for fact-checkers, discovering known misinformation, comparing fact-checking workflows, producing debunks or rebuttals, controlled claim editing, metacognitive intervention agents, or measuring whether explanations and interventions reduce misinformation belief.","exclusion":"Use evidence-grounded verification when an already-selected claim is judged against retrieved evidence; use textual content detection or social propagation modeling when the goal is simply assigning veracity; use LLM-mediated detection when LLM capability rather than intervention or workflow outcome is primary.","nearest_boundary":"Evidence-grounded verification","representative_topics":["multilingual check-worthiness classification of tweets using transformer models","crowdsourced versus expert fact-checking comparison on social media misinformation","abstractive summarization-based query generation for automated fact-checked claim detection"]}],"split":"dev"}
{"episode_id":"F156_T2024-07","field_index":156,"field":"misinformation detection","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Textual content detection","description":"Detect misinformation primarily from linguistic content, document representations, stylistic cues, narratives, sentiment, or text-classification architectures, including domain-specific textual settings such as health misinformation when the research objective remains content classification.","inclusion":"Assign work whose main objective is veracity, fake-news, disinformation, deception, propaganda, misleading-narrative, false-review, or health-misinformation classification from intrinsic text or content signals without central reliance on external evidence, social propagation, multimodal consistency, language-resource scarcity, or LLM-era generation questions.","exclusion":"Use evidence-grounded verification when explicit claim-evidence retrieval or reasoning is central; use social propagation modeling when diffusion, users, or conversation structure are central; use generalization across domains and languages when transfer or low-resource language coverage is the main objective; use LLM-mediated detection when LLM generation or LLM capability is the object.","nearest_boundary":"Evidence-grounded verification","representative_topics":["fake news classification under class imbalance using pre-trained encoders and negation-based data augmentation","multi-policy deep learning framework for joint statement and article fake news detection","hybrid ensemble of classical and deep learning models for fake news detection"]},{"direction_index":1,"label":"D1","name":"Social propagation modeling","description":"Detect misinformation by modeling how information spreads through users, reports, conversations, temporal cascades, propagation trees, social graphs, heterogeneous networks, or interaction contexts.","inclusion":"Assign work whose primary detection signal is social context, user behavior, feedback reports, graph structure, temporal diffusion, thread structure, or propagation dynamics, including GNN, hypergraph, heterogeneous-graph, streaming-rumor, and online propagation models.","exclusion":"Use textual content detection when social context is absent or only auxiliary; use evidence-grounded verification when graphs represent evidence or knowledge rather than social spread; use multimodal semantic verification when image/video-text reasoning is central.","nearest_boundary":"Textual content detection","representative_topics":["diverse example-based explanations for streaming rumour graph detection","graph neural network fake news detection via social propagation context and transformer content features","user-report-driven classification of reported misinformation content across platforms and regions"]},{"direction_index":2,"label":"D2","name":"Generalization across domains and languages","description":"Improve misinformation detection or verification under distribution shift, temporal drift, emerging topics, scarce labels, dataset bias, cross-domain transfer, multilingual settings, or low-resource languages.","inclusion":"Assign work whose primary objective is robustness or transfer across domains, time periods, topics, modalities, or languages, including domain adaptation, adversarial or contrastive feature alignment, weak/self-supervised learning, temporal knowledgebase construction, language-specific dataset creation, multilingual modeling, cross-lingual transfer, and low-resource language adaptation.","exclusion":"Use textual content detection for single-domain or single-language classification without a transfer, drift, or resource-scarcity objective; use multimodal semantic verification when cross-modal reasoning is the main contribution; use evidence-grounded verification when claim-evidence reasoning dominates and language or domain transfer is secondary.","nearest_boundary":"Textual content detection","representative_topics":["multi-domain fake news detection via memory-guided multi-view domain adaptation","low-resource urdu fake news detection using transformer and traditional feature representations","low-resource urdu fake news detection using transformer-based neural classification"]},{"direction_index":3,"label":"D3","name":"Evidence-grounded verification","description":"Verify claims by retrieving, selecting, reasoning over, distilling, or explaining textual or structured evidence, including interpretable claim-evidence reasoning and evidence-aware fake-news detection.","inclusion":"Assign work whose main objective is grounding veracity judgments in explicit evidence or producing human-interpretable reasoning for verification, including evidence retrieval, evidence selection, multi-hop reasoning, rationale extraction, justification generation, table verification, evidence distillation, claim-evidence entailment, and evidence-aware explainable detection.","exclusion":"Use textual content detection when no explicit evidence comparison is central; use multimodal semantic verification when visual or cross-modal evidence is central; use fact-checking workflow and intervention when the goal is selecting claims to check or producing corrective responses rather than verifying a claim against evidence.","nearest_boundary":"Fact-checking workflow and intervention","representative_topics":["chinese evidence-based fact-checking with joint evidence retrieval and veracity prediction","multi-modal explainable ai for interpretable fake news detection on social media","interpretable fake news classification using bert embeddings and word-token activation mapping"]},{"direction_index":4,"label":"D4","name":"Multimodal semantic verification","description":"Detect or verify misinformation by jointly modeling text with images, video, captions, OCR, charts, or other non-textual signals to assess cross-modal consistency or multimodal veracity.","inclusion":"Assign work whose primary objective is multimodal veracity assessment, image-text consistency, cross-modal entailment, multimodal fusion, visual-textual grounding, multimodal benchmark construction, cross-modal manipulation localization, or interpretable multimodal reasoning for misinformation.","exclusion":"Use synthetic media forensics when the main question is whether visual media are manipulated or generated at the signal/artifact level; use generalization across domains and languages when multimodal data are mainly used to study domain shift; use evidence-grounded verification for primarily textual or structured evidence reasoning.","nearest_boundary":"Synthetic media forensics","representative_topics":["bootstrapping multi-view representations for cross-modal consistency-aware fake news detection","multimodal fusion and ensemble learning for social media rumor veracity classification","parameter-efficient multimodal fact verification with cross-modal multi-type fusion and ensemble methods"]},{"direction_index":5,"label":"D5","name":"Synthetic media forensics","description":"Detect manipulated, generated, or inauthentic visual media by identifying forensic traces, generation artifacts, signal inconsistencies, physiological anomalies, facial dynamics, frequency-domain cues, or robustness failures in images, faces, or videos.","inclusion":"Assign work focused on authenticity detection for deepfakes, synthetic images, face forgeries, manipulated images, or AI-generated visual media using artifact analysis, forensic networks, physiological signals, landmark dynamics, degradation robustness, generator generalization, or visual-forensic prompts.","exclusion":"Use multimodal semantic verification when the central task is semantic truthfulness or image-text claim consistency; use LLM-mediated detection when the central artifact is generated text or LLM reasoning; use textual content detection for non-visual misinformation classification.","nearest_boundary":"Multimodal semantic verification","representative_topics":["human-centered deepfake detection using artifact attention and visual caricature amplification","robustness analysis of deepfake detectors under quality-degrading in-the-wild corruptions","multimodal fake video detection via physiological signal visual representations and graph fusion"]},{"direction_index":6,"label":"D6","name":"LLM-mediated detection","description":"Study misinformation detection problems whose primary object is LLM-generated content, LLM detector behavior, or LLM-based reasoning, prompting, and labeling for misinformation tasks.","inclusion":"Assign work where large language models are either the object of detection or the central methodological tool, including detecting machine-generated fake reviews or news, evaluating LLMs as misinformation detectors, prompt-engineered LLM classification, LLM-based false-statement judgment, detector bias in LLM-era settings, and LLM-assisted conspiracy or narrative identification.","exclusion":"Use textual content detection for ordinary BERT/RoBERTa-style text classifiers where LLM generation or LLM capability is not central; use multimodal semantic verification for image-text veracity unless LLM prompting/capability is the primary question; use synthetic media forensics for visual deepfake detection.","nearest_boundary":"Textual content detection","representative_topics":["fine-tuned language model detection of machine-generated fake restaurant reviews","llm-based prompt engineering for out-of-context image-caption coherence detection","fine-tuned transformer models for ai-generated fake news detection"]},{"direction_index":7,"label":"D7","name":"Fact-checking workflow and intervention","description":"Support the upstream and downstream human-facing fact-checking pipeline by identifying check-worthy claims, locating claim spans, retrieving prior fact checks, rewriting discovery queries, generating or recommending rebuttals, and evaluating corrective interventions.","inclusion":"Assign work whose primary objective is deciding what should be checked, extracting or reformulating candidate claims for fact-checkers, discovering known misinformation, comparing fact-checking workflows, producing debunks or rebuttals, controlled claim editing, metacognitive intervention agents, or measuring whether explanations and interventions reduce misinformation belief.","exclusion":"Use evidence-grounded verification when an already-selected claim is judged against retrieved evidence; use textual content detection or social propagation modeling when the goal is simply assigning veracity; use LLM-mediated detection when LLM capability rather than intervention or workflow outcome is primary.","nearest_boundary":"Evidence-grounded verification","representative_topics":["multilingual check-worthiness classification of tweets using transformer models","crowdsourced versus expert fact-checking comparison on social media misinformation","abstractive summarization-based query generation for automated fact-checked claim detection"]}],"split":"dev"}
{"episode_id":"F156_T2025-01","field_index":156,"field":"misinformation detection","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Textual content detection","description":"Detect misinformation primarily from linguistic content, document representations, stylistic cues, narratives, sentiment, or text-classification architectures, including domain-specific textual settings such as health misinformation when the research objective remains content classification.","inclusion":"Assign work whose main objective is veracity, fake-news, disinformation, deception, propaganda, misleading-narrative, false-review, or health-misinformation classification from intrinsic text or content signals without central reliance on external evidence, social propagation, multimodal consistency, language-resource scarcity, or LLM-era generation questions.","exclusion":"Use evidence-grounded verification when explicit claim-evidence retrieval or reasoning is central; use social propagation modeling when diffusion, users, or conversation structure are central; use generalization across domains and languages when transfer or low-resource language coverage is the main objective; use LLM-mediated detection when LLM generation or LLM capability is the object.","nearest_boundary":"Evidence-grounded verification","representative_topics":["fake news classification under class imbalance using pre-trained encoders and negation-based data augmentation","multi-policy deep learning framework for joint statement and article fake news detection","hybrid ensemble of classical and deep learning models for fake news detection"]},{"direction_index":1,"label":"D1","name":"Social propagation modeling","description":"Detect misinformation by modeling how information spreads through users, reports, conversations, temporal cascades, propagation trees, social graphs, heterogeneous networks, or interaction contexts.","inclusion":"Assign work whose primary detection signal is social context, user behavior, feedback reports, graph structure, temporal diffusion, thread structure, or propagation dynamics, including GNN, hypergraph, heterogeneous-graph, streaming-rumor, and online propagation models.","exclusion":"Use textual content detection when social context is absent or only auxiliary; use evidence-grounded verification when graphs represent evidence or knowledge rather than social spread; use multimodal semantic verification when image/video-text reasoning is central.","nearest_boundary":"Textual content detection","representative_topics":["diverse example-based explanations for streaming rumour graph detection","graph neural network fake news detection via social propagation context and transformer content features","user-report-driven classification of reported misinformation content across platforms and regions"]},{"direction_index":2,"label":"D2","name":"Generalization across domains and languages","description":"Improve misinformation detection or verification under distribution shift, temporal drift, emerging topics, scarce labels, dataset bias, cross-domain transfer, multilingual settings, or low-resource languages.","inclusion":"Assign work whose primary objective is robustness or transfer across domains, time periods, topics, modalities, or languages, including domain adaptation, adversarial or contrastive feature alignment, weak/self-supervised learning, temporal knowledgebase construction, language-specific dataset creation, multilingual modeling, cross-lingual transfer, and low-resource language adaptation.","exclusion":"Use textual content detection for single-domain or single-language classification without a transfer, drift, or resource-scarcity objective; use multimodal semantic verification when cross-modal reasoning is the main contribution; use evidence-grounded verification when claim-evidence reasoning dominates and language or domain transfer is secondary.","nearest_boundary":"Textual content detection","representative_topics":["multi-domain fake news detection via memory-guided multi-view domain adaptation","low-resource urdu fake news detection using transformer and traditional feature representations","low-resource urdu fake news detection using transformer-based neural classification"]},{"direction_index":3,"label":"D3","name":"Evidence-grounded verification","description":"Verify claims by retrieving, selecting, reasoning over, distilling, or explaining textual or structured evidence, including interpretable claim-evidence reasoning and evidence-aware fake-news detection.","inclusion":"Assign work whose main objective is grounding veracity judgments in explicit evidence or producing human-interpretable reasoning for verification, including evidence retrieval, evidence selection, multi-hop reasoning, rationale extraction, justification generation, table verification, evidence distillation, claim-evidence entailment, and evidence-aware explainable detection.","exclusion":"Use textual content detection when no explicit evidence comparison is central; use multimodal semantic verification when visual or cross-modal evidence is central; use fact-checking workflow and intervention when the goal is selecting claims to check or producing corrective responses rather than verifying a claim against evidence.","nearest_boundary":"Fact-checking workflow and intervention","representative_topics":["chinese evidence-based fact-checking with joint evidence retrieval and veracity prediction","multi-modal explainable ai for interpretable fake news detection on social media","interpretable fake news classification using bert embeddings and word-token activation mapping"]},{"direction_index":4,"label":"D4","name":"Multimodal semantic verification","description":"Detect or verify misinformation by jointly modeling text with images, video, captions, OCR, charts, or other non-textual signals to assess cross-modal consistency or multimodal veracity.","inclusion":"Assign work whose primary objective is multimodal veracity assessment, image-text consistency, cross-modal entailment, multimodal fusion, visual-textual grounding, multimodal benchmark construction, cross-modal manipulation localization, or interpretable multimodal reasoning for misinformation.","exclusion":"Use synthetic media forensics when the main question is whether visual media are manipulated or generated at the signal/artifact level; use generalization across domains and languages when multimodal data are mainly used to study domain shift; use evidence-grounded verification for primarily textual or structured evidence reasoning.","nearest_boundary":"Synthetic media forensics","representative_topics":["bootstrapping multi-view representations for cross-modal consistency-aware fake news detection","multimodal fusion and ensemble learning for social media rumor veracity classification","parameter-efficient multimodal fact verification with cross-modal multi-type fusion and ensemble methods"]},{"direction_index":5,"label":"D5","name":"Synthetic media forensics","description":"Detect manipulated, generated, or inauthentic visual media by identifying forensic traces, generation artifacts, signal inconsistencies, physiological anomalies, facial dynamics, frequency-domain cues, or robustness failures in images, faces, or videos.","inclusion":"Assign work focused on authenticity detection for deepfakes, synthetic images, face forgeries, manipulated images, or AI-generated visual media using artifact analysis, forensic networks, physiological signals, landmark dynamics, degradation robustness, generator generalization, or visual-forensic prompts.","exclusion":"Use multimodal semantic verification when the central task is semantic truthfulness or image-text claim consistency; use LLM-mediated detection when the central artifact is generated text or LLM reasoning; use textual content detection for non-visual misinformation classification.","nearest_boundary":"Multimodal semantic verification","representative_topics":["human-centered deepfake detection using artifact attention and visual caricature amplification","robustness analysis of deepfake detectors under quality-degrading in-the-wild corruptions","multimodal fake video detection via physiological signal visual representations and graph fusion"]},{"direction_index":6,"label":"D6","name":"LLM-mediated detection","description":"Study misinformation detection problems whose primary object is LLM-generated content, LLM detector behavior, or LLM-based reasoning, prompting, and labeling for misinformation tasks.","inclusion":"Assign work where large language models are either the object of detection or the central methodological tool, including detecting machine-generated fake reviews or news, evaluating LLMs as misinformation detectors, prompt-engineered LLM classification, LLM-based false-statement judgment, detector bias in LLM-era settings, and LLM-assisted conspiracy or narrative identification.","exclusion":"Use textual content detection for ordinary BERT/RoBERTa-style text classifiers where LLM generation or LLM capability is not central; use multimodal semantic verification for image-text veracity unless LLM prompting/capability is the primary question; use synthetic media forensics for visual deepfake detection.","nearest_boundary":"Textual content detection","representative_topics":["fine-tuned language model detection of machine-generated fake restaurant reviews","llm-based prompt engineering for out-of-context image-caption coherence detection","fine-tuned transformer models for ai-generated fake news detection"]},{"direction_index":7,"label":"D7","name":"Fact-checking workflow and intervention","description":"Support the upstream and downstream human-facing fact-checking pipeline by identifying check-worthy claims, locating claim spans, retrieving prior fact checks, rewriting discovery queries, generating or recommending rebuttals, and evaluating corrective interventions.","inclusion":"Assign work whose primary objective is deciding what should be checked, extracting or reformulating candidate claims for fact-checkers, discovering known misinformation, comparing fact-checking workflows, producing debunks or rebuttals, controlled claim editing, metacognitive intervention agents, or measuring whether explanations and interventions reduce misinformation belief.","exclusion":"Use evidence-grounded verification when an already-selected claim is judged against retrieved evidence; use textual content detection or social propagation modeling when the goal is simply assigning veracity; use LLM-mediated detection when LLM capability rather than intervention or workflow outcome is primary.","nearest_boundary":"Evidence-grounded verification","representative_topics":["multilingual check-worthiness classification of tweets using transformer models","crowdsourced versus expert fact-checking comparison on social media misinformation","abstractive summarization-based query generation for automated fact-checked claim detection"]}],"split":"dev"}
{"episode_id":"F156_T2025-07","field_index":156,"field":"misinformation detection","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Textual content detection","description":"Detect misinformation primarily from linguistic content, document representations, stylistic cues, narratives, sentiment, or text-classification architectures, including domain-specific textual settings such as health misinformation when the research objective remains content classification.","inclusion":"Assign work whose main objective is veracity, fake-news, disinformation, deception, propaganda, misleading-narrative, false-review, or health-misinformation classification from intrinsic text or content signals without central reliance on external evidence, social propagation, multimodal consistency, language-resource scarcity, or LLM-era generation questions.","exclusion":"Use evidence-grounded verification when explicit claim-evidence retrieval or reasoning is central; use social propagation modeling when diffusion, users, or conversation structure are central; use generalization across domains and languages when transfer or low-resource language coverage is the main objective; use LLM-mediated detection when LLM generation or LLM capability is the object.","nearest_boundary":"Evidence-grounded verification","representative_topics":["fake news classification under class imbalance using pre-trained encoders and negation-based data augmentation","multi-policy deep learning framework for joint statement and article fake news detection","hybrid ensemble of classical and deep learning models for fake news detection"]},{"direction_index":1,"label":"D1","name":"Social propagation modeling","description":"Detect misinformation by modeling how information spreads through users, reports, conversations, temporal cascades, propagation trees, social graphs, heterogeneous networks, or interaction contexts.","inclusion":"Assign work whose primary detection signal is social context, user behavior, feedback reports, graph structure, temporal diffusion, thread structure, or propagation dynamics, including GNN, hypergraph, heterogeneous-graph, streaming-rumor, and online propagation models.","exclusion":"Use textual content detection when social context is absent or only auxiliary; use evidence-grounded verification when graphs represent evidence or knowledge rather than social spread; use multimodal semantic verification when image/video-text reasoning is central.","nearest_boundary":"Textual content detection","representative_topics":["diverse example-based explanations for streaming rumour graph detection","graph neural network fake news detection via social propagation context and transformer content features","user-report-driven classification of reported misinformation content across platforms and regions"]},{"direction_index":2,"label":"D2","name":"Generalization across domains and languages","description":"Improve misinformation detection or verification under distribution shift, temporal drift, emerging topics, scarce labels, dataset bias, cross-domain transfer, multilingual settings, or low-resource languages.","inclusion":"Assign work whose primary objective is robustness or transfer across domains, time periods, topics, modalities, or languages, including domain adaptation, adversarial or contrastive feature alignment, weak/self-supervised learning, temporal knowledgebase construction, language-specific dataset creation, multilingual modeling, cross-lingual transfer, and low-resource language adaptation.","exclusion":"Use textual content detection for single-domain or single-language classification without a transfer, drift, or resource-scarcity objective; use multimodal semantic verification when cross-modal reasoning is the main contribution; use evidence-grounded verification when claim-evidence reasoning dominates and language or domain transfer is secondary.","nearest_boundary":"Textual content detection","representative_topics":["multi-domain fake news detection via memory-guided multi-view domain adaptation","low-resource urdu fake news detection using transformer and traditional feature representations","low-resource urdu fake news detection using transformer-based neural classification"]},{"direction_index":3,"label":"D3","name":"Evidence-grounded verification","description":"Verify claims by retrieving, selecting, reasoning over, distilling, or explaining textual or structured evidence, including interpretable claim-evidence reasoning and evidence-aware fake-news detection.","inclusion":"Assign work whose main objective is grounding veracity judgments in explicit evidence or producing human-interpretable reasoning for verification, including evidence retrieval, evidence selection, multi-hop reasoning, rationale extraction, justification generation, table verification, evidence distillation, claim-evidence entailment, and evidence-aware explainable detection.","exclusion":"Use textual content detection when no explicit evidence comparison is central; use multimodal semantic verification when visual or cross-modal evidence is central; use fact-checking workflow and intervention when the goal is selecting claims to check or producing corrective responses rather than verifying a claim against evidence.","nearest_boundary":"Fact-checking workflow and intervention","representative_topics":["chinese evidence-based fact-checking with joint evidence retrieval and veracity prediction","multi-modal explainable ai for interpretable fake news detection on social media","interpretable fake news classification using bert embeddings and word-token activation mapping"]},{"direction_index":4,"label":"D4","name":"Multimodal semantic verification","description":"Detect or verify misinformation by jointly modeling text with images, video, captions, OCR, charts, or other non-textual signals to assess cross-modal consistency or multimodal veracity.","inclusion":"Assign work whose primary objective is multimodal veracity assessment, image-text consistency, cross-modal entailment, multimodal fusion, visual-textual grounding, multimodal benchmark construction, cross-modal manipulation localization, or interpretable multimodal reasoning for misinformation.","exclusion":"Use synthetic media forensics when the main question is whether visual media are manipulated or generated at the signal/artifact level; use generalization across domains and languages when multimodal data are mainly used to study domain shift; use evidence-grounded verification for primarily textual or structured evidence reasoning.","nearest_boundary":"Synthetic media forensics","representative_topics":["bootstrapping multi-view representations for cross-modal consistency-aware fake news detection","multimodal fusion and ensemble learning for social media rumor veracity classification","parameter-efficient multimodal fact verification with cross-modal multi-type fusion and ensemble methods"]},{"direction_index":5,"label":"D5","name":"Synthetic media forensics","description":"Detect manipulated, generated, or inauthentic visual media by identifying forensic traces, generation artifacts, signal inconsistencies, physiological anomalies, facial dynamics, frequency-domain cues, or robustness failures in images, faces, or videos.","inclusion":"Assign work focused on authenticity detection for deepfakes, synthetic images, face forgeries, manipulated images, or AI-generated visual media using artifact analysis, forensic networks, physiological signals, landmark dynamics, degradation robustness, generator generalization, or visual-forensic prompts.","exclusion":"Use multimodal semantic verification when the central task is semantic truthfulness or image-text claim consistency; use LLM-mediated detection when the central artifact is generated text or LLM reasoning; use textual content detection for non-visual misinformation classification.","nearest_boundary":"Multimodal semantic verification","representative_topics":["human-centered deepfake detection using artifact attention and visual caricature amplification","robustness analysis of deepfake detectors under quality-degrading in-the-wild corruptions","multimodal fake video detection via physiological signal visual representations and graph fusion"]},{"direction_index":6,"label":"D6","name":"LLM-mediated detection","description":"Study misinformation detection problems whose primary object is LLM-generated content, LLM detector behavior, or LLM-based reasoning, prompting, and labeling for misinformation tasks.","inclusion":"Assign work where large language models are either the object of detection or the central methodological tool, including detecting machine-generated fake reviews or news, evaluating LLMs as misinformation detectors, prompt-engineered LLM classification, LLM-based false-statement judgment, detector bias in LLM-era settings, and LLM-assisted conspiracy or narrative identification.","exclusion":"Use textual content detection for ordinary BERT/RoBERTa-style text classifiers where LLM generation or LLM capability is not central; use multimodal semantic verification for image-text veracity unless LLM prompting/capability is the primary question; use synthetic media forensics for visual deepfake detection.","nearest_boundary":"Textual content detection","representative_topics":["fine-tuned language model detection of machine-generated fake restaurant reviews","llm-based prompt engineering for out-of-context image-caption coherence detection","fine-tuned transformer models for ai-generated fake news detection"]},{"direction_index":7,"label":"D7","name":"Fact-checking workflow and intervention","description":"Support the upstream and downstream human-facing fact-checking pipeline by identifying check-worthy claims, locating claim spans, retrieving prior fact checks, rewriting discovery queries, generating or recommending rebuttals, and evaluating corrective interventions.","inclusion":"Assign work whose primary objective is deciding what should be checked, extracting or reformulating candidate claims for fact-checkers, discovering known misinformation, comparing fact-checking workflows, producing debunks or rebuttals, controlled claim editing, metacognitive intervention agents, or measuring whether explanations and interventions reduce misinformation belief.","exclusion":"Use evidence-grounded verification when an already-selected claim is judged against retrieved evidence; use textual content detection or social propagation modeling when the goal is simply assigning veracity; use LLM-mediated detection when LLM capability rather than intervention or workflow outcome is primary.","nearest_boundary":"Evidence-grounded verification","representative_topics":["multilingual check-worthiness classification of tweets using transformer models","crowdsourced versus expert fact-checking comparison on social media misinformation","abstractive summarization-based query generation for automated fact-checked claim detection"]}],"split":"dev"}
{"episode_id":"F156_T2026-01","field_index":156,"field":"misinformation detection","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Textual content detection","description":"Detect misinformation primarily from linguistic content, document representations, stylistic cues, narratives, sentiment, or text-classification architectures, including domain-specific textual settings such as health misinformation when the research objective remains content classification.","inclusion":"Assign work whose main objective is veracity, fake-news, disinformation, deception, propaganda, misleading-narrative, false-review, or health-misinformation classification from intrinsic text or content signals without central reliance on external evidence, social propagation, multimodal consistency, language-resource scarcity, or LLM-era generation questions.","exclusion":"Use evidence-grounded verification when explicit claim-evidence retrieval or reasoning is central; use social propagation modeling when diffusion, users, or conversation structure are central; use generalization across domains and languages when transfer or low-resource language coverage is the main objective; use LLM-mediated detection when LLM generation or LLM capability is the object.","nearest_boundary":"Evidence-grounded verification","representative_topics":["fake news classification under class imbalance using pre-trained encoders and negation-based data augmentation","multi-policy deep learning framework for joint statement and article fake news detection","hybrid ensemble of classical and deep learning models for fake news detection"]},{"direction_index":1,"label":"D1","name":"Social propagation modeling","description":"Detect misinformation by modeling how information spreads through users, reports, conversations, temporal cascades, propagation trees, social graphs, heterogeneous networks, or interaction contexts.","inclusion":"Assign work whose primary detection signal is social context, user behavior, feedback reports, graph structure, temporal diffusion, thread structure, or propagation dynamics, including GNN, hypergraph, heterogeneous-graph, streaming-rumor, and online propagation models.","exclusion":"Use textual content detection when social context is absent or only auxiliary; use evidence-grounded verification when graphs represent evidence or knowledge rather than social spread; use multimodal semantic verification when image/video-text reasoning is central.","nearest_boundary":"Textual content detection","representative_topics":["diverse example-based explanations for streaming rumour graph detection","graph neural network fake news detection via social propagation context and transformer content features","user-report-driven classification of reported misinformation content across platforms and regions"]},{"direction_index":2,"label":"D2","name":"Generalization across domains and languages","description":"Improve misinformation detection or verification under distribution shift, temporal drift, emerging topics, scarce labels, dataset bias, cross-domain transfer, multilingual settings, or low-resource languages.","inclusion":"Assign work whose primary objective is robustness or transfer across domains, time periods, topics, modalities, or languages, including domain adaptation, adversarial or contrastive feature alignment, weak/self-supervised learning, temporal knowledgebase construction, language-specific dataset creation, multilingual modeling, cross-lingual transfer, and low-resource language adaptation.","exclusion":"Use textual content detection for single-domain or single-language classification without a transfer, drift, or resource-scarcity objective; use multimodal semantic verification when cross-modal reasoning is the main contribution; use evidence-grounded verification when claim-evidence reasoning dominates and language or domain transfer is secondary.","nearest_boundary":"Textual content detection","representative_topics":["multi-domain fake news detection via memory-guided multi-view domain adaptation","low-resource urdu fake news detection using transformer and traditional feature representations","low-resource urdu fake news detection using transformer-based neural classification"]},{"direction_index":3,"label":"D3","name":"Evidence-grounded verification","description":"Verify claims by retrieving, selecting, reasoning over, distilling, or explaining textual or structured evidence, including interpretable claim-evidence reasoning and evidence-aware fake-news detection.","inclusion":"Assign work whose main objective is grounding veracity judgments in explicit evidence or producing human-interpretable reasoning for verification, including evidence retrieval, evidence selection, multi-hop reasoning, rationale extraction, justification generation, table verification, evidence distillation, claim-evidence entailment, and evidence-aware explainable detection.","exclusion":"Use textual content detection when no explicit evidence comparison is central; use multimodal semantic verification when visual or cross-modal evidence is central; use fact-checking workflow and intervention when the goal is selecting claims to check or producing corrective responses rather than verifying a claim against evidence.","nearest_boundary":"Fact-checking workflow and intervention","representative_topics":["chinese evidence-based fact-checking with joint evidence retrieval and veracity prediction","multi-modal explainable ai for interpretable fake news detection on social media","interpretable fake news classification using bert embeddings and word-token activation mapping"]},{"direction_index":4,"label":"D4","name":"Multimodal semantic verification","description":"Detect or verify misinformation by jointly modeling text with images, video, captions, OCR, charts, or other non-textual signals to assess cross-modal consistency or multimodal veracity.","inclusion":"Assign work whose primary objective is multimodal veracity assessment, image-text consistency, cross-modal entailment, multimodal fusion, visual-textual grounding, multimodal benchmark construction, cross-modal manipulation localization, or interpretable multimodal reasoning for misinformation.","exclusion":"Use synthetic media forensics when the main question is whether visual media are manipulated or generated at the signal/artifact level; use generalization across domains and languages when multimodal data are mainly used to study domain shift; use evidence-grounded verification for primarily textual or structured evidence reasoning.","nearest_boundary":"Synthetic media forensics","representative_topics":["bootstrapping multi-view representations for cross-modal consistency-aware fake news detection","multimodal fusion and ensemble learning for social media rumor veracity classification","parameter-efficient multimodal fact verification with cross-modal multi-type fusion and ensemble methods"]},{"direction_index":5,"label":"D5","name":"Synthetic media forensics","description":"Detect manipulated, generated, or inauthentic visual media by identifying forensic traces, generation artifacts, signal inconsistencies, physiological anomalies, facial dynamics, frequency-domain cues, or robustness failures in images, faces, or videos.","inclusion":"Assign work focused on authenticity detection for deepfakes, synthetic images, face forgeries, manipulated images, or AI-generated visual media using artifact analysis, forensic networks, physiological signals, landmark dynamics, degradation robustness, generator generalization, or visual-forensic prompts.","exclusion":"Use multimodal semantic verification when the central task is semantic truthfulness or image-text claim consistency; use LLM-mediated detection when the central artifact is generated text or LLM reasoning; use textual content detection for non-visual misinformation classification.","nearest_boundary":"Multimodal semantic verification","representative_topics":["human-centered deepfake detection using artifact attention and visual caricature amplification","robustness analysis of deepfake detectors under quality-degrading in-the-wild corruptions","multimodal fake video detection via physiological signal visual representations and graph fusion"]},{"direction_index":6,"label":"D6","name":"LLM-mediated detection","description":"Study misinformation detection problems whose primary object is LLM-generated content, LLM detector behavior, or LLM-based reasoning, prompting, and labeling for misinformation tasks.","inclusion":"Assign work where large language models are either the object of detection or the central methodological tool, including detecting machine-generated fake reviews or news, evaluating LLMs as misinformation detectors, prompt-engineered LLM classification, LLM-based false-statement judgment, detector bias in LLM-era settings, and LLM-assisted conspiracy or narrative identification.","exclusion":"Use textual content detection for ordinary BERT/RoBERTa-style text classifiers where LLM generation or LLM capability is not central; use multimodal semantic verification for image-text veracity unless LLM prompting/capability is the primary question; use synthetic media forensics for visual deepfake detection.","nearest_boundary":"Textual content detection","representative_topics":["fine-tuned language model detection of machine-generated fake restaurant reviews","llm-based prompt engineering for out-of-context image-caption coherence detection","fine-tuned transformer models for ai-generated fake news detection"]},{"direction_index":7,"label":"D7","name":"Fact-checking workflow and intervention","description":"Support the upstream and downstream human-facing fact-checking pipeline by identifying check-worthy claims, locating claim spans, retrieving prior fact checks, rewriting discovery queries, generating or recommending rebuttals, and evaluating corrective interventions.","inclusion":"Assign work whose primary objective is deciding what should be checked, extracting or reformulating candidate claims for fact-checkers, discovering known misinformation, comparing fact-checking workflows, producing debunks or rebuttals, controlled claim editing, metacognitive intervention agents, or measuring whether explanations and interventions reduce misinformation belief.","exclusion":"Use evidence-grounded verification when an already-selected claim is judged against retrieved evidence; use textual content detection or social propagation modeling when the goal is simply assigning veracity; use LLM-mediated detection when LLM capability rather than intervention or workflow outcome is primary.","nearest_boundary":"Evidence-grounded verification","representative_topics":["multilingual check-worthiness classification of tweets using transformer models","crowdsourced versus expert fact-checking comparison on social media misinformation","abstractive summarization-based query generation for automated fact-checked claim detection"]}],"split":"dev"}
{"episode_id":"F160_T2024-01","field_index":160,"field":"knowledge graph construction","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Ontology-guided domain KG construction","description":"Build domain- or institution-specific knowledge graphs whose main objective is semantic modeling, domain schema design, ontology use, and interoperable representation for a vertical setting.","inclusion":"Include work primarily constructing or harmonizing a KG around domain ontologies, controlled vocabularies, schemas, domain rules, or interoperability requirements, even when extraction or analytics are supporting components.","exclusion":"If the primary contribution is a transferable text-to-KG extraction method, assign to text-KG transduction; if the main problem is reconciling multiple existing KGs, assign to KG integration and alignment.","nearest_boundary":"Text-KG transduction","representative_topics":["open knowledge graph construction for metal-organic framework discovery and synthesis","knowledge graph induction from heterogeneous sources for recommendation and trend analysis","multilingual knowledge graph construction for cross-lingual quote alignment and attribution verification"]},{"direction_index":1,"label":"D1","name":"Text-KG transduction","description":"Transform between natural language and KG content, covering text-to-KG extraction/population and KG-conditioned verbalization, simplification, or generation.","inclusion":"Include work whose primary objective is converting unstructured or semi-structured text into KG assertions, entities, relations, events, or schema-linked facts, or conversely generating faithful natural language from KG structures.","exclusion":"If the main objective is domain ontology design rather than transduction, assign to ontology-guided domain KG construction; if language models are used mainly to complete missing facts in an existing KG, assign to KG representation and completion.","nearest_boundary":"Ontology-guided domain KG construction","representative_topics":["harvesting symbolic knowledge graphs of arbitrary relations via few-shot prompting of pretrained language models","syntax-controlled knowledge graph-to-text generation with order optimization and semantic regularization","retrieval-augmented schema-aware prompting for data-efficient knowledge graph construction"]},{"direction_index":2,"label":"D2","name":"KG curation and quality enrichment","description":"Improve existing or partially built KGs through cleaning, canonicalization, deduplication, predicate consolidation, quality evaluation, and coverage or precision enrichment.","inclusion":"Include work primarily focused on post-construction KG quality, error detection, duplicate removal, canonicalization, single-KG linking, enrichment, or quality assessment.","exclusion":"If the primary objective is matching or fusing distinct KGs across sources, schemas, or languages, assign to KG integration and alignment; if the objective is embedding-based missing-link prediction, assign to KG representation and completion.","nearest_boundary":"KG integration and alignment","representative_topics":["iterative knowledge graph enrichment via collaborative join event extraction and relation alignment","knowledge graph curation framework for quality assessment, cleaning, and enrichment","predicate clustering for terminology convergence in open research knowledge graphs"]},{"direction_index":3,"label":"D3","name":"KG integration and alignment","description":"Align, merge, map, deduplicate, or reconcile heterogeneous KGs across sources, languages, schemas, ontologies, domains, or graph instances.","inclusion":"Include work whose central objective is cross-graph interoperability, entity alignment, ontology integration, schema mapping, cross-lingual grounding, multi-source matching, or collective reconciliation of multiple KGs.","exclusion":"If alignment is only a component of building one domain KG, assign to ontology-guided domain KG construction; if the work cleans or canonicalizes within a single KG, assign to KG curation and quality enrichment.","nearest_boundary":"KG curation and quality enrichment","representative_topics":["ontology integration for node deduplication and hierarchy connectivity in heterogeneous knowledge graphs","unsupervised cross-lingual sememe knowledge base construction via deep clustering of word embeddings","large-scale gold standard benchmarking for multi-source knowledge graph matching and alignment"]},{"direction_index":4,"label":"D4","name":"KG representation and completion","description":"Learn representations or language-aware models to predict missing KG facts, label relations, complete triples, or analyze completion behavior over existing graphs.","inclusion":"Include embedding-based, GNN-based, attention-based, contrastive, efficient, PLM-integrated, retrieval-augmented, or language-semantics-driven methods whose primary objective is KG completion or link/relation prediction.","exclusion":"If the primary objective is complex logical query answering, explicit rule reasoning, or first-order reasoning over a KG, assign to KG reasoning and query answering; if the output is extracted from raw text rather than inferred over an existing KG, assign to text-KG transduction.","nearest_boundary":"KG reasoning and query answering","representative_topics":["attention-based entity representation learning with semantic path context for knowledge graph relation prediction","knowledge graph embedding via relational prototype entities for global semantic similarity","negative sampling-free knowledge graph embedding via cross-correlation matrix contrastive learning"]},{"direction_index":5,"label":"D5","name":"KG reasoning and query answering","description":"Perform explicit reasoning, fact verification, complex query answering, or rule-guided inference over KGs using logical, neural-symbolic, or LLM-based reasoning mechanisms.","inclusion":"Include work primarily targeting multi-hop logical queries, negation, union, first-order logic operators, soft-rule grounding, fuzzy inference, explainable reasoning, fact verification, or LLM-driven reasoning over existing KGs.","exclusion":"If the work performs ordinary missing-link prediction without explicit logical/query structure, assign to KG representation and completion; if the work constructs the KG rather than reasons over it, assign to the relevant KG construction direction.","nearest_boundary":"KG representation and completion","representative_topics":["knowledge representation learning for logical query answering over multi-view knowledge graphs","probabilistic gamma embeddings for multi-hop logical query answering on knowledge graphs","knowledge graph completion via joint rule grounding and iterative embedding refinement"]},{"direction_index":6,"label":"D6","name":"Open-world KG construction","description":"Construct KGs where the main scientific objective is open, evolving, commonsense, schema-inducing, multimodal, temporal, causal, or dynamic representation rather than fixed-domain ontology instantiation.","inclusion":"Include commonsense KG construction and densification, open-schema KB construction, taxonomy or concept-hierarchy induction, sememe or synset construction, multimodal KG construction, dynamic or event-centric KG construction, and causal or temporal KG building.","exclusion":"If the work mainly instantiates a fixed domain ontology for a vertical application, assign to ontology-guided domain KG construction; if the work primarily extracts fixed-schema facts from text, assign to text-KG transduction.","nearest_boundary":"Ontology-guided domain KG construction","representative_topics":["informative negative commonsense statement generation using comparable concept seeds and closed-world pruning","densifying commonsense knowledge graphs via relation completion for multi-hop path coverage","automatic chinese open-domain knowledge graph with fine-grained hypernym-hyponym hierarchy and path disambiguation"]},{"direction_index":7,"label":"D7","name":"KG methodology and infrastructure","description":"Systematize, benchmark, survey, taxonomize, or provide reusable methodological infrastructure for KG construction, curation, alignment, embedding, querying, ontology learning, and evaluation.","inclusion":"Include surveys, taxonomies, benchmark datasets, gold standards, evaluation protocols, reusable libraries, systematic reviews, and methodological resources whose main contribution is organizing or enabling a KG research area rather than solving one construction or reasoning task directly.","exclusion":"If the paper primarily proposes an objective-specific construction, extraction, alignment, completion, reasoning, or curation method, assign to that objective-specific direction.","nearest_boundary":"KG curation and quality enrichment","representative_topics":["taxonomy and survey of healthcare knowledge graph construction methods and open issues","cross-lingual wordnet synset construction via translation ranking for low-resource languages","large-scale gold standard benchmarking for multi-source knowledge graph matching and alignment"]}],"split":"dev"}
{"episode_id":"F160_T2024-07","field_index":160,"field":"knowledge graph construction","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Ontology-guided domain KG construction","description":"Build domain- or institution-specific knowledge graphs whose main objective is semantic modeling, domain schema design, ontology use, and interoperable representation for a vertical setting.","inclusion":"Include work primarily constructing or harmonizing a KG around domain ontologies, controlled vocabularies, schemas, domain rules, or interoperability requirements, even when extraction or analytics are supporting components.","exclusion":"If the primary contribution is a transferable text-to-KG extraction method, assign to text-KG transduction; if the main problem is reconciling multiple existing KGs, assign to KG integration and alignment.","nearest_boundary":"Text-KG transduction","representative_topics":["open knowledge graph construction for metal-organic framework discovery and synthesis","knowledge graph induction from heterogeneous sources for recommendation and trend analysis","multilingual knowledge graph construction for cross-lingual quote alignment and attribution verification"]},{"direction_index":1,"label":"D1","name":"Text-KG transduction","description":"Transform between natural language and KG content, covering text-to-KG extraction/population and KG-conditioned verbalization, simplification, or generation.","inclusion":"Include work whose primary objective is converting unstructured or semi-structured text into KG assertions, entities, relations, events, or schema-linked facts, or conversely generating faithful natural language from KG structures.","exclusion":"If the main objective is domain ontology design rather than transduction, assign to ontology-guided domain KG construction; if language models are used mainly to complete missing facts in an existing KG, assign to KG representation and completion.","nearest_boundary":"Ontology-guided domain KG construction","representative_topics":["harvesting symbolic knowledge graphs of arbitrary relations via few-shot prompting of pretrained language models","syntax-controlled knowledge graph-to-text generation with order optimization and semantic regularization","retrieval-augmented schema-aware prompting for data-efficient knowledge graph construction"]},{"direction_index":2,"label":"D2","name":"KG curation and quality enrichment","description":"Improve existing or partially built KGs through cleaning, canonicalization, deduplication, predicate consolidation, quality evaluation, and coverage or precision enrichment.","inclusion":"Include work primarily focused on post-construction KG quality, error detection, duplicate removal, canonicalization, single-KG linking, enrichment, or quality assessment.","exclusion":"If the primary objective is matching or fusing distinct KGs across sources, schemas, or languages, assign to KG integration and alignment; if the objective is embedding-based missing-link prediction, assign to KG representation and completion.","nearest_boundary":"KG integration and alignment","representative_topics":["iterative knowledge graph enrichment via collaborative join event extraction and relation alignment","knowledge graph curation framework for quality assessment, cleaning, and enrichment","predicate clustering for terminology convergence in open research knowledge graphs"]},{"direction_index":3,"label":"D3","name":"KG integration and alignment","description":"Align, merge, map, deduplicate, or reconcile heterogeneous KGs across sources, languages, schemas, ontologies, domains, or graph instances.","inclusion":"Include work whose central objective is cross-graph interoperability, entity alignment, ontology integration, schema mapping, cross-lingual grounding, multi-source matching, or collective reconciliation of multiple KGs.","exclusion":"If alignment is only a component of building one domain KG, assign to ontology-guided domain KG construction; if the work cleans or canonicalizes within a single KG, assign to KG curation and quality enrichment.","nearest_boundary":"KG curation and quality enrichment","representative_topics":["ontology integration for node deduplication and hierarchy connectivity in heterogeneous knowledge graphs","unsupervised cross-lingual sememe knowledge base construction via deep clustering of word embeddings","large-scale gold standard benchmarking for multi-source knowledge graph matching and alignment"]},{"direction_index":4,"label":"D4","name":"KG representation and completion","description":"Learn representations or language-aware models to predict missing KG facts, label relations, complete triples, or analyze completion behavior over existing graphs.","inclusion":"Include embedding-based, GNN-based, attention-based, contrastive, efficient, PLM-integrated, retrieval-augmented, or language-semantics-driven methods whose primary objective is KG completion or link/relation prediction.","exclusion":"If the primary objective is complex logical query answering, explicit rule reasoning, or first-order reasoning over a KG, assign to KG reasoning and query answering; if the output is extracted from raw text rather than inferred over an existing KG, assign to text-KG transduction.","nearest_boundary":"KG reasoning and query answering","representative_topics":["attention-based entity representation learning with semantic path context for knowledge graph relation prediction","knowledge graph embedding via relational prototype entities for global semantic similarity","negative sampling-free knowledge graph embedding via cross-correlation matrix contrastive learning"]},{"direction_index":5,"label":"D5","name":"KG reasoning and query answering","description":"Perform explicit reasoning, fact verification, complex query answering, or rule-guided inference over KGs using logical, neural-symbolic, or LLM-based reasoning mechanisms.","inclusion":"Include work primarily targeting multi-hop logical queries, negation, union, first-order logic operators, soft-rule grounding, fuzzy inference, explainable reasoning, fact verification, or LLM-driven reasoning over existing KGs.","exclusion":"If the work performs ordinary missing-link prediction without explicit logical/query structure, assign to KG representation and completion; if the work constructs the KG rather than reasons over it, assign to the relevant KG construction direction.","nearest_boundary":"KG representation and completion","representative_topics":["knowledge representation learning for logical query answering over multi-view knowledge graphs","probabilistic gamma embeddings for multi-hop logical query answering on knowledge graphs","knowledge graph completion via joint rule grounding and iterative embedding refinement"]},{"direction_index":6,"label":"D6","name":"Open-world KG construction","description":"Construct KGs where the main scientific objective is open, evolving, commonsense, schema-inducing, multimodal, temporal, causal, or dynamic representation rather than fixed-domain ontology instantiation.","inclusion":"Include commonsense KG construction and densification, open-schema KB construction, taxonomy or concept-hierarchy induction, sememe or synset construction, multimodal KG construction, dynamic or event-centric KG construction, and causal or temporal KG building.","exclusion":"If the work mainly instantiates a fixed domain ontology for a vertical application, assign to ontology-guided domain KG construction; if the work primarily extracts fixed-schema facts from text, assign to text-KG transduction.","nearest_boundary":"Ontology-guided domain KG construction","representative_topics":["informative negative commonsense statement generation using comparable concept seeds and closed-world pruning","densifying commonsense knowledge graphs via relation completion for multi-hop path coverage","automatic chinese open-domain knowledge graph with fine-grained hypernym-hyponym hierarchy and path disambiguation"]},{"direction_index":7,"label":"D7","name":"KG methodology and infrastructure","description":"Systematize, benchmark, survey, taxonomize, or provide reusable methodological infrastructure for KG construction, curation, alignment, embedding, querying, ontology learning, and evaluation.","inclusion":"Include surveys, taxonomies, benchmark datasets, gold standards, evaluation protocols, reusable libraries, systematic reviews, and methodological resources whose main contribution is organizing or enabling a KG research area rather than solving one construction or reasoning task directly.","exclusion":"If the paper primarily proposes an objective-specific construction, extraction, alignment, completion, reasoning, or curation method, assign to that objective-specific direction.","nearest_boundary":"KG curation and quality enrichment","representative_topics":["taxonomy and survey of healthcare knowledge graph construction methods and open issues","cross-lingual wordnet synset construction via translation ranking for low-resource languages","large-scale gold standard benchmarking for multi-source knowledge graph matching and alignment"]}],"split":"dev"}
{"episode_id":"F160_T2025-01","field_index":160,"field":"knowledge graph construction","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Ontology-guided domain KG construction","description":"Build domain- or institution-specific knowledge graphs whose main objective is semantic modeling, domain schema design, ontology use, and interoperable representation for a vertical setting.","inclusion":"Include work primarily constructing or harmonizing a KG around domain ontologies, controlled vocabularies, schemas, domain rules, or interoperability requirements, even when extraction or analytics are supporting components.","exclusion":"If the primary contribution is a transferable text-to-KG extraction method, assign to text-KG transduction; if the main problem is reconciling multiple existing KGs, assign to KG integration and alignment.","nearest_boundary":"Text-KG transduction","representative_topics":["open knowledge graph construction for metal-organic framework discovery and synthesis","knowledge graph induction from heterogeneous sources for recommendation and trend analysis","multilingual knowledge graph construction for cross-lingual quote alignment and attribution verification"]},{"direction_index":1,"label":"D1","name":"Text-KG transduction","description":"Transform between natural language and KG content, covering text-to-KG extraction/population and KG-conditioned verbalization, simplification, or generation.","inclusion":"Include work whose primary objective is converting unstructured or semi-structured text into KG assertions, entities, relations, events, or schema-linked facts, or conversely generating faithful natural language from KG structures.","exclusion":"If the main objective is domain ontology design rather than transduction, assign to ontology-guided domain KG construction; if language models are used mainly to complete missing facts in an existing KG, assign to KG representation and completion.","nearest_boundary":"Ontology-guided domain KG construction","representative_topics":["harvesting symbolic knowledge graphs of arbitrary relations via few-shot prompting of pretrained language models","syntax-controlled knowledge graph-to-text generation with order optimization and semantic regularization","retrieval-augmented schema-aware prompting for data-efficient knowledge graph construction"]},{"direction_index":2,"label":"D2","name":"KG curation and quality enrichment","description":"Improve existing or partially built KGs through cleaning, canonicalization, deduplication, predicate consolidation, quality evaluation, and coverage or precision enrichment.","inclusion":"Include work primarily focused on post-construction KG quality, error detection, duplicate removal, canonicalization, single-KG linking, enrichment, or quality assessment.","exclusion":"If the primary objective is matching or fusing distinct KGs across sources, schemas, or languages, assign to KG integration and alignment; if the objective is embedding-based missing-link prediction, assign to KG representation and completion.","nearest_boundary":"KG integration and alignment","representative_topics":["iterative knowledge graph enrichment via collaborative join event extraction and relation alignment","knowledge graph curation framework for quality assessment, cleaning, and enrichment","predicate clustering for terminology convergence in open research knowledge graphs"]},{"direction_index":3,"label":"D3","name":"KG integration and alignment","description":"Align, merge, map, deduplicate, or reconcile heterogeneous KGs across sources, languages, schemas, ontologies, domains, or graph instances.","inclusion":"Include work whose central objective is cross-graph interoperability, entity alignment, ontology integration, schema mapping, cross-lingual grounding, multi-source matching, or collective reconciliation of multiple KGs.","exclusion":"If alignment is only a component of building one domain KG, assign to ontology-guided domain KG construction; if the work cleans or canonicalizes within a single KG, assign to KG curation and quality enrichment.","nearest_boundary":"KG curation and quality enrichment","representative_topics":["ontology integration for node deduplication and hierarchy connectivity in heterogeneous knowledge graphs","unsupervised cross-lingual sememe knowledge base construction via deep clustering of word embeddings","large-scale gold standard benchmarking for multi-source knowledge graph matching and alignment"]},{"direction_index":4,"label":"D4","name":"KG representation and completion","description":"Learn representations or language-aware models to predict missing KG facts, label relations, complete triples, or analyze completion behavior over existing graphs.","inclusion":"Include embedding-based, GNN-based, attention-based, contrastive, efficient, PLM-integrated, retrieval-augmented, or language-semantics-driven methods whose primary objective is KG completion or link/relation prediction.","exclusion":"If the primary objective is complex logical query answering, explicit rule reasoning, or first-order reasoning over a KG, assign to KG reasoning and query answering; if the output is extracted from raw text rather than inferred over an existing KG, assign to text-KG transduction.","nearest_boundary":"KG reasoning and query answering","representative_topics":["attention-based entity representation learning with semantic path context for knowledge graph relation prediction","knowledge graph embedding via relational prototype entities for global semantic similarity","negative sampling-free knowledge graph embedding via cross-correlation matrix contrastive learning"]},{"direction_index":5,"label":"D5","name":"KG reasoning and query answering","description":"Perform explicit reasoning, fact verification, complex query answering, or rule-guided inference over KGs using logical, neural-symbolic, or LLM-based reasoning mechanisms.","inclusion":"Include work primarily targeting multi-hop logical queries, negation, union, first-order logic operators, soft-rule grounding, fuzzy inference, explainable reasoning, fact verification, or LLM-driven reasoning over existing KGs.","exclusion":"If the work performs ordinary missing-link prediction without explicit logical/query structure, assign to KG representation and completion; if the work constructs the KG rather than reasons over it, assign to the relevant KG construction direction.","nearest_boundary":"KG representation and completion","representative_topics":["knowledge representation learning for logical query answering over multi-view knowledge graphs","probabilistic gamma embeddings for multi-hop logical query answering on knowledge graphs","knowledge graph completion via joint rule grounding and iterative embedding refinement"]},{"direction_index":6,"label":"D6","name":"Open-world KG construction","description":"Construct KGs where the main scientific objective is open, evolving, commonsense, schema-inducing, multimodal, temporal, causal, or dynamic representation rather than fixed-domain ontology instantiation.","inclusion":"Include commonsense KG construction and densification, open-schema KB construction, taxonomy or concept-hierarchy induction, sememe or synset construction, multimodal KG construction, dynamic or event-centric KG construction, and causal or temporal KG building.","exclusion":"If the work mainly instantiates a fixed domain ontology for a vertical application, assign to ontology-guided domain KG construction; if the work primarily extracts fixed-schema facts from text, assign to text-KG transduction.","nearest_boundary":"Ontology-guided domain KG construction","representative_topics":["informative negative commonsense statement generation using comparable concept seeds and closed-world pruning","densifying commonsense knowledge graphs via relation completion for multi-hop path coverage","automatic chinese open-domain knowledge graph with fine-grained hypernym-hyponym hierarchy and path disambiguation"]},{"direction_index":7,"label":"D7","name":"KG methodology and infrastructure","description":"Systematize, benchmark, survey, taxonomize, or provide reusable methodological infrastructure for KG construction, curation, alignment, embedding, querying, ontology learning, and evaluation.","inclusion":"Include surveys, taxonomies, benchmark datasets, gold standards, evaluation protocols, reusable libraries, systematic reviews, and methodological resources whose main contribution is organizing or enabling a KG research area rather than solving one construction or reasoning task directly.","exclusion":"If the paper primarily proposes an objective-specific construction, extraction, alignment, completion, reasoning, or curation method, assign to that objective-specific direction.","nearest_boundary":"KG curation and quality enrichment","representative_topics":["taxonomy and survey of healthcare knowledge graph construction methods and open issues","cross-lingual wordnet synset construction via translation ranking for low-resource languages","large-scale gold standard benchmarking for multi-source knowledge graph matching and alignment"]}],"split":"dev"}
{"episode_id":"F160_T2025-07","field_index":160,"field":"knowledge graph construction","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Ontology-guided domain KG construction","description":"Build domain- or institution-specific knowledge graphs whose main objective is semantic modeling, domain schema design, ontology use, and interoperable representation for a vertical setting.","inclusion":"Include work primarily constructing or harmonizing a KG around domain ontologies, controlled vocabularies, schemas, domain rules, or interoperability requirements, even when extraction or analytics are supporting components.","exclusion":"If the primary contribution is a transferable text-to-KG extraction method, assign to text-KG transduction; if the main problem is reconciling multiple existing KGs, assign to KG integration and alignment.","nearest_boundary":"Text-KG transduction","representative_topics":["open knowledge graph construction for metal-organic framework discovery and synthesis","knowledge graph induction from heterogeneous sources for recommendation and trend analysis","multilingual knowledge graph construction for cross-lingual quote alignment and attribution verification"]},{"direction_index":1,"label":"D1","name":"Text-KG transduction","description":"Transform between natural language and KG content, covering text-to-KG extraction/population and KG-conditioned verbalization, simplification, or generation.","inclusion":"Include work whose primary objective is converting unstructured or semi-structured text into KG assertions, entities, relations, events, or schema-linked facts, or conversely generating faithful natural language from KG structures.","exclusion":"If the main objective is domain ontology design rather than transduction, assign to ontology-guided domain KG construction; if language models are used mainly to complete missing facts in an existing KG, assign to KG representation and completion.","nearest_boundary":"Ontology-guided domain KG construction","representative_topics":["harvesting symbolic knowledge graphs of arbitrary relations via few-shot prompting of pretrained language models","syntax-controlled knowledge graph-to-text generation with order optimization and semantic regularization","retrieval-augmented schema-aware prompting for data-efficient knowledge graph construction"]},{"direction_index":2,"label":"D2","name":"KG curation and quality enrichment","description":"Improve existing or partially built KGs through cleaning, canonicalization, deduplication, predicate consolidation, quality evaluation, and coverage or precision enrichment.","inclusion":"Include work primarily focused on post-construction KG quality, error detection, duplicate removal, canonicalization, single-KG linking, enrichment, or quality assessment.","exclusion":"If the primary objective is matching or fusing distinct KGs across sources, schemas, or languages, assign to KG integration and alignment; if the objective is embedding-based missing-link prediction, assign to KG representation and completion.","nearest_boundary":"KG integration and alignment","representative_topics":["iterative knowledge graph enrichment via collaborative join event extraction and relation alignment","knowledge graph curation framework for quality assessment, cleaning, and enrichment","predicate clustering for terminology convergence in open research knowledge graphs"]},{"direction_index":3,"label":"D3","name":"KG integration and alignment","description":"Align, merge, map, deduplicate, or reconcile heterogeneous KGs across sources, languages, schemas, ontologies, domains, or graph instances.","inclusion":"Include work whose central objective is cross-graph interoperability, entity alignment, ontology integration, schema mapping, cross-lingual grounding, multi-source matching, or collective reconciliation of multiple KGs.","exclusion":"If alignment is only a component of building one domain KG, assign to ontology-guided domain KG construction; if the work cleans or canonicalizes within a single KG, assign to KG curation and quality enrichment.","nearest_boundary":"KG curation and quality enrichment","representative_topics":["ontology integration for node deduplication and hierarchy connectivity in heterogeneous knowledge graphs","unsupervised cross-lingual sememe knowledge base construction via deep clustering of word embeddings","large-scale gold standard benchmarking for multi-source knowledge graph matching and alignment"]},{"direction_index":4,"label":"D4","name":"KG representation and completion","description":"Learn representations or language-aware models to predict missing KG facts, label relations, complete triples, or analyze completion behavior over existing graphs.","inclusion":"Include embedding-based, GNN-based, attention-based, contrastive, efficient, PLM-integrated, retrieval-augmented, or language-semantics-driven methods whose primary objective is KG completion or link/relation prediction.","exclusion":"If the primary objective is complex logical query answering, explicit rule reasoning, or first-order reasoning over a KG, assign to KG reasoning and query answering; if the output is extracted from raw text rather than inferred over an existing KG, assign to text-KG transduction.","nearest_boundary":"KG reasoning and query answering","representative_topics":["attention-based entity representation learning with semantic path context for knowledge graph relation prediction","knowledge graph embedding via relational prototype entities for global semantic similarity","negative sampling-free knowledge graph embedding via cross-correlation matrix contrastive learning"]},{"direction_index":5,"label":"D5","name":"KG reasoning and query answering","description":"Perform explicit reasoning, fact verification, complex query answering, or rule-guided inference over KGs using logical, neural-symbolic, or LLM-based reasoning mechanisms.","inclusion":"Include work primarily targeting multi-hop logical queries, negation, union, first-order logic operators, soft-rule grounding, fuzzy inference, explainable reasoning, fact verification, or LLM-driven reasoning over existing KGs.","exclusion":"If the work performs ordinary missing-link prediction without explicit logical/query structure, assign to KG representation and completion; if the work constructs the KG rather than reasons over it, assign to the relevant KG construction direction.","nearest_boundary":"KG representation and completion","representative_topics":["knowledge representation learning for logical query answering over multi-view knowledge graphs","probabilistic gamma embeddings for multi-hop logical query answering on knowledge graphs","knowledge graph completion via joint rule grounding and iterative embedding refinement"]},{"direction_index":6,"label":"D6","name":"Open-world KG construction","description":"Construct KGs where the main scientific objective is open, evolving, commonsense, schema-inducing, multimodal, temporal, causal, or dynamic representation rather than fixed-domain ontology instantiation.","inclusion":"Include commonsense KG construction and densification, open-schema KB construction, taxonomy or concept-hierarchy induction, sememe or synset construction, multimodal KG construction, dynamic or event-centric KG construction, and causal or temporal KG building.","exclusion":"If the work mainly instantiates a fixed domain ontology for a vertical application, assign to ontology-guided domain KG construction; if the work primarily extracts fixed-schema facts from text, assign to text-KG transduction.","nearest_boundary":"Ontology-guided domain KG construction","representative_topics":["informative negative commonsense statement generation using comparable concept seeds and closed-world pruning","densifying commonsense knowledge graphs via relation completion for multi-hop path coverage","automatic chinese open-domain knowledge graph with fine-grained hypernym-hyponym hierarchy and path disambiguation"]},{"direction_index":7,"label":"D7","name":"KG methodology and infrastructure","description":"Systematize, benchmark, survey, taxonomize, or provide reusable methodological infrastructure for KG construction, curation, alignment, embedding, querying, ontology learning, and evaluation.","inclusion":"Include surveys, taxonomies, benchmark datasets, gold standards, evaluation protocols, reusable libraries, systematic reviews, and methodological resources whose main contribution is organizing or enabling a KG research area rather than solving one construction or reasoning task directly.","exclusion":"If the paper primarily proposes an objective-specific construction, extraction, alignment, completion, reasoning, or curation method, assign to that objective-specific direction.","nearest_boundary":"KG curation and quality enrichment","representative_topics":["taxonomy and survey of healthcare knowledge graph construction methods and open issues","cross-lingual wordnet synset construction via translation ranking for low-resource languages","large-scale gold standard benchmarking for multi-source knowledge graph matching and alignment"]}],"split":"dev"}
{"episode_id":"F160_T2026-01","field_index":160,"field":"knowledge graph construction","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Ontology-guided domain KG construction","description":"Build domain- or institution-specific knowledge graphs whose main objective is semantic modeling, domain schema design, ontology use, and interoperable representation for a vertical setting.","inclusion":"Include work primarily constructing or harmonizing a KG around domain ontologies, controlled vocabularies, schemas, domain rules, or interoperability requirements, even when extraction or analytics are supporting components.","exclusion":"If the primary contribution is a transferable text-to-KG extraction method, assign to text-KG transduction; if the main problem is reconciling multiple existing KGs, assign to KG integration and alignment.","nearest_boundary":"Text-KG transduction","representative_topics":["open knowledge graph construction for metal-organic framework discovery and synthesis","knowledge graph induction from heterogeneous sources for recommendation and trend analysis","multilingual knowledge graph construction for cross-lingual quote alignment and attribution verification"]},{"direction_index":1,"label":"D1","name":"Text-KG transduction","description":"Transform between natural language and KG content, covering text-to-KG extraction/population and KG-conditioned verbalization, simplification, or generation.","inclusion":"Include work whose primary objective is converting unstructured or semi-structured text into KG assertions, entities, relations, events, or schema-linked facts, or conversely generating faithful natural language from KG structures.","exclusion":"If the main objective is domain ontology design rather than transduction, assign to ontology-guided domain KG construction; if language models are used mainly to complete missing facts in an existing KG, assign to KG representation and completion.","nearest_boundary":"Ontology-guided domain KG construction","representative_topics":["harvesting symbolic knowledge graphs of arbitrary relations via few-shot prompting of pretrained language models","syntax-controlled knowledge graph-to-text generation with order optimization and semantic regularization","retrieval-augmented schema-aware prompting for data-efficient knowledge graph construction"]},{"direction_index":2,"label":"D2","name":"KG curation and quality enrichment","description":"Improve existing or partially built KGs through cleaning, canonicalization, deduplication, predicate consolidation, quality evaluation, and coverage or precision enrichment.","inclusion":"Include work primarily focused on post-construction KG quality, error detection, duplicate removal, canonicalization, single-KG linking, enrichment, or quality assessment.","exclusion":"If the primary objective is matching or fusing distinct KGs across sources, schemas, or languages, assign to KG integration and alignment; if the objective is embedding-based missing-link prediction, assign to KG representation and completion.","nearest_boundary":"KG integration and alignment","representative_topics":["iterative knowledge graph enrichment via collaborative join event extraction and relation alignment","knowledge graph curation framework for quality assessment, cleaning, and enrichment","predicate clustering for terminology convergence in open research knowledge graphs"]},{"direction_index":3,"label":"D3","name":"KG integration and alignment","description":"Align, merge, map, deduplicate, or reconcile heterogeneous KGs across sources, languages, schemas, ontologies, domains, or graph instances.","inclusion":"Include work whose central objective is cross-graph interoperability, entity alignment, ontology integration, schema mapping, cross-lingual grounding, multi-source matching, or collective reconciliation of multiple KGs.","exclusion":"If alignment is only a component of building one domain KG, assign to ontology-guided domain KG construction; if the work cleans or canonicalizes within a single KG, assign to KG curation and quality enrichment.","nearest_boundary":"KG curation and quality enrichment","representative_topics":["ontology integration for node deduplication and hierarchy connectivity in heterogeneous knowledge graphs","unsupervised cross-lingual sememe knowledge base construction via deep clustering of word embeddings","large-scale gold standard benchmarking for multi-source knowledge graph matching and alignment"]},{"direction_index":4,"label":"D4","name":"KG representation and completion","description":"Learn representations or language-aware models to predict missing KG facts, label relations, complete triples, or analyze completion behavior over existing graphs.","inclusion":"Include embedding-based, GNN-based, attention-based, contrastive, efficient, PLM-integrated, retrieval-augmented, or language-semantics-driven methods whose primary objective is KG completion or link/relation prediction.","exclusion":"If the primary objective is complex logical query answering, explicit rule reasoning, or first-order reasoning over a KG, assign to KG reasoning and query answering; if the output is extracted from raw text rather than inferred over an existing KG, assign to text-KG transduction.","nearest_boundary":"KG reasoning and query answering","representative_topics":["attention-based entity representation learning with semantic path context for knowledge graph relation prediction","knowledge graph embedding via relational prototype entities for global semantic similarity","negative sampling-free knowledge graph embedding via cross-correlation matrix contrastive learning"]},{"direction_index":5,"label":"D5","name":"KG reasoning and query answering","description":"Perform explicit reasoning, fact verification, complex query answering, or rule-guided inference over KGs using logical, neural-symbolic, or LLM-based reasoning mechanisms.","inclusion":"Include work primarily targeting multi-hop logical queries, negation, union, first-order logic operators, soft-rule grounding, fuzzy inference, explainable reasoning, fact verification, or LLM-driven reasoning over existing KGs.","exclusion":"If the work performs ordinary missing-link prediction without explicit logical/query structure, assign to KG representation and completion; if the work constructs the KG rather than reasons over it, assign to the relevant KG construction direction.","nearest_boundary":"KG representation and completion","representative_topics":["knowledge representation learning for logical query answering over multi-view knowledge graphs","probabilistic gamma embeddings for multi-hop logical query answering on knowledge graphs","knowledge graph completion via joint rule grounding and iterative embedding refinement"]},{"direction_index":6,"label":"D6","name":"Open-world KG construction","description":"Construct KGs where the main scientific objective is open, evolving, commonsense, schema-inducing, multimodal, temporal, causal, or dynamic representation rather than fixed-domain ontology instantiation.","inclusion":"Include commonsense KG construction and densification, open-schema KB construction, taxonomy or concept-hierarchy induction, sememe or synset construction, multimodal KG construction, dynamic or event-centric KG construction, and causal or temporal KG building.","exclusion":"If the work mainly instantiates a fixed domain ontology for a vertical application, assign to ontology-guided domain KG construction; if the work primarily extracts fixed-schema facts from text, assign to text-KG transduction.","nearest_boundary":"Ontology-guided domain KG construction","representative_topics":["informative negative commonsense statement generation using comparable concept seeds and closed-world pruning","densifying commonsense knowledge graphs via relation completion for multi-hop path coverage","automatic chinese open-domain knowledge graph with fine-grained hypernym-hyponym hierarchy and path disambiguation"]},{"direction_index":7,"label":"D7","name":"KG methodology and infrastructure","description":"Systematize, benchmark, survey, taxonomize, or provide reusable methodological infrastructure for KG construction, curation, alignment, embedding, querying, ontology learning, and evaluation.","inclusion":"Include surveys, taxonomies, benchmark datasets, gold standards, evaluation protocols, reusable libraries, systematic reviews, and methodological resources whose main contribution is organizing or enabling a KG research area rather than solving one construction or reasoning task directly.","exclusion":"If the paper primarily proposes an objective-specific construction, extraction, alignment, completion, reasoning, or curation method, assign to that objective-specific direction.","nearest_boundary":"KG curation and quality enrichment","representative_topics":["taxonomy and survey of healthcare knowledge graph construction methods and open issues","cross-lingual wordnet synset construction via translation ranking for low-resource languages","large-scale gold standard benchmarking for multi-source knowledge graph matching and alignment"]}],"split":"dev"}
{"episode_id":"F164_T2024-01","field_index":164,"field":"mental health nlp","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Depression and distress screening","description":"Text-based prediction, classification, or severity estimation for depression, depressive symptoms, stress, anxiety-distress, or closely paired affective distress states from online or dialogue text.","inclusion":"Include work whose primary objective is screening or estimating depression/distress status or severity from user-generated text, regardless of model architecture, supervision regime, feature engineering, or dataset construction for that predictive task.","exclusion":"If the primary endpoint is suicide or self-harm risk, assign to D3; if the primary objective is explaining or causally grounding the inference rather than detection, assign to D2; if multiple disorders are the central framing, assign to D1.","nearest_boundary":"D2","representative_topics":["recurrent convolutional networks with stacked embeddings for depression detection on sparse social media","multi-task hierarchical modeling for depression detection in sparse dialog","semi-supervised depression symptom detection from social media using iterative retraining and zero-shot learning"]},{"direction_index":1,"label":"D1","name":"Cross-disorder online detection","description":"General-purpose, multi-class, multi-label, or cross-condition mental-health classification from social media or other online text.","inclusion":"Include work framed around detecting more than one mental-health condition, comparing disorders, or building broad mental-illness classifiers from online text, including non-depression/non-suicide condition-specific online classifiers when part of a broader disorder-detection program.","exclusion":"If depression/distress alone is the target, assign to D0; if suicidality or self-harm is the target, assign to D3; if the data source and task are clinical-document or questionnaire workflow support, assign to D5.","nearest_boundary":"D0","representative_topics":["mental illness classification from social media text using deep and transfer learning models","emotion-annotated mental health corpus for condition detection from social media text","ontology-driven self-supervised concept embeddings for low-resource adverse childhood experience identification"]},{"direction_index":2,"label":"D2","name":"Explainable and causal inference","description":"Mental-health NLP whose primary research objective is interpretability, causal analysis, rationale production, clinical-marker grounding, or transparent inference.","inclusion":"Include work centered on causal categorization, causal explanation, interpretable psychiatric prediction, rationale-centered mental-health inference, symptom- or marker-grounded explanations, or interactive explanation of predictions.","exclusion":"If explanation is incidental to optimizing screening performance, assign to D0 or D1; if the central concern is bias, robustness, privacy, hallucination, stigma, or deployment risk rather than explanation of an inference, assign to D7.","nearest_boundary":"D0","representative_topics":["causal annotation and categorization of mental health issues in social media posts","explainable depression detection on twitter using hierarchical attention and metaphor concept mappings","explainable causal classification of mental health issues with inconsistency analysis on social media"]},{"direction_index":3,"label":"D3","name":"Suicide and self-harm risk","description":"Detection, assessment, corpus construction, or model evaluation for suicidal ideation, suicide intent, self-harm, or acute alarming-risk language.","inclusion":"Include work whose primary target is suicidality, self-harm, suicide-risk level, suicidal intent, or immediate self-harm/threat language from social media, clinical notes, student responses, or other text.","exclusion":"If the target is depression or general distress without suicidality as the explicit endpoint, assign to D0; if suicidality is only one extraction target inside a broader clinical-document workflow, assign to D5.","nearest_boundary":"D0","representative_topics":["deep learning model comparison for suicidal ideation detection in social media text","large-scale corpus and classification for nuanced suicidal ideation detection in online text","zero-shot learning for suicidality detection in sparse clinical notes via contextual relatedness"]},{"direction_index":4,"label":"D4","name":"Multimodal and non-text sensing","description":"Mental-health assessment using speech, audio, video, images, physiological signals, passive sensors, or explicit fusion of text with non-text modalities.","inclusion":"Include work where the central scientific contribution is non-text signal analysis or cross-modal fusion for detecting, estimating, or monitoring mental-health states, symptoms, emotion, alliance, or disorder severity.","exclusion":"If the work uses text-only social media or dialogue for disorder screening, assign to D0 or D1; if the work processes clinical transcripts or records without acoustic, visual, physiological, or sensor signals as central evidence, assign to D5.","nearest_boundary":"D5","representative_topics":["speech and language feature extraction for automated working alliance assessment in psychotherapy","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","bias mitigation in ecg-based mental health prediction via multitask learning and epistemic uncertainty"]},{"direction_index":5,"label":"D5","name":"Clinical documentation workflows","description":"NLP for clinical, counseling, questionnaire, therapy-transcript, or health-record workflows, including documentation, summarization, extraction, scoring, and diagnosis-support tasks.","inclusion":"Include work operating on clinical records, therapy or counseling transcripts, questionnaires, clinical interviews, or help-seeking texts when the goal is documentation support, information extraction, summarization, scoring, or clinical diagnostic workflow assistance.","exclusion":"If the system generates therapeutic responses or recommends interventions as an agent, assign to D6; if the primary evidence is non-text audio/video/sensor data, assign to D4; if the task is passive online disorder screening, assign to D0 or D1.","nearest_boundary":"D6","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","transformer-based working alliance scoring from psychotherapy dialogue using clinical inventory projection","curriculum-guided abstractive summarization of online mental health posts using reweighted training samples"]},{"direction_index":6,"label":"D6","name":"Counseling agents and interventions","description":"Systems that generate, guide, recommend, or evaluate therapeutic dialogue, counseling responses, psychological-support interactions, or intervention-oriented decision support.","inclusion":"Include work building, fine-tuning, surveying, or evaluating mental-health chatbots, counseling dialogue systems, empathetic response generators, response-act-guided systems, therapeutic strategy recommenders, reinforcement-learning intervention policies, or dialogue-data generation for counseling agents.","exclusion":"If the task summarizes or scores existing sessions without response generation or intervention recommendation, assign to D5; if the main objective is auditing safety, hallucination, bias, privacy, or deployment risk rather than agent/intervention construction, assign to D7.","nearest_boundary":"D5","representative_topics":["fine-tuned seq2seq transformer for empathetic person-centered therapy dialogue with post-processing","dialogue-act guided reinforcement learning for mental health counseling response generation","deep reinforcement learning for disorder-specific multi-objective psychotherapy topic recommendation with interpretable policy dynamics"]},{"direction_index":7,"label":"D7","name":"Responsible measurement and analytics","description":"NLP for descriptive measurement of mental-health language, therapeutic processes, communities, stigma, semantic change, annotation validity, and for auditing reliability, robustness, fairness, privacy, hallucination, or social harms of mental-health NLP systems.","inclusion":"Include work whose primary objective is understanding or measuring mental-health-related language, topics, emotions, populations, therapeutic behaviors, stigma, semantic shift, annotation validity, or deployment trustworthiness, including audits or mitigations of bias, privacy, robustness, safety, hallucination, and LLM reliability.","exclusion":"If the main output is a disorder classifier, assign to D0, D1, or D3; if the main output is an interpretable or causal prediction mechanism, assign to D2; if the main output is a clinical extraction or summarization workflow, assign to D5; if the main output is a counseling response generator or intervention agent, assign to D6.","nearest_boundary":"D2","representative_topics":["longitudinal validity and selection bias in self-disclosed mental health diagnosis annotations","detecting and mitigating semantic shift in longitudinal social media mental health analysis","unsupervised emotion and topic modeling for analyzing challenges in disability transition to adulthood"]}],"split":"dev"}
{"episode_id":"F164_T2024-07","field_index":164,"field":"mental health nlp","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Depression and distress screening","description":"Text-based prediction, classification, or severity estimation for depression, depressive symptoms, stress, anxiety-distress, or closely paired affective distress states from online or dialogue text.","inclusion":"Include work whose primary objective is screening or estimating depression/distress status or severity from user-generated text, regardless of model architecture, supervision regime, feature engineering, or dataset construction for that predictive task.","exclusion":"If the primary endpoint is suicide or self-harm risk, assign to D3; if the primary objective is explaining or causally grounding the inference rather than detection, assign to D2; if multiple disorders are the central framing, assign to D1.","nearest_boundary":"D2","representative_topics":["recurrent convolutional networks with stacked embeddings for depression detection on sparse social media","multi-task hierarchical modeling for depression detection in sparse dialog","semi-supervised depression symptom detection from social media using iterative retraining and zero-shot learning"]},{"direction_index":1,"label":"D1","name":"Cross-disorder online detection","description":"General-purpose, multi-class, multi-label, or cross-condition mental-health classification from social media or other online text.","inclusion":"Include work framed around detecting more than one mental-health condition, comparing disorders, or building broad mental-illness classifiers from online text, including non-depression/non-suicide condition-specific online classifiers when part of a broader disorder-detection program.","exclusion":"If depression/distress alone is the target, assign to D0; if suicidality or self-harm is the target, assign to D3; if the data source and task are clinical-document or questionnaire workflow support, assign to D5.","nearest_boundary":"D0","representative_topics":["mental illness classification from social media text using deep and transfer learning models","emotion-annotated mental health corpus for condition detection from social media text","ontology-driven self-supervised concept embeddings for low-resource adverse childhood experience identification"]},{"direction_index":2,"label":"D2","name":"Explainable and causal inference","description":"Mental-health NLP whose primary research objective is interpretability, causal analysis, rationale production, clinical-marker grounding, or transparent inference.","inclusion":"Include work centered on causal categorization, causal explanation, interpretable psychiatric prediction, rationale-centered mental-health inference, symptom- or marker-grounded explanations, or interactive explanation of predictions.","exclusion":"If explanation is incidental to optimizing screening performance, assign to D0 or D1; if the central concern is bias, robustness, privacy, hallucination, stigma, or deployment risk rather than explanation of an inference, assign to D7.","nearest_boundary":"D0","representative_topics":["causal annotation and categorization of mental health issues in social media posts","explainable depression detection on twitter using hierarchical attention and metaphor concept mappings","explainable causal classification of mental health issues with inconsistency analysis on social media"]},{"direction_index":3,"label":"D3","name":"Suicide and self-harm risk","description":"Detection, assessment, corpus construction, or model evaluation for suicidal ideation, suicide intent, self-harm, or acute alarming-risk language.","inclusion":"Include work whose primary target is suicidality, self-harm, suicide-risk level, suicidal intent, or immediate self-harm/threat language from social media, clinical notes, student responses, or other text.","exclusion":"If the target is depression or general distress without suicidality as the explicit endpoint, assign to D0; if suicidality is only one extraction target inside a broader clinical-document workflow, assign to D5.","nearest_boundary":"D0","representative_topics":["deep learning model comparison for suicidal ideation detection in social media text","large-scale corpus and classification for nuanced suicidal ideation detection in online text","zero-shot learning for suicidality detection in sparse clinical notes via contextual relatedness"]},{"direction_index":4,"label":"D4","name":"Multimodal and non-text sensing","description":"Mental-health assessment using speech, audio, video, images, physiological signals, passive sensors, or explicit fusion of text with non-text modalities.","inclusion":"Include work where the central scientific contribution is non-text signal analysis or cross-modal fusion for detecting, estimating, or monitoring mental-health states, symptoms, emotion, alliance, or disorder severity.","exclusion":"If the work uses text-only social media or dialogue for disorder screening, assign to D0 or D1; if the work processes clinical transcripts or records without acoustic, visual, physiological, or sensor signals as central evidence, assign to D5.","nearest_boundary":"D5","representative_topics":["speech and language feature extraction for automated working alliance assessment in psychotherapy","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","bias mitigation in ecg-based mental health prediction via multitask learning and epistemic uncertainty"]},{"direction_index":5,"label":"D5","name":"Clinical documentation workflows","description":"NLP for clinical, counseling, questionnaire, therapy-transcript, or health-record workflows, including documentation, summarization, extraction, scoring, and diagnosis-support tasks.","inclusion":"Include work operating on clinical records, therapy or counseling transcripts, questionnaires, clinical interviews, or help-seeking texts when the goal is documentation support, information extraction, summarization, scoring, or clinical diagnostic workflow assistance.","exclusion":"If the system generates therapeutic responses or recommends interventions as an agent, assign to D6; if the primary evidence is non-text audio/video/sensor data, assign to D4; if the task is passive online disorder screening, assign to D0 or D1.","nearest_boundary":"D6","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","transformer-based working alliance scoring from psychotherapy dialogue using clinical inventory projection","curriculum-guided abstractive summarization of online mental health posts using reweighted training samples"]},{"direction_index":6,"label":"D6","name":"Counseling agents and interventions","description":"Systems that generate, guide, recommend, or evaluate therapeutic dialogue, counseling responses, psychological-support interactions, or intervention-oriented decision support.","inclusion":"Include work building, fine-tuning, surveying, or evaluating mental-health chatbots, counseling dialogue systems, empathetic response generators, response-act-guided systems, therapeutic strategy recommenders, reinforcement-learning intervention policies, or dialogue-data generation for counseling agents.","exclusion":"If the task summarizes or scores existing sessions without response generation or intervention recommendation, assign to D5; if the main objective is auditing safety, hallucination, bias, privacy, or deployment risk rather than agent/intervention construction, assign to D7.","nearest_boundary":"D5","representative_topics":["fine-tuned seq2seq transformer for empathetic person-centered therapy dialogue with post-processing","dialogue-act guided reinforcement learning for mental health counseling response generation","deep reinforcement learning for disorder-specific multi-objective psychotherapy topic recommendation with interpretable policy dynamics"]},{"direction_index":7,"label":"D7","name":"Responsible measurement and analytics","description":"NLP for descriptive measurement of mental-health language, therapeutic processes, communities, stigma, semantic change, annotation validity, and for auditing reliability, robustness, fairness, privacy, hallucination, or social harms of mental-health NLP systems.","inclusion":"Include work whose primary objective is understanding or measuring mental-health-related language, topics, emotions, populations, therapeutic behaviors, stigma, semantic shift, annotation validity, or deployment trustworthiness, including audits or mitigations of bias, privacy, robustness, safety, hallucination, and LLM reliability.","exclusion":"If the main output is a disorder classifier, assign to D0, D1, or D3; if the main output is an interpretable or causal prediction mechanism, assign to D2; if the main output is a clinical extraction or summarization workflow, assign to D5; if the main output is a counseling response generator or intervention agent, assign to D6.","nearest_boundary":"D2","representative_topics":["longitudinal validity and selection bias in self-disclosed mental health diagnosis annotations","detecting and mitigating semantic shift in longitudinal social media mental health analysis","unsupervised emotion and topic modeling for analyzing challenges in disability transition to adulthood"]}],"split":"dev"}
{"episode_id":"F164_T2025-01","field_index":164,"field":"mental health nlp","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Depression and distress screening","description":"Text-based prediction, classification, or severity estimation for depression, depressive symptoms, stress, anxiety-distress, or closely paired affective distress states from online or dialogue text.","inclusion":"Include work whose primary objective is screening or estimating depression/distress status or severity from user-generated text, regardless of model architecture, supervision regime, feature engineering, or dataset construction for that predictive task.","exclusion":"If the primary endpoint is suicide or self-harm risk, assign to D3; if the primary objective is explaining or causally grounding the inference rather than detection, assign to D2; if multiple disorders are the central framing, assign to D1.","nearest_boundary":"D2","representative_topics":["recurrent convolutional networks with stacked embeddings for depression detection on sparse social media","multi-task hierarchical modeling for depression detection in sparse dialog","semi-supervised depression symptom detection from social media using iterative retraining and zero-shot learning"]},{"direction_index":1,"label":"D1","name":"Cross-disorder online detection","description":"General-purpose, multi-class, multi-label, or cross-condition mental-health classification from social media or other online text.","inclusion":"Include work framed around detecting more than one mental-health condition, comparing disorders, or building broad mental-illness classifiers from online text, including non-depression/non-suicide condition-specific online classifiers when part of a broader disorder-detection program.","exclusion":"If depression/distress alone is the target, assign to D0; if suicidality or self-harm is the target, assign to D3; if the data source and task are clinical-document or questionnaire workflow support, assign to D5.","nearest_boundary":"D0","representative_topics":["mental illness classification from social media text using deep and transfer learning models","emotion-annotated mental health corpus for condition detection from social media text","ontology-driven self-supervised concept embeddings for low-resource adverse childhood experience identification"]},{"direction_index":2,"label":"D2","name":"Explainable and causal inference","description":"Mental-health NLP whose primary research objective is interpretability, causal analysis, rationale production, clinical-marker grounding, or transparent inference.","inclusion":"Include work centered on causal categorization, causal explanation, interpretable psychiatric prediction, rationale-centered mental-health inference, symptom- or marker-grounded explanations, or interactive explanation of predictions.","exclusion":"If explanation is incidental to optimizing screening performance, assign to D0 or D1; if the central concern is bias, robustness, privacy, hallucination, stigma, or deployment risk rather than explanation of an inference, assign to D7.","nearest_boundary":"D0","representative_topics":["causal annotation and categorization of mental health issues in social media posts","explainable depression detection on twitter using hierarchical attention and metaphor concept mappings","explainable causal classification of mental health issues with inconsistency analysis on social media"]},{"direction_index":3,"label":"D3","name":"Suicide and self-harm risk","description":"Detection, assessment, corpus construction, or model evaluation for suicidal ideation, suicide intent, self-harm, or acute alarming-risk language.","inclusion":"Include work whose primary target is suicidality, self-harm, suicide-risk level, suicidal intent, or immediate self-harm/threat language from social media, clinical notes, student responses, or other text.","exclusion":"If the target is depression or general distress without suicidality as the explicit endpoint, assign to D0; if suicidality is only one extraction target inside a broader clinical-document workflow, assign to D5.","nearest_boundary":"D0","representative_topics":["deep learning model comparison for suicidal ideation detection in social media text","large-scale corpus and classification for nuanced suicidal ideation detection in online text","zero-shot learning for suicidality detection in sparse clinical notes via contextual relatedness"]},{"direction_index":4,"label":"D4","name":"Multimodal and non-text sensing","description":"Mental-health assessment using speech, audio, video, images, physiological signals, passive sensors, or explicit fusion of text with non-text modalities.","inclusion":"Include work where the central scientific contribution is non-text signal analysis or cross-modal fusion for detecting, estimating, or monitoring mental-health states, symptoms, emotion, alliance, or disorder severity.","exclusion":"If the work uses text-only social media or dialogue for disorder screening, assign to D0 or D1; if the work processes clinical transcripts or records without acoustic, visual, physiological, or sensor signals as central evidence, assign to D5.","nearest_boundary":"D5","representative_topics":["speech and language feature extraction for automated working alliance assessment in psychotherapy","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","bias mitigation in ecg-based mental health prediction via multitask learning and epistemic uncertainty"]},{"direction_index":5,"label":"D5","name":"Clinical documentation workflows","description":"NLP for clinical, counseling, questionnaire, therapy-transcript, or health-record workflows, including documentation, summarization, extraction, scoring, and diagnosis-support tasks.","inclusion":"Include work operating on clinical records, therapy or counseling transcripts, questionnaires, clinical interviews, or help-seeking texts when the goal is documentation support, information extraction, summarization, scoring, or clinical diagnostic workflow assistance.","exclusion":"If the system generates therapeutic responses or recommends interventions as an agent, assign to D6; if the primary evidence is non-text audio/video/sensor data, assign to D4; if the task is passive online disorder screening, assign to D0 or D1.","nearest_boundary":"D6","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","transformer-based working alliance scoring from psychotherapy dialogue using clinical inventory projection","curriculum-guided abstractive summarization of online mental health posts using reweighted training samples"]},{"direction_index":6,"label":"D6","name":"Counseling agents and interventions","description":"Systems that generate, guide, recommend, or evaluate therapeutic dialogue, counseling responses, psychological-support interactions, or intervention-oriented decision support.","inclusion":"Include work building, fine-tuning, surveying, or evaluating mental-health chatbots, counseling dialogue systems, empathetic response generators, response-act-guided systems, therapeutic strategy recommenders, reinforcement-learning intervention policies, or dialogue-data generation for counseling agents.","exclusion":"If the task summarizes or scores existing sessions without response generation or intervention recommendation, assign to D5; if the main objective is auditing safety, hallucination, bias, privacy, or deployment risk rather than agent/intervention construction, assign to D7.","nearest_boundary":"D5","representative_topics":["fine-tuned seq2seq transformer for empathetic person-centered therapy dialogue with post-processing","dialogue-act guided reinforcement learning for mental health counseling response generation","deep reinforcement learning for disorder-specific multi-objective psychotherapy topic recommendation with interpretable policy dynamics"]},{"direction_index":7,"label":"D7","name":"Responsible measurement and analytics","description":"NLP for descriptive measurement of mental-health language, therapeutic processes, communities, stigma, semantic change, annotation validity, and for auditing reliability, robustness, fairness, privacy, hallucination, or social harms of mental-health NLP systems.","inclusion":"Include work whose primary objective is understanding or measuring mental-health-related language, topics, emotions, populations, therapeutic behaviors, stigma, semantic shift, annotation validity, or deployment trustworthiness, including audits or mitigations of bias, privacy, robustness, safety, hallucination, and LLM reliability.","exclusion":"If the main output is a disorder classifier, assign to D0, D1, or D3; if the main output is an interpretable or causal prediction mechanism, assign to D2; if the main output is a clinical extraction or summarization workflow, assign to D5; if the main output is a counseling response generator or intervention agent, assign to D6.","nearest_boundary":"D2","representative_topics":["longitudinal validity and selection bias in self-disclosed mental health diagnosis annotations","detecting and mitigating semantic shift in longitudinal social media mental health analysis","unsupervised emotion and topic modeling for analyzing challenges in disability transition to adulthood"]}],"split":"dev"}
{"episode_id":"F164_T2025-07","field_index":164,"field":"mental health nlp","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Depression and distress screening","description":"Text-based prediction, classification, or severity estimation for depression, depressive symptoms, stress, anxiety-distress, or closely paired affective distress states from online or dialogue text.","inclusion":"Include work whose primary objective is screening or estimating depression/distress status or severity from user-generated text, regardless of model architecture, supervision regime, feature engineering, or dataset construction for that predictive task.","exclusion":"If the primary endpoint is suicide or self-harm risk, assign to D3; if the primary objective is explaining or causally grounding the inference rather than detection, assign to D2; if multiple disorders are the central framing, assign to D1.","nearest_boundary":"D2","representative_topics":["recurrent convolutional networks with stacked embeddings for depression detection on sparse social media","multi-task hierarchical modeling for depression detection in sparse dialog","semi-supervised depression symptom detection from social media using iterative retraining and zero-shot learning"]},{"direction_index":1,"label":"D1","name":"Cross-disorder online detection","description":"General-purpose, multi-class, multi-label, or cross-condition mental-health classification from social media or other online text.","inclusion":"Include work framed around detecting more than one mental-health condition, comparing disorders, or building broad mental-illness classifiers from online text, including non-depression/non-suicide condition-specific online classifiers when part of a broader disorder-detection program.","exclusion":"If depression/distress alone is the target, assign to D0; if suicidality or self-harm is the target, assign to D3; if the data source and task are clinical-document or questionnaire workflow support, assign to D5.","nearest_boundary":"D0","representative_topics":["mental illness classification from social media text using deep and transfer learning models","emotion-annotated mental health corpus for condition detection from social media text","ontology-driven self-supervised concept embeddings for low-resource adverse childhood experience identification"]},{"direction_index":2,"label":"D2","name":"Explainable and causal inference","description":"Mental-health NLP whose primary research objective is interpretability, causal analysis, rationale production, clinical-marker grounding, or transparent inference.","inclusion":"Include work centered on causal categorization, causal explanation, interpretable psychiatric prediction, rationale-centered mental-health inference, symptom- or marker-grounded explanations, or interactive explanation of predictions.","exclusion":"If explanation is incidental to optimizing screening performance, assign to D0 or D1; if the central concern is bias, robustness, privacy, hallucination, stigma, or deployment risk rather than explanation of an inference, assign to D7.","nearest_boundary":"D0","representative_topics":["causal annotation and categorization of mental health issues in social media posts","explainable depression detection on twitter using hierarchical attention and metaphor concept mappings","explainable causal classification of mental health issues with inconsistency analysis on social media"]},{"direction_index":3,"label":"D3","name":"Suicide and self-harm risk","description":"Detection, assessment, corpus construction, or model evaluation for suicidal ideation, suicide intent, self-harm, or acute alarming-risk language.","inclusion":"Include work whose primary target is suicidality, self-harm, suicide-risk level, suicidal intent, or immediate self-harm/threat language from social media, clinical notes, student responses, or other text.","exclusion":"If the target is depression or general distress without suicidality as the explicit endpoint, assign to D0; if suicidality is only one extraction target inside a broader clinical-document workflow, assign to D5.","nearest_boundary":"D0","representative_topics":["deep learning model comparison for suicidal ideation detection in social media text","large-scale corpus and classification for nuanced suicidal ideation detection in online text","zero-shot learning for suicidality detection in sparse clinical notes via contextual relatedness"]},{"direction_index":4,"label":"D4","name":"Multimodal and non-text sensing","description":"Mental-health assessment using speech, audio, video, images, physiological signals, passive sensors, or explicit fusion of text with non-text modalities.","inclusion":"Include work where the central scientific contribution is non-text signal analysis or cross-modal fusion for detecting, estimating, or monitoring mental-health states, symptoms, emotion, alliance, or disorder severity.","exclusion":"If the work uses text-only social media or dialogue for disorder screening, assign to D0 or D1; if the work processes clinical transcripts or records without acoustic, visual, physiological, or sensor signals as central evidence, assign to D5.","nearest_boundary":"D5","representative_topics":["speech and language feature extraction for automated working alliance assessment in psychotherapy","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","bias mitigation in ecg-based mental health prediction via multitask learning and epistemic uncertainty"]},{"direction_index":5,"label":"D5","name":"Clinical documentation workflows","description":"NLP for clinical, counseling, questionnaire, therapy-transcript, or health-record workflows, including documentation, summarization, extraction, scoring, and diagnosis-support tasks.","inclusion":"Include work operating on clinical records, therapy or counseling transcripts, questionnaires, clinical interviews, or help-seeking texts when the goal is documentation support, information extraction, summarization, scoring, or clinical diagnostic workflow assistance.","exclusion":"If the system generates therapeutic responses or recommends interventions as an agent, assign to D6; if the primary evidence is non-text audio/video/sensor data, assign to D4; if the task is passive online disorder screening, assign to D0 or D1.","nearest_boundary":"D6","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","transformer-based working alliance scoring from psychotherapy dialogue using clinical inventory projection","curriculum-guided abstractive summarization of online mental health posts using reweighted training samples"]},{"direction_index":6,"label":"D6","name":"Counseling agents and interventions","description":"Systems that generate, guide, recommend, or evaluate therapeutic dialogue, counseling responses, psychological-support interactions, or intervention-oriented decision support.","inclusion":"Include work building, fine-tuning, surveying, or evaluating mental-health chatbots, counseling dialogue systems, empathetic response generators, response-act-guided systems, therapeutic strategy recommenders, reinforcement-learning intervention policies, or dialogue-data generation for counseling agents.","exclusion":"If the task summarizes or scores existing sessions without response generation or intervention recommendation, assign to D5; if the main objective is auditing safety, hallucination, bias, privacy, or deployment risk rather than agent/intervention construction, assign to D7.","nearest_boundary":"D5","representative_topics":["fine-tuned seq2seq transformer for empathetic person-centered therapy dialogue with post-processing","dialogue-act guided reinforcement learning for mental health counseling response generation","deep reinforcement learning for disorder-specific multi-objective psychotherapy topic recommendation with interpretable policy dynamics"]},{"direction_index":7,"label":"D7","name":"Responsible measurement and analytics","description":"NLP for descriptive measurement of mental-health language, therapeutic processes, communities, stigma, semantic change, annotation validity, and for auditing reliability, robustness, fairness, privacy, hallucination, or social harms of mental-health NLP systems.","inclusion":"Include work whose primary objective is understanding or measuring mental-health-related language, topics, emotions, populations, therapeutic behaviors, stigma, semantic shift, annotation validity, or deployment trustworthiness, including audits or mitigations of bias, privacy, robustness, safety, hallucination, and LLM reliability.","exclusion":"If the main output is a disorder classifier, assign to D0, D1, or D3; if the main output is an interpretable or causal prediction mechanism, assign to D2; if the main output is a clinical extraction or summarization workflow, assign to D5; if the main output is a counseling response generator or intervention agent, assign to D6.","nearest_boundary":"D2","representative_topics":["longitudinal validity and selection bias in self-disclosed mental health diagnosis annotations","detecting and mitigating semantic shift in longitudinal social media mental health analysis","unsupervised emotion and topic modeling for analyzing challenges in disability transition to adulthood"]}],"split":"dev"}
{"episode_id":"F164_T2026-01","field_index":164,"field":"mental health nlp","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Depression and distress screening","description":"Text-based prediction, classification, or severity estimation for depression, depressive symptoms, stress, anxiety-distress, or closely paired affective distress states from online or dialogue text.","inclusion":"Include work whose primary objective is screening or estimating depression/distress status or severity from user-generated text, regardless of model architecture, supervision regime, feature engineering, or dataset construction for that predictive task.","exclusion":"If the primary endpoint is suicide or self-harm risk, assign to D3; if the primary objective is explaining or causally grounding the inference rather than detection, assign to D2; if multiple disorders are the central framing, assign to D1.","nearest_boundary":"D2","representative_topics":["recurrent convolutional networks with stacked embeddings for depression detection on sparse social media","multi-task hierarchical modeling for depression detection in sparse dialog","semi-supervised depression symptom detection from social media using iterative retraining and zero-shot learning"]},{"direction_index":1,"label":"D1","name":"Cross-disorder online detection","description":"General-purpose, multi-class, multi-label, or cross-condition mental-health classification from social media or other online text.","inclusion":"Include work framed around detecting more than one mental-health condition, comparing disorders, or building broad mental-illness classifiers from online text, including non-depression/non-suicide condition-specific online classifiers when part of a broader disorder-detection program.","exclusion":"If depression/distress alone is the target, assign to D0; if suicidality or self-harm is the target, assign to D3; if the data source and task are clinical-document or questionnaire workflow support, assign to D5.","nearest_boundary":"D0","representative_topics":["mental illness classification from social media text using deep and transfer learning models","emotion-annotated mental health corpus for condition detection from social media text","ontology-driven self-supervised concept embeddings for low-resource adverse childhood experience identification"]},{"direction_index":2,"label":"D2","name":"Explainable and causal inference","description":"Mental-health NLP whose primary research objective is interpretability, causal analysis, rationale production, clinical-marker grounding, or transparent inference.","inclusion":"Include work centered on causal categorization, causal explanation, interpretable psychiatric prediction, rationale-centered mental-health inference, symptom- or marker-grounded explanations, or interactive explanation of predictions.","exclusion":"If explanation is incidental to optimizing screening performance, assign to D0 or D1; if the central concern is bias, robustness, privacy, hallucination, stigma, or deployment risk rather than explanation of an inference, assign to D7.","nearest_boundary":"D0","representative_topics":["causal annotation and categorization of mental health issues in social media posts","explainable depression detection on twitter using hierarchical attention and metaphor concept mappings","explainable causal classification of mental health issues with inconsistency analysis on social media"]},{"direction_index":3,"label":"D3","name":"Suicide and self-harm risk","description":"Detection, assessment, corpus construction, or model evaluation for suicidal ideation, suicide intent, self-harm, or acute alarming-risk language.","inclusion":"Include work whose primary target is suicidality, self-harm, suicide-risk level, suicidal intent, or immediate self-harm/threat language from social media, clinical notes, student responses, or other text.","exclusion":"If the target is depression or general distress without suicidality as the explicit endpoint, assign to D0; if suicidality is only one extraction target inside a broader clinical-document workflow, assign to D5.","nearest_boundary":"D0","representative_topics":["deep learning model comparison for suicidal ideation detection in social media text","large-scale corpus and classification for nuanced suicidal ideation detection in online text","zero-shot learning for suicidality detection in sparse clinical notes via contextual relatedness"]},{"direction_index":4,"label":"D4","name":"Multimodal and non-text sensing","description":"Mental-health assessment using speech, audio, video, images, physiological signals, passive sensors, or explicit fusion of text with non-text modalities.","inclusion":"Include work where the central scientific contribution is non-text signal analysis or cross-modal fusion for detecting, estimating, or monitoring mental-health states, symptoms, emotion, alliance, or disorder severity.","exclusion":"If the work uses text-only social media or dialogue for disorder screening, assign to D0 or D1; if the work processes clinical transcripts or records without acoustic, visual, physiological, or sensor signals as central evidence, assign to D5.","nearest_boundary":"D5","representative_topics":["speech and language feature extraction for automated working alliance assessment in psychotherapy","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","bias mitigation in ecg-based mental health prediction via multitask learning and epistemic uncertainty"]},{"direction_index":5,"label":"D5","name":"Clinical documentation workflows","description":"NLP for clinical, counseling, questionnaire, therapy-transcript, or health-record workflows, including documentation, summarization, extraction, scoring, and diagnosis-support tasks.","inclusion":"Include work operating on clinical records, therapy or counseling transcripts, questionnaires, clinical interviews, or help-seeking texts when the goal is documentation support, information extraction, summarization, scoring, or clinical diagnostic workflow assistance.","exclusion":"If the system generates therapeutic responses or recommends interventions as an agent, assign to D6; if the primary evidence is non-text audio/video/sensor data, assign to D4; if the task is passive online disorder screening, assign to D0 or D1.","nearest_boundary":"D6","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","transformer-based working alliance scoring from psychotherapy dialogue using clinical inventory projection","curriculum-guided abstractive summarization of online mental health posts using reweighted training samples"]},{"direction_index":6,"label":"D6","name":"Counseling agents and interventions","description":"Systems that generate, guide, recommend, or evaluate therapeutic dialogue, counseling responses, psychological-support interactions, or intervention-oriented decision support.","inclusion":"Include work building, fine-tuning, surveying, or evaluating mental-health chatbots, counseling dialogue systems, empathetic response generators, response-act-guided systems, therapeutic strategy recommenders, reinforcement-learning intervention policies, or dialogue-data generation for counseling agents.","exclusion":"If the task summarizes or scores existing sessions without response generation or intervention recommendation, assign to D5; if the main objective is auditing safety, hallucination, bias, privacy, or deployment risk rather than agent/intervention construction, assign to D7.","nearest_boundary":"D5","representative_topics":["fine-tuned seq2seq transformer for empathetic person-centered therapy dialogue with post-processing","dialogue-act guided reinforcement learning for mental health counseling response generation","deep reinforcement learning for disorder-specific multi-objective psychotherapy topic recommendation with interpretable policy dynamics"]},{"direction_index":7,"label":"D7","name":"Responsible measurement and analytics","description":"NLP for descriptive measurement of mental-health language, therapeutic processes, communities, stigma, semantic change, annotation validity, and for auditing reliability, robustness, fairness, privacy, hallucination, or social harms of mental-health NLP systems.","inclusion":"Include work whose primary objective is understanding or measuring mental-health-related language, topics, emotions, populations, therapeutic behaviors, stigma, semantic shift, annotation validity, or deployment trustworthiness, including audits or mitigations of bias, privacy, robustness, safety, hallucination, and LLM reliability.","exclusion":"If the main output is a disorder classifier, assign to D0, D1, or D3; if the main output is an interpretable or causal prediction mechanism, assign to D2; if the main output is a clinical extraction or summarization workflow, assign to D5; if the main output is a counseling response generator or intervention agent, assign to D6.","nearest_boundary":"D2","representative_topics":["longitudinal validity and selection bias in self-disclosed mental health diagnosis annotations","detecting and mitigating semantic shift in longitudinal social media mental health analysis","unsupervised emotion and topic modeling for analyzing challenges in disability transition to adulthood"]}],"split":"dev"}
{"episode_id":"F170_T2024-01","field_index":170,"field":"knowledge graph embedding","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Generalizable and temporal link reasoning","description":"Predicting links or facts beyond a fixed static transductive graph, including unseen entities, unseen relations, emerging subgraphs, evolving facts, temporal validity, and future-event extrapolation.","inclusion":"Include work whose primary objective is generalization across entity/relation sets, graph batches, disconnected emerging graphs, or time-indexed fact histories.","exclusion":"Standard missing-triple prediction on a fixed graph belongs to core completion modeling; methods centered on geometric space design or language fusion belong to those respective programs unless unseen/time-varying prediction is the main target.","nearest_boundary":"Core completion modeling and training","representative_topics":["unifying factorisation-based and message-passing models for inductive knowledge graph completion","inductive link prediction via mutual information maximization on complete subgraph neighboring relations","inductive knowledge graph reasoning for continuously emerging multi-batch entities via walk-based aggregation"]},{"direction_index":1,"label":"D1","name":"Language-KG semantic integration","description":"Coupling knowledge graphs with pretrained or large language models for KG completion, factual grounding, graph-aware generation, and KG-guided language-model reasoning.","inclusion":"Include work where the primary objective is to transfer semantic evidence between text/language models and KGs, either to improve KG completion with textual semantics or to improve language-model factuality and reasoning with KG evidence.","exclusion":"Building or populating a KG from text or model internals belongs to KG construction and ontology engineering; complex formal query answering belongs to logical and explainable reasoning.","nearest_boundary":"KG construction and ontology engineering","representative_topics":["knowledge graph semantics compression and reconstruction via transformer language models","sparse knowledge graph completion via variational em fusion of text and structure knowledge","hypercomplex space integration of knowledge graph and language model embeddings for link prediction"]},{"direction_index":2,"label":"D2","name":"KG construction and ontology engineering","description":"Creating, populating, organizing, and semantically grounding knowledge graphs, schemas, ontologies, and construction pipelines from text, language models, or heterogeneous domain data.","inclusion":"Include work whose primary objective is KG construction, entity/relation extraction for graph population, ontology design or embedding, schema and quality management, or LLM-assisted KG generation.","exclusion":"Using existing textual or language-model representations to score missing links belongs to language-KG semantic integration; matching equivalent entities across existing graphs belongs to heterogeneous KG representation and alignment.","nearest_boundary":"Language-KG semantic integration","representative_topics":["harvesting symbolic knowledge graphs of arbitrary relations via few-shot prompting of pretrained language models","incremental knowledge graph construction pipelines integrating ontology and quality management","llm-based joint entity and relation extraction for knowledge graph construction from unstructured text"]},{"direction_index":3,"label":"D3","name":"Geometric relational representation","description":"Designing embedding spaces and relation operators whose main purpose is to encode relational structure through geometry, including translation, rotation, scaling, curvature, hyperbolic or ultrahyperbolic spaces, and interacting geometric views.","inclusion":"Include work where the primary scientific contribution is the choice, derivation, or combination of geometric spaces or transformations for entity-relation representation.","exclusion":"Neural, convolutional, attention, transformer, or distillation architectures belong to core completion modeling unless the geometric space or operator is the central objective; temporal geometric models belong to generalizable and temporal link reasoning when temporal evolution is primary.","nearest_boundary":"Core completion modeling and training","representative_topics":["ultrahyperbolic geometry embeddings for heterogeneous knowledge graph topological structures","spatio-functional knowledge graph embedding for joint composition and hierarchy inference patterns","multi-geometry interaction embeddings for knowledge graph completion with coexisting spatial structures"]},{"direction_index":4,"label":"D4","name":"Core completion modeling and training","description":"Improving standard transductive KG completion through scoring architectures, neural interaction mechanisms, regularization, self-distillation, contrastive objectives, negative sampling, pruning, and scalable training strategies.","inclusion":"Include work whose primary objective is better or cheaper missing-triple prediction on a largely fixed graph through model architecture, training objective, sampling, compression, hyperparameter optimization, or low-dimensional learning.","exclusion":"Methods primarily about unseen entities, temporal evolution, explicit geometry, language fusion, multimodal or hyper-relational structure, privacy, robustness, or benchmark synthesis belong to the corresponding neighboring programs.","nearest_boundary":"Geometric relational representation","representative_topics":["confidence-aware self-distillation for low-dimensional knowledge graph embedding","attention-based entity representation learning with semantic path context for knowledge graph relation prediction","equivariance-based regularization for generalizing knowledge graph completion models"]},{"direction_index":5,"label":"D5","name":"Operational evaluation and trustworthy KGE","description":"Assessing, benchmarking, explaining at the field level, scaling, securing, privatizing, and auditing KGE methods and infrastructure for reliable use.","inclusion":"Include work whose primary objective is evaluation methodology, benchmark or library construction, survey synthesis, hyperparameter sensitivity analysis, adversarial robustness, bias mitigation, privacy-preserving or federated KGE, or operational reliability rather than a new task-specific predictor.","exclusion":"A new completion architecture with evaluation as a secondary contribution belongs to core completion modeling; post-hoc explanation of individual predictions belongs to logical and explainable reasoning.","nearest_boundary":"Core completion modeling and training","representative_topics":["hyperparameter sensitivity analysis for knowledge graph embedding quality assessment","benchmarking knowledge graph embeddings for description logic class representation learning","unified library for multi-source knowledge graph embedding across link prediction, entity alignment, and question answering"]},{"direction_index":6,"label":"D6","name":"Logical and explainable KG reasoning","description":"Reasoning over KGs with explicit logical structure, rules, symbolic constraints, complex query forms, prototypes, interpretable mechanisms, or post-hoc explanations of embedding behavior.","inclusion":"Include work whose primary objective is rule integration, soft-rule modeling, complex logical query answering, multi-view logical reasoning, interpretable rule- or instance-based completion, influential-example explanation, rule-mining diagnostics, or semantic substructure explanation.","exclusion":"Plain one-hop completion without explicit logical or explanatory objectives belongs to core completion modeling; KG-guided LLM factual grounding without formal query structure belongs to language-KG semantic integration.","nearest_boundary":"Language-KG semantic integration","representative_topics":["transformer-based masked pre-training for complex logical query reasoning over knowledge graphs","knowledge representation learning for logical query answering over multi-view knowledge graphs","knowledge graph reasoning via joint rule and entity-relation embedding with soft logical inference"]},{"direction_index":7,"label":"D7","name":"Heterogeneous KG representation and alignment","description":"Representing, matching, and reasoning over KGs with heterogeneity beyond ordinary single-graph binary triples, including cross-graph entity alignment, multimodal evidence, qualifiers, hyper-relations, n-ary facts, numeric literals, and bi-level fact structures.","inclusion":"Include work whose primary objective is entity matching across KGs, alignment under heterogeneity or growth, multimodal fusion, missing-modality handling, multimodal reasoning, hyper-relational or qualified fact modeling, or specialized representations for non-binary fact structures.","exclusion":"Text-only language-model fusion for ordinary completion belongs to language-KG semantic integration; ontology-level construction without entity matching belongs to KG construction and ontology engineering; ordinary fixed-graph triple scoring belongs to core completion modeling.","nearest_boundary":"KG construction and ontology engineering","representative_topics":["dual-view hyper-relational knowledge graph embedding for link prediction and entity typing","continual entity alignment for growing knowledge graphs via inductive neighbor-based representation learning","scalable knowledge graph entity alignment via subgraph partitioning, reconstruction, and embedding merging"]}],"split":"dev"}
{"episode_id":"F170_T2024-07","field_index":170,"field":"knowledge graph embedding","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Generalizable and temporal link reasoning","description":"Predicting links or facts beyond a fixed static transductive graph, including unseen entities, unseen relations, emerging subgraphs, evolving facts, temporal validity, and future-event extrapolation.","inclusion":"Include work whose primary objective is generalization across entity/relation sets, graph batches, disconnected emerging graphs, or time-indexed fact histories.","exclusion":"Standard missing-triple prediction on a fixed graph belongs to core completion modeling; methods centered on geometric space design or language fusion belong to those respective programs unless unseen/time-varying prediction is the main target.","nearest_boundary":"Core completion modeling and training","representative_topics":["unifying factorisation-based and message-passing models for inductive knowledge graph completion","inductive link prediction via mutual information maximization on complete subgraph neighboring relations","inductive knowledge graph reasoning for continuously emerging multi-batch entities via walk-based aggregation"]},{"direction_index":1,"label":"D1","name":"Language-KG semantic integration","description":"Coupling knowledge graphs with pretrained or large language models for KG completion, factual grounding, graph-aware generation, and KG-guided language-model reasoning.","inclusion":"Include work where the primary objective is to transfer semantic evidence between text/language models and KGs, either to improve KG completion with textual semantics or to improve language-model factuality and reasoning with KG evidence.","exclusion":"Building or populating a KG from text or model internals belongs to KG construction and ontology engineering; complex formal query answering belongs to logical and explainable reasoning.","nearest_boundary":"KG construction and ontology engineering","representative_topics":["knowledge graph semantics compression and reconstruction via transformer language models","sparse knowledge graph completion via variational em fusion of text and structure knowledge","hypercomplex space integration of knowledge graph and language model embeddings for link prediction"]},{"direction_index":2,"label":"D2","name":"KG construction and ontology engineering","description":"Creating, populating, organizing, and semantically grounding knowledge graphs, schemas, ontologies, and construction pipelines from text, language models, or heterogeneous domain data.","inclusion":"Include work whose primary objective is KG construction, entity/relation extraction for graph population, ontology design or embedding, schema and quality management, or LLM-assisted KG generation.","exclusion":"Using existing textual or language-model representations to score missing links belongs to language-KG semantic integration; matching equivalent entities across existing graphs belongs to heterogeneous KG representation and alignment.","nearest_boundary":"Language-KG semantic integration","representative_topics":["harvesting symbolic knowledge graphs of arbitrary relations via few-shot prompting of pretrained language models","incremental knowledge graph construction pipelines integrating ontology and quality management","llm-based joint entity and relation extraction for knowledge graph construction from unstructured text"]},{"direction_index":3,"label":"D3","name":"Geometric relational representation","description":"Designing embedding spaces and relation operators whose main purpose is to encode relational structure through geometry, including translation, rotation, scaling, curvature, hyperbolic or ultrahyperbolic spaces, and interacting geometric views.","inclusion":"Include work where the primary scientific contribution is the choice, derivation, or combination of geometric spaces or transformations for entity-relation representation.","exclusion":"Neural, convolutional, attention, transformer, or distillation architectures belong to core completion modeling unless the geometric space or operator is the central objective; temporal geometric models belong to generalizable and temporal link reasoning when temporal evolution is primary.","nearest_boundary":"Core completion modeling and training","representative_topics":["ultrahyperbolic geometry embeddings for heterogeneous knowledge graph topological structures","spatio-functional knowledge graph embedding for joint composition and hierarchy inference patterns","multi-geometry interaction embeddings for knowledge graph completion with coexisting spatial structures"]},{"direction_index":4,"label":"D4","name":"Core completion modeling and training","description":"Improving standard transductive KG completion through scoring architectures, neural interaction mechanisms, regularization, self-distillation, contrastive objectives, negative sampling, pruning, and scalable training strategies.","inclusion":"Include work whose primary objective is better or cheaper missing-triple prediction on a largely fixed graph through model architecture, training objective, sampling, compression, hyperparameter optimization, or low-dimensional learning.","exclusion":"Methods primarily about unseen entities, temporal evolution, explicit geometry, language fusion, multimodal or hyper-relational structure, privacy, robustness, or benchmark synthesis belong to the corresponding neighboring programs.","nearest_boundary":"Geometric relational representation","representative_topics":["confidence-aware self-distillation for low-dimensional knowledge graph embedding","attention-based entity representation learning with semantic path context for knowledge graph relation prediction","equivariance-based regularization for generalizing knowledge graph completion models"]},{"direction_index":5,"label":"D5","name":"Operational evaluation and trustworthy KGE","description":"Assessing, benchmarking, explaining at the field level, scaling, securing, privatizing, and auditing KGE methods and infrastructure for reliable use.","inclusion":"Include work whose primary objective is evaluation methodology, benchmark or library construction, survey synthesis, hyperparameter sensitivity analysis, adversarial robustness, bias mitigation, privacy-preserving or federated KGE, or operational reliability rather than a new task-specific predictor.","exclusion":"A new completion architecture with evaluation as a secondary contribution belongs to core completion modeling; post-hoc explanation of individual predictions belongs to logical and explainable reasoning.","nearest_boundary":"Core completion modeling and training","representative_topics":["hyperparameter sensitivity analysis for knowledge graph embedding quality assessment","benchmarking knowledge graph embeddings for description logic class representation learning","unified library for multi-source knowledge graph embedding across link prediction, entity alignment, and question answering"]},{"direction_index":6,"label":"D6","name":"Logical and explainable KG reasoning","description":"Reasoning over KGs with explicit logical structure, rules, symbolic constraints, complex query forms, prototypes, interpretable mechanisms, or post-hoc explanations of embedding behavior.","inclusion":"Include work whose primary objective is rule integration, soft-rule modeling, complex logical query answering, multi-view logical reasoning, interpretable rule- or instance-based completion, influential-example explanation, rule-mining diagnostics, or semantic substructure explanation.","exclusion":"Plain one-hop completion without explicit logical or explanatory objectives belongs to core completion modeling; KG-guided LLM factual grounding without formal query structure belongs to language-KG semantic integration.","nearest_boundary":"Language-KG semantic integration","representative_topics":["transformer-based masked pre-training for complex logical query reasoning over knowledge graphs","knowledge representation learning for logical query answering over multi-view knowledge graphs","knowledge graph reasoning via joint rule and entity-relation embedding with soft logical inference"]},{"direction_index":7,"label":"D7","name":"Heterogeneous KG representation and alignment","description":"Representing, matching, and reasoning over KGs with heterogeneity beyond ordinary single-graph binary triples, including cross-graph entity alignment, multimodal evidence, qualifiers, hyper-relations, n-ary facts, numeric literals, and bi-level fact structures.","inclusion":"Include work whose primary objective is entity matching across KGs, alignment under heterogeneity or growth, multimodal fusion, missing-modality handling, multimodal reasoning, hyper-relational or qualified fact modeling, or specialized representations for non-binary fact structures.","exclusion":"Text-only language-model fusion for ordinary completion belongs to language-KG semantic integration; ontology-level construction without entity matching belongs to KG construction and ontology engineering; ordinary fixed-graph triple scoring belongs to core completion modeling.","nearest_boundary":"KG construction and ontology engineering","representative_topics":["dual-view hyper-relational knowledge graph embedding for link prediction and entity typing","continual entity alignment for growing knowledge graphs via inductive neighbor-based representation learning","scalable knowledge graph entity alignment via subgraph partitioning, reconstruction, and embedding merging"]}],"split":"dev"}
{"episode_id":"F170_T2025-01","field_index":170,"field":"knowledge graph embedding","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Generalizable and temporal link reasoning","description":"Predicting links or facts beyond a fixed static transductive graph, including unseen entities, unseen relations, emerging subgraphs, evolving facts, temporal validity, and future-event extrapolation.","inclusion":"Include work whose primary objective is generalization across entity/relation sets, graph batches, disconnected emerging graphs, or time-indexed fact histories.","exclusion":"Standard missing-triple prediction on a fixed graph belongs to core completion modeling; methods centered on geometric space design or language fusion belong to those respective programs unless unseen/time-varying prediction is the main target.","nearest_boundary":"Core completion modeling and training","representative_topics":["unifying factorisation-based and message-passing models for inductive knowledge graph completion","inductive link prediction via mutual information maximization on complete subgraph neighboring relations","inductive knowledge graph reasoning for continuously emerging multi-batch entities via walk-based aggregation"]},{"direction_index":1,"label":"D1","name":"Language-KG semantic integration","description":"Coupling knowledge graphs with pretrained or large language models for KG completion, factual grounding, graph-aware generation, and KG-guided language-model reasoning.","inclusion":"Include work where the primary objective is to transfer semantic evidence between text/language models and KGs, either to improve KG completion with textual semantics or to improve language-model factuality and reasoning with KG evidence.","exclusion":"Building or populating a KG from text or model internals belongs to KG construction and ontology engineering; complex formal query answering belongs to logical and explainable reasoning.","nearest_boundary":"KG construction and ontology engineering","representative_topics":["knowledge graph semantics compression and reconstruction via transformer language models","sparse knowledge graph completion via variational em fusion of text and structure knowledge","hypercomplex space integration of knowledge graph and language model embeddings for link prediction"]},{"direction_index":2,"label":"D2","name":"KG construction and ontology engineering","description":"Creating, populating, organizing, and semantically grounding knowledge graphs, schemas, ontologies, and construction pipelines from text, language models, or heterogeneous domain data.","inclusion":"Include work whose primary objective is KG construction, entity/relation extraction for graph population, ontology design or embedding, schema and quality management, or LLM-assisted KG generation.","exclusion":"Using existing textual or language-model representations to score missing links belongs to language-KG semantic integration; matching equivalent entities across existing graphs belongs to heterogeneous KG representation and alignment.","nearest_boundary":"Language-KG semantic integration","representative_topics":["harvesting symbolic knowledge graphs of arbitrary relations via few-shot prompting of pretrained language models","incremental knowledge graph construction pipelines integrating ontology and quality management","llm-based joint entity and relation extraction for knowledge graph construction from unstructured text"]},{"direction_index":3,"label":"D3","name":"Geometric relational representation","description":"Designing embedding spaces and relation operators whose main purpose is to encode relational structure through geometry, including translation, rotation, scaling, curvature, hyperbolic or ultrahyperbolic spaces, and interacting geometric views.","inclusion":"Include work where the primary scientific contribution is the choice, derivation, or combination of geometric spaces or transformations for entity-relation representation.","exclusion":"Neural, convolutional, attention, transformer, or distillation architectures belong to core completion modeling unless the geometric space or operator is the central objective; temporal geometric models belong to generalizable and temporal link reasoning when temporal evolution is primary.","nearest_boundary":"Core completion modeling and training","representative_topics":["ultrahyperbolic geometry embeddings for heterogeneous knowledge graph topological structures","spatio-functional knowledge graph embedding for joint composition and hierarchy inference patterns","multi-geometry interaction embeddings for knowledge graph completion with coexisting spatial structures"]},{"direction_index":4,"label":"D4","name":"Core completion modeling and training","description":"Improving standard transductive KG completion through scoring architectures, neural interaction mechanisms, regularization, self-distillation, contrastive objectives, negative sampling, pruning, and scalable training strategies.","inclusion":"Include work whose primary objective is better or cheaper missing-triple prediction on a largely fixed graph through model architecture, training objective, sampling, compression, hyperparameter optimization, or low-dimensional learning.","exclusion":"Methods primarily about unseen entities, temporal evolution, explicit geometry, language fusion, multimodal or hyper-relational structure, privacy, robustness, or benchmark synthesis belong to the corresponding neighboring programs.","nearest_boundary":"Geometric relational representation","representative_topics":["confidence-aware self-distillation for low-dimensional knowledge graph embedding","attention-based entity representation learning with semantic path context for knowledge graph relation prediction","equivariance-based regularization for generalizing knowledge graph completion models"]},{"direction_index":5,"label":"D5","name":"Operational evaluation and trustworthy KGE","description":"Assessing, benchmarking, explaining at the field level, scaling, securing, privatizing, and auditing KGE methods and infrastructure for reliable use.","inclusion":"Include work whose primary objective is evaluation methodology, benchmark or library construction, survey synthesis, hyperparameter sensitivity analysis, adversarial robustness, bias mitigation, privacy-preserving or federated KGE, or operational reliability rather than a new task-specific predictor.","exclusion":"A new completion architecture with evaluation as a secondary contribution belongs to core completion modeling; post-hoc explanation of individual predictions belongs to logical and explainable reasoning.","nearest_boundary":"Core completion modeling and training","representative_topics":["hyperparameter sensitivity analysis for knowledge graph embedding quality assessment","benchmarking knowledge graph embeddings for description logic class representation learning","unified library for multi-source knowledge graph embedding across link prediction, entity alignment, and question answering"]},{"direction_index":6,"label":"D6","name":"Logical and explainable KG reasoning","description":"Reasoning over KGs with explicit logical structure, rules, symbolic constraints, complex query forms, prototypes, interpretable mechanisms, or post-hoc explanations of embedding behavior.","inclusion":"Include work whose primary objective is rule integration, soft-rule modeling, complex logical query answering, multi-view logical reasoning, interpretable rule- or instance-based completion, influential-example explanation, rule-mining diagnostics, or semantic substructure explanation.","exclusion":"Plain one-hop completion without explicit logical or explanatory objectives belongs to core completion modeling; KG-guided LLM factual grounding without formal query structure belongs to language-KG semantic integration.","nearest_boundary":"Language-KG semantic integration","representative_topics":["transformer-based masked pre-training for complex logical query reasoning over knowledge graphs","knowledge representation learning for logical query answering over multi-view knowledge graphs","knowledge graph reasoning via joint rule and entity-relation embedding with soft logical inference"]},{"direction_index":7,"label":"D7","name":"Heterogeneous KG representation and alignment","description":"Representing, matching, and reasoning over KGs with heterogeneity beyond ordinary single-graph binary triples, including cross-graph entity alignment, multimodal evidence, qualifiers, hyper-relations, n-ary facts, numeric literals, and bi-level fact structures.","inclusion":"Include work whose primary objective is entity matching across KGs, alignment under heterogeneity or growth, multimodal fusion, missing-modality handling, multimodal reasoning, hyper-relational or qualified fact modeling, or specialized representations for non-binary fact structures.","exclusion":"Text-only language-model fusion for ordinary completion belongs to language-KG semantic integration; ontology-level construction without entity matching belongs to KG construction and ontology engineering; ordinary fixed-graph triple scoring belongs to core completion modeling.","nearest_boundary":"KG construction and ontology engineering","representative_topics":["dual-view hyper-relational knowledge graph embedding for link prediction and entity typing","continual entity alignment for growing knowledge graphs via inductive neighbor-based representation learning","scalable knowledge graph entity alignment via subgraph partitioning, reconstruction, and embedding merging"]}],"split":"dev"}
{"episode_id":"F170_T2025-07","field_index":170,"field":"knowledge graph embedding","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Generalizable and temporal link reasoning","description":"Predicting links or facts beyond a fixed static transductive graph, including unseen entities, unseen relations, emerging subgraphs, evolving facts, temporal validity, and future-event extrapolation.","inclusion":"Include work whose primary objective is generalization across entity/relation sets, graph batches, disconnected emerging graphs, or time-indexed fact histories.","exclusion":"Standard missing-triple prediction on a fixed graph belongs to core completion modeling; methods centered on geometric space design or language fusion belong to those respective programs unless unseen/time-varying prediction is the main target.","nearest_boundary":"Core completion modeling and training","representative_topics":["unifying factorisation-based and message-passing models for inductive knowledge graph completion","inductive link prediction via mutual information maximization on complete subgraph neighboring relations","inductive knowledge graph reasoning for continuously emerging multi-batch entities via walk-based aggregation"]},{"direction_index":1,"label":"D1","name":"Language-KG semantic integration","description":"Coupling knowledge graphs with pretrained or large language models for KG completion, factual grounding, graph-aware generation, and KG-guided language-model reasoning.","inclusion":"Include work where the primary objective is to transfer semantic evidence between text/language models and KGs, either to improve KG completion with textual semantics or to improve language-model factuality and reasoning with KG evidence.","exclusion":"Building or populating a KG from text or model internals belongs to KG construction and ontology engineering; complex formal query answering belongs to logical and explainable reasoning.","nearest_boundary":"KG construction and ontology engineering","representative_topics":["knowledge graph semantics compression and reconstruction via transformer language models","sparse knowledge graph completion via variational em fusion of text and structure knowledge","hypercomplex space integration of knowledge graph and language model embeddings for link prediction"]},{"direction_index":2,"label":"D2","name":"KG construction and ontology engineering","description":"Creating, populating, organizing, and semantically grounding knowledge graphs, schemas, ontologies, and construction pipelines from text, language models, or heterogeneous domain data.","inclusion":"Include work whose primary objective is KG construction, entity/relation extraction for graph population, ontology design or embedding, schema and quality management, or LLM-assisted KG generation.","exclusion":"Using existing textual or language-model representations to score missing links belongs to language-KG semantic integration; matching equivalent entities across existing graphs belongs to heterogeneous KG representation and alignment.","nearest_boundary":"Language-KG semantic integration","representative_topics":["harvesting symbolic knowledge graphs of arbitrary relations via few-shot prompting of pretrained language models","incremental knowledge graph construction pipelines integrating ontology and quality management","llm-based joint entity and relation extraction for knowledge graph construction from unstructured text"]},{"direction_index":3,"label":"D3","name":"Geometric relational representation","description":"Designing embedding spaces and relation operators whose main purpose is to encode relational structure through geometry, including translation, rotation, scaling, curvature, hyperbolic or ultrahyperbolic spaces, and interacting geometric views.","inclusion":"Include work where the primary scientific contribution is the choice, derivation, or combination of geometric spaces or transformations for entity-relation representation.","exclusion":"Neural, convolutional, attention, transformer, or distillation architectures belong to core completion modeling unless the geometric space or operator is the central objective; temporal geometric models belong to generalizable and temporal link reasoning when temporal evolution is primary.","nearest_boundary":"Core completion modeling and training","representative_topics":["ultrahyperbolic geometry embeddings for heterogeneous knowledge graph topological structures","spatio-functional knowledge graph embedding for joint composition and hierarchy inference patterns","multi-geometry interaction embeddings for knowledge graph completion with coexisting spatial structures"]},{"direction_index":4,"label":"D4","name":"Core completion modeling and training","description":"Improving standard transductive KG completion through scoring architectures, neural interaction mechanisms, regularization, self-distillation, contrastive objectives, negative sampling, pruning, and scalable training strategies.","inclusion":"Include work whose primary objective is better or cheaper missing-triple prediction on a largely fixed graph through model architecture, training objective, sampling, compression, hyperparameter optimization, or low-dimensional learning.","exclusion":"Methods primarily about unseen entities, temporal evolution, explicit geometry, language fusion, multimodal or hyper-relational structure, privacy, robustness, or benchmark synthesis belong to the corresponding neighboring programs.","nearest_boundary":"Geometric relational representation","representative_topics":["confidence-aware self-distillation for low-dimensional knowledge graph embedding","attention-based entity representation learning with semantic path context for knowledge graph relation prediction","equivariance-based regularization for generalizing knowledge graph completion models"]},{"direction_index":5,"label":"D5","name":"Operational evaluation and trustworthy KGE","description":"Assessing, benchmarking, explaining at the field level, scaling, securing, privatizing, and auditing KGE methods and infrastructure for reliable use.","inclusion":"Include work whose primary objective is evaluation methodology, benchmark or library construction, survey synthesis, hyperparameter sensitivity analysis, adversarial robustness, bias mitigation, privacy-preserving or federated KGE, or operational reliability rather than a new task-specific predictor.","exclusion":"A new completion architecture with evaluation as a secondary contribution belongs to core completion modeling; post-hoc explanation of individual predictions belongs to logical and explainable reasoning.","nearest_boundary":"Core completion modeling and training","representative_topics":["hyperparameter sensitivity analysis for knowledge graph embedding quality assessment","benchmarking knowledge graph embeddings for description logic class representation learning","unified library for multi-source knowledge graph embedding across link prediction, entity alignment, and question answering"]},{"direction_index":6,"label":"D6","name":"Logical and explainable KG reasoning","description":"Reasoning over KGs with explicit logical structure, rules, symbolic constraints, complex query forms, prototypes, interpretable mechanisms, or post-hoc explanations of embedding behavior.","inclusion":"Include work whose primary objective is rule integration, soft-rule modeling, complex logical query answering, multi-view logical reasoning, interpretable rule- or instance-based completion, influential-example explanation, rule-mining diagnostics, or semantic substructure explanation.","exclusion":"Plain one-hop completion without explicit logical or explanatory objectives belongs to core completion modeling; KG-guided LLM factual grounding without formal query structure belongs to language-KG semantic integration.","nearest_boundary":"Language-KG semantic integration","representative_topics":["transformer-based masked pre-training for complex logical query reasoning over knowledge graphs","knowledge representation learning for logical query answering over multi-view knowledge graphs","knowledge graph reasoning via joint rule and entity-relation embedding with soft logical inference"]},{"direction_index":7,"label":"D7","name":"Heterogeneous KG representation and alignment","description":"Representing, matching, and reasoning over KGs with heterogeneity beyond ordinary single-graph binary triples, including cross-graph entity alignment, multimodal evidence, qualifiers, hyper-relations, n-ary facts, numeric literals, and bi-level fact structures.","inclusion":"Include work whose primary objective is entity matching across KGs, alignment under heterogeneity or growth, multimodal fusion, missing-modality handling, multimodal reasoning, hyper-relational or qualified fact modeling, or specialized representations for non-binary fact structures.","exclusion":"Text-only language-model fusion for ordinary completion belongs to language-KG semantic integration; ontology-level construction without entity matching belongs to KG construction and ontology engineering; ordinary fixed-graph triple scoring belongs to core completion modeling.","nearest_boundary":"KG construction and ontology engineering","representative_topics":["dual-view hyper-relational knowledge graph embedding for link prediction and entity typing","continual entity alignment for growing knowledge graphs via inductive neighbor-based representation learning","scalable knowledge graph entity alignment via subgraph partitioning, reconstruction, and embedding merging"]}],"split":"dev"}
{"episode_id":"F170_T2026-01","field_index":170,"field":"knowledge graph embedding","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Generalizable and temporal link reasoning","description":"Predicting links or facts beyond a fixed static transductive graph, including unseen entities, unseen relations, emerging subgraphs, evolving facts, temporal validity, and future-event extrapolation.","inclusion":"Include work whose primary objective is generalization across entity/relation sets, graph batches, disconnected emerging graphs, or time-indexed fact histories.","exclusion":"Standard missing-triple prediction on a fixed graph belongs to core completion modeling; methods centered on geometric space design or language fusion belong to those respective programs unless unseen/time-varying prediction is the main target.","nearest_boundary":"Core completion modeling and training","representative_topics":["unifying factorisation-based and message-passing models for inductive knowledge graph completion","inductive link prediction via mutual information maximization on complete subgraph neighboring relations","inductive knowledge graph reasoning for continuously emerging multi-batch entities via walk-based aggregation"]},{"direction_index":1,"label":"D1","name":"Language-KG semantic integration","description":"Coupling knowledge graphs with pretrained or large language models for KG completion, factual grounding, graph-aware generation, and KG-guided language-model reasoning.","inclusion":"Include work where the primary objective is to transfer semantic evidence between text/language models and KGs, either to improve KG completion with textual semantics or to improve language-model factuality and reasoning with KG evidence.","exclusion":"Building or populating a KG from text or model internals belongs to KG construction and ontology engineering; complex formal query answering belongs to logical and explainable reasoning.","nearest_boundary":"KG construction and ontology engineering","representative_topics":["knowledge graph semantics compression and reconstruction via transformer language models","sparse knowledge graph completion via variational em fusion of text and structure knowledge","hypercomplex space integration of knowledge graph and language model embeddings for link prediction"]},{"direction_index":2,"label":"D2","name":"KG construction and ontology engineering","description":"Creating, populating, organizing, and semantically grounding knowledge graphs, schemas, ontologies, and construction pipelines from text, language models, or heterogeneous domain data.","inclusion":"Include work whose primary objective is KG construction, entity/relation extraction for graph population, ontology design or embedding, schema and quality management, or LLM-assisted KG generation.","exclusion":"Using existing textual or language-model representations to score missing links belongs to language-KG semantic integration; matching equivalent entities across existing graphs belongs to heterogeneous KG representation and alignment.","nearest_boundary":"Language-KG semantic integration","representative_topics":["harvesting symbolic knowledge graphs of arbitrary relations via few-shot prompting of pretrained language models","incremental knowledge graph construction pipelines integrating ontology and quality management","llm-based joint entity and relation extraction for knowledge graph construction from unstructured text"]},{"direction_index":3,"label":"D3","name":"Geometric relational representation","description":"Designing embedding spaces and relation operators whose main purpose is to encode relational structure through geometry, including translation, rotation, scaling, curvature, hyperbolic or ultrahyperbolic spaces, and interacting geometric views.","inclusion":"Include work where the primary scientific contribution is the choice, derivation, or combination of geometric spaces or transformations for entity-relation representation.","exclusion":"Neural, convolutional, attention, transformer, or distillation architectures belong to core completion modeling unless the geometric space or operator is the central objective; temporal geometric models belong to generalizable and temporal link reasoning when temporal evolution is primary.","nearest_boundary":"Core completion modeling and training","representative_topics":["ultrahyperbolic geometry embeddings for heterogeneous knowledge graph topological structures","spatio-functional knowledge graph embedding for joint composition and hierarchy inference patterns","multi-geometry interaction embeddings for knowledge graph completion with coexisting spatial structures"]},{"direction_index":4,"label":"D4","name":"Core completion modeling and training","description":"Improving standard transductive KG completion through scoring architectures, neural interaction mechanisms, regularization, self-distillation, contrastive objectives, negative sampling, pruning, and scalable training strategies.","inclusion":"Include work whose primary objective is better or cheaper missing-triple prediction on a largely fixed graph through model architecture, training objective, sampling, compression, hyperparameter optimization, or low-dimensional learning.","exclusion":"Methods primarily about unseen entities, temporal evolution, explicit geometry, language fusion, multimodal or hyper-relational structure, privacy, robustness, or benchmark synthesis belong to the corresponding neighboring programs.","nearest_boundary":"Geometric relational representation","representative_topics":["confidence-aware self-distillation for low-dimensional knowledge graph embedding","attention-based entity representation learning with semantic path context for knowledge graph relation prediction","equivariance-based regularization for generalizing knowledge graph completion models"]},{"direction_index":5,"label":"D5","name":"Operational evaluation and trustworthy KGE","description":"Assessing, benchmarking, explaining at the field level, scaling, securing, privatizing, and auditing KGE methods and infrastructure for reliable use.","inclusion":"Include work whose primary objective is evaluation methodology, benchmark or library construction, survey synthesis, hyperparameter sensitivity analysis, adversarial robustness, bias mitigation, privacy-preserving or federated KGE, or operational reliability rather than a new task-specific predictor.","exclusion":"A new completion architecture with evaluation as a secondary contribution belongs to core completion modeling; post-hoc explanation of individual predictions belongs to logical and explainable reasoning.","nearest_boundary":"Core completion modeling and training","representative_topics":["hyperparameter sensitivity analysis for knowledge graph embedding quality assessment","benchmarking knowledge graph embeddings for description logic class representation learning","unified library for multi-source knowledge graph embedding across link prediction, entity alignment, and question answering"]},{"direction_index":6,"label":"D6","name":"Logical and explainable KG reasoning","description":"Reasoning over KGs with explicit logical structure, rules, symbolic constraints, complex query forms, prototypes, interpretable mechanisms, or post-hoc explanations of embedding behavior.","inclusion":"Include work whose primary objective is rule integration, soft-rule modeling, complex logical query answering, multi-view logical reasoning, interpretable rule- or instance-based completion, influential-example explanation, rule-mining diagnostics, or semantic substructure explanation.","exclusion":"Plain one-hop completion without explicit logical or explanatory objectives belongs to core completion modeling; KG-guided LLM factual grounding without formal query structure belongs to language-KG semantic integration.","nearest_boundary":"Language-KG semantic integration","representative_topics":["transformer-based masked pre-training for complex logical query reasoning over knowledge graphs","knowledge representation learning for logical query answering over multi-view knowledge graphs","knowledge graph reasoning via joint rule and entity-relation embedding with soft logical inference"]},{"direction_index":7,"label":"D7","name":"Heterogeneous KG representation and alignment","description":"Representing, matching, and reasoning over KGs with heterogeneity beyond ordinary single-graph binary triples, including cross-graph entity alignment, multimodal evidence, qualifiers, hyper-relations, n-ary facts, numeric literals, and bi-level fact structures.","inclusion":"Include work whose primary objective is entity matching across KGs, alignment under heterogeneity or growth, multimodal fusion, missing-modality handling, multimodal reasoning, hyper-relational or qualified fact modeling, or specialized representations for non-binary fact structures.","exclusion":"Text-only language-model fusion for ordinary completion belongs to language-KG semantic integration; ontology-level construction without entity matching belongs to KG construction and ontology engineering; ordinary fixed-graph triple scoring belongs to core completion modeling.","nearest_boundary":"KG construction and ontology engineering","representative_topics":["dual-view hyper-relational knowledge graph embedding for link prediction and entity typing","continual entity alignment for growing knowledge graphs via inductive neighbor-based representation learning","scalable knowledge graph entity alignment via subgraph partitioning, reconstruction, and embedding merging"]}],"split":"dev"}
{"episode_id":"F174_T2024-01","field_index":174,"field":"multimodal emotion recognition","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Dialogue relational context modeling","description":"Model emotion in multi-turn or multi-speaker interaction by representing utterance order, speaker dependencies, dialogue context, and explicit relational structure across modalities.","inclusion":"Include work whose main objective is conversational emotion recognition through speaker-aware context encoders, inter-utterance dependency modeling, dialogue graphs, directed or relational graph propagation, or conversation-specific multimodal complementation.","exclusion":"Use D1 when the main objective is a general fusion architecture not centered on explicit dialogue relations; use D2 when the main objective is a representation-learning loss evaluated on conversations.","nearest_boundary":"D1","representative_topics":["speaker-guided encoder-decoder for dynamic intra- and inter-speaker dependency modeling in conversation emotion recognition","graph-based cross-modal feature complementation for multimodal conversational emotion recognition","speaker-localised multimodal emotion recognition via realignment of facial expressions in conversations"]},{"direction_index":1,"label":"D1","name":"Adaptive multimodal fusion architectures","description":"Design neural mechanisms that combine affective evidence across text, audio, visual, spoken, or other modalities, emphasizing how modalities interact, are weighted, or are decomposed inside the prediction model.","inclusion":"Include work whose primary objective is a cross-modal fusion architecture, gated or attention-based interaction mechanism, additive or interpretable fusion decomposition, prompt or transformer fusion, or modality-contribution mechanism for emotion or sentiment prediction.","exclusion":"Use D0 when the defining contribution is dialogue-relational or graph-structured context; use D6 when the main objective is field-level benchmarking or evaluation rather than a new fusion mechanism.","nearest_boundary":"D0","representative_topics":["multi-modal fusion network for emotion recognition in conversational video","interpretable multimodal emotion recognition via additive feature attribution and hybrid fusion","interpretable multimodal speech-image emotion recognition using hybrid fusion and shapley attribution"]},{"direction_index":2,"label":"D2","name":"Representation learning objectives","description":"Learn aligned, discriminative, uncertainty-aware, distilled, contrastive, self-supervised, or otherwise optimized multimodal affect representations.","inclusion":"Include work whose main scientific contribution is a learning objective or representation program such as contrastive alignment, cluster-level affect-space learning, masked or self-supervised pretraining, cross-modal distillation, label-distribution modeling, or modality-invariant representation learning when not primarily framed as robustness to failed inputs.","exclusion":"Use D4 when the objective is explicitly reliable prediction under missing, noisy, adversarial, irrelevant, or incongruent modalities; use D1 when the loss is auxiliary to a fusion architecture.","nearest_boundary":"D4","representative_topics":["unsupervised multimodal emotion recognition via pairwise modality contrastive feature learning","distribution-based sequence-to-sequence emotion recognition with bayesian uncertainty estimation in conversation","cluster-level contrastive learning in low-dimensional affect space for emotion recognition in conversations"]},{"direction_index":3,"label":"D3","name":"General video and speech fusion","description":"Fuse audio, visual, text, subtitle, speech-language, or video streams for non-dialogue-centered emotion, sentiment, affect intensity, temporal localization, or continuous affect estimation.","inclusion":"Include work whose main objective is multimodal fusion for video-based, audio-visual, speech-centered, or single-instance affect recognition outside explicit multi-speaker conversational dependency modeling, including temporal transformers, TCNs, end-to-end audio-video-text interaction, AV facial affect, and temporal localization.","exclusion":"Use D0 for multi-turn speaker-context modeling; use D4 for methods primarily designed around missing, noisy, or incongruent modalities; use D5 when the task/domain phenomenon rather than the fusion method is the main objective.","nearest_boundary":"D0","representative_topics":["multi-granularity fusion of pre-trained speech and language models for multimodal emotion recognition","temporal emotion localization in untrimmed video via cross-modal subtitle consensus and multi-granularity context integration","cross-modal gated attention fusion for multimodal sentiment analysis in opinion videos"]},{"direction_index":4,"label":"D4","name":"Robustness to unreliable modalities","description":"Make multimodal emotion recognition reliable when modalities are absent, noisy, adversarially corrupted, semantically irrelevant, weak, incongruent, uncertain, or unevenly informative.","inclusion":"Include work whose primary objective is missing-modality learning, modality reconstruction or invariance, relevance estimation, adversarial or noise-resistant training, uncertainty or distribution-based robustness, incongruity-aware fusion, or time-continuous recognition under absent streams.","exclusion":"Use D2 for representation learning not centered on modality failure or degraded conditions; use D1 for ordinary adaptive attention or gating without an explicit robustness objective.","nearest_boundary":"D2","representative_topics":["robust multimodal emotion recognition with modality-invariant feature learning under missing modalities","distribution-based sequence-to-sequence emotion recognition with bayesian uncertainty estimation in conversation","multimodal relevance estimation for bridging semantic gaps in heterogeneous emotion modalities"]},{"direction_index":5,"label":"D5","name":"Specialized affect domains","description":"Adapt multimodal emotion recognition to specific affective phenomena, populations, data sources, deployment settings, or non-standard task definitions.","inclusion":"Include work whose main contribution is the domain or task formulation, such as sarcasm, humor, emotional reactions, stress, chronic-disease daily life, short video, emergency or social-media affect, code-mixed or cross-language settings, specialized user groups, or unusual modality sets when the setting drives the research objective.","exclusion":"Use D3 when the same data type is mainly used to demonstrate a general fusion method; use D7 when physiological or wearable signal fusion itself is the central modeling problem.","nearest_boundary":"D3","representative_topics":["multimodal emotion recognition in sarcasm with valence and arousal annotation","audio-visual multimodal recognition of humor, in-the-wild emotions, and stress","multimodal fusion for humor, reaction, and stress detection using hybrid feature mining"]},{"direction_index":6,"label":"D6","name":"Evaluation and unification frameworks","description":"Produce comparative, benchmark, survey, explanatory-evaluation, challenge, or unified-task knowledge about multimodal emotion and sentiment recognition.","inclusion":"Include work whose main objective is benchmarking, surveying, challenge or shared-task design, unified sentiment-emotion formulation, modality influence analysis, model comparison, affective-capability evaluation, or field-level interpretability/evaluation infrastructure rather than a single task-specific recognition model.","exclusion":"Use D5 for papers primarily proposing a model for a specialized domain; use D1 for new fusion architectures; use D4 for specific robustness methods rather than robustness benchmarks.","nearest_boundary":"D5","representative_topics":["unified multimodal sentiment and emotion recognition via cross-modal contrastive knowledge sharing","multimodal emotion recognition benchmarking for noise robustness and semi-supervised learning","multimodal affect and sentiment analysis across mimicked emotion, cross-cultural humour, and personalisation challenges"]},{"direction_index":7,"label":"D7","name":"Physiological and wearable fusion","description":"Recognize emotion from physiological, neural, peripheral, body, or wearable signals by learning specialized fusion or representation mechanisms for biosignal modalities.","inclusion":"Include work where the primary modalities are EEG, EDA, PPG, smartwatch, peripheral physiological, facial-physiological, or related wearable/body signals and the contribution centers on their fusion, cross-modal attention, transformer pretraining, self-supervised signal learning, or token/channel representation.","exclusion":"Use D5 when wearable data mainly serve a specialized deployment-domain study and biosignal fusion is not central; use D2 for general self-supervised multimodal pretraining not specific to physiological or wearable signals.","nearest_boundary":"D5","representative_topics":["cross-modal attentive fusion of wearable signals for multimodal emotion recognition","multimodal eda and ppg fusion for emotion recognition using low-cost wearable devices","multimodal physiological emotion recognition via pre-trained transformer pre-training"]}],"split":"dev"}
{"episode_id":"F174_T2024-07","field_index":174,"field":"multimodal emotion recognition","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Dialogue relational context modeling","description":"Model emotion in multi-turn or multi-speaker interaction by representing utterance order, speaker dependencies, dialogue context, and explicit relational structure across modalities.","inclusion":"Include work whose main objective is conversational emotion recognition through speaker-aware context encoders, inter-utterance dependency modeling, dialogue graphs, directed or relational graph propagation, or conversation-specific multimodal complementation.","exclusion":"Use D1 when the main objective is a general fusion architecture not centered on explicit dialogue relations; use D2 when the main objective is a representation-learning loss evaluated on conversations.","nearest_boundary":"D1","representative_topics":["speaker-guided encoder-decoder for dynamic intra- and inter-speaker dependency modeling in conversation emotion recognition","graph-based cross-modal feature complementation for multimodal conversational emotion recognition","speaker-localised multimodal emotion recognition via realignment of facial expressions in conversations"]},{"direction_index":1,"label":"D1","name":"Adaptive multimodal fusion architectures","description":"Design neural mechanisms that combine affective evidence across text, audio, visual, spoken, or other modalities, emphasizing how modalities interact, are weighted, or are decomposed inside the prediction model.","inclusion":"Include work whose primary objective is a cross-modal fusion architecture, gated or attention-based interaction mechanism, additive or interpretable fusion decomposition, prompt or transformer fusion, or modality-contribution mechanism for emotion or sentiment prediction.","exclusion":"Use D0 when the defining contribution is dialogue-relational or graph-structured context; use D6 when the main objective is field-level benchmarking or evaluation rather than a new fusion mechanism.","nearest_boundary":"D0","representative_topics":["multi-modal fusion network for emotion recognition in conversational video","interpretable multimodal emotion recognition via additive feature attribution and hybrid fusion","interpretable multimodal speech-image emotion recognition using hybrid fusion and shapley attribution"]},{"direction_index":2,"label":"D2","name":"Representation learning objectives","description":"Learn aligned, discriminative, uncertainty-aware, distilled, contrastive, self-supervised, or otherwise optimized multimodal affect representations.","inclusion":"Include work whose main scientific contribution is a learning objective or representation program such as contrastive alignment, cluster-level affect-space learning, masked or self-supervised pretraining, cross-modal distillation, label-distribution modeling, or modality-invariant representation learning when not primarily framed as robustness to failed inputs.","exclusion":"Use D4 when the objective is explicitly reliable prediction under missing, noisy, adversarial, irrelevant, or incongruent modalities; use D1 when the loss is auxiliary to a fusion architecture.","nearest_boundary":"D4","representative_topics":["unsupervised multimodal emotion recognition via pairwise modality contrastive feature learning","distribution-based sequence-to-sequence emotion recognition with bayesian uncertainty estimation in conversation","cluster-level contrastive learning in low-dimensional affect space for emotion recognition in conversations"]},{"direction_index":3,"label":"D3","name":"General video and speech fusion","description":"Fuse audio, visual, text, subtitle, speech-language, or video streams for non-dialogue-centered emotion, sentiment, affect intensity, temporal localization, or continuous affect estimation.","inclusion":"Include work whose main objective is multimodal fusion for video-based, audio-visual, speech-centered, or single-instance affect recognition outside explicit multi-speaker conversational dependency modeling, including temporal transformers, TCNs, end-to-end audio-video-text interaction, AV facial affect, and temporal localization.","exclusion":"Use D0 for multi-turn speaker-context modeling; use D4 for methods primarily designed around missing, noisy, or incongruent modalities; use D5 when the task/domain phenomenon rather than the fusion method is the main objective.","nearest_boundary":"D0","representative_topics":["multi-granularity fusion of pre-trained speech and language models for multimodal emotion recognition","temporal emotion localization in untrimmed video via cross-modal subtitle consensus and multi-granularity context integration","cross-modal gated attention fusion for multimodal sentiment analysis in opinion videos"]},{"direction_index":4,"label":"D4","name":"Robustness to unreliable modalities","description":"Make multimodal emotion recognition reliable when modalities are absent, noisy, adversarially corrupted, semantically irrelevant, weak, incongruent, uncertain, or unevenly informative.","inclusion":"Include work whose primary objective is missing-modality learning, modality reconstruction or invariance, relevance estimation, adversarial or noise-resistant training, uncertainty or distribution-based robustness, incongruity-aware fusion, or time-continuous recognition under absent streams.","exclusion":"Use D2 for representation learning not centered on modality failure or degraded conditions; use D1 for ordinary adaptive attention or gating without an explicit robustness objective.","nearest_boundary":"D2","representative_topics":["robust multimodal emotion recognition with modality-invariant feature learning under missing modalities","distribution-based sequence-to-sequence emotion recognition with bayesian uncertainty estimation in conversation","multimodal relevance estimation for bridging semantic gaps in heterogeneous emotion modalities"]},{"direction_index":5,"label":"D5","name":"Specialized affect domains","description":"Adapt multimodal emotion recognition to specific affective phenomena, populations, data sources, deployment settings, or non-standard task definitions.","inclusion":"Include work whose main contribution is the domain or task formulation, such as sarcasm, humor, emotional reactions, stress, chronic-disease daily life, short video, emergency or social-media affect, code-mixed or cross-language settings, specialized user groups, or unusual modality sets when the setting drives the research objective.","exclusion":"Use D3 when the same data type is mainly used to demonstrate a general fusion method; use D7 when physiological or wearable signal fusion itself is the central modeling problem.","nearest_boundary":"D3","representative_topics":["multimodal emotion recognition in sarcasm with valence and arousal annotation","audio-visual multimodal recognition of humor, in-the-wild emotions, and stress","multimodal fusion for humor, reaction, and stress detection using hybrid feature mining"]},{"direction_index":6,"label":"D6","name":"Evaluation and unification frameworks","description":"Produce comparative, benchmark, survey, explanatory-evaluation, challenge, or unified-task knowledge about multimodal emotion and sentiment recognition.","inclusion":"Include work whose main objective is benchmarking, surveying, challenge or shared-task design, unified sentiment-emotion formulation, modality influence analysis, model comparison, affective-capability evaluation, or field-level interpretability/evaluation infrastructure rather than a single task-specific recognition model.","exclusion":"Use D5 for papers primarily proposing a model for a specialized domain; use D1 for new fusion architectures; use D4 for specific robustness methods rather than robustness benchmarks.","nearest_boundary":"D5","representative_topics":["unified multimodal sentiment and emotion recognition via cross-modal contrastive knowledge sharing","multimodal emotion recognition benchmarking for noise robustness and semi-supervised learning","multimodal affect and sentiment analysis across mimicked emotion, cross-cultural humour, and personalisation challenges"]},{"direction_index":7,"label":"D7","name":"Physiological and wearable fusion","description":"Recognize emotion from physiological, neural, peripheral, body, or wearable signals by learning specialized fusion or representation mechanisms for biosignal modalities.","inclusion":"Include work where the primary modalities are EEG, EDA, PPG, smartwatch, peripheral physiological, facial-physiological, or related wearable/body signals and the contribution centers on their fusion, cross-modal attention, transformer pretraining, self-supervised signal learning, or token/channel representation.","exclusion":"Use D5 when wearable data mainly serve a specialized deployment-domain study and biosignal fusion is not central; use D2 for general self-supervised multimodal pretraining not specific to physiological or wearable signals.","nearest_boundary":"D5","representative_topics":["cross-modal attentive fusion of wearable signals for multimodal emotion recognition","multimodal eda and ppg fusion for emotion recognition using low-cost wearable devices","multimodal physiological emotion recognition via pre-trained transformer pre-training"]}],"split":"dev"}
{"episode_id":"F174_T2025-01","field_index":174,"field":"multimodal emotion recognition","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Dialogue relational context modeling","description":"Model emotion in multi-turn or multi-speaker interaction by representing utterance order, speaker dependencies, dialogue context, and explicit relational structure across modalities.","inclusion":"Include work whose main objective is conversational emotion recognition through speaker-aware context encoders, inter-utterance dependency modeling, dialogue graphs, directed or relational graph propagation, or conversation-specific multimodal complementation.","exclusion":"Use D1 when the main objective is a general fusion architecture not centered on explicit dialogue relations; use D2 when the main objective is a representation-learning loss evaluated on conversations.","nearest_boundary":"D1","representative_topics":["speaker-guided encoder-decoder for dynamic intra- and inter-speaker dependency modeling in conversation emotion recognition","graph-based cross-modal feature complementation for multimodal conversational emotion recognition","speaker-localised multimodal emotion recognition via realignment of facial expressions in conversations"]},{"direction_index":1,"label":"D1","name":"Adaptive multimodal fusion architectures","description":"Design neural mechanisms that combine affective evidence across text, audio, visual, spoken, or other modalities, emphasizing how modalities interact, are weighted, or are decomposed inside the prediction model.","inclusion":"Include work whose primary objective is a cross-modal fusion architecture, gated or attention-based interaction mechanism, additive or interpretable fusion decomposition, prompt or transformer fusion, or modality-contribution mechanism for emotion or sentiment prediction.","exclusion":"Use D0 when the defining contribution is dialogue-relational or graph-structured context; use D6 when the main objective is field-level benchmarking or evaluation rather than a new fusion mechanism.","nearest_boundary":"D0","representative_topics":["multi-modal fusion network for emotion recognition in conversational video","interpretable multimodal emotion recognition via additive feature attribution and hybrid fusion","interpretable multimodal speech-image emotion recognition using hybrid fusion and shapley attribution"]},{"direction_index":2,"label":"D2","name":"Representation learning objectives","description":"Learn aligned, discriminative, uncertainty-aware, distilled, contrastive, self-supervised, or otherwise optimized multimodal affect representations.","inclusion":"Include work whose main scientific contribution is a learning objective or representation program such as contrastive alignment, cluster-level affect-space learning, masked or self-supervised pretraining, cross-modal distillation, label-distribution modeling, or modality-invariant representation learning when not primarily framed as robustness to failed inputs.","exclusion":"Use D4 when the objective is explicitly reliable prediction under missing, noisy, adversarial, irrelevant, or incongruent modalities; use D1 when the loss is auxiliary to a fusion architecture.","nearest_boundary":"D4","representative_topics":["unsupervised multimodal emotion recognition via pairwise modality contrastive feature learning","distribution-based sequence-to-sequence emotion recognition with bayesian uncertainty estimation in conversation","cluster-level contrastive learning in low-dimensional affect space for emotion recognition in conversations"]},{"direction_index":3,"label":"D3","name":"General video and speech fusion","description":"Fuse audio, visual, text, subtitle, speech-language, or video streams for non-dialogue-centered emotion, sentiment, affect intensity, temporal localization, or continuous affect estimation.","inclusion":"Include work whose main objective is multimodal fusion for video-based, audio-visual, speech-centered, or single-instance affect recognition outside explicit multi-speaker conversational dependency modeling, including temporal transformers, TCNs, end-to-end audio-video-text interaction, AV facial affect, and temporal localization.","exclusion":"Use D0 for multi-turn speaker-context modeling; use D4 for methods primarily designed around missing, noisy, or incongruent modalities; use D5 when the task/domain phenomenon rather than the fusion method is the main objective.","nearest_boundary":"D0","representative_topics":["multi-granularity fusion of pre-trained speech and language models for multimodal emotion recognition","temporal emotion localization in untrimmed video via cross-modal subtitle consensus and multi-granularity context integration","cross-modal gated attention fusion for multimodal sentiment analysis in opinion videos"]},{"direction_index":4,"label":"D4","name":"Robustness to unreliable modalities","description":"Make multimodal emotion recognition reliable when modalities are absent, noisy, adversarially corrupted, semantically irrelevant, weak, incongruent, uncertain, or unevenly informative.","inclusion":"Include work whose primary objective is missing-modality learning, modality reconstruction or invariance, relevance estimation, adversarial or noise-resistant training, uncertainty or distribution-based robustness, incongruity-aware fusion, or time-continuous recognition under absent streams.","exclusion":"Use D2 for representation learning not centered on modality failure or degraded conditions; use D1 for ordinary adaptive attention or gating without an explicit robustness objective.","nearest_boundary":"D2","representative_topics":["robust multimodal emotion recognition with modality-invariant feature learning under missing modalities","distribution-based sequence-to-sequence emotion recognition with bayesian uncertainty estimation in conversation","multimodal relevance estimation for bridging semantic gaps in heterogeneous emotion modalities"]},{"direction_index":5,"label":"D5","name":"Specialized affect domains","description":"Adapt multimodal emotion recognition to specific affective phenomena, populations, data sources, deployment settings, or non-standard task definitions.","inclusion":"Include work whose main contribution is the domain or task formulation, such as sarcasm, humor, emotional reactions, stress, chronic-disease daily life, short video, emergency or social-media affect, code-mixed or cross-language settings, specialized user groups, or unusual modality sets when the setting drives the research objective.","exclusion":"Use D3 when the same data type is mainly used to demonstrate a general fusion method; use D7 when physiological or wearable signal fusion itself is the central modeling problem.","nearest_boundary":"D3","representative_topics":["multimodal emotion recognition in sarcasm with valence and arousal annotation","audio-visual multimodal recognition of humor, in-the-wild emotions, and stress","multimodal fusion for humor, reaction, and stress detection using hybrid feature mining"]},{"direction_index":6,"label":"D6","name":"Evaluation and unification frameworks","description":"Produce comparative, benchmark, survey, explanatory-evaluation, challenge, or unified-task knowledge about multimodal emotion and sentiment recognition.","inclusion":"Include work whose main objective is benchmarking, surveying, challenge or shared-task design, unified sentiment-emotion formulation, modality influence analysis, model comparison, affective-capability evaluation, or field-level interpretability/evaluation infrastructure rather than a single task-specific recognition model.","exclusion":"Use D5 for papers primarily proposing a model for a specialized domain; use D1 for new fusion architectures; use D4 for specific robustness methods rather than robustness benchmarks.","nearest_boundary":"D5","representative_topics":["unified multimodal sentiment and emotion recognition via cross-modal contrastive knowledge sharing","multimodal emotion recognition benchmarking for noise robustness and semi-supervised learning","multimodal affect and sentiment analysis across mimicked emotion, cross-cultural humour, and personalisation challenges"]},{"direction_index":7,"label":"D7","name":"Physiological and wearable fusion","description":"Recognize emotion from physiological, neural, peripheral, body, or wearable signals by learning specialized fusion or representation mechanisms for biosignal modalities.","inclusion":"Include work where the primary modalities are EEG, EDA, PPG, smartwatch, peripheral physiological, facial-physiological, or related wearable/body signals and the contribution centers on their fusion, cross-modal attention, transformer pretraining, self-supervised signal learning, or token/channel representation.","exclusion":"Use D5 when wearable data mainly serve a specialized deployment-domain study and biosignal fusion is not central; use D2 for general self-supervised multimodal pretraining not specific to physiological or wearable signals.","nearest_boundary":"D5","representative_topics":["cross-modal attentive fusion of wearable signals for multimodal emotion recognition","multimodal eda and ppg fusion for emotion recognition using low-cost wearable devices","multimodal physiological emotion recognition via pre-trained transformer pre-training"]}],"split":"dev"}
{"episode_id":"F174_T2025-07","field_index":174,"field":"multimodal emotion recognition","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Dialogue relational context modeling","description":"Model emotion in multi-turn or multi-speaker interaction by representing utterance order, speaker dependencies, dialogue context, and explicit relational structure across modalities.","inclusion":"Include work whose main objective is conversational emotion recognition through speaker-aware context encoders, inter-utterance dependency modeling, dialogue graphs, directed or relational graph propagation, or conversation-specific multimodal complementation.","exclusion":"Use D1 when the main objective is a general fusion architecture not centered on explicit dialogue relations; use D2 when the main objective is a representation-learning loss evaluated on conversations.","nearest_boundary":"D1","representative_topics":["speaker-guided encoder-decoder for dynamic intra- and inter-speaker dependency modeling in conversation emotion recognition","graph-based cross-modal feature complementation for multimodal conversational emotion recognition","speaker-localised multimodal emotion recognition via realignment of facial expressions in conversations"]},{"direction_index":1,"label":"D1","name":"Adaptive multimodal fusion architectures","description":"Design neural mechanisms that combine affective evidence across text, audio, visual, spoken, or other modalities, emphasizing how modalities interact, are weighted, or are decomposed inside the prediction model.","inclusion":"Include work whose primary objective is a cross-modal fusion architecture, gated or attention-based interaction mechanism, additive or interpretable fusion decomposition, prompt or transformer fusion, or modality-contribution mechanism for emotion or sentiment prediction.","exclusion":"Use D0 when the defining contribution is dialogue-relational or graph-structured context; use D6 when the main objective is field-level benchmarking or evaluation rather than a new fusion mechanism.","nearest_boundary":"D0","representative_topics":["multi-modal fusion network for emotion recognition in conversational video","interpretable multimodal emotion recognition via additive feature attribution and hybrid fusion","interpretable multimodal speech-image emotion recognition using hybrid fusion and shapley attribution"]},{"direction_index":2,"label":"D2","name":"Representation learning objectives","description":"Learn aligned, discriminative, uncertainty-aware, distilled, contrastive, self-supervised, or otherwise optimized multimodal affect representations.","inclusion":"Include work whose main scientific contribution is a learning objective or representation program such as contrastive alignment, cluster-level affect-space learning, masked or self-supervised pretraining, cross-modal distillation, label-distribution modeling, or modality-invariant representation learning when not primarily framed as robustness to failed inputs.","exclusion":"Use D4 when the objective is explicitly reliable prediction under missing, noisy, adversarial, irrelevant, or incongruent modalities; use D1 when the loss is auxiliary to a fusion architecture.","nearest_boundary":"D4","representative_topics":["unsupervised multimodal emotion recognition via pairwise modality contrastive feature learning","distribution-based sequence-to-sequence emotion recognition with bayesian uncertainty estimation in conversation","cluster-level contrastive learning in low-dimensional affect space for emotion recognition in conversations"]},{"direction_index":3,"label":"D3","name":"General video and speech fusion","description":"Fuse audio, visual, text, subtitle, speech-language, or video streams for non-dialogue-centered emotion, sentiment, affect intensity, temporal localization, or continuous affect estimation.","inclusion":"Include work whose main objective is multimodal fusion for video-based, audio-visual, speech-centered, or single-instance affect recognition outside explicit multi-speaker conversational dependency modeling, including temporal transformers, TCNs, end-to-end audio-video-text interaction, AV facial affect, and temporal localization.","exclusion":"Use D0 for multi-turn speaker-context modeling; use D4 for methods primarily designed around missing, noisy, or incongruent modalities; use D5 when the task/domain phenomenon rather than the fusion method is the main objective.","nearest_boundary":"D0","representative_topics":["multi-granularity fusion of pre-trained speech and language models for multimodal emotion recognition","temporal emotion localization in untrimmed video via cross-modal subtitle consensus and multi-granularity context integration","cross-modal gated attention fusion for multimodal sentiment analysis in opinion videos"]},{"direction_index":4,"label":"D4","name":"Robustness to unreliable modalities","description":"Make multimodal emotion recognition reliable when modalities are absent, noisy, adversarially corrupted, semantically irrelevant, weak, incongruent, uncertain, or unevenly informative.","inclusion":"Include work whose primary objective is missing-modality learning, modality reconstruction or invariance, relevance estimation, adversarial or noise-resistant training, uncertainty or distribution-based robustness, incongruity-aware fusion, or time-continuous recognition under absent streams.","exclusion":"Use D2 for representation learning not centered on modality failure or degraded conditions; use D1 for ordinary adaptive attention or gating without an explicit robustness objective.","nearest_boundary":"D2","representative_topics":["robust multimodal emotion recognition with modality-invariant feature learning under missing modalities","distribution-based sequence-to-sequence emotion recognition with bayesian uncertainty estimation in conversation","multimodal relevance estimation for bridging semantic gaps in heterogeneous emotion modalities"]},{"direction_index":5,"label":"D5","name":"Specialized affect domains","description":"Adapt multimodal emotion recognition to specific affective phenomena, populations, data sources, deployment settings, or non-standard task definitions.","inclusion":"Include work whose main contribution is the domain or task formulation, such as sarcasm, humor, emotional reactions, stress, chronic-disease daily life, short video, emergency or social-media affect, code-mixed or cross-language settings, specialized user groups, or unusual modality sets when the setting drives the research objective.","exclusion":"Use D3 when the same data type is mainly used to demonstrate a general fusion method; use D7 when physiological or wearable signal fusion itself is the central modeling problem.","nearest_boundary":"D3","representative_topics":["multimodal emotion recognition in sarcasm with valence and arousal annotation","audio-visual multimodal recognition of humor, in-the-wild emotions, and stress","multimodal fusion for humor, reaction, and stress detection using hybrid feature mining"]},{"direction_index":6,"label":"D6","name":"Evaluation and unification frameworks","description":"Produce comparative, benchmark, survey, explanatory-evaluation, challenge, or unified-task knowledge about multimodal emotion and sentiment recognition.","inclusion":"Include work whose main objective is benchmarking, surveying, challenge or shared-task design, unified sentiment-emotion formulation, modality influence analysis, model comparison, affective-capability evaluation, or field-level interpretability/evaluation infrastructure rather than a single task-specific recognition model.","exclusion":"Use D5 for papers primarily proposing a model for a specialized domain; use D1 for new fusion architectures; use D4 for specific robustness methods rather than robustness benchmarks.","nearest_boundary":"D5","representative_topics":["unified multimodal sentiment and emotion recognition via cross-modal contrastive knowledge sharing","multimodal emotion recognition benchmarking for noise robustness and semi-supervised learning","multimodal affect and sentiment analysis across mimicked emotion, cross-cultural humour, and personalisation challenges"]},{"direction_index":7,"label":"D7","name":"Physiological and wearable fusion","description":"Recognize emotion from physiological, neural, peripheral, body, or wearable signals by learning specialized fusion or representation mechanisms for biosignal modalities.","inclusion":"Include work where the primary modalities are EEG, EDA, PPG, smartwatch, peripheral physiological, facial-physiological, or related wearable/body signals and the contribution centers on their fusion, cross-modal attention, transformer pretraining, self-supervised signal learning, or token/channel representation.","exclusion":"Use D5 when wearable data mainly serve a specialized deployment-domain study and biosignal fusion is not central; use D2 for general self-supervised multimodal pretraining not specific to physiological or wearable signals.","nearest_boundary":"D5","representative_topics":["cross-modal attentive fusion of wearable signals for multimodal emotion recognition","multimodal eda and ppg fusion for emotion recognition using low-cost wearable devices","multimodal physiological emotion recognition via pre-trained transformer pre-training"]}],"split":"dev"}
{"episode_id":"F174_T2026-01","field_index":174,"field":"multimodal emotion recognition","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Dialogue relational context modeling","description":"Model emotion in multi-turn or multi-speaker interaction by representing utterance order, speaker dependencies, dialogue context, and explicit relational structure across modalities.","inclusion":"Include work whose main objective is conversational emotion recognition through speaker-aware context encoders, inter-utterance dependency modeling, dialogue graphs, directed or relational graph propagation, or conversation-specific multimodal complementation.","exclusion":"Use D1 when the main objective is a general fusion architecture not centered on explicit dialogue relations; use D2 when the main objective is a representation-learning loss evaluated on conversations.","nearest_boundary":"D1","representative_topics":["speaker-guided encoder-decoder for dynamic intra- and inter-speaker dependency modeling in conversation emotion recognition","graph-based cross-modal feature complementation for multimodal conversational emotion recognition","speaker-localised multimodal emotion recognition via realignment of facial expressions in conversations"]},{"direction_index":1,"label":"D1","name":"Adaptive multimodal fusion architectures","description":"Design neural mechanisms that combine affective evidence across text, audio, visual, spoken, or other modalities, emphasizing how modalities interact, are weighted, or are decomposed inside the prediction model.","inclusion":"Include work whose primary objective is a cross-modal fusion architecture, gated or attention-based interaction mechanism, additive or interpretable fusion decomposition, prompt or transformer fusion, or modality-contribution mechanism for emotion or sentiment prediction.","exclusion":"Use D0 when the defining contribution is dialogue-relational or graph-structured context; use D6 when the main objective is field-level benchmarking or evaluation rather than a new fusion mechanism.","nearest_boundary":"D0","representative_topics":["multi-modal fusion network for emotion recognition in conversational video","interpretable multimodal emotion recognition via additive feature attribution and hybrid fusion","interpretable multimodal speech-image emotion recognition using hybrid fusion and shapley attribution"]},{"direction_index":2,"label":"D2","name":"Representation learning objectives","description":"Learn aligned, discriminative, uncertainty-aware, distilled, contrastive, self-supervised, or otherwise optimized multimodal affect representations.","inclusion":"Include work whose main scientific contribution is a learning objective or representation program such as contrastive alignment, cluster-level affect-space learning, masked or self-supervised pretraining, cross-modal distillation, label-distribution modeling, or modality-invariant representation learning when not primarily framed as robustness to failed inputs.","exclusion":"Use D4 when the objective is explicitly reliable prediction under missing, noisy, adversarial, irrelevant, or incongruent modalities; use D1 when the loss is auxiliary to a fusion architecture.","nearest_boundary":"D4","representative_topics":["unsupervised multimodal emotion recognition via pairwise modality contrastive feature learning","distribution-based sequence-to-sequence emotion recognition with bayesian uncertainty estimation in conversation","cluster-level contrastive learning in low-dimensional affect space for emotion recognition in conversations"]},{"direction_index":3,"label":"D3","name":"General video and speech fusion","description":"Fuse audio, visual, text, subtitle, speech-language, or video streams for non-dialogue-centered emotion, sentiment, affect intensity, temporal localization, or continuous affect estimation.","inclusion":"Include work whose main objective is multimodal fusion for video-based, audio-visual, speech-centered, or single-instance affect recognition outside explicit multi-speaker conversational dependency modeling, including temporal transformers, TCNs, end-to-end audio-video-text interaction, AV facial affect, and temporal localization.","exclusion":"Use D0 for multi-turn speaker-context modeling; use D4 for methods primarily designed around missing, noisy, or incongruent modalities; use D5 when the task/domain phenomenon rather than the fusion method is the main objective.","nearest_boundary":"D0","representative_topics":["multi-granularity fusion of pre-trained speech and language models for multimodal emotion recognition","temporal emotion localization in untrimmed video via cross-modal subtitle consensus and multi-granularity context integration","cross-modal gated attention fusion for multimodal sentiment analysis in opinion videos"]},{"direction_index":4,"label":"D4","name":"Robustness to unreliable modalities","description":"Make multimodal emotion recognition reliable when modalities are absent, noisy, adversarially corrupted, semantically irrelevant, weak, incongruent, uncertain, or unevenly informative.","inclusion":"Include work whose primary objective is missing-modality learning, modality reconstruction or invariance, relevance estimation, adversarial or noise-resistant training, uncertainty or distribution-based robustness, incongruity-aware fusion, or time-continuous recognition under absent streams.","exclusion":"Use D2 for representation learning not centered on modality failure or degraded conditions; use D1 for ordinary adaptive attention or gating without an explicit robustness objective.","nearest_boundary":"D2","representative_topics":["robust multimodal emotion recognition with modality-invariant feature learning under missing modalities","distribution-based sequence-to-sequence emotion recognition with bayesian uncertainty estimation in conversation","multimodal relevance estimation for bridging semantic gaps in heterogeneous emotion modalities"]},{"direction_index":5,"label":"D5","name":"Specialized affect domains","description":"Adapt multimodal emotion recognition to specific affective phenomena, populations, data sources, deployment settings, or non-standard task definitions.","inclusion":"Include work whose main contribution is the domain or task formulation, such as sarcasm, humor, emotional reactions, stress, chronic-disease daily life, short video, emergency or social-media affect, code-mixed or cross-language settings, specialized user groups, or unusual modality sets when the setting drives the research objective.","exclusion":"Use D3 when the same data type is mainly used to demonstrate a general fusion method; use D7 when physiological or wearable signal fusion itself is the central modeling problem.","nearest_boundary":"D3","representative_topics":["multimodal emotion recognition in sarcasm with valence and arousal annotation","audio-visual multimodal recognition of humor, in-the-wild emotions, and stress","multimodal fusion for humor, reaction, and stress detection using hybrid feature mining"]},{"direction_index":6,"label":"D6","name":"Evaluation and unification frameworks","description":"Produce comparative, benchmark, survey, explanatory-evaluation, challenge, or unified-task knowledge about multimodal emotion and sentiment recognition.","inclusion":"Include work whose main objective is benchmarking, surveying, challenge or shared-task design, unified sentiment-emotion formulation, modality influence analysis, model comparison, affective-capability evaluation, or field-level interpretability/evaluation infrastructure rather than a single task-specific recognition model.","exclusion":"Use D5 for papers primarily proposing a model for a specialized domain; use D1 for new fusion architectures; use D4 for specific robustness methods rather than robustness benchmarks.","nearest_boundary":"D5","representative_topics":["unified multimodal sentiment and emotion recognition via cross-modal contrastive knowledge sharing","multimodal emotion recognition benchmarking for noise robustness and semi-supervised learning","multimodal affect and sentiment analysis across mimicked emotion, cross-cultural humour, and personalisation challenges"]},{"direction_index":7,"label":"D7","name":"Physiological and wearable fusion","description":"Recognize emotion from physiological, neural, peripheral, body, or wearable signals by learning specialized fusion or representation mechanisms for biosignal modalities.","inclusion":"Include work where the primary modalities are EEG, EDA, PPG, smartwatch, peripheral physiological, facial-physiological, or related wearable/body signals and the contribution centers on their fusion, cross-modal attention, transformer pretraining, self-supervised signal learning, or token/channel representation.","exclusion":"Use D5 when wearable data mainly serve a specialized deployment-domain study and biosignal fusion is not central; use D2 for general self-supervised multimodal pretraining not specific to physiological or wearable signals.","nearest_boundary":"D5","representative_topics":["cross-modal attentive fusion of wearable signals for multimodal emotion recognition","multimodal eda and ppg fusion for emotion recognition using low-cost wearable devices","multimodal physiological emotion recognition via pre-trained transformer pre-training"]}],"split":"dev"}
{"episode_id":"F180_T2024-01","field_index":180,"field":"adversarial machine learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Backdoor attacks and defenses","description":"Research whose primary objective is implanting, detecting, analyzing, removing, or hardening against hidden trojan/backdoor behavior in models, data, architectures, or training pipelines.","inclusion":"Include work centered on backdoor trigger design, clean-label or architectural trojans, poisoned-sample selection, stealth and persistence of implanted behavior, trigger inversion, poisoned-data identification, model repair, pruning, unlearning, and backdoor-specific hardening.","exclusion":"Federated or graph papers belong to structured/collaborative integrity when the federated protocol or graph structure is the central research object rather than the backdoor mechanism itself.","nearest_boundary":"Structured and collaborative integrity","representative_topics":["clean-label textual backdoor attacks via adversarial perturbation and trigger reliance","architectural backdoor attacks exploiting inductive bias in reusable neural network code","backdoor erasure dynamics in continuously fine-tuned evolving neural network models"]},{"direction_index":1,"label":"D1","name":"Evasion attack generation and evaluation","description":"Research whose primary objective is creating, improving, transferring, querying, or systematically evaluating test-time adversarial evasion attacks and their threat models.","inclusion":"Include digital adversarial example generation, transferability mechanisms, surrogate attacks, query-efficient black-box attacks, score or decision based attack optimization, side-channel aided evasion, general attack benchmarks, robustness evaluation protocols, taxonomies, and detector-evaluation frameworks.","exclusion":"Domain-constrained attacks belong to deployed modality robustness when the contribution is primarily adapting attacks or defenses to physical sensing, language, biometrics, cybersecurity, or sequential decision systems.","nearest_boundary":"Deployed modality robustness","representative_topics":["early-stage transferability of adversarial perturbations across independently trained networks","dark knowledge-augmented surrogate models for transferable adversarial example generation","generalization bounds on substitute classifier design for transferable black-box adversarial examples"]},{"direction_index":2,"label":"D2","name":"Robust learning and representations","description":"Research whose primary objective is improving robustness through training procedures, optimization choices, loss design, data selection, feature learning, or representation-level constraints.","inclusion":"Include adversarial training optimization, fast or efficient adversarial training, robust-overfitting and catastrophic-overfitting mitigation, accuracy-robustness trade-off methods, adaptive perturbation budgets, robust feature distillation, contrastive robust learning, information bottleneck defenses, feature disentanglement, latent consistency, and representation analyses of robust versus non-robust features.","exclusion":"Methods belong to certified robustness when the central contribution is a formal or probabilistic guarantee rather than empirical robust training or representation learning.","nearest_boundary":"Certified and theoretical robustness","representative_topics":["confidence threshold reduction for simultaneous adversarial robustness and generalization improvement","adaptive-margin adversarial training for robustness of medical image segmentation and detection models","soft adversarial training via certified input subsets to preserve natural accuracy"]},{"direction_index":3,"label":"D3","name":"Certified and theoretical robustness","description":"Research whose primary objective is proving guarantees, limits, or foundational properties of adversarial robustness, poisoning robustness, verification, or robust learnability.","inclusion":"Include randomized smoothing, interval bound propagation, branch-and-bound verification, formal verification surveys, certified training, certified purification when guarantees are central, PAC or minimax robust learnability, theoretical vulnerability analysis, adversarial surrogate-risk theory, poisoning sample-complexity theory, and abstract robustness trade-off analysis.","exclusion":"Empirical defenses without guarantees belong to robust learning or inference-time defense depending on whether the intervention occurs during training or testing.","nearest_boundary":"Robust learning and representations","representative_topics":["classifier-specific certification of randomized smoothing to overcome dimensionality-induced robustness limits","certified adversarial robustness via pretrained diffusion denoising without retraining","guided diffusion purification for certified adversarial robustness against l-infinity attacks"]},{"direction_index":4,"label":"D4","name":"Inference-time adversarial defense","description":"Research whose primary objective is detecting, rejecting, transforming, denoising, purifying, localizing, or repairing adversarial inputs at test time without making robust training the central program.","inclusion":"Include input preprocessing, nonlocal denoising, diffusion or score-based purification without a primary certification claim, adversarial example detection, patch localization and removal, multimodal inconsistency detection, feature-based correction, test-time randomization, selective rejection, and inference-time semantic repair.","exclusion":"Training-loop changes belong to robust learning and representations; certified purification belongs to certified and theoretical robustness when the guarantee is the main contribution.","nearest_boundary":"Robust learning and representations","representative_topics":["adversarial correction via intermediate layer feature analysis and conditional distribution matching","data-centric adversarial defense via nonlocal denoising and luminance-based example generation","defending multimodal fusion models against single-source adversarial perturbations via cross-modal inconsistency detection"]},{"direction_index":5,"label":"D5","name":"Privacy leakage and model theft","description":"Research whose primary objective is extracting or protecting private training information, membership facts, model functionality, parameters, architecture, provenance, or intellectual property.","inclusion":"Include gradient inversion, model inversion, training-data reconstruction, membership inference, privacy defenses against those attacks, model extraction, model stealing, encoder or architecture recovery, side-channel model theft, watermarking, provenance verification, and active defenses against functionality cloning.","exclusion":"Federated poisoning, Byzantine aggregation, and malicious-client integrity belong to structured/collaborative integrity unless the main objective is privacy leakage through gradients or updates.","nearest_boundary":"Structured and collaborative integrity","representative_topics":["subject-level membership inference attacks in cross-silo federated learning","semantic data reconstruction attacks bypassing gradient obfuscation in federated learning","black-box deep feature reconstruction attacks using gan generator as distribution constraint"]},{"direction_index":6,"label":"D6","name":"Structured and collaborative integrity","description":"Research whose primary objective is attacking or preserving model integrity under adversarial constraints imposed by collaborative learning protocols or graph-structured data.","inclusion":"Include federated or distributed poisoning, Byzantine behavior, malicious-client filtering, robust aggregation, federated backdoors when the protocol is central, federated adversarial training, graph structure or feature perturbation attacks, graph poisoning, graph backdoors, GNN defenses, graph sanitization, and graph-specific robustness analysis.","exclusion":"General non-graph evasion and non-federated attack algorithms belong to evasion attack generation and evaluation; privacy leakage from federated gradients belongs to privacy leakage and model theft.","nearest_boundary":"Privacy leakage and model theft","representative_topics":["durable backdoor attacks in federated learning via low-magnitude parameter targeting","randomized byzantine gradient aggregation defense against tailored adversarial attacks in federated learning","defending centered clipping robust aggregation against byzantine attacks in federated learning"]},{"direction_index":7,"label":"D7","name":"Deployed modality robustness","description":"Research whose primary objective is adapting adversarial attacks or defenses to the constraints of deployed modalities, sensors, human-facing systems, cybersecurity detectors, language/multimodal models, biometric identity systems, or sequential decision agents.","inclusion":"Include physical-world patches, optical or sensor attacks, autonomous-driving perception attacks, aerial or infrared attacks, adversarial text and vision-language manipulation, biometric evasion or impersonation, speaker or face recognition attacks and defenses, malware or intrusion-detector robustness, cyber-physical security ML robustness, reinforcement-learning policy attacks and defenses, multi-agent robustness, reward poisoning, and bandit or online decision-system attacks.","exclusion":"Place work in evasion attack generation and evaluation when the main contribution is a general attack, transfer, query, or benchmark mechanism not tied to a deployment constraint or modality-specific objective.","nearest_boundary":"Evasion attack generation and evaluation","representative_topics":["light-based physical adversarial attacks on vision dnn models using laser spot optimization","physical adversarial patches for attacking monocular depth estimation in autonomous driving","multimodal adversarial attacks on vision-language pre-training models via joint image-text perturbation"]}],"split":"dev"}
{"episode_id":"F180_T2024-07","field_index":180,"field":"adversarial machine learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Backdoor attacks and defenses","description":"Research whose primary objective is implanting, detecting, analyzing, removing, or hardening against hidden trojan/backdoor behavior in models, data, architectures, or training pipelines.","inclusion":"Include work centered on backdoor trigger design, clean-label or architectural trojans, poisoned-sample selection, stealth and persistence of implanted behavior, trigger inversion, poisoned-data identification, model repair, pruning, unlearning, and backdoor-specific hardening.","exclusion":"Federated or graph papers belong to structured/collaborative integrity when the federated protocol or graph structure is the central research object rather than the backdoor mechanism itself.","nearest_boundary":"Structured and collaborative integrity","representative_topics":["clean-label textual backdoor attacks via adversarial perturbation and trigger reliance","architectural backdoor attacks exploiting inductive bias in reusable neural network code","backdoor erasure dynamics in continuously fine-tuned evolving neural network models"]},{"direction_index":1,"label":"D1","name":"Evasion attack generation and evaluation","description":"Research whose primary objective is creating, improving, transferring, querying, or systematically evaluating test-time adversarial evasion attacks and their threat models.","inclusion":"Include digital adversarial example generation, transferability mechanisms, surrogate attacks, query-efficient black-box attacks, score or decision based attack optimization, side-channel aided evasion, general attack benchmarks, robustness evaluation protocols, taxonomies, and detector-evaluation frameworks.","exclusion":"Domain-constrained attacks belong to deployed modality robustness when the contribution is primarily adapting attacks or defenses to physical sensing, language, biometrics, cybersecurity, or sequential decision systems.","nearest_boundary":"Deployed modality robustness","representative_topics":["early-stage transferability of adversarial perturbations across independently trained networks","dark knowledge-augmented surrogate models for transferable adversarial example generation","generalization bounds on substitute classifier design for transferable black-box adversarial examples"]},{"direction_index":2,"label":"D2","name":"Robust learning and representations","description":"Research whose primary objective is improving robustness through training procedures, optimization choices, loss design, data selection, feature learning, or representation-level constraints.","inclusion":"Include adversarial training optimization, fast or efficient adversarial training, robust-overfitting and catastrophic-overfitting mitigation, accuracy-robustness trade-off methods, adaptive perturbation budgets, robust feature distillation, contrastive robust learning, information bottleneck defenses, feature disentanglement, latent consistency, and representation analyses of robust versus non-robust features.","exclusion":"Methods belong to certified robustness when the central contribution is a formal or probabilistic guarantee rather than empirical robust training or representation learning.","nearest_boundary":"Certified and theoretical robustness","representative_topics":["confidence threshold reduction for simultaneous adversarial robustness and generalization improvement","adaptive-margin adversarial training for robustness of medical image segmentation and detection models","soft adversarial training via certified input subsets to preserve natural accuracy"]},{"direction_index":3,"label":"D3","name":"Certified and theoretical robustness","description":"Research whose primary objective is proving guarantees, limits, or foundational properties of adversarial robustness, poisoning robustness, verification, or robust learnability.","inclusion":"Include randomized smoothing, interval bound propagation, branch-and-bound verification, formal verification surveys, certified training, certified purification when guarantees are central, PAC or minimax robust learnability, theoretical vulnerability analysis, adversarial surrogate-risk theory, poisoning sample-complexity theory, and abstract robustness trade-off analysis.","exclusion":"Empirical defenses without guarantees belong to robust learning or inference-time defense depending on whether the intervention occurs during training or testing.","nearest_boundary":"Robust learning and representations","representative_topics":["classifier-specific certification of randomized smoothing to overcome dimensionality-induced robustness limits","certified adversarial robustness via pretrained diffusion denoising without retraining","guided diffusion purification for certified adversarial robustness against l-infinity attacks"]},{"direction_index":4,"label":"D4","name":"Inference-time adversarial defense","description":"Research whose primary objective is detecting, rejecting, transforming, denoising, purifying, localizing, or repairing adversarial inputs at test time without making robust training the central program.","inclusion":"Include input preprocessing, nonlocal denoising, diffusion or score-based purification without a primary certification claim, adversarial example detection, patch localization and removal, multimodal inconsistency detection, feature-based correction, test-time randomization, selective rejection, and inference-time semantic repair.","exclusion":"Training-loop changes belong to robust learning and representations; certified purification belongs to certified and theoretical robustness when the guarantee is the main contribution.","nearest_boundary":"Robust learning and representations","representative_topics":["adversarial correction via intermediate layer feature analysis and conditional distribution matching","data-centric adversarial defense via nonlocal denoising and luminance-based example generation","defending multimodal fusion models against single-source adversarial perturbations via cross-modal inconsistency detection"]},{"direction_index":5,"label":"D5","name":"Privacy leakage and model theft","description":"Research whose primary objective is extracting or protecting private training information, membership facts, model functionality, parameters, architecture, provenance, or intellectual property.","inclusion":"Include gradient inversion, model inversion, training-data reconstruction, membership inference, privacy defenses against those attacks, model extraction, model stealing, encoder or architecture recovery, side-channel model theft, watermarking, provenance verification, and active defenses against functionality cloning.","exclusion":"Federated poisoning, Byzantine aggregation, and malicious-client integrity belong to structured/collaborative integrity unless the main objective is privacy leakage through gradients or updates.","nearest_boundary":"Structured and collaborative integrity","representative_topics":["subject-level membership inference attacks in cross-silo federated learning","semantic data reconstruction attacks bypassing gradient obfuscation in federated learning","black-box deep feature reconstruction attacks using gan generator as distribution constraint"]},{"direction_index":6,"label":"D6","name":"Structured and collaborative integrity","description":"Research whose primary objective is attacking or preserving model integrity under adversarial constraints imposed by collaborative learning protocols or graph-structured data.","inclusion":"Include federated or distributed poisoning, Byzantine behavior, malicious-client filtering, robust aggregation, federated backdoors when the protocol is central, federated adversarial training, graph structure or feature perturbation attacks, graph poisoning, graph backdoors, GNN defenses, graph sanitization, and graph-specific robustness analysis.","exclusion":"General non-graph evasion and non-federated attack algorithms belong to evasion attack generation and evaluation; privacy leakage from federated gradients belongs to privacy leakage and model theft.","nearest_boundary":"Privacy leakage and model theft","representative_topics":["durable backdoor attacks in federated learning via low-magnitude parameter targeting","randomized byzantine gradient aggregation defense against tailored adversarial attacks in federated learning","defending centered clipping robust aggregation against byzantine attacks in federated learning"]},{"direction_index":7,"label":"D7","name":"Deployed modality robustness","description":"Research whose primary objective is adapting adversarial attacks or defenses to the constraints of deployed modalities, sensors, human-facing systems, cybersecurity detectors, language/multimodal models, biometric identity systems, or sequential decision agents.","inclusion":"Include physical-world patches, optical or sensor attacks, autonomous-driving perception attacks, aerial or infrared attacks, adversarial text and vision-language manipulation, biometric evasion or impersonation, speaker or face recognition attacks and defenses, malware or intrusion-detector robustness, cyber-physical security ML robustness, reinforcement-learning policy attacks and defenses, multi-agent robustness, reward poisoning, and bandit or online decision-system attacks.","exclusion":"Place work in evasion attack generation and evaluation when the main contribution is a general attack, transfer, query, or benchmark mechanism not tied to a deployment constraint or modality-specific objective.","nearest_boundary":"Evasion attack generation and evaluation","representative_topics":["light-based physical adversarial attacks on vision dnn models using laser spot optimization","physical adversarial patches for attacking monocular depth estimation in autonomous driving","multimodal adversarial attacks on vision-language pre-training models via joint image-text perturbation"]}],"split":"dev"}
{"episode_id":"F180_T2025-01","field_index":180,"field":"adversarial machine learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Backdoor attacks and defenses","description":"Research whose primary objective is implanting, detecting, analyzing, removing, or hardening against hidden trojan/backdoor behavior in models, data, architectures, or training pipelines.","inclusion":"Include work centered on backdoor trigger design, clean-label or architectural trojans, poisoned-sample selection, stealth and persistence of implanted behavior, trigger inversion, poisoned-data identification, model repair, pruning, unlearning, and backdoor-specific hardening.","exclusion":"Federated or graph papers belong to structured/collaborative integrity when the federated protocol or graph structure is the central research object rather than the backdoor mechanism itself.","nearest_boundary":"Structured and collaborative integrity","representative_topics":["clean-label textual backdoor attacks via adversarial perturbation and trigger reliance","architectural backdoor attacks exploiting inductive bias in reusable neural network code","backdoor erasure dynamics in continuously fine-tuned evolving neural network models"]},{"direction_index":1,"label":"D1","name":"Evasion attack generation and evaluation","description":"Research whose primary objective is creating, improving, transferring, querying, or systematically evaluating test-time adversarial evasion attacks and their threat models.","inclusion":"Include digital adversarial example generation, transferability mechanisms, surrogate attacks, query-efficient black-box attacks, score or decision based attack optimization, side-channel aided evasion, general attack benchmarks, robustness evaluation protocols, taxonomies, and detector-evaluation frameworks.","exclusion":"Domain-constrained attacks belong to deployed modality robustness when the contribution is primarily adapting attacks or defenses to physical sensing, language, biometrics, cybersecurity, or sequential decision systems.","nearest_boundary":"Deployed modality robustness","representative_topics":["early-stage transferability of adversarial perturbations across independently trained networks","dark knowledge-augmented surrogate models for transferable adversarial example generation","generalization bounds on substitute classifier design for transferable black-box adversarial examples"]},{"direction_index":2,"label":"D2","name":"Robust learning and representations","description":"Research whose primary objective is improving robustness through training procedures, optimization choices, loss design, data selection, feature learning, or representation-level constraints.","inclusion":"Include adversarial training optimization, fast or efficient adversarial training, robust-overfitting and catastrophic-overfitting mitigation, accuracy-robustness trade-off methods, adaptive perturbation budgets, robust feature distillation, contrastive robust learning, information bottleneck defenses, feature disentanglement, latent consistency, and representation analyses of robust versus non-robust features.","exclusion":"Methods belong to certified robustness when the central contribution is a formal or probabilistic guarantee rather than empirical robust training or representation learning.","nearest_boundary":"Certified and theoretical robustness","representative_topics":["confidence threshold reduction for simultaneous adversarial robustness and generalization improvement","adaptive-margin adversarial training for robustness of medical image segmentation and detection models","soft adversarial training via certified input subsets to preserve natural accuracy"]},{"direction_index":3,"label":"D3","name":"Certified and theoretical robustness","description":"Research whose primary objective is proving guarantees, limits, or foundational properties of adversarial robustness, poisoning robustness, verification, or robust learnability.","inclusion":"Include randomized smoothing, interval bound propagation, branch-and-bound verification, formal verification surveys, certified training, certified purification when guarantees are central, PAC or minimax robust learnability, theoretical vulnerability analysis, adversarial surrogate-risk theory, poisoning sample-complexity theory, and abstract robustness trade-off analysis.","exclusion":"Empirical defenses without guarantees belong to robust learning or inference-time defense depending on whether the intervention occurs during training or testing.","nearest_boundary":"Robust learning and representations","representative_topics":["classifier-specific certification of randomized smoothing to overcome dimensionality-induced robustness limits","certified adversarial robustness via pretrained diffusion denoising without retraining","guided diffusion purification for certified adversarial robustness against l-infinity attacks"]},{"direction_index":4,"label":"D4","name":"Inference-time adversarial defense","description":"Research whose primary objective is detecting, rejecting, transforming, denoising, purifying, localizing, or repairing adversarial inputs at test time without making robust training the central program.","inclusion":"Include input preprocessing, nonlocal denoising, diffusion or score-based purification without a primary certification claim, adversarial example detection, patch localization and removal, multimodal inconsistency detection, feature-based correction, test-time randomization, selective rejection, and inference-time semantic repair.","exclusion":"Training-loop changes belong to robust learning and representations; certified purification belongs to certified and theoretical robustness when the guarantee is the main contribution.","nearest_boundary":"Robust learning and representations","representative_topics":["adversarial correction via intermediate layer feature analysis and conditional distribution matching","data-centric adversarial defense via nonlocal denoising and luminance-based example generation","defending multimodal fusion models against single-source adversarial perturbations via cross-modal inconsistency detection"]},{"direction_index":5,"label":"D5","name":"Privacy leakage and model theft","description":"Research whose primary objective is extracting or protecting private training information, membership facts, model functionality, parameters, architecture, provenance, or intellectual property.","inclusion":"Include gradient inversion, model inversion, training-data reconstruction, membership inference, privacy defenses against those attacks, model extraction, model stealing, encoder or architecture recovery, side-channel model theft, watermarking, provenance verification, and active defenses against functionality cloning.","exclusion":"Federated poisoning, Byzantine aggregation, and malicious-client integrity belong to structured/collaborative integrity unless the main objective is privacy leakage through gradients or updates.","nearest_boundary":"Structured and collaborative integrity","representative_topics":["subject-level membership inference attacks in cross-silo federated learning","semantic data reconstruction attacks bypassing gradient obfuscation in federated learning","black-box deep feature reconstruction attacks using gan generator as distribution constraint"]},{"direction_index":6,"label":"D6","name":"Structured and collaborative integrity","description":"Research whose primary objective is attacking or preserving model integrity under adversarial constraints imposed by collaborative learning protocols or graph-structured data.","inclusion":"Include federated or distributed poisoning, Byzantine behavior, malicious-client filtering, robust aggregation, federated backdoors when the protocol is central, federated adversarial training, graph structure or feature perturbation attacks, graph poisoning, graph backdoors, GNN defenses, graph sanitization, and graph-specific robustness analysis.","exclusion":"General non-graph evasion and non-federated attack algorithms belong to evasion attack generation and evaluation; privacy leakage from federated gradients belongs to privacy leakage and model theft.","nearest_boundary":"Privacy leakage and model theft","representative_topics":["durable backdoor attacks in federated learning via low-magnitude parameter targeting","randomized byzantine gradient aggregation defense against tailored adversarial attacks in federated learning","defending centered clipping robust aggregation against byzantine attacks in federated learning"]},{"direction_index":7,"label":"D7","name":"Deployed modality robustness","description":"Research whose primary objective is adapting adversarial attacks or defenses to the constraints of deployed modalities, sensors, human-facing systems, cybersecurity detectors, language/multimodal models, biometric identity systems, or sequential decision agents.","inclusion":"Include physical-world patches, optical or sensor attacks, autonomous-driving perception attacks, aerial or infrared attacks, adversarial text and vision-language manipulation, biometric evasion or impersonation, speaker or face recognition attacks and defenses, malware or intrusion-detector robustness, cyber-physical security ML robustness, reinforcement-learning policy attacks and defenses, multi-agent robustness, reward poisoning, and bandit or online decision-system attacks.","exclusion":"Place work in evasion attack generation and evaluation when the main contribution is a general attack, transfer, query, or benchmark mechanism not tied to a deployment constraint or modality-specific objective.","nearest_boundary":"Evasion attack generation and evaluation","representative_topics":["light-based physical adversarial attacks on vision dnn models using laser spot optimization","physical adversarial patches for attacking monocular depth estimation in autonomous driving","multimodal adversarial attacks on vision-language pre-training models via joint image-text perturbation"]}],"split":"dev"}
{"episode_id":"F180_T2025-07","field_index":180,"field":"adversarial machine learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Backdoor attacks and defenses","description":"Research whose primary objective is implanting, detecting, analyzing, removing, or hardening against hidden trojan/backdoor behavior in models, data, architectures, or training pipelines.","inclusion":"Include work centered on backdoor trigger design, clean-label or architectural trojans, poisoned-sample selection, stealth and persistence of implanted behavior, trigger inversion, poisoned-data identification, model repair, pruning, unlearning, and backdoor-specific hardening.","exclusion":"Federated or graph papers belong to structured/collaborative integrity when the federated protocol or graph structure is the central research object rather than the backdoor mechanism itself.","nearest_boundary":"Structured and collaborative integrity","representative_topics":["clean-label textual backdoor attacks via adversarial perturbation and trigger reliance","architectural backdoor attacks exploiting inductive bias in reusable neural network code","backdoor erasure dynamics in continuously fine-tuned evolving neural network models"]},{"direction_index":1,"label":"D1","name":"Evasion attack generation and evaluation","description":"Research whose primary objective is creating, improving, transferring, querying, or systematically evaluating test-time adversarial evasion attacks and their threat models.","inclusion":"Include digital adversarial example generation, transferability mechanisms, surrogate attacks, query-efficient black-box attacks, score or decision based attack optimization, side-channel aided evasion, general attack benchmarks, robustness evaluation protocols, taxonomies, and detector-evaluation frameworks.","exclusion":"Domain-constrained attacks belong to deployed modality robustness when the contribution is primarily adapting attacks or defenses to physical sensing, language, biometrics, cybersecurity, or sequential decision systems.","nearest_boundary":"Deployed modality robustness","representative_topics":["early-stage transferability of adversarial perturbations across independently trained networks","dark knowledge-augmented surrogate models for transferable adversarial example generation","generalization bounds on substitute classifier design for transferable black-box adversarial examples"]},{"direction_index":2,"label":"D2","name":"Robust learning and representations","description":"Research whose primary objective is improving robustness through training procedures, optimization choices, loss design, data selection, feature learning, or representation-level constraints.","inclusion":"Include adversarial training optimization, fast or efficient adversarial training, robust-overfitting and catastrophic-overfitting mitigation, accuracy-robustness trade-off methods, adaptive perturbation budgets, robust feature distillation, contrastive robust learning, information bottleneck defenses, feature disentanglement, latent consistency, and representation analyses of robust versus non-robust features.","exclusion":"Methods belong to certified robustness when the central contribution is a formal or probabilistic guarantee rather than empirical robust training or representation learning.","nearest_boundary":"Certified and theoretical robustness","representative_topics":["confidence threshold reduction for simultaneous adversarial robustness and generalization improvement","adaptive-margin adversarial training for robustness of medical image segmentation and detection models","soft adversarial training via certified input subsets to preserve natural accuracy"]},{"direction_index":3,"label":"D3","name":"Certified and theoretical robustness","description":"Research whose primary objective is proving guarantees, limits, or foundational properties of adversarial robustness, poisoning robustness, verification, or robust learnability.","inclusion":"Include randomized smoothing, interval bound propagation, branch-and-bound verification, formal verification surveys, certified training, certified purification when guarantees are central, PAC or minimax robust learnability, theoretical vulnerability analysis, adversarial surrogate-risk theory, poisoning sample-complexity theory, and abstract robustness trade-off analysis.","exclusion":"Empirical defenses without guarantees belong to robust learning or inference-time defense depending on whether the intervention occurs during training or testing.","nearest_boundary":"Robust learning and representations","representative_topics":["classifier-specific certification of randomized smoothing to overcome dimensionality-induced robustness limits","certified adversarial robustness via pretrained diffusion denoising without retraining","guided diffusion purification for certified adversarial robustness against l-infinity attacks"]},{"direction_index":4,"label":"D4","name":"Inference-time adversarial defense","description":"Research whose primary objective is detecting, rejecting, transforming, denoising, purifying, localizing, or repairing adversarial inputs at test time without making robust training the central program.","inclusion":"Include input preprocessing, nonlocal denoising, diffusion or score-based purification without a primary certification claim, adversarial example detection, patch localization and removal, multimodal inconsistency detection, feature-based correction, test-time randomization, selective rejection, and inference-time semantic repair.","exclusion":"Training-loop changes belong to robust learning and representations; certified purification belongs to certified and theoretical robustness when the guarantee is the main contribution.","nearest_boundary":"Robust learning and representations","representative_topics":["adversarial correction via intermediate layer feature analysis and conditional distribution matching","data-centric adversarial defense via nonlocal denoising and luminance-based example generation","defending multimodal fusion models against single-source adversarial perturbations via cross-modal inconsistency detection"]},{"direction_index":5,"label":"D5","name":"Privacy leakage and model theft","description":"Research whose primary objective is extracting or protecting private training information, membership facts, model functionality, parameters, architecture, provenance, or intellectual property.","inclusion":"Include gradient inversion, model inversion, training-data reconstruction, membership inference, privacy defenses against those attacks, model extraction, model stealing, encoder or architecture recovery, side-channel model theft, watermarking, provenance verification, and active defenses against functionality cloning.","exclusion":"Federated poisoning, Byzantine aggregation, and malicious-client integrity belong to structured/collaborative integrity unless the main objective is privacy leakage through gradients or updates.","nearest_boundary":"Structured and collaborative integrity","representative_topics":["subject-level membership inference attacks in cross-silo federated learning","semantic data reconstruction attacks bypassing gradient obfuscation in federated learning","black-box deep feature reconstruction attacks using gan generator as distribution constraint"]},{"direction_index":6,"label":"D6","name":"Structured and collaborative integrity","description":"Research whose primary objective is attacking or preserving model integrity under adversarial constraints imposed by collaborative learning protocols or graph-structured data.","inclusion":"Include federated or distributed poisoning, Byzantine behavior, malicious-client filtering, robust aggregation, federated backdoors when the protocol is central, federated adversarial training, graph structure or feature perturbation attacks, graph poisoning, graph backdoors, GNN defenses, graph sanitization, and graph-specific robustness analysis.","exclusion":"General non-graph evasion and non-federated attack algorithms belong to evasion attack generation and evaluation; privacy leakage from federated gradients belongs to privacy leakage and model theft.","nearest_boundary":"Privacy leakage and model theft","representative_topics":["durable backdoor attacks in federated learning via low-magnitude parameter targeting","randomized byzantine gradient aggregation defense against tailored adversarial attacks in federated learning","defending centered clipping robust aggregation against byzantine attacks in federated learning"]},{"direction_index":7,"label":"D7","name":"Deployed modality robustness","description":"Research whose primary objective is adapting adversarial attacks or defenses to the constraints of deployed modalities, sensors, human-facing systems, cybersecurity detectors, language/multimodal models, biometric identity systems, or sequential decision agents.","inclusion":"Include physical-world patches, optical or sensor attacks, autonomous-driving perception attacks, aerial or infrared attacks, adversarial text and vision-language manipulation, biometric evasion or impersonation, speaker or face recognition attacks and defenses, malware or intrusion-detector robustness, cyber-physical security ML robustness, reinforcement-learning policy attacks and defenses, multi-agent robustness, reward poisoning, and bandit or online decision-system attacks.","exclusion":"Place work in evasion attack generation and evaluation when the main contribution is a general attack, transfer, query, or benchmark mechanism not tied to a deployment constraint or modality-specific objective.","nearest_boundary":"Evasion attack generation and evaluation","representative_topics":["light-based physical adversarial attacks on vision dnn models using laser spot optimization","physical adversarial patches for attacking monocular depth estimation in autonomous driving","multimodal adversarial attacks on vision-language pre-training models via joint image-text perturbation"]}],"split":"dev"}
{"episode_id":"F180_T2026-01","field_index":180,"field":"adversarial machine learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Backdoor attacks and defenses","description":"Research whose primary objective is implanting, detecting, analyzing, removing, or hardening against hidden trojan/backdoor behavior in models, data, architectures, or training pipelines.","inclusion":"Include work centered on backdoor trigger design, clean-label or architectural trojans, poisoned-sample selection, stealth and persistence of implanted behavior, trigger inversion, poisoned-data identification, model repair, pruning, unlearning, and backdoor-specific hardening.","exclusion":"Federated or graph papers belong to structured/collaborative integrity when the federated protocol or graph structure is the central research object rather than the backdoor mechanism itself.","nearest_boundary":"Structured and collaborative integrity","representative_topics":["clean-label textual backdoor attacks via adversarial perturbation and trigger reliance","architectural backdoor attacks exploiting inductive bias in reusable neural network code","backdoor erasure dynamics in continuously fine-tuned evolving neural network models"]},{"direction_index":1,"label":"D1","name":"Evasion attack generation and evaluation","description":"Research whose primary objective is creating, improving, transferring, querying, or systematically evaluating test-time adversarial evasion attacks and their threat models.","inclusion":"Include digital adversarial example generation, transferability mechanisms, surrogate attacks, query-efficient black-box attacks, score or decision based attack optimization, side-channel aided evasion, general attack benchmarks, robustness evaluation protocols, taxonomies, and detector-evaluation frameworks.","exclusion":"Domain-constrained attacks belong to deployed modality robustness when the contribution is primarily adapting attacks or defenses to physical sensing, language, biometrics, cybersecurity, or sequential decision systems.","nearest_boundary":"Deployed modality robustness","representative_topics":["early-stage transferability of adversarial perturbations across independently trained networks","dark knowledge-augmented surrogate models for transferable adversarial example generation","generalization bounds on substitute classifier design for transferable black-box adversarial examples"]},{"direction_index":2,"label":"D2","name":"Robust learning and representations","description":"Research whose primary objective is improving robustness through training procedures, optimization choices, loss design, data selection, feature learning, or representation-level constraints.","inclusion":"Include adversarial training optimization, fast or efficient adversarial training, robust-overfitting and catastrophic-overfitting mitigation, accuracy-robustness trade-off methods, adaptive perturbation budgets, robust feature distillation, contrastive robust learning, information bottleneck defenses, feature disentanglement, latent consistency, and representation analyses of robust versus non-robust features.","exclusion":"Methods belong to certified robustness when the central contribution is a formal or probabilistic guarantee rather than empirical robust training or representation learning.","nearest_boundary":"Certified and theoretical robustness","representative_topics":["confidence threshold reduction for simultaneous adversarial robustness and generalization improvement","adaptive-margin adversarial training for robustness of medical image segmentation and detection models","soft adversarial training via certified input subsets to preserve natural accuracy"]},{"direction_index":3,"label":"D3","name":"Certified and theoretical robustness","description":"Research whose primary objective is proving guarantees, limits, or foundational properties of adversarial robustness, poisoning robustness, verification, or robust learnability.","inclusion":"Include randomized smoothing, interval bound propagation, branch-and-bound verification, formal verification surveys, certified training, certified purification when guarantees are central, PAC or minimax robust learnability, theoretical vulnerability analysis, adversarial surrogate-risk theory, poisoning sample-complexity theory, and abstract robustness trade-off analysis.","exclusion":"Empirical defenses without guarantees belong to robust learning or inference-time defense depending on whether the intervention occurs during training or testing.","nearest_boundary":"Robust learning and representations","representative_topics":["classifier-specific certification of randomized smoothing to overcome dimensionality-induced robustness limits","certified adversarial robustness via pretrained diffusion denoising without retraining","guided diffusion purification for certified adversarial robustness against l-infinity attacks"]},{"direction_index":4,"label":"D4","name":"Inference-time adversarial defense","description":"Research whose primary objective is detecting, rejecting, transforming, denoising, purifying, localizing, or repairing adversarial inputs at test time without making robust training the central program.","inclusion":"Include input preprocessing, nonlocal denoising, diffusion or score-based purification without a primary certification claim, adversarial example detection, patch localization and removal, multimodal inconsistency detection, feature-based correction, test-time randomization, selective rejection, and inference-time semantic repair.","exclusion":"Training-loop changes belong to robust learning and representations; certified purification belongs to certified and theoretical robustness when the guarantee is the main contribution.","nearest_boundary":"Robust learning and representations","representative_topics":["adversarial correction via intermediate layer feature analysis and conditional distribution matching","data-centric adversarial defense via nonlocal denoising and luminance-based example generation","defending multimodal fusion models against single-source adversarial perturbations via cross-modal inconsistency detection"]},{"direction_index":5,"label":"D5","name":"Privacy leakage and model theft","description":"Research whose primary objective is extracting or protecting private training information, membership facts, model functionality, parameters, architecture, provenance, or intellectual property.","inclusion":"Include gradient inversion, model inversion, training-data reconstruction, membership inference, privacy defenses against those attacks, model extraction, model stealing, encoder or architecture recovery, side-channel model theft, watermarking, provenance verification, and active defenses against functionality cloning.","exclusion":"Federated poisoning, Byzantine aggregation, and malicious-client integrity belong to structured/collaborative integrity unless the main objective is privacy leakage through gradients or updates.","nearest_boundary":"Structured and collaborative integrity","representative_topics":["subject-level membership inference attacks in cross-silo federated learning","semantic data reconstruction attacks bypassing gradient obfuscation in federated learning","black-box deep feature reconstruction attacks using gan generator as distribution constraint"]},{"direction_index":6,"label":"D6","name":"Structured and collaborative integrity","description":"Research whose primary objective is attacking or preserving model integrity under adversarial constraints imposed by collaborative learning protocols or graph-structured data.","inclusion":"Include federated or distributed poisoning, Byzantine behavior, malicious-client filtering, robust aggregation, federated backdoors when the protocol is central, federated adversarial training, graph structure or feature perturbation attacks, graph poisoning, graph backdoors, GNN defenses, graph sanitization, and graph-specific robustness analysis.","exclusion":"General non-graph evasion and non-federated attack algorithms belong to evasion attack generation and evaluation; privacy leakage from federated gradients belongs to privacy leakage and model theft.","nearest_boundary":"Privacy leakage and model theft","representative_topics":["durable backdoor attacks in federated learning via low-magnitude parameter targeting","randomized byzantine gradient aggregation defense against tailored adversarial attacks in federated learning","defending centered clipping robust aggregation against byzantine attacks in federated learning"]},{"direction_index":7,"label":"D7","name":"Deployed modality robustness","description":"Research whose primary objective is adapting adversarial attacks or defenses to the constraints of deployed modalities, sensors, human-facing systems, cybersecurity detectors, language/multimodal models, biometric identity systems, or sequential decision agents.","inclusion":"Include physical-world patches, optical or sensor attacks, autonomous-driving perception attacks, aerial or infrared attacks, adversarial text and vision-language manipulation, biometric evasion or impersonation, speaker or face recognition attacks and defenses, malware or intrusion-detector robustness, cyber-physical security ML robustness, reinforcement-learning policy attacks and defenses, multi-agent robustness, reward poisoning, and bandit or online decision-system attacks.","exclusion":"Place work in evasion attack generation and evaluation when the main contribution is a general attack, transfer, query, or benchmark mechanism not tied to a deployment constraint or modality-specific objective.","nearest_boundary":"Evasion attack generation and evaluation","representative_topics":["light-based physical adversarial attacks on vision dnn models using laser spot optimization","physical adversarial patches for attacking monocular depth estimation in autonomous driving","multimodal adversarial attacks on vision-language pre-training models via joint image-text perturbation"]}],"split":"dev"}
{"episode_id":"F183_T2024-01","field_index":183,"field":"vision transformer","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Token-set efficiency","description":"Reducing Vision Transformer computation or memory by changing the number, scale, grouping, or routing of visual tokens processed by the model.","inclusion":"Include work whose primary objective is compute reduction through token or patch pruning, dropping, merging, clustering, sparsification, adaptive sampling, mixed-scale tokenization, early exiting, or conditional routing.","exclusion":"If the token set is retained and the main contribution is a cheaper attention operator or connectivity pattern, assign to efficient attention operators instead; if the main goal is whole-model compression or hardware deployment, assign to compression and deployment.","nearest_boundary":"D1","representative_topics":["dynamic spatial sparsification for accelerating vision transformers and cnns via asymmetric token pruning","patch dropout for efficient high-resolution vision transformer training","patch sampling schedule for accelerating vision transformer training and inference throughput"]},{"direction_index":1,"label":"D1","name":"Efficient attention operators","description":"Reducing the computational cost of visual context modeling by replacing, approximating, factorizing, masking, or structurally constraining the self-attention operator.","inclusion":"Include work whose primary objective is attention-mechanism efficiency through linear attention, local-global or windowed attention, sparse or mask-constrained attention, low-rank or softmax-free approximations, separable attention, spectral/frequency attention, or cross-axis factorization.","exclusion":"If efficiency is primarily achieved by removing, merging, or routing tokens, assign to token-set efficiency; if the attention change is part of a general-purpose backbone design rather than an operator-efficiency program, assign to backbone architecture and spatial priors.","nearest_boundary":"D0","representative_topics":["linear-complexity separable self-attention for efficient mobile vision transformers","linear-complexity key-only attention for local-global vision transformer efficiency","attention mask-based spatial locality encoding for efficient image transformers"]},{"direction_index":2,"label":"D2","name":"Compression and deployment","description":"Making Vision Transformers smaller, lower precision, lower latency, or practically deployable under mobile, edge, accelerator, memory, or training-resource constraints.","inclusion":"Include work whose primary objective is whole-model footprint or deployment efficiency through quantization, binarization, integer inference, structured pruning, factorization, knowledge distillation, compact/mobile architecture design, hardware-aware co-design, NAS for compact ViTs, or deployment-centered surveys.","exclusion":"If the main mechanism is token-count reduction, assign to token-set efficiency; if the main mechanism is an efficient attention operator studied independently of deployment constraints, assign to efficient attention operators.","nearest_boundary":"D0","representative_topics":["efficient pure vision transformer design for low-latency mobile deployment","integer-only quantization of vision transformers via dyadic arithmetic and bit-shifting non-linearity approximation","efficient hybrid vision transformer architecture for industrial edge deployment"]},{"direction_index":3,"label":"D3","name":"Pre-training and transfer","description":"Learning reusable ViT representations before, between, or during downstream use through self-supervised pre-training, synthetic pre-training, parameter-efficient adaptation, domain transfer, or low-label transfer.","inclusion":"Include work whose primary objective is masked image modeling, self-supervised relation or positional objectives, synthetic/formula-driven pre-training, prompt tuning, adapters, side or bypass modules, domain adaptation/generalization, further pre-training for target domains, or efficient fine-tuning of pre-trained ViTs.","exclusion":"If the main objective is compressing a pre-trained model for deployment, assign to compression and deployment; if the method is a task-specific architecture rather than a reusable pre-training or adaptation program, assign to task-specialized ViT modeling.","nearest_boundary":"D2","representative_topics":["self-supervised vision transformer pretraining using patch relation tasks for small-dataset adaptation","self-supervised positional label training for spatial structure learning in vision transformers","self-supervised vision transformer training via spatial and channel feature self-relations"]},{"direction_index":4,"label":"D4","name":"Backbones and spatial priors","description":"Designing general-purpose ViT backbones, token mixers, hybrid CNN-transformer structures, and spatial or geometric priors for stronger visual representation learning.","inclusion":"Include work whose primary objective is reusable backbone or token-mixer design, MetaFormer-style unification, MLP/convolution/spectral/high-order mixers, CNN-transformer dual paths, cross-scale fusion, convolutional embeddings, locality priors, positional encodings, equivariance, biologically inspired encodings, or scale-adaptive inductive bias.","exclusion":"If the architectural change is primarily an attention-complexity reduction, assign to efficient attention operators; if it is optimized mainly for mobile or constrained deployment, assign to compression and deployment; if it is primarily for a particular downstream task or domain, assign to task-specialized ViT modeling.","nearest_boundary":"D1","representative_topics":["biologically inspired peripheral vision encoding in vision transformers for image classification","locality-guided dual-task training for vision transformers on tiny datasets","convolutional embedding for inductive bias in hierarchical vision transformer backbones"]},{"direction_index":5,"label":"D5","name":"Robustness and security","description":"Understanding or improving Vision Transformer resistance to adversarial perturbations, backdoors, evasion attacks, spurious features, distribution shifts, and other security or robustness failures.","inclusion":"Include work whose primary objective is adversarial training, efficient adversarial training, adversarial attack or patch generation, backdoor attack or defense, certified or empirical adversarial robustness, or architectural/training changes motivated primarily by security resilience.","exclusion":"If robustness is incidental to pre-training, adaptation, or architecture design, assign to that primary program; if the work only analyzes behavior without an attack, defense, or robustness intervention, assign to learning behavior and explanation.","nearest_boundary":"D6","representative_topics":["patch-based defense of vision transformers against backdoor attacks via positional encoding manipulation","efficient adversarial training for vision transformers via attention-guided self-attention pruning","adversarial robustness of vision transformers via self-ensembling intermediate feature representations"]},{"direction_index":6,"label":"D6","name":"Learning behavior and explanation","description":"Improving, explaining, visualizing, benchmarking, or surveying ViT learning behavior, attention quality, token representations, training dynamics, scaling, and empirical properties.","inclusion":"Include work whose primary objective is attention or token regularization, token-level augmentation, token-label alignment, attention diversity or density, dropout in attention, trivial-attention suppression, fairness-aware attention, interpretability, visual analytics, representation analysis, supervision or transfer studies, broad surveys, scaling analysis, or holistic empirical benchmarking.","exclusion":"If attention modification is mainly for computational efficiency, assign to efficient attention operators; if analysis is packaged around a robustness attack or defense, assign to robustness and security; if the objective is adapting a pre-trained model, assign to pre-training and transfer.","nearest_boundary":"D5","representative_topics":["survey of vision transformer architectures, training tricks, and self-supervised learning methods","token-level image mixing with activation-guided labels for vision transformer augmentation","key-based dropout with decreasing schedule for self-attention regularization in vision transformers"]},{"direction_index":7,"label":"D7","name":"Task-specialized ViT modeling","description":"Adapting or designing Vision Transformer models for particular downstream tasks, domains, modalities, or application pipelines where task suitability is the main research objective.","inclusion":"Include work whose primary objective is ViT modeling for semantic segmentation, dense prediction, object detection, video or action recognition, remote sensing, medical imaging, image restoration, retrieval, denoising, diffusion generation, joint image-video learning, or other specified visual domains.","exclusion":"If the contribution is a task-agnostic backbone or token mixer, assign to backbones and spatial priors; if it is a reusable parameter-efficient transfer method, assign to pre-training and transfer; if it is a broad survey or benchmark without task-specific modeling, assign to learning behavior and explanation.","nearest_boundary":"D4","representative_topics":["self-attention-based transformer backbone for object detection replacing convolutional operations","intrinsically interpretable vision transformer for weakly supervised semantic segmentation via attribute-guided explanation","hierarchical local-global vision transformers for semantic segmentation and dense prediction"]}],"split":"dev"}
{"episode_id":"F183_T2024-07","field_index":183,"field":"vision transformer","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Token-set efficiency","description":"Reducing Vision Transformer computation or memory by changing the number, scale, grouping, or routing of visual tokens processed by the model.","inclusion":"Include work whose primary objective is compute reduction through token or patch pruning, dropping, merging, clustering, sparsification, adaptive sampling, mixed-scale tokenization, early exiting, or conditional routing.","exclusion":"If the token set is retained and the main contribution is a cheaper attention operator or connectivity pattern, assign to efficient attention operators instead; if the main goal is whole-model compression or hardware deployment, assign to compression and deployment.","nearest_boundary":"D1","representative_topics":["dynamic spatial sparsification for accelerating vision transformers and cnns via asymmetric token pruning","patch dropout for efficient high-resolution vision transformer training","patch sampling schedule for accelerating vision transformer training and inference throughput"]},{"direction_index":1,"label":"D1","name":"Efficient attention operators","description":"Reducing the computational cost of visual context modeling by replacing, approximating, factorizing, masking, or structurally constraining the self-attention operator.","inclusion":"Include work whose primary objective is attention-mechanism efficiency through linear attention, local-global or windowed attention, sparse or mask-constrained attention, low-rank or softmax-free approximations, separable attention, spectral/frequency attention, or cross-axis factorization.","exclusion":"If efficiency is primarily achieved by removing, merging, or routing tokens, assign to token-set efficiency; if the attention change is part of a general-purpose backbone design rather than an operator-efficiency program, assign to backbone architecture and spatial priors.","nearest_boundary":"D0","representative_topics":["linear-complexity separable self-attention for efficient mobile vision transformers","linear-complexity key-only attention for local-global vision transformer efficiency","attention mask-based spatial locality encoding for efficient image transformers"]},{"direction_index":2,"label":"D2","name":"Compression and deployment","description":"Making Vision Transformers smaller, lower precision, lower latency, or practically deployable under mobile, edge, accelerator, memory, or training-resource constraints.","inclusion":"Include work whose primary objective is whole-model footprint or deployment efficiency through quantization, binarization, integer inference, structured pruning, factorization, knowledge distillation, compact/mobile architecture design, hardware-aware co-design, NAS for compact ViTs, or deployment-centered surveys.","exclusion":"If the main mechanism is token-count reduction, assign to token-set efficiency; if the main mechanism is an efficient attention operator studied independently of deployment constraints, assign to efficient attention operators.","nearest_boundary":"D0","representative_topics":["efficient pure vision transformer design for low-latency mobile deployment","integer-only quantization of vision transformers via dyadic arithmetic and bit-shifting non-linearity approximation","efficient hybrid vision transformer architecture for industrial edge deployment"]},{"direction_index":3,"label":"D3","name":"Pre-training and transfer","description":"Learning reusable ViT representations before, between, or during downstream use through self-supervised pre-training, synthetic pre-training, parameter-efficient adaptation, domain transfer, or low-label transfer.","inclusion":"Include work whose primary objective is masked image modeling, self-supervised relation or positional objectives, synthetic/formula-driven pre-training, prompt tuning, adapters, side or bypass modules, domain adaptation/generalization, further pre-training for target domains, or efficient fine-tuning of pre-trained ViTs.","exclusion":"If the main objective is compressing a pre-trained model for deployment, assign to compression and deployment; if the method is a task-specific architecture rather than a reusable pre-training or adaptation program, assign to task-specialized ViT modeling.","nearest_boundary":"D2","representative_topics":["self-supervised vision transformer pretraining using patch relation tasks for small-dataset adaptation","self-supervised positional label training for spatial structure learning in vision transformers","self-supervised vision transformer training via spatial and channel feature self-relations"]},{"direction_index":4,"label":"D4","name":"Backbones and spatial priors","description":"Designing general-purpose ViT backbones, token mixers, hybrid CNN-transformer structures, and spatial or geometric priors for stronger visual representation learning.","inclusion":"Include work whose primary objective is reusable backbone or token-mixer design, MetaFormer-style unification, MLP/convolution/spectral/high-order mixers, CNN-transformer dual paths, cross-scale fusion, convolutional embeddings, locality priors, positional encodings, equivariance, biologically inspired encodings, or scale-adaptive inductive bias.","exclusion":"If the architectural change is primarily an attention-complexity reduction, assign to efficient attention operators; if it is optimized mainly for mobile or constrained deployment, assign to compression and deployment; if it is primarily for a particular downstream task or domain, assign to task-specialized ViT modeling.","nearest_boundary":"D1","representative_topics":["biologically inspired peripheral vision encoding in vision transformers for image classification","locality-guided dual-task training for vision transformers on tiny datasets","convolutional embedding for inductive bias in hierarchical vision transformer backbones"]},{"direction_index":5,"label":"D5","name":"Robustness and security","description":"Understanding or improving Vision Transformer resistance to adversarial perturbations, backdoors, evasion attacks, spurious features, distribution shifts, and other security or robustness failures.","inclusion":"Include work whose primary objective is adversarial training, efficient adversarial training, adversarial attack or patch generation, backdoor attack or defense, certified or empirical adversarial robustness, or architectural/training changes motivated primarily by security resilience.","exclusion":"If robustness is incidental to pre-training, adaptation, or architecture design, assign to that primary program; if the work only analyzes behavior without an attack, defense, or robustness intervention, assign to learning behavior and explanation.","nearest_boundary":"D6","representative_topics":["patch-based defense of vision transformers against backdoor attacks via positional encoding manipulation","efficient adversarial training for vision transformers via attention-guided self-attention pruning","adversarial robustness of vision transformers via self-ensembling intermediate feature representations"]},{"direction_index":6,"label":"D6","name":"Learning behavior and explanation","description":"Improving, explaining, visualizing, benchmarking, or surveying ViT learning behavior, attention quality, token representations, training dynamics, scaling, and empirical properties.","inclusion":"Include work whose primary objective is attention or token regularization, token-level augmentation, token-label alignment, attention diversity or density, dropout in attention, trivial-attention suppression, fairness-aware attention, interpretability, visual analytics, representation analysis, supervision or transfer studies, broad surveys, scaling analysis, or holistic empirical benchmarking.","exclusion":"If attention modification is mainly for computational efficiency, assign to efficient attention operators; if analysis is packaged around a robustness attack or defense, assign to robustness and security; if the objective is adapting a pre-trained model, assign to pre-training and transfer.","nearest_boundary":"D5","representative_topics":["survey of vision transformer architectures, training tricks, and self-supervised learning methods","token-level image mixing with activation-guided labels for vision transformer augmentation","key-based dropout with decreasing schedule for self-attention regularization in vision transformers"]},{"direction_index":7,"label":"D7","name":"Task-specialized ViT modeling","description":"Adapting or designing Vision Transformer models for particular downstream tasks, domains, modalities, or application pipelines where task suitability is the main research objective.","inclusion":"Include work whose primary objective is ViT modeling for semantic segmentation, dense prediction, object detection, video or action recognition, remote sensing, medical imaging, image restoration, retrieval, denoising, diffusion generation, joint image-video learning, or other specified visual domains.","exclusion":"If the contribution is a task-agnostic backbone or token mixer, assign to backbones and spatial priors; if it is a reusable parameter-efficient transfer method, assign to pre-training and transfer; if it is a broad survey or benchmark without task-specific modeling, assign to learning behavior and explanation.","nearest_boundary":"D4","representative_topics":["self-attention-based transformer backbone for object detection replacing convolutional operations","intrinsically interpretable vision transformer for weakly supervised semantic segmentation via attribute-guided explanation","hierarchical local-global vision transformers for semantic segmentation and dense prediction"]}],"split":"dev"}
{"episode_id":"F183_T2025-01","field_index":183,"field":"vision transformer","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Token-set efficiency","description":"Reducing Vision Transformer computation or memory by changing the number, scale, grouping, or routing of visual tokens processed by the model.","inclusion":"Include work whose primary objective is compute reduction through token or patch pruning, dropping, merging, clustering, sparsification, adaptive sampling, mixed-scale tokenization, early exiting, or conditional routing.","exclusion":"If the token set is retained and the main contribution is a cheaper attention operator or connectivity pattern, assign to efficient attention operators instead; if the main goal is whole-model compression or hardware deployment, assign to compression and deployment.","nearest_boundary":"D1","representative_topics":["dynamic spatial sparsification for accelerating vision transformers and cnns via asymmetric token pruning","patch dropout for efficient high-resolution vision transformer training","patch sampling schedule for accelerating vision transformer training and inference throughput"]},{"direction_index":1,"label":"D1","name":"Efficient attention operators","description":"Reducing the computational cost of visual context modeling by replacing, approximating, factorizing, masking, or structurally constraining the self-attention operator.","inclusion":"Include work whose primary objective is attention-mechanism efficiency through linear attention, local-global or windowed attention, sparse or mask-constrained attention, low-rank or softmax-free approximations, separable attention, spectral/frequency attention, or cross-axis factorization.","exclusion":"If efficiency is primarily achieved by removing, merging, or routing tokens, assign to token-set efficiency; if the attention change is part of a general-purpose backbone design rather than an operator-efficiency program, assign to backbone architecture and spatial priors.","nearest_boundary":"D0","representative_topics":["linear-complexity separable self-attention for efficient mobile vision transformers","linear-complexity key-only attention for local-global vision transformer efficiency","attention mask-based spatial locality encoding for efficient image transformers"]},{"direction_index":2,"label":"D2","name":"Compression and deployment","description":"Making Vision Transformers smaller, lower precision, lower latency, or practically deployable under mobile, edge, accelerator, memory, or training-resource constraints.","inclusion":"Include work whose primary objective is whole-model footprint or deployment efficiency through quantization, binarization, integer inference, structured pruning, factorization, knowledge distillation, compact/mobile architecture design, hardware-aware co-design, NAS for compact ViTs, or deployment-centered surveys.","exclusion":"If the main mechanism is token-count reduction, assign to token-set efficiency; if the main mechanism is an efficient attention operator studied independently of deployment constraints, assign to efficient attention operators.","nearest_boundary":"D0","representative_topics":["efficient pure vision transformer design for low-latency mobile deployment","integer-only quantization of vision transformers via dyadic arithmetic and bit-shifting non-linearity approximation","efficient hybrid vision transformer architecture for industrial edge deployment"]},{"direction_index":3,"label":"D3","name":"Pre-training and transfer","description":"Learning reusable ViT representations before, between, or during downstream use through self-supervised pre-training, synthetic pre-training, parameter-efficient adaptation, domain transfer, or low-label transfer.","inclusion":"Include work whose primary objective is masked image modeling, self-supervised relation or positional objectives, synthetic/formula-driven pre-training, prompt tuning, adapters, side or bypass modules, domain adaptation/generalization, further pre-training for target domains, or efficient fine-tuning of pre-trained ViTs.","exclusion":"If the main objective is compressing a pre-trained model for deployment, assign to compression and deployment; if the method is a task-specific architecture rather than a reusable pre-training or adaptation program, assign to task-specialized ViT modeling.","nearest_boundary":"D2","representative_topics":["self-supervised vision transformer pretraining using patch relation tasks for small-dataset adaptation","self-supervised positional label training for spatial structure learning in vision transformers","self-supervised vision transformer training via spatial and channel feature self-relations"]},{"direction_index":4,"label":"D4","name":"Backbones and spatial priors","description":"Designing general-purpose ViT backbones, token mixers, hybrid CNN-transformer structures, and spatial or geometric priors for stronger visual representation learning.","inclusion":"Include work whose primary objective is reusable backbone or token-mixer design, MetaFormer-style unification, MLP/convolution/spectral/high-order mixers, CNN-transformer dual paths, cross-scale fusion, convolutional embeddings, locality priors, positional encodings, equivariance, biologically inspired encodings, or scale-adaptive inductive bias.","exclusion":"If the architectural change is primarily an attention-complexity reduction, assign to efficient attention operators; if it is optimized mainly for mobile or constrained deployment, assign to compression and deployment; if it is primarily for a particular downstream task or domain, assign to task-specialized ViT modeling.","nearest_boundary":"D1","representative_topics":["biologically inspired peripheral vision encoding in vision transformers for image classification","locality-guided dual-task training for vision transformers on tiny datasets","convolutional embedding for inductive bias in hierarchical vision transformer backbones"]},{"direction_index":5,"label":"D5","name":"Robustness and security","description":"Understanding or improving Vision Transformer resistance to adversarial perturbations, backdoors, evasion attacks, spurious features, distribution shifts, and other security or robustness failures.","inclusion":"Include work whose primary objective is adversarial training, efficient adversarial training, adversarial attack or patch generation, backdoor attack or defense, certified or empirical adversarial robustness, or architectural/training changes motivated primarily by security resilience.","exclusion":"If robustness is incidental to pre-training, adaptation, or architecture design, assign to that primary program; if the work only analyzes behavior without an attack, defense, or robustness intervention, assign to learning behavior and explanation.","nearest_boundary":"D6","representative_topics":["patch-based defense of vision transformers against backdoor attacks via positional encoding manipulation","efficient adversarial training for vision transformers via attention-guided self-attention pruning","adversarial robustness of vision transformers via self-ensembling intermediate feature representations"]},{"direction_index":6,"label":"D6","name":"Learning behavior and explanation","description":"Improving, explaining, visualizing, benchmarking, or surveying ViT learning behavior, attention quality, token representations, training dynamics, scaling, and empirical properties.","inclusion":"Include work whose primary objective is attention or token regularization, token-level augmentation, token-label alignment, attention diversity or density, dropout in attention, trivial-attention suppression, fairness-aware attention, interpretability, visual analytics, representation analysis, supervision or transfer studies, broad surveys, scaling analysis, or holistic empirical benchmarking.","exclusion":"If attention modification is mainly for computational efficiency, assign to efficient attention operators; if analysis is packaged around a robustness attack or defense, assign to robustness and security; if the objective is adapting a pre-trained model, assign to pre-training and transfer.","nearest_boundary":"D5","representative_topics":["survey of vision transformer architectures, training tricks, and self-supervised learning methods","token-level image mixing with activation-guided labels for vision transformer augmentation","key-based dropout with decreasing schedule for self-attention regularization in vision transformers"]},{"direction_index":7,"label":"D7","name":"Task-specialized ViT modeling","description":"Adapting or designing Vision Transformer models for particular downstream tasks, domains, modalities, or application pipelines where task suitability is the main research objective.","inclusion":"Include work whose primary objective is ViT modeling for semantic segmentation, dense prediction, object detection, video or action recognition, remote sensing, medical imaging, image restoration, retrieval, denoising, diffusion generation, joint image-video learning, or other specified visual domains.","exclusion":"If the contribution is a task-agnostic backbone or token mixer, assign to backbones and spatial priors; if it is a reusable parameter-efficient transfer method, assign to pre-training and transfer; if it is a broad survey or benchmark without task-specific modeling, assign to learning behavior and explanation.","nearest_boundary":"D4","representative_topics":["self-attention-based transformer backbone for object detection replacing convolutional operations","intrinsically interpretable vision transformer for weakly supervised semantic segmentation via attribute-guided explanation","hierarchical local-global vision transformers for semantic segmentation and dense prediction"]}],"split":"dev"}
{"episode_id":"F183_T2025-07","field_index":183,"field":"vision transformer","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Token-set efficiency","description":"Reducing Vision Transformer computation or memory by changing the number, scale, grouping, or routing of visual tokens processed by the model.","inclusion":"Include work whose primary objective is compute reduction through token or patch pruning, dropping, merging, clustering, sparsification, adaptive sampling, mixed-scale tokenization, early exiting, or conditional routing.","exclusion":"If the token set is retained and the main contribution is a cheaper attention operator or connectivity pattern, assign to efficient attention operators instead; if the main goal is whole-model compression or hardware deployment, assign to compression and deployment.","nearest_boundary":"D1","representative_topics":["dynamic spatial sparsification for accelerating vision transformers and cnns via asymmetric token pruning","patch dropout for efficient high-resolution vision transformer training","patch sampling schedule for accelerating vision transformer training and inference throughput"]},{"direction_index":1,"label":"D1","name":"Efficient attention operators","description":"Reducing the computational cost of visual context modeling by replacing, approximating, factorizing, masking, or structurally constraining the self-attention operator.","inclusion":"Include work whose primary objective is attention-mechanism efficiency through linear attention, local-global or windowed attention, sparse or mask-constrained attention, low-rank or softmax-free approximations, separable attention, spectral/frequency attention, or cross-axis factorization.","exclusion":"If efficiency is primarily achieved by removing, merging, or routing tokens, assign to token-set efficiency; if the attention change is part of a general-purpose backbone design rather than an operator-efficiency program, assign to backbone architecture and spatial priors.","nearest_boundary":"D0","representative_topics":["linear-complexity separable self-attention for efficient mobile vision transformers","linear-complexity key-only attention for local-global vision transformer efficiency","attention mask-based spatial locality encoding for efficient image transformers"]},{"direction_index":2,"label":"D2","name":"Compression and deployment","description":"Making Vision Transformers smaller, lower precision, lower latency, or practically deployable under mobile, edge, accelerator, memory, or training-resource constraints.","inclusion":"Include work whose primary objective is whole-model footprint or deployment efficiency through quantization, binarization, integer inference, structured pruning, factorization, knowledge distillation, compact/mobile architecture design, hardware-aware co-design, NAS for compact ViTs, or deployment-centered surveys.","exclusion":"If the main mechanism is token-count reduction, assign to token-set efficiency; if the main mechanism is an efficient attention operator studied independently of deployment constraints, assign to efficient attention operators.","nearest_boundary":"D0","representative_topics":["efficient pure vision transformer design for low-latency mobile deployment","integer-only quantization of vision transformers via dyadic arithmetic and bit-shifting non-linearity approximation","efficient hybrid vision transformer architecture for industrial edge deployment"]},{"direction_index":3,"label":"D3","name":"Pre-training and transfer","description":"Learning reusable ViT representations before, between, or during downstream use through self-supervised pre-training, synthetic pre-training, parameter-efficient adaptation, domain transfer, or low-label transfer.","inclusion":"Include work whose primary objective is masked image modeling, self-supervised relation or positional objectives, synthetic/formula-driven pre-training, prompt tuning, adapters, side or bypass modules, domain adaptation/generalization, further pre-training for target domains, or efficient fine-tuning of pre-trained ViTs.","exclusion":"If the main objective is compressing a pre-trained model for deployment, assign to compression and deployment; if the method is a task-specific architecture rather than a reusable pre-training or adaptation program, assign to task-specialized ViT modeling.","nearest_boundary":"D2","representative_topics":["self-supervised vision transformer pretraining using patch relation tasks for small-dataset adaptation","self-supervised positional label training for spatial structure learning in vision transformers","self-supervised vision transformer training via spatial and channel feature self-relations"]},{"direction_index":4,"label":"D4","name":"Backbones and spatial priors","description":"Designing general-purpose ViT backbones, token mixers, hybrid CNN-transformer structures, and spatial or geometric priors for stronger visual representation learning.","inclusion":"Include work whose primary objective is reusable backbone or token-mixer design, MetaFormer-style unification, MLP/convolution/spectral/high-order mixers, CNN-transformer dual paths, cross-scale fusion, convolutional embeddings, locality priors, positional encodings, equivariance, biologically inspired encodings, or scale-adaptive inductive bias.","exclusion":"If the architectural change is primarily an attention-complexity reduction, assign to efficient attention operators; if it is optimized mainly for mobile or constrained deployment, assign to compression and deployment; if it is primarily for a particular downstream task or domain, assign to task-specialized ViT modeling.","nearest_boundary":"D1","representative_topics":["biologically inspired peripheral vision encoding in vision transformers for image classification","locality-guided dual-task training for vision transformers on tiny datasets","convolutional embedding for inductive bias in hierarchical vision transformer backbones"]},{"direction_index":5,"label":"D5","name":"Robustness and security","description":"Understanding or improving Vision Transformer resistance to adversarial perturbations, backdoors, evasion attacks, spurious features, distribution shifts, and other security or robustness failures.","inclusion":"Include work whose primary objective is adversarial training, efficient adversarial training, adversarial attack or patch generation, backdoor attack or defense, certified or empirical adversarial robustness, or architectural/training changes motivated primarily by security resilience.","exclusion":"If robustness is incidental to pre-training, adaptation, or architecture design, assign to that primary program; if the work only analyzes behavior without an attack, defense, or robustness intervention, assign to learning behavior and explanation.","nearest_boundary":"D6","representative_topics":["patch-based defense of vision transformers against backdoor attacks via positional encoding manipulation","efficient adversarial training for vision transformers via attention-guided self-attention pruning","adversarial robustness of vision transformers via self-ensembling intermediate feature representations"]},{"direction_index":6,"label":"D6","name":"Learning behavior and explanation","description":"Improving, explaining, visualizing, benchmarking, or surveying ViT learning behavior, attention quality, token representations, training dynamics, scaling, and empirical properties.","inclusion":"Include work whose primary objective is attention or token regularization, token-level augmentation, token-label alignment, attention diversity or density, dropout in attention, trivial-attention suppression, fairness-aware attention, interpretability, visual analytics, representation analysis, supervision or transfer studies, broad surveys, scaling analysis, or holistic empirical benchmarking.","exclusion":"If attention modification is mainly for computational efficiency, assign to efficient attention operators; if analysis is packaged around a robustness attack or defense, assign to robustness and security; if the objective is adapting a pre-trained model, assign to pre-training and transfer.","nearest_boundary":"D5","representative_topics":["survey of vision transformer architectures, training tricks, and self-supervised learning methods","token-level image mixing with activation-guided labels for vision transformer augmentation","key-based dropout with decreasing schedule for self-attention regularization in vision transformers"]},{"direction_index":7,"label":"D7","name":"Task-specialized ViT modeling","description":"Adapting or designing Vision Transformer models for particular downstream tasks, domains, modalities, or application pipelines where task suitability is the main research objective.","inclusion":"Include work whose primary objective is ViT modeling for semantic segmentation, dense prediction, object detection, video or action recognition, remote sensing, medical imaging, image restoration, retrieval, denoising, diffusion generation, joint image-video learning, or other specified visual domains.","exclusion":"If the contribution is a task-agnostic backbone or token mixer, assign to backbones and spatial priors; if it is a reusable parameter-efficient transfer method, assign to pre-training and transfer; if it is a broad survey or benchmark without task-specific modeling, assign to learning behavior and explanation.","nearest_boundary":"D4","representative_topics":["self-attention-based transformer backbone for object detection replacing convolutional operations","intrinsically interpretable vision transformer for weakly supervised semantic segmentation via attribute-guided explanation","hierarchical local-global vision transformers for semantic segmentation and dense prediction"]}],"split":"dev"}
{"episode_id":"F183_T2026-01","field_index":183,"field":"vision transformer","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Token-set efficiency","description":"Reducing Vision Transformer computation or memory by changing the number, scale, grouping, or routing of visual tokens processed by the model.","inclusion":"Include work whose primary objective is compute reduction through token or patch pruning, dropping, merging, clustering, sparsification, adaptive sampling, mixed-scale tokenization, early exiting, or conditional routing.","exclusion":"If the token set is retained and the main contribution is a cheaper attention operator or connectivity pattern, assign to efficient attention operators instead; if the main goal is whole-model compression or hardware deployment, assign to compression and deployment.","nearest_boundary":"D1","representative_topics":["dynamic spatial sparsification for accelerating vision transformers and cnns via asymmetric token pruning","patch dropout for efficient high-resolution vision transformer training","patch sampling schedule for accelerating vision transformer training and inference throughput"]},{"direction_index":1,"label":"D1","name":"Efficient attention operators","description":"Reducing the computational cost of visual context modeling by replacing, approximating, factorizing, masking, or structurally constraining the self-attention operator.","inclusion":"Include work whose primary objective is attention-mechanism efficiency through linear attention, local-global or windowed attention, sparse or mask-constrained attention, low-rank or softmax-free approximations, separable attention, spectral/frequency attention, or cross-axis factorization.","exclusion":"If efficiency is primarily achieved by removing, merging, or routing tokens, assign to token-set efficiency; if the attention change is part of a general-purpose backbone design rather than an operator-efficiency program, assign to backbone architecture and spatial priors.","nearest_boundary":"D0","representative_topics":["linear-complexity separable self-attention for efficient mobile vision transformers","linear-complexity key-only attention for local-global vision transformer efficiency","attention mask-based spatial locality encoding for efficient image transformers"]},{"direction_index":2,"label":"D2","name":"Compression and deployment","description":"Making Vision Transformers smaller, lower precision, lower latency, or practically deployable under mobile, edge, accelerator, memory, or training-resource constraints.","inclusion":"Include work whose primary objective is whole-model footprint or deployment efficiency through quantization, binarization, integer inference, structured pruning, factorization, knowledge distillation, compact/mobile architecture design, hardware-aware co-design, NAS for compact ViTs, or deployment-centered surveys.","exclusion":"If the main mechanism is token-count reduction, assign to token-set efficiency; if the main mechanism is an efficient attention operator studied independently of deployment constraints, assign to efficient attention operators.","nearest_boundary":"D0","representative_topics":["efficient pure vision transformer design for low-latency mobile deployment","integer-only quantization of vision transformers via dyadic arithmetic and bit-shifting non-linearity approximation","efficient hybrid vision transformer architecture for industrial edge deployment"]},{"direction_index":3,"label":"D3","name":"Pre-training and transfer","description":"Learning reusable ViT representations before, between, or during downstream use through self-supervised pre-training, synthetic pre-training, parameter-efficient adaptation, domain transfer, or low-label transfer.","inclusion":"Include work whose primary objective is masked image modeling, self-supervised relation or positional objectives, synthetic/formula-driven pre-training, prompt tuning, adapters, side or bypass modules, domain adaptation/generalization, further pre-training for target domains, or efficient fine-tuning of pre-trained ViTs.","exclusion":"If the main objective is compressing a pre-trained model for deployment, assign to compression and deployment; if the method is a task-specific architecture rather than a reusable pre-training or adaptation program, assign to task-specialized ViT modeling.","nearest_boundary":"D2","representative_topics":["self-supervised vision transformer pretraining using patch relation tasks for small-dataset adaptation","self-supervised positional label training for spatial structure learning in vision transformers","self-supervised vision transformer training via spatial and channel feature self-relations"]},{"direction_index":4,"label":"D4","name":"Backbones and spatial priors","description":"Designing general-purpose ViT backbones, token mixers, hybrid CNN-transformer structures, and spatial or geometric priors for stronger visual representation learning.","inclusion":"Include work whose primary objective is reusable backbone or token-mixer design, MetaFormer-style unification, MLP/convolution/spectral/high-order mixers, CNN-transformer dual paths, cross-scale fusion, convolutional embeddings, locality priors, positional encodings, equivariance, biologically inspired encodings, or scale-adaptive inductive bias.","exclusion":"If the architectural change is primarily an attention-complexity reduction, assign to efficient attention operators; if it is optimized mainly for mobile or constrained deployment, assign to compression and deployment; if it is primarily for a particular downstream task or domain, assign to task-specialized ViT modeling.","nearest_boundary":"D1","representative_topics":["biologically inspired peripheral vision encoding in vision transformers for image classification","locality-guided dual-task training for vision transformers on tiny datasets","convolutional embedding for inductive bias in hierarchical vision transformer backbones"]},{"direction_index":5,"label":"D5","name":"Robustness and security","description":"Understanding or improving Vision Transformer resistance to adversarial perturbations, backdoors, evasion attacks, spurious features, distribution shifts, and other security or robustness failures.","inclusion":"Include work whose primary objective is adversarial training, efficient adversarial training, adversarial attack or patch generation, backdoor attack or defense, certified or empirical adversarial robustness, or architectural/training changes motivated primarily by security resilience.","exclusion":"If robustness is incidental to pre-training, adaptation, or architecture design, assign to that primary program; if the work only analyzes behavior without an attack, defense, or robustness intervention, assign to learning behavior and explanation.","nearest_boundary":"D6","representative_topics":["patch-based defense of vision transformers against backdoor attacks via positional encoding manipulation","efficient adversarial training for vision transformers via attention-guided self-attention pruning","adversarial robustness of vision transformers via self-ensembling intermediate feature representations"]},{"direction_index":6,"label":"D6","name":"Learning behavior and explanation","description":"Improving, explaining, visualizing, benchmarking, or surveying ViT learning behavior, attention quality, token representations, training dynamics, scaling, and empirical properties.","inclusion":"Include work whose primary objective is attention or token regularization, token-level augmentation, token-label alignment, attention diversity or density, dropout in attention, trivial-attention suppression, fairness-aware attention, interpretability, visual analytics, representation analysis, supervision or transfer studies, broad surveys, scaling analysis, or holistic empirical benchmarking.","exclusion":"If attention modification is mainly for computational efficiency, assign to efficient attention operators; if analysis is packaged around a robustness attack or defense, assign to robustness and security; if the objective is adapting a pre-trained model, assign to pre-training and transfer.","nearest_boundary":"D5","representative_topics":["survey of vision transformer architectures, training tricks, and self-supervised learning methods","token-level image mixing with activation-guided labels for vision transformer augmentation","key-based dropout with decreasing schedule for self-attention regularization in vision transformers"]},{"direction_index":7,"label":"D7","name":"Task-specialized ViT modeling","description":"Adapting or designing Vision Transformer models for particular downstream tasks, domains, modalities, or application pipelines where task suitability is the main research objective.","inclusion":"Include work whose primary objective is ViT modeling for semantic segmentation, dense prediction, object detection, video or action recognition, remote sensing, medical imaging, image restoration, retrieval, denoising, diffusion generation, joint image-video learning, or other specified visual domains.","exclusion":"If the contribution is a task-agnostic backbone or token mixer, assign to backbones and spatial priors; if it is a reusable parameter-efficient transfer method, assign to pre-training and transfer; if it is a broad survey or benchmark without task-specific modeling, assign to learning behavior and explanation.","nearest_boundary":"D4","representative_topics":["self-attention-based transformer backbone for object detection replacing convolutional operations","intrinsically interpretable vision transformer for weakly supervised semantic segmentation via attribute-guided explanation","hierarchical local-global vision transformers for semantic segmentation and dense prediction"]}],"split":"dev"}
{"episode_id":"F186_T2024-01","field_index":186,"field":"backdoor attacks","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Federated backdoor security","description":"Backdoor attack and defense programs whose primary objective is to exploit or protect federated-learning training, aggregation, client updates, and distributed population effects.","inclusion":"Include source atoms focused on FL-specific malicious update construction, persistent or single-shot FL poisoning, vertical or horizontal FL triggers, robust aggregation, client-update filtering, clustering, Bayesian or gradient inspection, and mitigation under FL constraints.","exclusion":"Centralized data poisoning, post-training repair, or generic model-level detection belongs to the corresponding centralized attack or defense direction unless federated aggregation or client behavior is central.","nearest_boundary":"Centralized defense and repair","representative_topics":["durable backdoor attacks in federated learning via low-magnitude parameter targeting","backdoor federated learning via population distribution alignment to overcome dilution effects","knowledge distillation-based backdoor attacks bypassing federated learning defenses via model smoothing"]},{"direction_index":1,"label":"D1","name":"Textual backdoor security","description":"Backdoor attack and defense programs whose primary objective is to create, detect, or mitigate backdoors in NLP, language-model, prompt, code, or sequence-generation systems using text-specific properties.","inclusion":"Include textual trigger construction, prompt-as-trigger attacks, lexical or style triggers, generative trigger attacks, sequence-to-sequence backdoors, masked-LM scoring defenses, paraphrase or perturbation defenses, mutation testing, textual feature anomaly detection, and contrastive defenses for NLP models.","exclusion":"Foundation-model or multimodal pretraining backdoors belong to foundation-model backdoor pipelines when transfer, pretraining, diffusion, or vision-language alignment is the central objective; domain-agnostic post-training detection belongs to centralized defense and repair.","nearest_boundary":"Foundation-model backdoor pipelines","representative_topics":["training-free lexical backdoor attacks on language models via tokenizer embedding manipulation","input-unique backdoor attacks using language model-generated triggers in nlp","black-box generative model-based stealthy textual backdoor attacks on classifiers"]},{"direction_index":2,"label":"D2","name":"Graph backdoor security","description":"Backdoor attack and defense programs whose primary objective is to exploit or protect graph neural networks and graph-structured data through topology-aware triggers, motifs, node or edge injection, graph embeddings, or graph explainability.","inclusion":"Include graph-specific trigger design, motif or semantic node backdoors, trigger placement in graph topology, graph embedding analysis, subgraph attribution, and explanation-guided detection or mitigation for GNN backdoors.","exclusion":"Non-graph structured tasks such as point clouds, object detection, power systems, trajectories, tabular data, or time series belong to specialized task and physical deployment attacks unless graph topology itself is the central object.","nearest_boundary":"Specialized task and physical deployment attacks","representative_topics":["explainability-driven backdoor detection and defense in graph neural networks","motif-driven backdoor trigger design for graph neural network attacks","semantic backdoor attacks on graph convolutional networks using node-level triggers"]},{"direction_index":3,"label":"D3","name":"Foundation-model backdoor pipelines","description":"Backdoor programs whose primary objective depends on pretraining, fine-tuning, self-supervised learning, contrastive learning, diffusion modeling, multimodal alignment, prompt tuning, or reuse of large pretrained models.","inclusion":"Include backdoor injection, transfer, detection, or mitigation where the central issue is foundation-model training or adaptation, including contrastive encoders, text-to-image systems, diffusion models, masked image modeling, self-supervised encoders, vision-language models, and fine-tuned language models.","exclusion":"Ordinary task-specific text, vision, or graph backdoors belong to their modality or attack-design directions when pretraining transfer or foundation-model adaptation is not central; supply-chain or hardware insertion without a foundation-model pipeline belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Production-pipeline risk and field infrastructure","representative_topics":["backdoor attack threat modeling across masked image modeling supply chain phases","backdoor attacks on text encoders for text-to-image synthesis via trigger-based injection","data poisoning backdoor attacks on contrastive learning via theory-guided poisoned input generation"]},{"direction_index":4,"label":"D4","name":"Stealthy digital trigger attacks","description":"Centralized non-FL backdoor attack design whose primary objective is improving digital trigger stealth, invisibility, clean-label behavior, poisoning efficiency, transform-domain operation, or attack optimization in standard vision-like classification settings.","inclusion":"Include frequency-domain triggers, clean-label image backdoors, generative or optimized trigger synthesis, low-poisoning-rate attacks, spatial transformation triggers, submodular sample selection, feature-repair triggers, and detection-evasive centralized poisoning when the main contribution is trigger or poison optimization.","exclusion":"Attacks whose main novelty is language syntax, FL aggregation, graph topology, foundation-model transfer, physical deployment, or task-specific dense/structured prediction belong to those specialized directions.","nearest_boundary":"Specialized task and physical deployment attacks","representative_topics":["image-specific two-phase triggers for stealthy clean-label backdoor attacks","frequency domain data poisoning for invisible backdoor attacks without mislabeling","submodular optimization of trojan trigger injection and strategic evasion of detection"]},{"direction_index":5,"label":"D5","name":"Specialized task and physical deployment attacks","description":"Backdoor attack programs whose primary objective is to make attacks work under nonstandard task structure, modality geometry, dense or regression outputs, sensor constraints, natural triggers, or real-world physical variation.","inclusion":"Include backdoor attacks on deep hashing, crowd counting, object detection, segmentation, 3D point clouds, power-system fault localization, autonomous-driving scenarios, audio or time-series tasks, and attacks centered on physical objects, natural features, capture effects, environmental conditions, or deployment-world transformations.","exclusion":"Generic image-classification digital trigger optimization belongs to stealthy digital trigger attacks; graph-topology triggers belong to graph backdoor security; benchmarking of physical datasets without attack construction belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Stealthy digital trigger attacks","representative_topics":["generative-based imperceptible backdoor attacks on deep hashing with clean label","backdoor attacks on crowd counting models via density map manipulation and data poisoning","imperceptible robust backdoor attacks on 3d point cloud via nonlinear local transformations"]},{"direction_index":6,"label":"D6","name":"Centralized defense and repair","description":"Non-FL, non-text-specialized defense programs whose primary objective is detecting, sanitizing, robustly training against, or removing backdoors in centralized models or datasets.","inclusion":"Include post-training model detection, trigger reverse-engineering, test-time alarms, confidence or activation-based detection, poisoned-sample filtering, robust training on poisoned data, self-training defenses, data clustering or auditing, model pruning, weight masking, fine-tuning, representation purification, data-free repair, and model hardening.","exclusion":"FL-specific detection or aggregation belongs to federated backdoor security; text-specific defense belongs to textual backdoor security; evaluation-only or field-level benchmarking belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Production-pipeline risk and field infrastructure","representative_topics":["backdoor detection via confidence-based distribution transfer and patch learning","black-box trojan backdoor defense using perturbation-based prediction confidence filtering","universal backdoor detection via adaptive adversarial probing of diverse trigger patterns"]},{"direction_index":7,"label":"D7","name":"Production-pipeline risk and field infrastructure","description":"Programs whose primary objective is to characterize, benchmark, tool, theorize, or operationalize backdoor risk at the system, supply-chain, hardware, compilation, evaluation, or field-infrastructure level rather than only proposing a standard attack or defense method.","inclusion":"Include hardware or compilation-stage insertion, weight or deployment-pathway manipulation, augmentation-pipeline poisoning, masked-image-model supply-chain analysis, benchmark suites, unified evaluation frameworks, toolboxes, natural backdoor datasets, systematic empirical or theoretical characterization, and backdoor-based auditing or benign repurposing such as watermarking or debiasing when grouped in methodological infrastructure atoms.","exclusion":"A paper whose primary contribution is a concrete centralized trigger attack belongs to stealthy digital trigger attacks; a concrete repair, sanitization, or detection algorithm belongs to centralized defense and repair; foundation-model pretraining attacks belong to foundation-model backdoor pipelines when transfer through pretrained models is central.","nearest_boundary":"Foundation-model backdoor pipelines","representative_topics":["hardware trojan insertion attacks on vision transformers via patch salience and dram bit flipping","compilation-stage backdoor insertion evading training and data preparation detection","backdoor attack threat modeling across masked image modeling supply chain phases"]}],"split":"dev"}
{"episode_id":"F186_T2024-07","field_index":186,"field":"backdoor attacks","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Federated backdoor security","description":"Backdoor attack and defense programs whose primary objective is to exploit or protect federated-learning training, aggregation, client updates, and distributed population effects.","inclusion":"Include source atoms focused on FL-specific malicious update construction, persistent or single-shot FL poisoning, vertical or horizontal FL triggers, robust aggregation, client-update filtering, clustering, Bayesian or gradient inspection, and mitigation under FL constraints.","exclusion":"Centralized data poisoning, post-training repair, or generic model-level detection belongs to the corresponding centralized attack or defense direction unless federated aggregation or client behavior is central.","nearest_boundary":"Centralized defense and repair","representative_topics":["durable backdoor attacks in federated learning via low-magnitude parameter targeting","backdoor federated learning via population distribution alignment to overcome dilution effects","knowledge distillation-based backdoor attacks bypassing federated learning defenses via model smoothing"]},{"direction_index":1,"label":"D1","name":"Textual backdoor security","description":"Backdoor attack and defense programs whose primary objective is to create, detect, or mitigate backdoors in NLP, language-model, prompt, code, or sequence-generation systems using text-specific properties.","inclusion":"Include textual trigger construction, prompt-as-trigger attacks, lexical or style triggers, generative trigger attacks, sequence-to-sequence backdoors, masked-LM scoring defenses, paraphrase or perturbation defenses, mutation testing, textual feature anomaly detection, and contrastive defenses for NLP models.","exclusion":"Foundation-model or multimodal pretraining backdoors belong to foundation-model backdoor pipelines when transfer, pretraining, diffusion, or vision-language alignment is the central objective; domain-agnostic post-training detection belongs to centralized defense and repair.","nearest_boundary":"Foundation-model backdoor pipelines","representative_topics":["training-free lexical backdoor attacks on language models via tokenizer embedding manipulation","input-unique backdoor attacks using language model-generated triggers in nlp","black-box generative model-based stealthy textual backdoor attacks on classifiers"]},{"direction_index":2,"label":"D2","name":"Graph backdoor security","description":"Backdoor attack and defense programs whose primary objective is to exploit or protect graph neural networks and graph-structured data through topology-aware triggers, motifs, node or edge injection, graph embeddings, or graph explainability.","inclusion":"Include graph-specific trigger design, motif or semantic node backdoors, trigger placement in graph topology, graph embedding analysis, subgraph attribution, and explanation-guided detection or mitigation for GNN backdoors.","exclusion":"Non-graph structured tasks such as point clouds, object detection, power systems, trajectories, tabular data, or time series belong to specialized task and physical deployment attacks unless graph topology itself is the central object.","nearest_boundary":"Specialized task and physical deployment attacks","representative_topics":["explainability-driven backdoor detection and defense in graph neural networks","motif-driven backdoor trigger design for graph neural network attacks","semantic backdoor attacks on graph convolutional networks using node-level triggers"]},{"direction_index":3,"label":"D3","name":"Foundation-model backdoor pipelines","description":"Backdoor programs whose primary objective depends on pretraining, fine-tuning, self-supervised learning, contrastive learning, diffusion modeling, multimodal alignment, prompt tuning, or reuse of large pretrained models.","inclusion":"Include backdoor injection, transfer, detection, or mitigation where the central issue is foundation-model training or adaptation, including contrastive encoders, text-to-image systems, diffusion models, masked image modeling, self-supervised encoders, vision-language models, and fine-tuned language models.","exclusion":"Ordinary task-specific text, vision, or graph backdoors belong to their modality or attack-design directions when pretraining transfer or foundation-model adaptation is not central; supply-chain or hardware insertion without a foundation-model pipeline belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Production-pipeline risk and field infrastructure","representative_topics":["backdoor attack threat modeling across masked image modeling supply chain phases","backdoor attacks on text encoders for text-to-image synthesis via trigger-based injection","data poisoning backdoor attacks on contrastive learning via theory-guided poisoned input generation"]},{"direction_index":4,"label":"D4","name":"Stealthy digital trigger attacks","description":"Centralized non-FL backdoor attack design whose primary objective is improving digital trigger stealth, invisibility, clean-label behavior, poisoning efficiency, transform-domain operation, or attack optimization in standard vision-like classification settings.","inclusion":"Include frequency-domain triggers, clean-label image backdoors, generative or optimized trigger synthesis, low-poisoning-rate attacks, spatial transformation triggers, submodular sample selection, feature-repair triggers, and detection-evasive centralized poisoning when the main contribution is trigger or poison optimization.","exclusion":"Attacks whose main novelty is language syntax, FL aggregation, graph topology, foundation-model transfer, physical deployment, or task-specific dense/structured prediction belong to those specialized directions.","nearest_boundary":"Specialized task and physical deployment attacks","representative_topics":["image-specific two-phase triggers for stealthy clean-label backdoor attacks","frequency domain data poisoning for invisible backdoor attacks without mislabeling","submodular optimization of trojan trigger injection and strategic evasion of detection"]},{"direction_index":5,"label":"D5","name":"Specialized task and physical deployment attacks","description":"Backdoor attack programs whose primary objective is to make attacks work under nonstandard task structure, modality geometry, dense or regression outputs, sensor constraints, natural triggers, or real-world physical variation.","inclusion":"Include backdoor attacks on deep hashing, crowd counting, object detection, segmentation, 3D point clouds, power-system fault localization, autonomous-driving scenarios, audio or time-series tasks, and attacks centered on physical objects, natural features, capture effects, environmental conditions, or deployment-world transformations.","exclusion":"Generic image-classification digital trigger optimization belongs to stealthy digital trigger attacks; graph-topology triggers belong to graph backdoor security; benchmarking of physical datasets without attack construction belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Stealthy digital trigger attacks","representative_topics":["generative-based imperceptible backdoor attacks on deep hashing with clean label","backdoor attacks on crowd counting models via density map manipulation and data poisoning","imperceptible robust backdoor attacks on 3d point cloud via nonlinear local transformations"]},{"direction_index":6,"label":"D6","name":"Centralized defense and repair","description":"Non-FL, non-text-specialized defense programs whose primary objective is detecting, sanitizing, robustly training against, or removing backdoors in centralized models or datasets.","inclusion":"Include post-training model detection, trigger reverse-engineering, test-time alarms, confidence or activation-based detection, poisoned-sample filtering, robust training on poisoned data, self-training defenses, data clustering or auditing, model pruning, weight masking, fine-tuning, representation purification, data-free repair, and model hardening.","exclusion":"FL-specific detection or aggregation belongs to federated backdoor security; text-specific defense belongs to textual backdoor security; evaluation-only or field-level benchmarking belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Production-pipeline risk and field infrastructure","representative_topics":["backdoor detection via confidence-based distribution transfer and patch learning","black-box trojan backdoor defense using perturbation-based prediction confidence filtering","universal backdoor detection via adaptive adversarial probing of diverse trigger patterns"]},{"direction_index":7,"label":"D7","name":"Production-pipeline risk and field infrastructure","description":"Programs whose primary objective is to characterize, benchmark, tool, theorize, or operationalize backdoor risk at the system, supply-chain, hardware, compilation, evaluation, or field-infrastructure level rather than only proposing a standard attack or defense method.","inclusion":"Include hardware or compilation-stage insertion, weight or deployment-pathway manipulation, augmentation-pipeline poisoning, masked-image-model supply-chain analysis, benchmark suites, unified evaluation frameworks, toolboxes, natural backdoor datasets, systematic empirical or theoretical characterization, and backdoor-based auditing or benign repurposing such as watermarking or debiasing when grouped in methodological infrastructure atoms.","exclusion":"A paper whose primary contribution is a concrete centralized trigger attack belongs to stealthy digital trigger attacks; a concrete repair, sanitization, or detection algorithm belongs to centralized defense and repair; foundation-model pretraining attacks belong to foundation-model backdoor pipelines when transfer through pretrained models is central.","nearest_boundary":"Foundation-model backdoor pipelines","representative_topics":["hardware trojan insertion attacks on vision transformers via patch salience and dram bit flipping","compilation-stage backdoor insertion evading training and data preparation detection","backdoor attack threat modeling across masked image modeling supply chain phases"]}],"split":"dev"}
{"episode_id":"F186_T2025-01","field_index":186,"field":"backdoor attacks","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Federated backdoor security","description":"Backdoor attack and defense programs whose primary objective is to exploit or protect federated-learning training, aggregation, client updates, and distributed population effects.","inclusion":"Include source atoms focused on FL-specific malicious update construction, persistent or single-shot FL poisoning, vertical or horizontal FL triggers, robust aggregation, client-update filtering, clustering, Bayesian or gradient inspection, and mitigation under FL constraints.","exclusion":"Centralized data poisoning, post-training repair, or generic model-level detection belongs to the corresponding centralized attack or defense direction unless federated aggregation or client behavior is central.","nearest_boundary":"Centralized defense and repair","representative_topics":["durable backdoor attacks in federated learning via low-magnitude parameter targeting","backdoor federated learning via population distribution alignment to overcome dilution effects","knowledge distillation-based backdoor attacks bypassing federated learning defenses via model smoothing"]},{"direction_index":1,"label":"D1","name":"Textual backdoor security","description":"Backdoor attack and defense programs whose primary objective is to create, detect, or mitigate backdoors in NLP, language-model, prompt, code, or sequence-generation systems using text-specific properties.","inclusion":"Include textual trigger construction, prompt-as-trigger attacks, lexical or style triggers, generative trigger attacks, sequence-to-sequence backdoors, masked-LM scoring defenses, paraphrase or perturbation defenses, mutation testing, textual feature anomaly detection, and contrastive defenses for NLP models.","exclusion":"Foundation-model or multimodal pretraining backdoors belong to foundation-model backdoor pipelines when transfer, pretraining, diffusion, or vision-language alignment is the central objective; domain-agnostic post-training detection belongs to centralized defense and repair.","nearest_boundary":"Foundation-model backdoor pipelines","representative_topics":["training-free lexical backdoor attacks on language models via tokenizer embedding manipulation","input-unique backdoor attacks using language model-generated triggers in nlp","black-box generative model-based stealthy textual backdoor attacks on classifiers"]},{"direction_index":2,"label":"D2","name":"Graph backdoor security","description":"Backdoor attack and defense programs whose primary objective is to exploit or protect graph neural networks and graph-structured data through topology-aware triggers, motifs, node or edge injection, graph embeddings, or graph explainability.","inclusion":"Include graph-specific trigger design, motif or semantic node backdoors, trigger placement in graph topology, graph embedding analysis, subgraph attribution, and explanation-guided detection or mitigation for GNN backdoors.","exclusion":"Non-graph structured tasks such as point clouds, object detection, power systems, trajectories, tabular data, or time series belong to specialized task and physical deployment attacks unless graph topology itself is the central object.","nearest_boundary":"Specialized task and physical deployment attacks","representative_topics":["explainability-driven backdoor detection and defense in graph neural networks","motif-driven backdoor trigger design for graph neural network attacks","semantic backdoor attacks on graph convolutional networks using node-level triggers"]},{"direction_index":3,"label":"D3","name":"Foundation-model backdoor pipelines","description":"Backdoor programs whose primary objective depends on pretraining, fine-tuning, self-supervised learning, contrastive learning, diffusion modeling, multimodal alignment, prompt tuning, or reuse of large pretrained models.","inclusion":"Include backdoor injection, transfer, detection, or mitigation where the central issue is foundation-model training or adaptation, including contrastive encoders, text-to-image systems, diffusion models, masked image modeling, self-supervised encoders, vision-language models, and fine-tuned language models.","exclusion":"Ordinary task-specific text, vision, or graph backdoors belong to their modality or attack-design directions when pretraining transfer or foundation-model adaptation is not central; supply-chain or hardware insertion without a foundation-model pipeline belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Production-pipeline risk and field infrastructure","representative_topics":["backdoor attack threat modeling across masked image modeling supply chain phases","backdoor attacks on text encoders for text-to-image synthesis via trigger-based injection","data poisoning backdoor attacks on contrastive learning via theory-guided poisoned input generation"]},{"direction_index":4,"label":"D4","name":"Stealthy digital trigger attacks","description":"Centralized non-FL backdoor attack design whose primary objective is improving digital trigger stealth, invisibility, clean-label behavior, poisoning efficiency, transform-domain operation, or attack optimization in standard vision-like classification settings.","inclusion":"Include frequency-domain triggers, clean-label image backdoors, generative or optimized trigger synthesis, low-poisoning-rate attacks, spatial transformation triggers, submodular sample selection, feature-repair triggers, and detection-evasive centralized poisoning when the main contribution is trigger or poison optimization.","exclusion":"Attacks whose main novelty is language syntax, FL aggregation, graph topology, foundation-model transfer, physical deployment, or task-specific dense/structured prediction belong to those specialized directions.","nearest_boundary":"Specialized task and physical deployment attacks","representative_topics":["image-specific two-phase triggers for stealthy clean-label backdoor attacks","frequency domain data poisoning for invisible backdoor attacks without mislabeling","submodular optimization of trojan trigger injection and strategic evasion of detection"]},{"direction_index":5,"label":"D5","name":"Specialized task and physical deployment attacks","description":"Backdoor attack programs whose primary objective is to make attacks work under nonstandard task structure, modality geometry, dense or regression outputs, sensor constraints, natural triggers, or real-world physical variation.","inclusion":"Include backdoor attacks on deep hashing, crowd counting, object detection, segmentation, 3D point clouds, power-system fault localization, autonomous-driving scenarios, audio or time-series tasks, and attacks centered on physical objects, natural features, capture effects, environmental conditions, or deployment-world transformations.","exclusion":"Generic image-classification digital trigger optimization belongs to stealthy digital trigger attacks; graph-topology triggers belong to graph backdoor security; benchmarking of physical datasets without attack construction belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Stealthy digital trigger attacks","representative_topics":["generative-based imperceptible backdoor attacks on deep hashing with clean label","backdoor attacks on crowd counting models via density map manipulation and data poisoning","imperceptible robust backdoor attacks on 3d point cloud via nonlinear local transformations"]},{"direction_index":6,"label":"D6","name":"Centralized defense and repair","description":"Non-FL, non-text-specialized defense programs whose primary objective is detecting, sanitizing, robustly training against, or removing backdoors in centralized models or datasets.","inclusion":"Include post-training model detection, trigger reverse-engineering, test-time alarms, confidence or activation-based detection, poisoned-sample filtering, robust training on poisoned data, self-training defenses, data clustering or auditing, model pruning, weight masking, fine-tuning, representation purification, data-free repair, and model hardening.","exclusion":"FL-specific detection or aggregation belongs to federated backdoor security; text-specific defense belongs to textual backdoor security; evaluation-only or field-level benchmarking belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Production-pipeline risk and field infrastructure","representative_topics":["backdoor detection via confidence-based distribution transfer and patch learning","black-box trojan backdoor defense using perturbation-based prediction confidence filtering","universal backdoor detection via adaptive adversarial probing of diverse trigger patterns"]},{"direction_index":7,"label":"D7","name":"Production-pipeline risk and field infrastructure","description":"Programs whose primary objective is to characterize, benchmark, tool, theorize, or operationalize backdoor risk at the system, supply-chain, hardware, compilation, evaluation, or field-infrastructure level rather than only proposing a standard attack or defense method.","inclusion":"Include hardware or compilation-stage insertion, weight or deployment-pathway manipulation, augmentation-pipeline poisoning, masked-image-model supply-chain analysis, benchmark suites, unified evaluation frameworks, toolboxes, natural backdoor datasets, systematic empirical or theoretical characterization, and backdoor-based auditing or benign repurposing such as watermarking or debiasing when grouped in methodological infrastructure atoms.","exclusion":"A paper whose primary contribution is a concrete centralized trigger attack belongs to stealthy digital trigger attacks; a concrete repair, sanitization, or detection algorithm belongs to centralized defense and repair; foundation-model pretraining attacks belong to foundation-model backdoor pipelines when transfer through pretrained models is central.","nearest_boundary":"Foundation-model backdoor pipelines","representative_topics":["hardware trojan insertion attacks on vision transformers via patch salience and dram bit flipping","compilation-stage backdoor insertion evading training and data preparation detection","backdoor attack threat modeling across masked image modeling supply chain phases"]}],"split":"dev"}
{"episode_id":"F186_T2025-07","field_index":186,"field":"backdoor attacks","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Federated backdoor security","description":"Backdoor attack and defense programs whose primary objective is to exploit or protect federated-learning training, aggregation, client updates, and distributed population effects.","inclusion":"Include source atoms focused on FL-specific malicious update construction, persistent or single-shot FL poisoning, vertical or horizontal FL triggers, robust aggregation, client-update filtering, clustering, Bayesian or gradient inspection, and mitigation under FL constraints.","exclusion":"Centralized data poisoning, post-training repair, or generic model-level detection belongs to the corresponding centralized attack or defense direction unless federated aggregation or client behavior is central.","nearest_boundary":"Centralized defense and repair","representative_topics":["durable backdoor attacks in federated learning via low-magnitude parameter targeting","backdoor federated learning via population distribution alignment to overcome dilution effects","knowledge distillation-based backdoor attacks bypassing federated learning defenses via model smoothing"]},{"direction_index":1,"label":"D1","name":"Textual backdoor security","description":"Backdoor attack and defense programs whose primary objective is to create, detect, or mitigate backdoors in NLP, language-model, prompt, code, or sequence-generation systems using text-specific properties.","inclusion":"Include textual trigger construction, prompt-as-trigger attacks, lexical or style triggers, generative trigger attacks, sequence-to-sequence backdoors, masked-LM scoring defenses, paraphrase or perturbation defenses, mutation testing, textual feature anomaly detection, and contrastive defenses for NLP models.","exclusion":"Foundation-model or multimodal pretraining backdoors belong to foundation-model backdoor pipelines when transfer, pretraining, diffusion, or vision-language alignment is the central objective; domain-agnostic post-training detection belongs to centralized defense and repair.","nearest_boundary":"Foundation-model backdoor pipelines","representative_topics":["training-free lexical backdoor attacks on language models via tokenizer embedding manipulation","input-unique backdoor attacks using language model-generated triggers in nlp","black-box generative model-based stealthy textual backdoor attacks on classifiers"]},{"direction_index":2,"label":"D2","name":"Graph backdoor security","description":"Backdoor attack and defense programs whose primary objective is to exploit or protect graph neural networks and graph-structured data through topology-aware triggers, motifs, node or edge injection, graph embeddings, or graph explainability.","inclusion":"Include graph-specific trigger design, motif or semantic node backdoors, trigger placement in graph topology, graph embedding analysis, subgraph attribution, and explanation-guided detection or mitigation for GNN backdoors.","exclusion":"Non-graph structured tasks such as point clouds, object detection, power systems, trajectories, tabular data, or time series belong to specialized task and physical deployment attacks unless graph topology itself is the central object.","nearest_boundary":"Specialized task and physical deployment attacks","representative_topics":["explainability-driven backdoor detection and defense in graph neural networks","motif-driven backdoor trigger design for graph neural network attacks","semantic backdoor attacks on graph convolutional networks using node-level triggers"]},{"direction_index":3,"label":"D3","name":"Foundation-model backdoor pipelines","description":"Backdoor programs whose primary objective depends on pretraining, fine-tuning, self-supervised learning, contrastive learning, diffusion modeling, multimodal alignment, prompt tuning, or reuse of large pretrained models.","inclusion":"Include backdoor injection, transfer, detection, or mitigation where the central issue is foundation-model training or adaptation, including contrastive encoders, text-to-image systems, diffusion models, masked image modeling, self-supervised encoders, vision-language models, and fine-tuned language models.","exclusion":"Ordinary task-specific text, vision, or graph backdoors belong to their modality or attack-design directions when pretraining transfer or foundation-model adaptation is not central; supply-chain or hardware insertion without a foundation-model pipeline belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Production-pipeline risk and field infrastructure","representative_topics":["backdoor attack threat modeling across masked image modeling supply chain phases","backdoor attacks on text encoders for text-to-image synthesis via trigger-based injection","data poisoning backdoor attacks on contrastive learning via theory-guided poisoned input generation"]},{"direction_index":4,"label":"D4","name":"Stealthy digital trigger attacks","description":"Centralized non-FL backdoor attack design whose primary objective is improving digital trigger stealth, invisibility, clean-label behavior, poisoning efficiency, transform-domain operation, or attack optimization in standard vision-like classification settings.","inclusion":"Include frequency-domain triggers, clean-label image backdoors, generative or optimized trigger synthesis, low-poisoning-rate attacks, spatial transformation triggers, submodular sample selection, feature-repair triggers, and detection-evasive centralized poisoning when the main contribution is trigger or poison optimization.","exclusion":"Attacks whose main novelty is language syntax, FL aggregation, graph topology, foundation-model transfer, physical deployment, or task-specific dense/structured prediction belong to those specialized directions.","nearest_boundary":"Specialized task and physical deployment attacks","representative_topics":["image-specific two-phase triggers for stealthy clean-label backdoor attacks","frequency domain data poisoning for invisible backdoor attacks without mislabeling","submodular optimization of trojan trigger injection and strategic evasion of detection"]},{"direction_index":5,"label":"D5","name":"Specialized task and physical deployment attacks","description":"Backdoor attack programs whose primary objective is to make attacks work under nonstandard task structure, modality geometry, dense or regression outputs, sensor constraints, natural triggers, or real-world physical variation.","inclusion":"Include backdoor attacks on deep hashing, crowd counting, object detection, segmentation, 3D point clouds, power-system fault localization, autonomous-driving scenarios, audio or time-series tasks, and attacks centered on physical objects, natural features, capture effects, environmental conditions, or deployment-world transformations.","exclusion":"Generic image-classification digital trigger optimization belongs to stealthy digital trigger attacks; graph-topology triggers belong to graph backdoor security; benchmarking of physical datasets without attack construction belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Stealthy digital trigger attacks","representative_topics":["generative-based imperceptible backdoor attacks on deep hashing with clean label","backdoor attacks on crowd counting models via density map manipulation and data poisoning","imperceptible robust backdoor attacks on 3d point cloud via nonlinear local transformations"]},{"direction_index":6,"label":"D6","name":"Centralized defense and repair","description":"Non-FL, non-text-specialized defense programs whose primary objective is detecting, sanitizing, robustly training against, or removing backdoors in centralized models or datasets.","inclusion":"Include post-training model detection, trigger reverse-engineering, test-time alarms, confidence or activation-based detection, poisoned-sample filtering, robust training on poisoned data, self-training defenses, data clustering or auditing, model pruning, weight masking, fine-tuning, representation purification, data-free repair, and model hardening.","exclusion":"FL-specific detection or aggregation belongs to federated backdoor security; text-specific defense belongs to textual backdoor security; evaluation-only or field-level benchmarking belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Production-pipeline risk and field infrastructure","representative_topics":["backdoor detection via confidence-based distribution transfer and patch learning","black-box trojan backdoor defense using perturbation-based prediction confidence filtering","universal backdoor detection via adaptive adversarial probing of diverse trigger patterns"]},{"direction_index":7,"label":"D7","name":"Production-pipeline risk and field infrastructure","description":"Programs whose primary objective is to characterize, benchmark, tool, theorize, or operationalize backdoor risk at the system, supply-chain, hardware, compilation, evaluation, or field-infrastructure level rather than only proposing a standard attack or defense method.","inclusion":"Include hardware or compilation-stage insertion, weight or deployment-pathway manipulation, augmentation-pipeline poisoning, masked-image-model supply-chain analysis, benchmark suites, unified evaluation frameworks, toolboxes, natural backdoor datasets, systematic empirical or theoretical characterization, and backdoor-based auditing or benign repurposing such as watermarking or debiasing when grouped in methodological infrastructure atoms.","exclusion":"A paper whose primary contribution is a concrete centralized trigger attack belongs to stealthy digital trigger attacks; a concrete repair, sanitization, or detection algorithm belongs to centralized defense and repair; foundation-model pretraining attacks belong to foundation-model backdoor pipelines when transfer through pretrained models is central.","nearest_boundary":"Foundation-model backdoor pipelines","representative_topics":["hardware trojan insertion attacks on vision transformers via patch salience and dram bit flipping","compilation-stage backdoor insertion evading training and data preparation detection","backdoor attack threat modeling across masked image modeling supply chain phases"]}],"split":"dev"}
{"episode_id":"F186_T2026-01","field_index":186,"field":"backdoor attacks","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Federated backdoor security","description":"Backdoor attack and defense programs whose primary objective is to exploit or protect federated-learning training, aggregation, client updates, and distributed population effects.","inclusion":"Include source atoms focused on FL-specific malicious update construction, persistent or single-shot FL poisoning, vertical or horizontal FL triggers, robust aggregation, client-update filtering, clustering, Bayesian or gradient inspection, and mitigation under FL constraints.","exclusion":"Centralized data poisoning, post-training repair, or generic model-level detection belongs to the corresponding centralized attack or defense direction unless federated aggregation or client behavior is central.","nearest_boundary":"Centralized defense and repair","representative_topics":["durable backdoor attacks in federated learning via low-magnitude parameter targeting","backdoor federated learning via population distribution alignment to overcome dilution effects","knowledge distillation-based backdoor attacks bypassing federated learning defenses via model smoothing"]},{"direction_index":1,"label":"D1","name":"Textual backdoor security","description":"Backdoor attack and defense programs whose primary objective is to create, detect, or mitigate backdoors in NLP, language-model, prompt, code, or sequence-generation systems using text-specific properties.","inclusion":"Include textual trigger construction, prompt-as-trigger attacks, lexical or style triggers, generative trigger attacks, sequence-to-sequence backdoors, masked-LM scoring defenses, paraphrase or perturbation defenses, mutation testing, textual feature anomaly detection, and contrastive defenses for NLP models.","exclusion":"Foundation-model or multimodal pretraining backdoors belong to foundation-model backdoor pipelines when transfer, pretraining, diffusion, or vision-language alignment is the central objective; domain-agnostic post-training detection belongs to centralized defense and repair.","nearest_boundary":"Foundation-model backdoor pipelines","representative_topics":["training-free lexical backdoor attacks on language models via tokenizer embedding manipulation","input-unique backdoor attacks using language model-generated triggers in nlp","black-box generative model-based stealthy textual backdoor attacks on classifiers"]},{"direction_index":2,"label":"D2","name":"Graph backdoor security","description":"Backdoor attack and defense programs whose primary objective is to exploit or protect graph neural networks and graph-structured data through topology-aware triggers, motifs, node or edge injection, graph embeddings, or graph explainability.","inclusion":"Include graph-specific trigger design, motif or semantic node backdoors, trigger placement in graph topology, graph embedding analysis, subgraph attribution, and explanation-guided detection or mitigation for GNN backdoors.","exclusion":"Non-graph structured tasks such as point clouds, object detection, power systems, trajectories, tabular data, or time series belong to specialized task and physical deployment attacks unless graph topology itself is the central object.","nearest_boundary":"Specialized task and physical deployment attacks","representative_topics":["explainability-driven backdoor detection and defense in graph neural networks","motif-driven backdoor trigger design for graph neural network attacks","semantic backdoor attacks on graph convolutional networks using node-level triggers"]},{"direction_index":3,"label":"D3","name":"Foundation-model backdoor pipelines","description":"Backdoor programs whose primary objective depends on pretraining, fine-tuning, self-supervised learning, contrastive learning, diffusion modeling, multimodal alignment, prompt tuning, or reuse of large pretrained models.","inclusion":"Include backdoor injection, transfer, detection, or mitigation where the central issue is foundation-model training or adaptation, including contrastive encoders, text-to-image systems, diffusion models, masked image modeling, self-supervised encoders, vision-language models, and fine-tuned language models.","exclusion":"Ordinary task-specific text, vision, or graph backdoors belong to their modality or attack-design directions when pretraining transfer or foundation-model adaptation is not central; supply-chain or hardware insertion without a foundation-model pipeline belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Production-pipeline risk and field infrastructure","representative_topics":["backdoor attack threat modeling across masked image modeling supply chain phases","backdoor attacks on text encoders for text-to-image synthesis via trigger-based injection","data poisoning backdoor attacks on contrastive learning via theory-guided poisoned input generation"]},{"direction_index":4,"label":"D4","name":"Stealthy digital trigger attacks","description":"Centralized non-FL backdoor attack design whose primary objective is improving digital trigger stealth, invisibility, clean-label behavior, poisoning efficiency, transform-domain operation, or attack optimization in standard vision-like classification settings.","inclusion":"Include frequency-domain triggers, clean-label image backdoors, generative or optimized trigger synthesis, low-poisoning-rate attacks, spatial transformation triggers, submodular sample selection, feature-repair triggers, and detection-evasive centralized poisoning when the main contribution is trigger or poison optimization.","exclusion":"Attacks whose main novelty is language syntax, FL aggregation, graph topology, foundation-model transfer, physical deployment, or task-specific dense/structured prediction belong to those specialized directions.","nearest_boundary":"Specialized task and physical deployment attacks","representative_topics":["image-specific two-phase triggers for stealthy clean-label backdoor attacks","frequency domain data poisoning for invisible backdoor attacks without mislabeling","submodular optimization of trojan trigger injection and strategic evasion of detection"]},{"direction_index":5,"label":"D5","name":"Specialized task and physical deployment attacks","description":"Backdoor attack programs whose primary objective is to make attacks work under nonstandard task structure, modality geometry, dense or regression outputs, sensor constraints, natural triggers, or real-world physical variation.","inclusion":"Include backdoor attacks on deep hashing, crowd counting, object detection, segmentation, 3D point clouds, power-system fault localization, autonomous-driving scenarios, audio or time-series tasks, and attacks centered on physical objects, natural features, capture effects, environmental conditions, or deployment-world transformations.","exclusion":"Generic image-classification digital trigger optimization belongs to stealthy digital trigger attacks; graph-topology triggers belong to graph backdoor security; benchmarking of physical datasets without attack construction belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Stealthy digital trigger attacks","representative_topics":["generative-based imperceptible backdoor attacks on deep hashing with clean label","backdoor attacks on crowd counting models via density map manipulation and data poisoning","imperceptible robust backdoor attacks on 3d point cloud via nonlinear local transformations"]},{"direction_index":6,"label":"D6","name":"Centralized defense and repair","description":"Non-FL, non-text-specialized defense programs whose primary objective is detecting, sanitizing, robustly training against, or removing backdoors in centralized models or datasets.","inclusion":"Include post-training model detection, trigger reverse-engineering, test-time alarms, confidence or activation-based detection, poisoned-sample filtering, robust training on poisoned data, self-training defenses, data clustering or auditing, model pruning, weight masking, fine-tuning, representation purification, data-free repair, and model hardening.","exclusion":"FL-specific detection or aggregation belongs to federated backdoor security; text-specific defense belongs to textual backdoor security; evaluation-only or field-level benchmarking belongs to production-pipeline risk and field infrastructure.","nearest_boundary":"Production-pipeline risk and field infrastructure","representative_topics":["backdoor detection via confidence-based distribution transfer and patch learning","black-box trojan backdoor defense using perturbation-based prediction confidence filtering","universal backdoor detection via adaptive adversarial probing of diverse trigger patterns"]},{"direction_index":7,"label":"D7","name":"Production-pipeline risk and field infrastructure","description":"Programs whose primary objective is to characterize, benchmark, tool, theorize, or operationalize backdoor risk at the system, supply-chain, hardware, compilation, evaluation, or field-infrastructure level rather than only proposing a standard attack or defense method.","inclusion":"Include hardware or compilation-stage insertion, weight or deployment-pathway manipulation, augmentation-pipeline poisoning, masked-image-model supply-chain analysis, benchmark suites, unified evaluation frameworks, toolboxes, natural backdoor datasets, systematic empirical or theoretical characterization, and backdoor-based auditing or benign repurposing such as watermarking or debiasing when grouped in methodological infrastructure atoms.","exclusion":"A paper whose primary contribution is a concrete centralized trigger attack belongs to stealthy digital trigger attacks; a concrete repair, sanitization, or detection algorithm belongs to centralized defense and repair; foundation-model pretraining attacks belong to foundation-model backdoor pipelines when transfer through pretrained models is central.","nearest_boundary":"Foundation-model backdoor pipelines","representative_topics":["hardware trojan insertion attacks on vision transformers via patch salience and dram bit flipping","compilation-stage backdoor insertion evading training and data preparation detection","backdoor attack threat modeling across masked image modeling supply chain phases"]}],"split":"dev"}
{"episode_id":"F195_T2024-01","field_index":195,"field":"talking head generation","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"General speech-driven synthesis","description":"End-to-end generation of realistic talking-head video or facial animation from speech or audio, emphasizing broad audiovisual coherence, identity preservation, natural motion, and generalization when no narrower objective dominates.","inclusion":"Include work whose primary objective is general audio-conditioned talking-face or portrait synthesis, including diffusion-based generators when diffusion is the implementation for broad synthesis rather than a separate scientific objective.","exclusion":"Assign to personalization, lip-sync editing, controllability, 3D representation, geometry transfer, contextual interaction, or deployment/restoration when one of those objectives is central.","nearest_boundary":"Sparse personalized avatar generation","representative_topics":["autoregressive diffusion models for reference-free talking face generation with head movement and expression","generalized audio-driven talking head synthesis via latent diffusion with identity-agnostic conditioning","diffusion autoencoder-based speech-driven talking face generation with data-driven latent representations and pose controllability"]},{"direction_index":1,"label":"D1","name":"Sparse personalized avatars","description":"Creating identity-preserving talking avatars from minimal person-specific evidence such as a single portrait, few-shot video, or fast adaptation data.","inclusion":"Include zero-, one-, or few-shot talking-head generation and rapid personalized adaptation where sparse identity acquisition is the main objective, regardless of backbone or rendering method.","exclusion":"Assign to general synthesis when identity transfer is not sparse-personalization-centered, and to 3D representation when the main contribution is the representation or renderer rather than minimal-data personalization.","nearest_boundary":"General speech-driven synthesis","representative_topics":["few-shot talking head synthesis via dynamic facial radiance fields conditioned on 2d appearance priors","one-shot audio-driven talking head generation with disentangled motion and lip synchronization","speech-driven portrait animation with controllable facial expressions and head pose from a single image"]},{"direction_index":2,"label":"D2","name":"Lip synchronization and dubbing","description":"Precise alignment, correction, generation, or editing of mouth and facial motion to match target speech, phonetic context, or dubbing audio.","inclusion":"Include work whose main objective is lip-audio synchrony, phoneme/viseme fidelity, dubbing, lip-region repair, synchronization losses, or lip-sync measurement in generated or existing videos.","exclusion":"Assign to general synthesis when lip sync is only one quality criterion, and to expression controllability when the main target is affective or semantic expression editing rather than mouth-speech alignment.","nearest_boundary":"General speech-driven synthesis","representative_topics":["audio-visual lip synchronization via phoneme-level memory retrieval in talking face generation","audio-driven lip synchronization for talking head video editing via sequential expression and face enhancement","phonetic context-aware lip synchronization for talking face generation using masked learning"]},{"direction_index":3,"label":"D3","name":"Disentangled expressive control","description":"Learning or using separable controls for facial motion factors such as pose, expression, emotion, speaking style, identity, and speech content to enable explicit manipulation of talking-head behavior.","inclusion":"Include work where factor separation, controllable latent spaces, expression or emotion control, and explicit manipulation of pose/style/expression are the primary research objective.","exclusion":"Assign to sparse personalized avatars when disentanglement mainly supports one-shot identity transfer, and to contextual multimodal interaction when the main objective is heterogeneous input fusion or conversational responsiveness.","nearest_boundary":"Contextual multimodal interaction","representative_topics":["progressive disentangled representation learning for multi-factor controllable talking head synthesis","disentangled head pose and expression control for audio-video-driven talking head generation","speech-driven 3d face animation with disentangled emotional and content embeddings"]},{"direction_index":4,"label":"D4","name":"Geometry-guided motion transfer","description":"Animating or reenacting talking heads by using explicit landmarks, keypoints, meshes, depth, displacement fields, or geometric priors to drive spatial warping and motion transfer.","inclusion":"Include work whose core objective is geometry- or keypoint-guided face warping, reenactment, displacement-field estimation, depth-aware transfer, or self-supervised 3D-aware motion transfer without making a persistent neural 3D scene representation the main claim.","exclusion":"Assign to 3D representation and animation when the main contribution is NeRF, neural volume rendering, implicit representation, tri-plane rendering, or mesh-sequence generation as the primary output.","nearest_boundary":"3D representation and animation","representative_topics":["geometry-guided progressive warping for one-shot face animation with displacement field estimation","audio-visual talking head reenactment using learnable keypoint motion fields and identity-aware generation","keypoint-based enhancement for naturalness in audio-driven free-viewpoint talking head synthesis"]},{"direction_index":5,"label":"D5","name":"3D representation and animation","description":"Modeling talking heads through 3D-aware representations or outputs, including neural radiance fields, neural volumes, tri-planes, implicit faces, 4D face models, meshes, blendshapes, and viseme curves.","inclusion":"Include work whose primary objective is view-consistent or high-fidelity 3D representation/rendering, neural scene modeling, or speech-driven 3D facial mesh/blendshape/viseme animation.","exclusion":"Assign to geometry-guided motion transfer when geometry is mainly used for 2D warping, and to deployment efficiency when the dominant claim is acceleration, compression, or real-time operation rather than the 3D representation itself.","nearest_boundary":"Geometry-guided motion transfer","representative_topics":["neural volume rendering for identity-preserving talking-head face animation","generalizable nerf-based talking face synthesis with out-of-domain audio and head-torso joint rendering","large-scale multi-modal 4d face dataset and non-autoregressive audio-driven 3d face animation"]},{"direction_index":6,"label":"D6","name":"Contextual multimodal interaction","description":"Generating talking-head behavior from semantic, affective, multimodal, stochastic, or conversational context beyond deterministic single-speaker audio-to-video mapping.","inclusion":"Include work whose main objective is text/speech/visual modality fusion, self-supervised cross-modal alignment, unified heterogeneous drivers, diverse stochastic motion, conversational head generation, listener response, turn-taking, or interactive nonverbal behavior.","exclusion":"Assign to disentangled expressive control when the main contribution is explicit manipulation of facial factors within a monologue setting, and to general synthesis when the system is standard audio-only generation without contextual, multimodal, stochastic, or interaction modeling.","nearest_boundary":"Disentangled expressive control","representative_topics":["unified facial landmark generation via shared text-to-speech latent representations","diffusion-based talking face generation with multimodal emotion and geometry-aware texture transfer","perceptual conversational head generation via regularized audio-to-head driver and enhanced rendering"]},{"direction_index":7,"label":"D7","name":"Refinement and deployability","description":"Improving generated talking-head systems for practical use through visual restoration, stabilization, super-resolution, relighting, stylized rendering, compression, acceleration, or real-time deployment.","inclusion":"Include work whose primary objective is post-generation or system-level improvement of quality, stability, identity fidelity, resolution, artifact repair, output-domain adaptation, model size, latency, training cost, or edge deployment.","exclusion":"Assign to general synthesis or 3D representation when quality or speed gains are incidental to the core generator, and to lip synchronization when the primary target is speech-mouth alignment rather than restoration or deployability.","nearest_boundary":"General speech-driven synthesis","representative_topics":["audio-visual prior-guided extreme-scale talking face video upsampling from 8x8 inputs","mitigating lip motion jitter in 3d-based speech-driven talking face generation","lightweight talking face generation via distillation and selective mixed-precision quantization"]}],"split":"dev"}
{"episode_id":"F195_T2024-07","field_index":195,"field":"talking head generation","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"General speech-driven synthesis","description":"End-to-end generation of realistic talking-head video or facial animation from speech or audio, emphasizing broad audiovisual coherence, identity preservation, natural motion, and generalization when no narrower objective dominates.","inclusion":"Include work whose primary objective is general audio-conditioned talking-face or portrait synthesis, including diffusion-based generators when diffusion is the implementation for broad synthesis rather than a separate scientific objective.","exclusion":"Assign to personalization, lip-sync editing, controllability, 3D representation, geometry transfer, contextual interaction, or deployment/restoration when one of those objectives is central.","nearest_boundary":"Sparse personalized avatar generation","representative_topics":["autoregressive diffusion models for reference-free talking face generation with head movement and expression","generalized audio-driven talking head synthesis via latent diffusion with identity-agnostic conditioning","diffusion autoencoder-based speech-driven talking face generation with data-driven latent representations and pose controllability"]},{"direction_index":1,"label":"D1","name":"Sparse personalized avatars","description":"Creating identity-preserving talking avatars from minimal person-specific evidence such as a single portrait, few-shot video, or fast adaptation data.","inclusion":"Include zero-, one-, or few-shot talking-head generation and rapid personalized adaptation where sparse identity acquisition is the main objective, regardless of backbone or rendering method.","exclusion":"Assign to general synthesis when identity transfer is not sparse-personalization-centered, and to 3D representation when the main contribution is the representation or renderer rather than minimal-data personalization.","nearest_boundary":"General speech-driven synthesis","representative_topics":["few-shot talking head synthesis via dynamic facial radiance fields conditioned on 2d appearance priors","one-shot audio-driven talking head generation with disentangled motion and lip synchronization","speech-driven portrait animation with controllable facial expressions and head pose from a single image"]},{"direction_index":2,"label":"D2","name":"Lip synchronization and dubbing","description":"Precise alignment, correction, generation, or editing of mouth and facial motion to match target speech, phonetic context, or dubbing audio.","inclusion":"Include work whose main objective is lip-audio synchrony, phoneme/viseme fidelity, dubbing, lip-region repair, synchronization losses, or lip-sync measurement in generated or existing videos.","exclusion":"Assign to general synthesis when lip sync is only one quality criterion, and to expression controllability when the main target is affective or semantic expression editing rather than mouth-speech alignment.","nearest_boundary":"General speech-driven synthesis","representative_topics":["audio-visual lip synchronization via phoneme-level memory retrieval in talking face generation","audio-driven lip synchronization for talking head video editing via sequential expression and face enhancement","phonetic context-aware lip synchronization for talking face generation using masked learning"]},{"direction_index":3,"label":"D3","name":"Disentangled expressive control","description":"Learning or using separable controls for facial motion factors such as pose, expression, emotion, speaking style, identity, and speech content to enable explicit manipulation of talking-head behavior.","inclusion":"Include work where factor separation, controllable latent spaces, expression or emotion control, and explicit manipulation of pose/style/expression are the primary research objective.","exclusion":"Assign to sparse personalized avatars when disentanglement mainly supports one-shot identity transfer, and to contextual multimodal interaction when the main objective is heterogeneous input fusion or conversational responsiveness.","nearest_boundary":"Contextual multimodal interaction","representative_topics":["progressive disentangled representation learning for multi-factor controllable talking head synthesis","disentangled head pose and expression control for audio-video-driven talking head generation","speech-driven 3d face animation with disentangled emotional and content embeddings"]},{"direction_index":4,"label":"D4","name":"Geometry-guided motion transfer","description":"Animating or reenacting talking heads by using explicit landmarks, keypoints, meshes, depth, displacement fields, or geometric priors to drive spatial warping and motion transfer.","inclusion":"Include work whose core objective is geometry- or keypoint-guided face warping, reenactment, displacement-field estimation, depth-aware transfer, or self-supervised 3D-aware motion transfer without making a persistent neural 3D scene representation the main claim.","exclusion":"Assign to 3D representation and animation when the main contribution is NeRF, neural volume rendering, implicit representation, tri-plane rendering, or mesh-sequence generation as the primary output.","nearest_boundary":"3D representation and animation","representative_topics":["geometry-guided progressive warping for one-shot face animation with displacement field estimation","audio-visual talking head reenactment using learnable keypoint motion fields and identity-aware generation","keypoint-based enhancement for naturalness in audio-driven free-viewpoint talking head synthesis"]},{"direction_index":5,"label":"D5","name":"3D representation and animation","description":"Modeling talking heads through 3D-aware representations or outputs, including neural radiance fields, neural volumes, tri-planes, implicit faces, 4D face models, meshes, blendshapes, and viseme curves.","inclusion":"Include work whose primary objective is view-consistent or high-fidelity 3D representation/rendering, neural scene modeling, or speech-driven 3D facial mesh/blendshape/viseme animation.","exclusion":"Assign to geometry-guided motion transfer when geometry is mainly used for 2D warping, and to deployment efficiency when the dominant claim is acceleration, compression, or real-time operation rather than the 3D representation itself.","nearest_boundary":"Geometry-guided motion transfer","representative_topics":["neural volume rendering for identity-preserving talking-head face animation","generalizable nerf-based talking face synthesis with out-of-domain audio and head-torso joint rendering","large-scale multi-modal 4d face dataset and non-autoregressive audio-driven 3d face animation"]},{"direction_index":6,"label":"D6","name":"Contextual multimodal interaction","description":"Generating talking-head behavior from semantic, affective, multimodal, stochastic, or conversational context beyond deterministic single-speaker audio-to-video mapping.","inclusion":"Include work whose main objective is text/speech/visual modality fusion, self-supervised cross-modal alignment, unified heterogeneous drivers, diverse stochastic motion, conversational head generation, listener response, turn-taking, or interactive nonverbal behavior.","exclusion":"Assign to disentangled expressive control when the main contribution is explicit manipulation of facial factors within a monologue setting, and to general synthesis when the system is standard audio-only generation without contextual, multimodal, stochastic, or interaction modeling.","nearest_boundary":"Disentangled expressive control","representative_topics":["unified facial landmark generation via shared text-to-speech latent representations","diffusion-based talking face generation with multimodal emotion and geometry-aware texture transfer","perceptual conversational head generation via regularized audio-to-head driver and enhanced rendering"]},{"direction_index":7,"label":"D7","name":"Refinement and deployability","description":"Improving generated talking-head systems for practical use through visual restoration, stabilization, super-resolution, relighting, stylized rendering, compression, acceleration, or real-time deployment.","inclusion":"Include work whose primary objective is post-generation or system-level improvement of quality, stability, identity fidelity, resolution, artifact repair, output-domain adaptation, model size, latency, training cost, or edge deployment.","exclusion":"Assign to general synthesis or 3D representation when quality or speed gains are incidental to the core generator, and to lip synchronization when the primary target is speech-mouth alignment rather than restoration or deployability.","nearest_boundary":"General speech-driven synthesis","representative_topics":["audio-visual prior-guided extreme-scale talking face video upsampling from 8x8 inputs","mitigating lip motion jitter in 3d-based speech-driven talking face generation","lightweight talking face generation via distillation and selective mixed-precision quantization"]}],"split":"dev"}
{"episode_id":"F195_T2025-01","field_index":195,"field":"talking head generation","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"General speech-driven synthesis","description":"End-to-end generation of realistic talking-head video or facial animation from speech or audio, emphasizing broad audiovisual coherence, identity preservation, natural motion, and generalization when no narrower objective dominates.","inclusion":"Include work whose primary objective is general audio-conditioned talking-face or portrait synthesis, including diffusion-based generators when diffusion is the implementation for broad synthesis rather than a separate scientific objective.","exclusion":"Assign to personalization, lip-sync editing, controllability, 3D representation, geometry transfer, contextual interaction, or deployment/restoration when one of those objectives is central.","nearest_boundary":"Sparse personalized avatar generation","representative_topics":["autoregressive diffusion models for reference-free talking face generation with head movement and expression","generalized audio-driven talking head synthesis via latent diffusion with identity-agnostic conditioning","diffusion autoencoder-based speech-driven talking face generation with data-driven latent representations and pose controllability"]},{"direction_index":1,"label":"D1","name":"Sparse personalized avatars","description":"Creating identity-preserving talking avatars from minimal person-specific evidence such as a single portrait, few-shot video, or fast adaptation data.","inclusion":"Include zero-, one-, or few-shot talking-head generation and rapid personalized adaptation where sparse identity acquisition is the main objective, regardless of backbone or rendering method.","exclusion":"Assign to general synthesis when identity transfer is not sparse-personalization-centered, and to 3D representation when the main contribution is the representation or renderer rather than minimal-data personalization.","nearest_boundary":"General speech-driven synthesis","representative_topics":["few-shot talking head synthesis via dynamic facial radiance fields conditioned on 2d appearance priors","one-shot audio-driven talking head generation with disentangled motion and lip synchronization","speech-driven portrait animation with controllable facial expressions and head pose from a single image"]},{"direction_index":2,"label":"D2","name":"Lip synchronization and dubbing","description":"Precise alignment, correction, generation, or editing of mouth and facial motion to match target speech, phonetic context, or dubbing audio.","inclusion":"Include work whose main objective is lip-audio synchrony, phoneme/viseme fidelity, dubbing, lip-region repair, synchronization losses, or lip-sync measurement in generated or existing videos.","exclusion":"Assign to general synthesis when lip sync is only one quality criterion, and to expression controllability when the main target is affective or semantic expression editing rather than mouth-speech alignment.","nearest_boundary":"General speech-driven synthesis","representative_topics":["audio-visual lip synchronization via phoneme-level memory retrieval in talking face generation","audio-driven lip synchronization for talking head video editing via sequential expression and face enhancement","phonetic context-aware lip synchronization for talking face generation using masked learning"]},{"direction_index":3,"label":"D3","name":"Disentangled expressive control","description":"Learning or using separable controls for facial motion factors such as pose, expression, emotion, speaking style, identity, and speech content to enable explicit manipulation of talking-head behavior.","inclusion":"Include work where factor separation, controllable latent spaces, expression or emotion control, and explicit manipulation of pose/style/expression are the primary research objective.","exclusion":"Assign to sparse personalized avatars when disentanglement mainly supports one-shot identity transfer, and to contextual multimodal interaction when the main objective is heterogeneous input fusion or conversational responsiveness.","nearest_boundary":"Contextual multimodal interaction","representative_topics":["progressive disentangled representation learning for multi-factor controllable talking head synthesis","disentangled head pose and expression control for audio-video-driven talking head generation","speech-driven 3d face animation with disentangled emotional and content embeddings"]},{"direction_index":4,"label":"D4","name":"Geometry-guided motion transfer","description":"Animating or reenacting talking heads by using explicit landmarks, keypoints, meshes, depth, displacement fields, or geometric priors to drive spatial warping and motion transfer.","inclusion":"Include work whose core objective is geometry- or keypoint-guided face warping, reenactment, displacement-field estimation, depth-aware transfer, or self-supervised 3D-aware motion transfer without making a persistent neural 3D scene representation the main claim.","exclusion":"Assign to 3D representation and animation when the main contribution is NeRF, neural volume rendering, implicit representation, tri-plane rendering, or mesh-sequence generation as the primary output.","nearest_boundary":"3D representation and animation","representative_topics":["geometry-guided progressive warping for one-shot face animation with displacement field estimation","audio-visual talking head reenactment using learnable keypoint motion fields and identity-aware generation","keypoint-based enhancement for naturalness in audio-driven free-viewpoint talking head synthesis"]},{"direction_index":5,"label":"D5","name":"3D representation and animation","description":"Modeling talking heads through 3D-aware representations or outputs, including neural radiance fields, neural volumes, tri-planes, implicit faces, 4D face models, meshes, blendshapes, and viseme curves.","inclusion":"Include work whose primary objective is view-consistent or high-fidelity 3D representation/rendering, neural scene modeling, or speech-driven 3D facial mesh/blendshape/viseme animation.","exclusion":"Assign to geometry-guided motion transfer when geometry is mainly used for 2D warping, and to deployment efficiency when the dominant claim is acceleration, compression, or real-time operation rather than the 3D representation itself.","nearest_boundary":"Geometry-guided motion transfer","representative_topics":["neural volume rendering for identity-preserving talking-head face animation","generalizable nerf-based talking face synthesis with out-of-domain audio and head-torso joint rendering","large-scale multi-modal 4d face dataset and non-autoregressive audio-driven 3d face animation"]},{"direction_index":6,"label":"D6","name":"Contextual multimodal interaction","description":"Generating talking-head behavior from semantic, affective, multimodal, stochastic, or conversational context beyond deterministic single-speaker audio-to-video mapping.","inclusion":"Include work whose main objective is text/speech/visual modality fusion, self-supervised cross-modal alignment, unified heterogeneous drivers, diverse stochastic motion, conversational head generation, listener response, turn-taking, or interactive nonverbal behavior.","exclusion":"Assign to disentangled expressive control when the main contribution is explicit manipulation of facial factors within a monologue setting, and to general synthesis when the system is standard audio-only generation without contextual, multimodal, stochastic, or interaction modeling.","nearest_boundary":"Disentangled expressive control","representative_topics":["unified facial landmark generation via shared text-to-speech latent representations","diffusion-based talking face generation with multimodal emotion and geometry-aware texture transfer","perceptual conversational head generation via regularized audio-to-head driver and enhanced rendering"]},{"direction_index":7,"label":"D7","name":"Refinement and deployability","description":"Improving generated talking-head systems for practical use through visual restoration, stabilization, super-resolution, relighting, stylized rendering, compression, acceleration, or real-time deployment.","inclusion":"Include work whose primary objective is post-generation or system-level improvement of quality, stability, identity fidelity, resolution, artifact repair, output-domain adaptation, model size, latency, training cost, or edge deployment.","exclusion":"Assign to general synthesis or 3D representation when quality or speed gains are incidental to the core generator, and to lip synchronization when the primary target is speech-mouth alignment rather than restoration or deployability.","nearest_boundary":"General speech-driven synthesis","representative_topics":["audio-visual prior-guided extreme-scale talking face video upsampling from 8x8 inputs","mitigating lip motion jitter in 3d-based speech-driven talking face generation","lightweight talking face generation via distillation and selective mixed-precision quantization"]}],"split":"dev"}
{"episode_id":"F195_T2025-07","field_index":195,"field":"talking head generation","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"General speech-driven synthesis","description":"End-to-end generation of realistic talking-head video or facial animation from speech or audio, emphasizing broad audiovisual coherence, identity preservation, natural motion, and generalization when no narrower objective dominates.","inclusion":"Include work whose primary objective is general audio-conditioned talking-face or portrait synthesis, including diffusion-based generators when diffusion is the implementation for broad synthesis rather than a separate scientific objective.","exclusion":"Assign to personalization, lip-sync editing, controllability, 3D representation, geometry transfer, contextual interaction, or deployment/restoration when one of those objectives is central.","nearest_boundary":"Sparse personalized avatar generation","representative_topics":["autoregressive diffusion models for reference-free talking face generation with head movement and expression","generalized audio-driven talking head synthesis via latent diffusion with identity-agnostic conditioning","diffusion autoencoder-based speech-driven talking face generation with data-driven latent representations and pose controllability"]},{"direction_index":1,"label":"D1","name":"Sparse personalized avatars","description":"Creating identity-preserving talking avatars from minimal person-specific evidence such as a single portrait, few-shot video, or fast adaptation data.","inclusion":"Include zero-, one-, or few-shot talking-head generation and rapid personalized adaptation where sparse identity acquisition is the main objective, regardless of backbone or rendering method.","exclusion":"Assign to general synthesis when identity transfer is not sparse-personalization-centered, and to 3D representation when the main contribution is the representation or renderer rather than minimal-data personalization.","nearest_boundary":"General speech-driven synthesis","representative_topics":["few-shot talking head synthesis via dynamic facial radiance fields conditioned on 2d appearance priors","one-shot audio-driven talking head generation with disentangled motion and lip synchronization","speech-driven portrait animation with controllable facial expressions and head pose from a single image"]},{"direction_index":2,"label":"D2","name":"Lip synchronization and dubbing","description":"Precise alignment, correction, generation, or editing of mouth and facial motion to match target speech, phonetic context, or dubbing audio.","inclusion":"Include work whose main objective is lip-audio synchrony, phoneme/viseme fidelity, dubbing, lip-region repair, synchronization losses, or lip-sync measurement in generated or existing videos.","exclusion":"Assign to general synthesis when lip sync is only one quality criterion, and to expression controllability when the main target is affective or semantic expression editing rather than mouth-speech alignment.","nearest_boundary":"General speech-driven synthesis","representative_topics":["audio-visual lip synchronization via phoneme-level memory retrieval in talking face generation","audio-driven lip synchronization for talking head video editing via sequential expression and face enhancement","phonetic context-aware lip synchronization for talking face generation using masked learning"]},{"direction_index":3,"label":"D3","name":"Disentangled expressive control","description":"Learning or using separable controls for facial motion factors such as pose, expression, emotion, speaking style, identity, and speech content to enable explicit manipulation of talking-head behavior.","inclusion":"Include work where factor separation, controllable latent spaces, expression or emotion control, and explicit manipulation of pose/style/expression are the primary research objective.","exclusion":"Assign to sparse personalized avatars when disentanglement mainly supports one-shot identity transfer, and to contextual multimodal interaction when the main objective is heterogeneous input fusion or conversational responsiveness.","nearest_boundary":"Contextual multimodal interaction","representative_topics":["progressive disentangled representation learning for multi-factor controllable talking head synthesis","disentangled head pose and expression control for audio-video-driven talking head generation","speech-driven 3d face animation with disentangled emotional and content embeddings"]},{"direction_index":4,"label":"D4","name":"Geometry-guided motion transfer","description":"Animating or reenacting talking heads by using explicit landmarks, keypoints, meshes, depth, displacement fields, or geometric priors to drive spatial warping and motion transfer.","inclusion":"Include work whose core objective is geometry- or keypoint-guided face warping, reenactment, displacement-field estimation, depth-aware transfer, or self-supervised 3D-aware motion transfer without making a persistent neural 3D scene representation the main claim.","exclusion":"Assign to 3D representation and animation when the main contribution is NeRF, neural volume rendering, implicit representation, tri-plane rendering, or mesh-sequence generation as the primary output.","nearest_boundary":"3D representation and animation","representative_topics":["geometry-guided progressive warping for one-shot face animation with displacement field estimation","audio-visual talking head reenactment using learnable keypoint motion fields and identity-aware generation","keypoint-based enhancement for naturalness in audio-driven free-viewpoint talking head synthesis"]},{"direction_index":5,"label":"D5","name":"3D representation and animation","description":"Modeling talking heads through 3D-aware representations or outputs, including neural radiance fields, neural volumes, tri-planes, implicit faces, 4D face models, meshes, blendshapes, and viseme curves.","inclusion":"Include work whose primary objective is view-consistent or high-fidelity 3D representation/rendering, neural scene modeling, or speech-driven 3D facial mesh/blendshape/viseme animation.","exclusion":"Assign to geometry-guided motion transfer when geometry is mainly used for 2D warping, and to deployment efficiency when the dominant claim is acceleration, compression, or real-time operation rather than the 3D representation itself.","nearest_boundary":"Geometry-guided motion transfer","representative_topics":["neural volume rendering for identity-preserving talking-head face animation","generalizable nerf-based talking face synthesis with out-of-domain audio and head-torso joint rendering","large-scale multi-modal 4d face dataset and non-autoregressive audio-driven 3d face animation"]},{"direction_index":6,"label":"D6","name":"Contextual multimodal interaction","description":"Generating talking-head behavior from semantic, affective, multimodal, stochastic, or conversational context beyond deterministic single-speaker audio-to-video mapping.","inclusion":"Include work whose main objective is text/speech/visual modality fusion, self-supervised cross-modal alignment, unified heterogeneous drivers, diverse stochastic motion, conversational head generation, listener response, turn-taking, or interactive nonverbal behavior.","exclusion":"Assign to disentangled expressive control when the main contribution is explicit manipulation of facial factors within a monologue setting, and to general synthesis when the system is standard audio-only generation without contextual, multimodal, stochastic, or interaction modeling.","nearest_boundary":"Disentangled expressive control","representative_topics":["unified facial landmark generation via shared text-to-speech latent representations","diffusion-based talking face generation with multimodal emotion and geometry-aware texture transfer","perceptual conversational head generation via regularized audio-to-head driver and enhanced rendering"]},{"direction_index":7,"label":"D7","name":"Refinement and deployability","description":"Improving generated talking-head systems for practical use through visual restoration, stabilization, super-resolution, relighting, stylized rendering, compression, acceleration, or real-time deployment.","inclusion":"Include work whose primary objective is post-generation or system-level improvement of quality, stability, identity fidelity, resolution, artifact repair, output-domain adaptation, model size, latency, training cost, or edge deployment.","exclusion":"Assign to general synthesis or 3D representation when quality or speed gains are incidental to the core generator, and to lip synchronization when the primary target is speech-mouth alignment rather than restoration or deployability.","nearest_boundary":"General speech-driven synthesis","representative_topics":["audio-visual prior-guided extreme-scale talking face video upsampling from 8x8 inputs","mitigating lip motion jitter in 3d-based speech-driven talking face generation","lightweight talking face generation via distillation and selective mixed-precision quantization"]}],"split":"dev"}
{"episode_id":"F195_T2026-01","field_index":195,"field":"talking head generation","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"General speech-driven synthesis","description":"End-to-end generation of realistic talking-head video or facial animation from speech or audio, emphasizing broad audiovisual coherence, identity preservation, natural motion, and generalization when no narrower objective dominates.","inclusion":"Include work whose primary objective is general audio-conditioned talking-face or portrait synthesis, including diffusion-based generators when diffusion is the implementation for broad synthesis rather than a separate scientific objective.","exclusion":"Assign to personalization, lip-sync editing, controllability, 3D representation, geometry transfer, contextual interaction, or deployment/restoration when one of those objectives is central.","nearest_boundary":"Sparse personalized avatar generation","representative_topics":["autoregressive diffusion models for reference-free talking face generation with head movement and expression","generalized audio-driven talking head synthesis via latent diffusion with identity-agnostic conditioning","diffusion autoencoder-based speech-driven talking face generation with data-driven latent representations and pose controllability"]},{"direction_index":1,"label":"D1","name":"Sparse personalized avatars","description":"Creating identity-preserving talking avatars from minimal person-specific evidence such as a single portrait, few-shot video, or fast adaptation data.","inclusion":"Include zero-, one-, or few-shot talking-head generation and rapid personalized adaptation where sparse identity acquisition is the main objective, regardless of backbone or rendering method.","exclusion":"Assign to general synthesis when identity transfer is not sparse-personalization-centered, and to 3D representation when the main contribution is the representation or renderer rather than minimal-data personalization.","nearest_boundary":"General speech-driven synthesis","representative_topics":["few-shot talking head synthesis via dynamic facial radiance fields conditioned on 2d appearance priors","one-shot audio-driven talking head generation with disentangled motion and lip synchronization","speech-driven portrait animation with controllable facial expressions and head pose from a single image"]},{"direction_index":2,"label":"D2","name":"Lip synchronization and dubbing","description":"Precise alignment, correction, generation, or editing of mouth and facial motion to match target speech, phonetic context, or dubbing audio.","inclusion":"Include work whose main objective is lip-audio synchrony, phoneme/viseme fidelity, dubbing, lip-region repair, synchronization losses, or lip-sync measurement in generated or existing videos.","exclusion":"Assign to general synthesis when lip sync is only one quality criterion, and to expression controllability when the main target is affective or semantic expression editing rather than mouth-speech alignment.","nearest_boundary":"General speech-driven synthesis","representative_topics":["audio-visual lip synchronization via phoneme-level memory retrieval in talking face generation","audio-driven lip synchronization for talking head video editing via sequential expression and face enhancement","phonetic context-aware lip synchronization for talking face generation using masked learning"]},{"direction_index":3,"label":"D3","name":"Disentangled expressive control","description":"Learning or using separable controls for facial motion factors such as pose, expression, emotion, speaking style, identity, and speech content to enable explicit manipulation of talking-head behavior.","inclusion":"Include work where factor separation, controllable latent spaces, expression or emotion control, and explicit manipulation of pose/style/expression are the primary research objective.","exclusion":"Assign to sparse personalized avatars when disentanglement mainly supports one-shot identity transfer, and to contextual multimodal interaction when the main objective is heterogeneous input fusion or conversational responsiveness.","nearest_boundary":"Contextual multimodal interaction","representative_topics":["progressive disentangled representation learning for multi-factor controllable talking head synthesis","disentangled head pose and expression control for audio-video-driven talking head generation","speech-driven 3d face animation with disentangled emotional and content embeddings"]},{"direction_index":4,"label":"D4","name":"Geometry-guided motion transfer","description":"Animating or reenacting talking heads by using explicit landmarks, keypoints, meshes, depth, displacement fields, or geometric priors to drive spatial warping and motion transfer.","inclusion":"Include work whose core objective is geometry- or keypoint-guided face warping, reenactment, displacement-field estimation, depth-aware transfer, or self-supervised 3D-aware motion transfer without making a persistent neural 3D scene representation the main claim.","exclusion":"Assign to 3D representation and animation when the main contribution is NeRF, neural volume rendering, implicit representation, tri-plane rendering, or mesh-sequence generation as the primary output.","nearest_boundary":"3D representation and animation","representative_topics":["geometry-guided progressive warping for one-shot face animation with displacement field estimation","audio-visual talking head reenactment using learnable keypoint motion fields and identity-aware generation","keypoint-based enhancement for naturalness in audio-driven free-viewpoint talking head synthesis"]},{"direction_index":5,"label":"D5","name":"3D representation and animation","description":"Modeling talking heads through 3D-aware representations or outputs, including neural radiance fields, neural volumes, tri-planes, implicit faces, 4D face models, meshes, blendshapes, and viseme curves.","inclusion":"Include work whose primary objective is view-consistent or high-fidelity 3D representation/rendering, neural scene modeling, or speech-driven 3D facial mesh/blendshape/viseme animation.","exclusion":"Assign to geometry-guided motion transfer when geometry is mainly used for 2D warping, and to deployment efficiency when the dominant claim is acceleration, compression, or real-time operation rather than the 3D representation itself.","nearest_boundary":"Geometry-guided motion transfer","representative_topics":["neural volume rendering for identity-preserving talking-head face animation","generalizable nerf-based talking face synthesis with out-of-domain audio and head-torso joint rendering","large-scale multi-modal 4d face dataset and non-autoregressive audio-driven 3d face animation"]},{"direction_index":6,"label":"D6","name":"Contextual multimodal interaction","description":"Generating talking-head behavior from semantic, affective, multimodal, stochastic, or conversational context beyond deterministic single-speaker audio-to-video mapping.","inclusion":"Include work whose main objective is text/speech/visual modality fusion, self-supervised cross-modal alignment, unified heterogeneous drivers, diverse stochastic motion, conversational head generation, listener response, turn-taking, or interactive nonverbal behavior.","exclusion":"Assign to disentangled expressive control when the main contribution is explicit manipulation of facial factors within a monologue setting, and to general synthesis when the system is standard audio-only generation without contextual, multimodal, stochastic, or interaction modeling.","nearest_boundary":"Disentangled expressive control","representative_topics":["unified facial landmark generation via shared text-to-speech latent representations","diffusion-based talking face generation with multimodal emotion and geometry-aware texture transfer","perceptual conversational head generation via regularized audio-to-head driver and enhanced rendering"]},{"direction_index":7,"label":"D7","name":"Refinement and deployability","description":"Improving generated talking-head systems for practical use through visual restoration, stabilization, super-resolution, relighting, stylized rendering, compression, acceleration, or real-time deployment.","inclusion":"Include work whose primary objective is post-generation or system-level improvement of quality, stability, identity fidelity, resolution, artifact repair, output-domain adaptation, model size, latency, training cost, or edge deployment.","exclusion":"Assign to general synthesis or 3D representation when quality or speed gains are incidental to the core generator, and to lip synchronization when the primary target is speech-mouth alignment rather than restoration or deployability.","nearest_boundary":"General speech-driven synthesis","representative_topics":["audio-visual prior-guided extreme-scale talking face video upsampling from 8x8 inputs","mitigating lip motion jitter in 3d-based speech-driven talking face generation","lightweight talking face generation via distillation and selective mixed-precision quantization"]}],"split":"dev"}
{"episode_id":"F196_T2024-01","field_index":196,"field":"clinical risk prediction","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structured cohort risk modeling","description":"Patient-level diagnosis, prognosis, survival, treatment-response, adherence, and outcome prediction from structured cohort, registry, claims, laboratory, demographic, genomic, or other tabular clinical variables.","inclusion":"Assign atoms whose main objective is building or evaluating risk models on structured/tabular patient data, including survival modeling, feature selection, classical machine learning, ensembles, missing-data handling, and imbalance handling when these serve the tabular cohort prediction program.","exclusion":"Assign to the modality, longitudinal EHR, acute-care, disease-program, multimodal, or trustworthiness direction when that objective is primary rather than general structured cohort prediction.","nearest_boundary":"Named disease and behavioral risk programs","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","sparse neural network variable selection for interval-censored accelerated failure time regression","class-imbalanced learning for adverse event severity prediction in immunization data"]},{"direction_index":1,"label":"D1","name":"Image-based clinical risk prediction","description":"Prediction of disease risk, severity, stage, prognosis, survival, treatment outcomes, or systemic biomarkers from medical images and image-derived representations.","inclusion":"Assign atoms where radiology, histopathology, ophthalmology, external-eye, X-ray, CT, MRI, fundus, echocardiographic, or related clinical images are the main evidentiary source for prediction, including radiomics, CNNs, transformers, self-supervised learning, relabeling, and image-centered image-plus-tabular models.","exclusion":"Assign to multimodal fusion when cross-modal integration is the central contribution, to signal-based prediction for ECG/EEG/wearable waveforms, or to disease-program directions when a named disease objective dominates over the imaging program.","nearest_boundary":"Multimodal fusion for outcome prediction","representative_topics":["multi-organ systemic biomarker detection in external eye photos via deep learning","deep learning-based opportunistic hip fracture risk prediction from plain radiographs and ct projections","mitigating supervision bias in lung cancer prediction through retrieval-based re-labeling"]},{"direction_index":2,"label":"D2","name":"Physiological and neurological signal prediction","description":"Prediction of cardiovascular, neurological, neuromotor, sleep, hemodynamic, and wearable-monitored health states from ECG, EEG, PPG, gait, smartphone, wearable, sleep, pump-flow, or other physiological time-series signals.","inclusion":"Assign atoms where biosignals or remote physiological measurements are the primary data source, including ECG cardiac risk, arrhythmia or atrial fibrillation prediction, EEG seizure or neurological prognosis, Parkinson's monitoring, gait-based disorder prediction, sleep-event detection, and wearable/sensor health prediction.","exclusion":"Assign to image-based prediction when the primary evidence is clinical imaging, to longitudinal EHR when the signal is secondary to patient-record sequences, or to named disease programs when the disease-screening objective dominates over the signal modality.","nearest_boundary":"Image-based clinical risk prediction","representative_topics":["feature engineering versus deep learning comparison for 12-lead ecg diagnosis and age estimation","zero-shot personalized ecg arrhythmia detection via sparse representation domain adaptation for wearable surveillance","atrial fibrillation recurrence risk prediction from pre- and post-ablation ecg morphological features"]},{"direction_index":3,"label":"D3","name":"Longitudinal EHR trajectory modeling","description":"Learning temporal patient representations from longitudinal EHRs, claims sequences, visit histories, clinical notes, event streams, and irregular patient journeys for clinical outcome prediction.","inclusion":"Assign atoms whose primary objective is modeling patient trajectories with recurrent, transformer, pretrained, generative, self-supervised, multitask, zero-shot, or time-aware architectures over longitudinal clinical records.","exclusion":"Assign to acute-care operations when the endpoint is hospital flow, ICU deterioration, sepsis, triage, or length of stay; to multimodal fusion when cross-modal integration is primary; or to privacy/trustworthiness when deployment validity, privacy, or uncertainty is the central objective.","nearest_boundary":"Acute-care and hospital operations prediction","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","generalizable multi-task clinical prediction via hierarchical ehr representation and multi-source self-supervised learning","transformer-based clinical note embedding for mortality prediction via umls concept mapping"]},{"direction_index":4,"label":"D4","name":"Acute-care and hospital operations prediction","description":"Early warning, sepsis prediction, ICU prognosis, emergency triage, hospital length of stay, discharge, readmission, acute-care use, and other resource-relevant predictions for acute-care decision support.","inclusion":"Assign atoms where the scientific objective is predicting deterioration, sepsis onset or outcome, critical-care prognosis, ED departure or triage, hospital flow, readmission, transfer, discharge placement, or length of stay for acute-care management.","exclusion":"Assign to COVID-specific programs for COVID acute outcomes, to longitudinal EHR trajectory modeling when the contribution is a general sequence architecture, or to trustworthiness when the main focus is leakage, causal validity, or fairness of acute-care models.","nearest_boundary":"Longitudinal EHR trajectory modeling","representative_topics":["early warning systems for critical health episodes using feature selection and gradient boosting","inherently interpretable bayesian survival modeling for discharge placement and readmission prediction","free-text clinical note analysis for acute care risk prediction in oncology patients"]},{"direction_index":5,"label":"D5","name":"Named disease and behavioral risk programs","description":"Condition-defined clinical prediction programs where the primary research objective is risk stratification, screening, diagnosis, prognosis, or treatment response for a named disease area or behavioral health outcome.","inclusion":"Assign atoms whose organizing objective is a specific disease or behavioral/psychiatric program, including COVID-19, post-COVID sequelae, mental health, autism, depression, suicide risk, food allergy, rheumatoid arthritis, breast cancer metastasis, and similar condition-tailored tools.","exclusion":"Assign to structured cohort, imaging, signal, multimodal, or EHR directions when the named disease is mainly a case study for a broader data-modality or method objective; assign COVID validity benchmarks to trustworthiness when methodological validity is primary.","nearest_boundary":"Structured cohort risk modeling","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]},{"direction_index":6,"label":"D6","name":"Multimodal fusion for outcome prediction","description":"Integration, alignment, imputation, or joint modeling of heterogeneous clinical modalities to improve patient-level prediction.","inclusion":"Assign atoms whose central scientific objective is fusing or completing multiple distinct data sources such as structured EHR, notes, imaging, laboratory values, motion, genetics, behavioral signals, or incomplete modalities for outcome prediction.","exclusion":"Assign to a single-modality direction when one modality supplies the main evidence; to named disease programs when the condition-specific objective dominates; or to trustworthiness when explanation, fairness, privacy, or uncertainty is the central contribution.","nearest_boundary":"Longitudinal EHR trajectory modeling","representative_topics":["multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","multimodal transformer for interpretable in-hospital mortality prediction from clinical notes and structured ehr"]},{"direction_index":7,"label":"D7","name":"Trustworthy and collaborative prediction","description":"Methods that make clinical prediction models interpretable, reliable, equitable, causally valid, portable, privacy-preserving, calibrated, uncertainty-aware, or safe for deployment.","inclusion":"Assign atoms whose primary objective is explanation, sparse risk scoring, clinician trust, uncertainty quantification, calibration, conformal or Bayesian reliability, fairness, bias mitigation, causal validation, leakage critique, cross-site portability, federated learning, secure computation, blockchain, or privacy-preserving collaboration.","exclusion":"Assign to the relevant disease, modality, EHR, acute-care, or multimodal program when explanation, uncertainty, fairness, privacy, or robustness is only an auxiliary analysis rather than the main scientific objective.","nearest_boundary":"Multimodal fusion for outcome prediction","representative_topics":["explainable meta-review-informed detection of cytokine release syndrome from multi-patient profiles","fast beam-search and star-ray rounding for interpretable sparse risk score learning","uncertainty estimation methods for deep learning models in clinical decision support"]}],"split":"dev"}
{"episode_id":"F196_T2024-07","field_index":196,"field":"clinical risk prediction","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structured cohort risk modeling","description":"Patient-level diagnosis, prognosis, survival, treatment-response, adherence, and outcome prediction from structured cohort, registry, claims, laboratory, demographic, genomic, or other tabular clinical variables.","inclusion":"Assign atoms whose main objective is building or evaluating risk models on structured/tabular patient data, including survival modeling, feature selection, classical machine learning, ensembles, missing-data handling, and imbalance handling when these serve the tabular cohort prediction program.","exclusion":"Assign to the modality, longitudinal EHR, acute-care, disease-program, multimodal, or trustworthiness direction when that objective is primary rather than general structured cohort prediction.","nearest_boundary":"Named disease and behavioral risk programs","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","sparse neural network variable selection for interval-censored accelerated failure time regression","class-imbalanced learning for adverse event severity prediction in immunization data"]},{"direction_index":1,"label":"D1","name":"Image-based clinical risk prediction","description":"Prediction of disease risk, severity, stage, prognosis, survival, treatment outcomes, or systemic biomarkers from medical images and image-derived representations.","inclusion":"Assign atoms where radiology, histopathology, ophthalmology, external-eye, X-ray, CT, MRI, fundus, echocardiographic, or related clinical images are the main evidentiary source for prediction, including radiomics, CNNs, transformers, self-supervised learning, relabeling, and image-centered image-plus-tabular models.","exclusion":"Assign to multimodal fusion when cross-modal integration is the central contribution, to signal-based prediction for ECG/EEG/wearable waveforms, or to disease-program directions when a named disease objective dominates over the imaging program.","nearest_boundary":"Multimodal fusion for outcome prediction","representative_topics":["multi-organ systemic biomarker detection in external eye photos via deep learning","deep learning-based opportunistic hip fracture risk prediction from plain radiographs and ct projections","mitigating supervision bias in lung cancer prediction through retrieval-based re-labeling"]},{"direction_index":2,"label":"D2","name":"Physiological and neurological signal prediction","description":"Prediction of cardiovascular, neurological, neuromotor, sleep, hemodynamic, and wearable-monitored health states from ECG, EEG, PPG, gait, smartphone, wearable, sleep, pump-flow, or other physiological time-series signals.","inclusion":"Assign atoms where biosignals or remote physiological measurements are the primary data source, including ECG cardiac risk, arrhythmia or atrial fibrillation prediction, EEG seizure or neurological prognosis, Parkinson's monitoring, gait-based disorder prediction, sleep-event detection, and wearable/sensor health prediction.","exclusion":"Assign to image-based prediction when the primary evidence is clinical imaging, to longitudinal EHR when the signal is secondary to patient-record sequences, or to named disease programs when the disease-screening objective dominates over the signal modality.","nearest_boundary":"Image-based clinical risk prediction","representative_topics":["feature engineering versus deep learning comparison for 12-lead ecg diagnosis and age estimation","zero-shot personalized ecg arrhythmia detection via sparse representation domain adaptation for wearable surveillance","atrial fibrillation recurrence risk prediction from pre- and post-ablation ecg morphological features"]},{"direction_index":3,"label":"D3","name":"Longitudinal EHR trajectory modeling","description":"Learning temporal patient representations from longitudinal EHRs, claims sequences, visit histories, clinical notes, event streams, and irregular patient journeys for clinical outcome prediction.","inclusion":"Assign atoms whose primary objective is modeling patient trajectories with recurrent, transformer, pretrained, generative, self-supervised, multitask, zero-shot, or time-aware architectures over longitudinal clinical records.","exclusion":"Assign to acute-care operations when the endpoint is hospital flow, ICU deterioration, sepsis, triage, or length of stay; to multimodal fusion when cross-modal integration is primary; or to privacy/trustworthiness when deployment validity, privacy, or uncertainty is the central objective.","nearest_boundary":"Acute-care and hospital operations prediction","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","generalizable multi-task clinical prediction via hierarchical ehr representation and multi-source self-supervised learning","transformer-based clinical note embedding for mortality prediction via umls concept mapping"]},{"direction_index":4,"label":"D4","name":"Acute-care and hospital operations prediction","description":"Early warning, sepsis prediction, ICU prognosis, emergency triage, hospital length of stay, discharge, readmission, acute-care use, and other resource-relevant predictions for acute-care decision support.","inclusion":"Assign atoms where the scientific objective is predicting deterioration, sepsis onset or outcome, critical-care prognosis, ED departure or triage, hospital flow, readmission, transfer, discharge placement, or length of stay for acute-care management.","exclusion":"Assign to COVID-specific programs for COVID acute outcomes, to longitudinal EHR trajectory modeling when the contribution is a general sequence architecture, or to trustworthiness when the main focus is leakage, causal validity, or fairness of acute-care models.","nearest_boundary":"Longitudinal EHR trajectory modeling","representative_topics":["early warning systems for critical health episodes using feature selection and gradient boosting","inherently interpretable bayesian survival modeling for discharge placement and readmission prediction","free-text clinical note analysis for acute care risk prediction in oncology patients"]},{"direction_index":5,"label":"D5","name":"Named disease and behavioral risk programs","description":"Condition-defined clinical prediction programs where the primary research objective is risk stratification, screening, diagnosis, prognosis, or treatment response for a named disease area or behavioral health outcome.","inclusion":"Assign atoms whose organizing objective is a specific disease or behavioral/psychiatric program, including COVID-19, post-COVID sequelae, mental health, autism, depression, suicide risk, food allergy, rheumatoid arthritis, breast cancer metastasis, and similar condition-tailored tools.","exclusion":"Assign to structured cohort, imaging, signal, multimodal, or EHR directions when the named disease is mainly a case study for a broader data-modality or method objective; assign COVID validity benchmarks to trustworthiness when methodological validity is primary.","nearest_boundary":"Structured cohort risk modeling","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]},{"direction_index":6,"label":"D6","name":"Multimodal fusion for outcome prediction","description":"Integration, alignment, imputation, or joint modeling of heterogeneous clinical modalities to improve patient-level prediction.","inclusion":"Assign atoms whose central scientific objective is fusing or completing multiple distinct data sources such as structured EHR, notes, imaging, laboratory values, motion, genetics, behavioral signals, or incomplete modalities for outcome prediction.","exclusion":"Assign to a single-modality direction when one modality supplies the main evidence; to named disease programs when the condition-specific objective dominates; or to trustworthiness when explanation, fairness, privacy, or uncertainty is the central contribution.","nearest_boundary":"Longitudinal EHR trajectory modeling","representative_topics":["multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","multimodal transformer for interpretable in-hospital mortality prediction from clinical notes and structured ehr"]},{"direction_index":7,"label":"D7","name":"Trustworthy and collaborative prediction","description":"Methods that make clinical prediction models interpretable, reliable, equitable, causally valid, portable, privacy-preserving, calibrated, uncertainty-aware, or safe for deployment.","inclusion":"Assign atoms whose primary objective is explanation, sparse risk scoring, clinician trust, uncertainty quantification, calibration, conformal or Bayesian reliability, fairness, bias mitigation, causal validation, leakage critique, cross-site portability, federated learning, secure computation, blockchain, or privacy-preserving collaboration.","exclusion":"Assign to the relevant disease, modality, EHR, acute-care, or multimodal program when explanation, uncertainty, fairness, privacy, or robustness is only an auxiliary analysis rather than the main scientific objective.","nearest_boundary":"Multimodal fusion for outcome prediction","representative_topics":["explainable meta-review-informed detection of cytokine release syndrome from multi-patient profiles","fast beam-search and star-ray rounding for interpretable sparse risk score learning","uncertainty estimation methods for deep learning models in clinical decision support"]}],"split":"dev"}
{"episode_id":"F196_T2025-01","field_index":196,"field":"clinical risk prediction","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structured cohort risk modeling","description":"Patient-level diagnosis, prognosis, survival, treatment-response, adherence, and outcome prediction from structured cohort, registry, claims, laboratory, demographic, genomic, or other tabular clinical variables.","inclusion":"Assign atoms whose main objective is building or evaluating risk models on structured/tabular patient data, including survival modeling, feature selection, classical machine learning, ensembles, missing-data handling, and imbalance handling when these serve the tabular cohort prediction program.","exclusion":"Assign to the modality, longitudinal EHR, acute-care, disease-program, multimodal, or trustworthiness direction when that objective is primary rather than general structured cohort prediction.","nearest_boundary":"Named disease and behavioral risk programs","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","sparse neural network variable selection for interval-censored accelerated failure time regression","class-imbalanced learning for adverse event severity prediction in immunization data"]},{"direction_index":1,"label":"D1","name":"Image-based clinical risk prediction","description":"Prediction of disease risk, severity, stage, prognosis, survival, treatment outcomes, or systemic biomarkers from medical images and image-derived representations.","inclusion":"Assign atoms where radiology, histopathology, ophthalmology, external-eye, X-ray, CT, MRI, fundus, echocardiographic, or related clinical images are the main evidentiary source for prediction, including radiomics, CNNs, transformers, self-supervised learning, relabeling, and image-centered image-plus-tabular models.","exclusion":"Assign to multimodal fusion when cross-modal integration is the central contribution, to signal-based prediction for ECG/EEG/wearable waveforms, or to disease-program directions when a named disease objective dominates over the imaging program.","nearest_boundary":"Multimodal fusion for outcome prediction","representative_topics":["multi-organ systemic biomarker detection in external eye photos via deep learning","deep learning-based opportunistic hip fracture risk prediction from plain radiographs and ct projections","mitigating supervision bias in lung cancer prediction through retrieval-based re-labeling"]},{"direction_index":2,"label":"D2","name":"Physiological and neurological signal prediction","description":"Prediction of cardiovascular, neurological, neuromotor, sleep, hemodynamic, and wearable-monitored health states from ECG, EEG, PPG, gait, smartphone, wearable, sleep, pump-flow, or other physiological time-series signals.","inclusion":"Assign atoms where biosignals or remote physiological measurements are the primary data source, including ECG cardiac risk, arrhythmia or atrial fibrillation prediction, EEG seizure or neurological prognosis, Parkinson's monitoring, gait-based disorder prediction, sleep-event detection, and wearable/sensor health prediction.","exclusion":"Assign to image-based prediction when the primary evidence is clinical imaging, to longitudinal EHR when the signal is secondary to patient-record sequences, or to named disease programs when the disease-screening objective dominates over the signal modality.","nearest_boundary":"Image-based clinical risk prediction","representative_topics":["feature engineering versus deep learning comparison for 12-lead ecg diagnosis and age estimation","zero-shot personalized ecg arrhythmia detection via sparse representation domain adaptation for wearable surveillance","atrial fibrillation recurrence risk prediction from pre- and post-ablation ecg morphological features"]},{"direction_index":3,"label":"D3","name":"Longitudinal EHR trajectory modeling","description":"Learning temporal patient representations from longitudinal EHRs, claims sequences, visit histories, clinical notes, event streams, and irregular patient journeys for clinical outcome prediction.","inclusion":"Assign atoms whose primary objective is modeling patient trajectories with recurrent, transformer, pretrained, generative, self-supervised, multitask, zero-shot, or time-aware architectures over longitudinal clinical records.","exclusion":"Assign to acute-care operations when the endpoint is hospital flow, ICU deterioration, sepsis, triage, or length of stay; to multimodal fusion when cross-modal integration is primary; or to privacy/trustworthiness when deployment validity, privacy, or uncertainty is the central objective.","nearest_boundary":"Acute-care and hospital operations prediction","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","generalizable multi-task clinical prediction via hierarchical ehr representation and multi-source self-supervised learning","transformer-based clinical note embedding for mortality prediction via umls concept mapping"]},{"direction_index":4,"label":"D4","name":"Acute-care and hospital operations prediction","description":"Early warning, sepsis prediction, ICU prognosis, emergency triage, hospital length of stay, discharge, readmission, acute-care use, and other resource-relevant predictions for acute-care decision support.","inclusion":"Assign atoms where the scientific objective is predicting deterioration, sepsis onset or outcome, critical-care prognosis, ED departure or triage, hospital flow, readmission, transfer, discharge placement, or length of stay for acute-care management.","exclusion":"Assign to COVID-specific programs for COVID acute outcomes, to longitudinal EHR trajectory modeling when the contribution is a general sequence architecture, or to trustworthiness when the main focus is leakage, causal validity, or fairness of acute-care models.","nearest_boundary":"Longitudinal EHR trajectory modeling","representative_topics":["early warning systems for critical health episodes using feature selection and gradient boosting","inherently interpretable bayesian survival modeling for discharge placement and readmission prediction","free-text clinical note analysis for acute care risk prediction in oncology patients"]},{"direction_index":5,"label":"D5","name":"Named disease and behavioral risk programs","description":"Condition-defined clinical prediction programs where the primary research objective is risk stratification, screening, diagnosis, prognosis, or treatment response for a named disease area or behavioral health outcome.","inclusion":"Assign atoms whose organizing objective is a specific disease or behavioral/psychiatric program, including COVID-19, post-COVID sequelae, mental health, autism, depression, suicide risk, food allergy, rheumatoid arthritis, breast cancer metastasis, and similar condition-tailored tools.","exclusion":"Assign to structured cohort, imaging, signal, multimodal, or EHR directions when the named disease is mainly a case study for a broader data-modality or method objective; assign COVID validity benchmarks to trustworthiness when methodological validity is primary.","nearest_boundary":"Structured cohort risk modeling","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]},{"direction_index":6,"label":"D6","name":"Multimodal fusion for outcome prediction","description":"Integration, alignment, imputation, or joint modeling of heterogeneous clinical modalities to improve patient-level prediction.","inclusion":"Assign atoms whose central scientific objective is fusing or completing multiple distinct data sources such as structured EHR, notes, imaging, laboratory values, motion, genetics, behavioral signals, or incomplete modalities for outcome prediction.","exclusion":"Assign to a single-modality direction when one modality supplies the main evidence; to named disease programs when the condition-specific objective dominates; or to trustworthiness when explanation, fairness, privacy, or uncertainty is the central contribution.","nearest_boundary":"Longitudinal EHR trajectory modeling","representative_topics":["multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","multimodal transformer for interpretable in-hospital mortality prediction from clinical notes and structured ehr"]},{"direction_index":7,"label":"D7","name":"Trustworthy and collaborative prediction","description":"Methods that make clinical prediction models interpretable, reliable, equitable, causally valid, portable, privacy-preserving, calibrated, uncertainty-aware, or safe for deployment.","inclusion":"Assign atoms whose primary objective is explanation, sparse risk scoring, clinician trust, uncertainty quantification, calibration, conformal or Bayesian reliability, fairness, bias mitigation, causal validation, leakage critique, cross-site portability, federated learning, secure computation, blockchain, or privacy-preserving collaboration.","exclusion":"Assign to the relevant disease, modality, EHR, acute-care, or multimodal program when explanation, uncertainty, fairness, privacy, or robustness is only an auxiliary analysis rather than the main scientific objective.","nearest_boundary":"Multimodal fusion for outcome prediction","representative_topics":["explainable meta-review-informed detection of cytokine release syndrome from multi-patient profiles","fast beam-search and star-ray rounding for interpretable sparse risk score learning","uncertainty estimation methods for deep learning models in clinical decision support"]}],"split":"dev"}
{"episode_id":"F196_T2025-07","field_index":196,"field":"clinical risk prediction","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structured cohort risk modeling","description":"Patient-level diagnosis, prognosis, survival, treatment-response, adherence, and outcome prediction from structured cohort, registry, claims, laboratory, demographic, genomic, or other tabular clinical variables.","inclusion":"Assign atoms whose main objective is building or evaluating risk models on structured/tabular patient data, including survival modeling, feature selection, classical machine learning, ensembles, missing-data handling, and imbalance handling when these serve the tabular cohort prediction program.","exclusion":"Assign to the modality, longitudinal EHR, acute-care, disease-program, multimodal, or trustworthiness direction when that objective is primary rather than general structured cohort prediction.","nearest_boundary":"Named disease and behavioral risk programs","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","sparse neural network variable selection for interval-censored accelerated failure time regression","class-imbalanced learning for adverse event severity prediction in immunization data"]},{"direction_index":1,"label":"D1","name":"Image-based clinical risk prediction","description":"Prediction of disease risk, severity, stage, prognosis, survival, treatment outcomes, or systemic biomarkers from medical images and image-derived representations.","inclusion":"Assign atoms where radiology, histopathology, ophthalmology, external-eye, X-ray, CT, MRI, fundus, echocardiographic, or related clinical images are the main evidentiary source for prediction, including radiomics, CNNs, transformers, self-supervised learning, relabeling, and image-centered image-plus-tabular models.","exclusion":"Assign to multimodal fusion when cross-modal integration is the central contribution, to signal-based prediction for ECG/EEG/wearable waveforms, or to disease-program directions when a named disease objective dominates over the imaging program.","nearest_boundary":"Multimodal fusion for outcome prediction","representative_topics":["multi-organ systemic biomarker detection in external eye photos via deep learning","deep learning-based opportunistic hip fracture risk prediction from plain radiographs and ct projections","mitigating supervision bias in lung cancer prediction through retrieval-based re-labeling"]},{"direction_index":2,"label":"D2","name":"Physiological and neurological signal prediction","description":"Prediction of cardiovascular, neurological, neuromotor, sleep, hemodynamic, and wearable-monitored health states from ECG, EEG, PPG, gait, smartphone, wearable, sleep, pump-flow, or other physiological time-series signals.","inclusion":"Assign atoms where biosignals or remote physiological measurements are the primary data source, including ECG cardiac risk, arrhythmia or atrial fibrillation prediction, EEG seizure or neurological prognosis, Parkinson's monitoring, gait-based disorder prediction, sleep-event detection, and wearable/sensor health prediction.","exclusion":"Assign to image-based prediction when the primary evidence is clinical imaging, to longitudinal EHR when the signal is secondary to patient-record sequences, or to named disease programs when the disease-screening objective dominates over the signal modality.","nearest_boundary":"Image-based clinical risk prediction","representative_topics":["feature engineering versus deep learning comparison for 12-lead ecg diagnosis and age estimation","zero-shot personalized ecg arrhythmia detection via sparse representation domain adaptation for wearable surveillance","atrial fibrillation recurrence risk prediction from pre- and post-ablation ecg morphological features"]},{"direction_index":3,"label":"D3","name":"Longitudinal EHR trajectory modeling","description":"Learning temporal patient representations from longitudinal EHRs, claims sequences, visit histories, clinical notes, event streams, and irregular patient journeys for clinical outcome prediction.","inclusion":"Assign atoms whose primary objective is modeling patient trajectories with recurrent, transformer, pretrained, generative, self-supervised, multitask, zero-shot, or time-aware architectures over longitudinal clinical records.","exclusion":"Assign to acute-care operations when the endpoint is hospital flow, ICU deterioration, sepsis, triage, or length of stay; to multimodal fusion when cross-modal integration is primary; or to privacy/trustworthiness when deployment validity, privacy, or uncertainty is the central objective.","nearest_boundary":"Acute-care and hospital operations prediction","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","generalizable multi-task clinical prediction via hierarchical ehr representation and multi-source self-supervised learning","transformer-based clinical note embedding for mortality prediction via umls concept mapping"]},{"direction_index":4,"label":"D4","name":"Acute-care and hospital operations prediction","description":"Early warning, sepsis prediction, ICU prognosis, emergency triage, hospital length of stay, discharge, readmission, acute-care use, and other resource-relevant predictions for acute-care decision support.","inclusion":"Assign atoms where the scientific objective is predicting deterioration, sepsis onset or outcome, critical-care prognosis, ED departure or triage, hospital flow, readmission, transfer, discharge placement, or length of stay for acute-care management.","exclusion":"Assign to COVID-specific programs for COVID acute outcomes, to longitudinal EHR trajectory modeling when the contribution is a general sequence architecture, or to trustworthiness when the main focus is leakage, causal validity, or fairness of acute-care models.","nearest_boundary":"Longitudinal EHR trajectory modeling","representative_topics":["early warning systems for critical health episodes using feature selection and gradient boosting","inherently interpretable bayesian survival modeling for discharge placement and readmission prediction","free-text clinical note analysis for acute care risk prediction in oncology patients"]},{"direction_index":5,"label":"D5","name":"Named disease and behavioral risk programs","description":"Condition-defined clinical prediction programs where the primary research objective is risk stratification, screening, diagnosis, prognosis, or treatment response for a named disease area or behavioral health outcome.","inclusion":"Assign atoms whose organizing objective is a specific disease or behavioral/psychiatric program, including COVID-19, post-COVID sequelae, mental health, autism, depression, suicide risk, food allergy, rheumatoid arthritis, breast cancer metastasis, and similar condition-tailored tools.","exclusion":"Assign to structured cohort, imaging, signal, multimodal, or EHR directions when the named disease is mainly a case study for a broader data-modality or method objective; assign COVID validity benchmarks to trustworthiness when methodological validity is primary.","nearest_boundary":"Structured cohort risk modeling","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]},{"direction_index":6,"label":"D6","name":"Multimodal fusion for outcome prediction","description":"Integration, alignment, imputation, or joint modeling of heterogeneous clinical modalities to improve patient-level prediction.","inclusion":"Assign atoms whose central scientific objective is fusing or completing multiple distinct data sources such as structured EHR, notes, imaging, laboratory values, motion, genetics, behavioral signals, or incomplete modalities for outcome prediction.","exclusion":"Assign to a single-modality direction when one modality supplies the main evidence; to named disease programs when the condition-specific objective dominates; or to trustworthiness when explanation, fairness, privacy, or uncertainty is the central contribution.","nearest_boundary":"Longitudinal EHR trajectory modeling","representative_topics":["multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","multimodal transformer for interpretable in-hospital mortality prediction from clinical notes and structured ehr"]},{"direction_index":7,"label":"D7","name":"Trustworthy and collaborative prediction","description":"Methods that make clinical prediction models interpretable, reliable, equitable, causally valid, portable, privacy-preserving, calibrated, uncertainty-aware, or safe for deployment.","inclusion":"Assign atoms whose primary objective is explanation, sparse risk scoring, clinician trust, uncertainty quantification, calibration, conformal or Bayesian reliability, fairness, bias mitigation, causal validation, leakage critique, cross-site portability, federated learning, secure computation, blockchain, or privacy-preserving collaboration.","exclusion":"Assign to the relevant disease, modality, EHR, acute-care, or multimodal program when explanation, uncertainty, fairness, privacy, or robustness is only an auxiliary analysis rather than the main scientific objective.","nearest_boundary":"Multimodal fusion for outcome prediction","representative_topics":["explainable meta-review-informed detection of cytokine release syndrome from multi-patient profiles","fast beam-search and star-ray rounding for interpretable sparse risk score learning","uncertainty estimation methods for deep learning models in clinical decision support"]}],"split":"dev"}
{"episode_id":"F196_T2026-01","field_index":196,"field":"clinical risk prediction","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structured cohort risk modeling","description":"Patient-level diagnosis, prognosis, survival, treatment-response, adherence, and outcome prediction from structured cohort, registry, claims, laboratory, demographic, genomic, or other tabular clinical variables.","inclusion":"Assign atoms whose main objective is building or evaluating risk models on structured/tabular patient data, including survival modeling, feature selection, classical machine learning, ensembles, missing-data handling, and imbalance handling when these serve the tabular cohort prediction program.","exclusion":"Assign to the modality, longitudinal EHR, acute-care, disease-program, multimodal, or trustworthiness direction when that objective is primary rather than general structured cohort prediction.","nearest_boundary":"Named disease and behavioral risk programs","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","sparse neural network variable selection for interval-censored accelerated failure time regression","class-imbalanced learning for adverse event severity prediction in immunization data"]},{"direction_index":1,"label":"D1","name":"Image-based clinical risk prediction","description":"Prediction of disease risk, severity, stage, prognosis, survival, treatment outcomes, or systemic biomarkers from medical images and image-derived representations.","inclusion":"Assign atoms where radiology, histopathology, ophthalmology, external-eye, X-ray, CT, MRI, fundus, echocardiographic, or related clinical images are the main evidentiary source for prediction, including radiomics, CNNs, transformers, self-supervised learning, relabeling, and image-centered image-plus-tabular models.","exclusion":"Assign to multimodal fusion when cross-modal integration is the central contribution, to signal-based prediction for ECG/EEG/wearable waveforms, or to disease-program directions when a named disease objective dominates over the imaging program.","nearest_boundary":"Multimodal fusion for outcome prediction","representative_topics":["multi-organ systemic biomarker detection in external eye photos via deep learning","deep learning-based opportunistic hip fracture risk prediction from plain radiographs and ct projections","mitigating supervision bias in lung cancer prediction through retrieval-based re-labeling"]},{"direction_index":2,"label":"D2","name":"Physiological and neurological signal prediction","description":"Prediction of cardiovascular, neurological, neuromotor, sleep, hemodynamic, and wearable-monitored health states from ECG, EEG, PPG, gait, smartphone, wearable, sleep, pump-flow, or other physiological time-series signals.","inclusion":"Assign atoms where biosignals or remote physiological measurements are the primary data source, including ECG cardiac risk, arrhythmia or atrial fibrillation prediction, EEG seizure or neurological prognosis, Parkinson's monitoring, gait-based disorder prediction, sleep-event detection, and wearable/sensor health prediction.","exclusion":"Assign to image-based prediction when the primary evidence is clinical imaging, to longitudinal EHR when the signal is secondary to patient-record sequences, or to named disease programs when the disease-screening objective dominates over the signal modality.","nearest_boundary":"Image-based clinical risk prediction","representative_topics":["feature engineering versus deep learning comparison for 12-lead ecg diagnosis and age estimation","zero-shot personalized ecg arrhythmia detection via sparse representation domain adaptation for wearable surveillance","atrial fibrillation recurrence risk prediction from pre- and post-ablation ecg morphological features"]},{"direction_index":3,"label":"D3","name":"Longitudinal EHR trajectory modeling","description":"Learning temporal patient representations from longitudinal EHRs, claims sequences, visit histories, clinical notes, event streams, and irregular patient journeys for clinical outcome prediction.","inclusion":"Assign atoms whose primary objective is modeling patient trajectories with recurrent, transformer, pretrained, generative, self-supervised, multitask, zero-shot, or time-aware architectures over longitudinal clinical records.","exclusion":"Assign to acute-care operations when the endpoint is hospital flow, ICU deterioration, sepsis, triage, or length of stay; to multimodal fusion when cross-modal integration is primary; or to privacy/trustworthiness when deployment validity, privacy, or uncertainty is the central objective.","nearest_boundary":"Acute-care and hospital operations prediction","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","generalizable multi-task clinical prediction via hierarchical ehr representation and multi-source self-supervised learning","transformer-based clinical note embedding for mortality prediction via umls concept mapping"]},{"direction_index":4,"label":"D4","name":"Acute-care and hospital operations prediction","description":"Early warning, sepsis prediction, ICU prognosis, emergency triage, hospital length of stay, discharge, readmission, acute-care use, and other resource-relevant predictions for acute-care decision support.","inclusion":"Assign atoms where the scientific objective is predicting deterioration, sepsis onset or outcome, critical-care prognosis, ED departure or triage, hospital flow, readmission, transfer, discharge placement, or length of stay for acute-care management.","exclusion":"Assign to COVID-specific programs for COVID acute outcomes, to longitudinal EHR trajectory modeling when the contribution is a general sequence architecture, or to trustworthiness when the main focus is leakage, causal validity, or fairness of acute-care models.","nearest_boundary":"Longitudinal EHR trajectory modeling","representative_topics":["early warning systems for critical health episodes using feature selection and gradient boosting","inherently interpretable bayesian survival modeling for discharge placement and readmission prediction","free-text clinical note analysis for acute care risk prediction in oncology patients"]},{"direction_index":5,"label":"D5","name":"Named disease and behavioral risk programs","description":"Condition-defined clinical prediction programs where the primary research objective is risk stratification, screening, diagnosis, prognosis, or treatment response for a named disease area or behavioral health outcome.","inclusion":"Assign atoms whose organizing objective is a specific disease or behavioral/psychiatric program, including COVID-19, post-COVID sequelae, mental health, autism, depression, suicide risk, food allergy, rheumatoid arthritis, breast cancer metastasis, and similar condition-tailored tools.","exclusion":"Assign to structured cohort, imaging, signal, multimodal, or EHR directions when the named disease is mainly a case study for a broader data-modality or method objective; assign COVID validity benchmarks to trustworthiness when methodological validity is primary.","nearest_boundary":"Structured cohort risk modeling","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]},{"direction_index":6,"label":"D6","name":"Multimodal fusion for outcome prediction","description":"Integration, alignment, imputation, or joint modeling of heterogeneous clinical modalities to improve patient-level prediction.","inclusion":"Assign atoms whose central scientific objective is fusing or completing multiple distinct data sources such as structured EHR, notes, imaging, laboratory values, motion, genetics, behavioral signals, or incomplete modalities for outcome prediction.","exclusion":"Assign to a single-modality direction when one modality supplies the main evidence; to named disease programs when the condition-specific objective dominates; or to trustworthiness when explanation, fairness, privacy, or uncertainty is the central contribution.","nearest_boundary":"Longitudinal EHR trajectory modeling","representative_topics":["multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","multimodal transformer for interpretable in-hospital mortality prediction from clinical notes and structured ehr"]},{"direction_index":7,"label":"D7","name":"Trustworthy and collaborative prediction","description":"Methods that make clinical prediction models interpretable, reliable, equitable, causally valid, portable, privacy-preserving, calibrated, uncertainty-aware, or safe for deployment.","inclusion":"Assign atoms whose primary objective is explanation, sparse risk scoring, clinician trust, uncertainty quantification, calibration, conformal or Bayesian reliability, fairness, bias mitigation, causal validation, leakage critique, cross-site portability, federated learning, secure computation, blockchain, or privacy-preserving collaboration.","exclusion":"Assign to the relevant disease, modality, EHR, acute-care, or multimodal program when explanation, uncertainty, fairness, privacy, or robustness is only an auxiliary analysis rather than the main scientific objective.","nearest_boundary":"Multimodal fusion for outcome prediction","representative_topics":["explainable meta-review-informed detection of cytokine release syndrome from multi-patient profiles","fast beam-search and star-ray rounding for interpretable sparse risk score learning","uncertainty estimation methods for deep learning models in clinical decision support"]}],"split":"dev"}
{"episode_id":"F197_T2024-01","field_index":197,"field":"spatiotemporal forecasting","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Aggregate transportation forecasting","description":"Forecasting aggregate states, demand, availability, flows, or service metrics for road networks and urban mobility systems.","inclusion":"Include papers whose primary objective is predicting road-segment or network traffic variables, travel time, incidents, OD demand, transit flow or arrivals, parking, EV-charging availability, construction activity, or other aggregate urban mobility service quantities.","exclusion":"Individual agent paths, poses, occupancy fields, or scene states belong to agent and scene motion forecasting; graph discovery as the main scientific objective belongs to dependency-structure learning; uncertainty-first formulations belong to probabilistic and data-limited forecasting.","nearest_boundary":"Agent and scene motion forecasting","representative_topics":["spatio-temporal road graph incident prediction using periodicity-aware traffic forecasting","decoupled diffusion-inherent signal separation for dynamic spatial-temporal traffic forecasting","social-aware transformer for long-term network-wide traffic forecasting using natural language"]},{"direction_index":1,"label":"D1","name":"Agent and scene motion forecasting","description":"Forecasting future motion of individual agents or dynamic scenes as trajectories, poses, multimodal futures, occupancy grids, flow fields, semantic maps, or complete scene states.","inclusion":"Include work centered on pedestrian, vehicle, driver, crowd, sports, industrial human-pose, autonomous-driving, evacuation, or scene-level future motion prediction, including generative or diversity-seeking multi-future trajectory models.","exclusion":"Aggregate road traffic flow, speed, demand, or service usage belongs to aggregate transportation forecasting; general video or PDE prediction without an explicit agent or scene-motion objective belongs to general architectures and infrastructure.","nearest_boundary":"Aggregate transportation forecasting","representative_topics":["multi-future pedestrian trajectory prediction via graph spatial transformer and memory replay smoothing","lane boundary-aware graph networks for motion forecasting with traffic rule constraints","spatio-temporal transformer networks for heterogeneous multi-agent trajectory prediction in autonomous driving"]},{"direction_index":2,"label":"D2","name":"Weather and precipitation forecasting","description":"Forecasting atmospheric, climate, earth-system, precipitation, radar, satellite, and meteorological fields from nowcasting through medium-range or broader weather-state prediction.","inclusion":"Include global or regional weather models, climate-variable forecasting, AI-NWP hybrids, precipitation-specific forecasting, radar or satellite nowcasting, weather benchmark datasets, and meteorological post-processing when the target is an atmospheric or climate state.","exclusion":"Air quality, wildfire, drought, hydrology, soil moisture, ecological, agricultural, or health outcomes belong to environmental and public-health forecasting; energy generation from wind or solar belongs to energy and engineered infrastructure forecasting when framed as system operation.","nearest_boundary":"Environmental and public-health forecasting","representative_topics":["deep learning bias correction for short-range global precipitation in hybrid nwp models","space-time transformer with cuboid attention for earth system forecasting","video transformer with shifted window cross-attention for spatiotemporal weather forecasting"]},{"direction_index":3,"label":"D3","name":"Environmental and public-health forecasting","description":"Forecasting non-weather environmental hazards, ecological or resource variables, and public-health or epidemic dynamics over space and time.","inclusion":"Include air pollution, drought, wildfire, soil moisture, hydrology, crop or ecological indicators, epidemic cases, COVID-19 dynamics, antibiotic resistance, metapopulation disease spread, and mobility-aware public-health forecasts.","exclusion":"Broad atmospheric weather and precipitation fields belong to weather and precipitation forecasting; engineered energy production or utility operation belongs to energy and engineered infrastructure forecasting; generic mobility demand without health outcome belongs to aggregate transportation forecasting.","nearest_boundary":"Weather and precipitation forecasting","representative_topics":["graph neural network epidemic forecasting with spatial transmission risk encoding and region-aware learning","spatio-temporal pm2.5 trend forecasting in african urban regions using neural network models","spatio-temporal neural network forecasting of regional drought severity with climate priors"]},{"direction_index":4,"label":"D4","name":"Energy and engineered infrastructure forecasting","description":"Forecasting operational quantities in energy systems and other specialized engineered or industrial spatiotemporal infrastructures outside ordinary road traffic and urban mobility services.","inclusion":"Include wind or solar power and irradiance for generation, power-grid states, energy load, flight delay propagation, railroad track geometry, reservoirs, telecommunications traffic, industrial infrastructure, and other specialized operational networks.","exclusion":"EV-charging, parking, taxi, transit, or road traffic service usage belongs to aggregate transportation forecasting when framed as urban mobility demand; environmental geophysical variables without infrastructure operation framing belong to environmental and public-health or weather forecasting.","nearest_boundary":"Aggregate transportation forecasting","representative_topics":["spatiotemporal embedding-based solar irradiance forecasting with robustness to missing data","spatiotemporal graph convolutional networks for network-wide flight delay propagation modeling","spatiotemporal wind power forecasting with dynamic turbine context and spatial distribution"]},{"direction_index":5,"label":"D5","name":"Dependency-structure learning","description":"Learning, adapting, discovering, sparsifying, or searching spatial, relational, causal, latent, hypergraph, or virtual dependency structures for spatiotemporal forecasting.","inclusion":"Include work whose primary scientific objective is the dependency structure itself, such as adaptive adjacency learning, graph search, dynamic graph construction, latent relation inference, causal graph learning, hypergraph learning, graph sparsification, or explicit-prior and learned-link fusion.","exclusion":"Using a learned or fixed graph mainly to improve a domain-specific target belongs to that target program; general efficient architectures without a graph-discovery objective belong to general architectures and infrastructure.","nearest_boundary":"General architectures and infrastructure","representative_topics":["regularized graph structure learning fusing explicit and implicit graphs for multivariate time-series forecasting","meta-graph structure learning for spatio-temporal graph convolutional recurrent forecasting","graph state-space models for joint latent relational structure and dynamics learning on time series"]},{"direction_index":6,"label":"D6","name":"Probabilistic and data-limited forecasting","description":"Forecasting under uncertainty, distributional objectives, sparse observations, missingness, irregular sampling, unobserved locations, or generative ensemble requirements for non-trajectory spatiotemporal targets.","inclusion":"Include calibrated predictive distributions, prediction intervals, uncertainty decomposition, Bayesian or variational models, diffusion or flow-based probabilistic forecasts, neural processes, sparse-sensor inference, kriging, imputation, matrix or tensor completion, and unobserved-node forecasting when uncertainty or incomplete observation is central.","exclusion":"Multimodal future paths for individual agents belong to agent and scene motion forecasting; ordinary point forecasting with incidental uncertainty or missing-data handling belongs to the relevant target-centered program.","nearest_boundary":"Agent and scene motion forecasting","representative_topics":["uncertainty quantification for spatio-temporal traffic forecasting via aleatoric and epistemic decomposition","uncertainty quantification for sparse od travel demand using zero-inflated negative binomial graph neural networks","non-consecutive spatiotemporal forecasting under missing observations via factor-decoupled hierarchical learning"]},{"direction_index":7,"label":"D7","name":"General architectures and infrastructure","description":"Developing broadly applicable spatiotemporal forecasting representations, architectures, training paradigms, transfer mechanisms, robustness methods, benchmarks, surveys, toolboxes, and evaluation infrastructure.","inclusion":"Include general-purpose transformers, recurrent-free or lightweight architectures, continuous-time or implicit representations, PDE dynamics models, self-supervised, contrastive, federated, meta-learning, cross-region or cross-task generalization, domain shift robustness, benchmarks, surveys, datasets, and systematic comparisons when the objective is not a single application target.","exclusion":"Adaptive graph discovery belongs to dependency-structure learning; probabilistic or sparse-observation objectives belong to probabilistic and data-limited forecasting; domain-specific architectures whose main aim is skill on one target belong to that target program.","nearest_boundary":"Dependency-structure learning","representative_topics":["multi-scale recurrent framework for memory-efficient spatiotemporal predictive learning","federated spatial-temporal graph learning with dynamic topology and relevance-driven aggregation","identity-aware baseline for multivariate time series forecasting via spatial-temporal feature tagging"]}],"split":"dev"}
{"episode_id":"F197_T2024-07","field_index":197,"field":"spatiotemporal forecasting","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Aggregate transportation forecasting","description":"Forecasting aggregate states, demand, availability, flows, or service metrics for road networks and urban mobility systems.","inclusion":"Include papers whose primary objective is predicting road-segment or network traffic variables, travel time, incidents, OD demand, transit flow or arrivals, parking, EV-charging availability, construction activity, or other aggregate urban mobility service quantities.","exclusion":"Individual agent paths, poses, occupancy fields, or scene states belong to agent and scene motion forecasting; graph discovery as the main scientific objective belongs to dependency-structure learning; uncertainty-first formulations belong to probabilistic and data-limited forecasting.","nearest_boundary":"Agent and scene motion forecasting","representative_topics":["spatio-temporal road graph incident prediction using periodicity-aware traffic forecasting","decoupled diffusion-inherent signal separation for dynamic spatial-temporal traffic forecasting","social-aware transformer for long-term network-wide traffic forecasting using natural language"]},{"direction_index":1,"label":"D1","name":"Agent and scene motion forecasting","description":"Forecasting future motion of individual agents or dynamic scenes as trajectories, poses, multimodal futures, occupancy grids, flow fields, semantic maps, or complete scene states.","inclusion":"Include work centered on pedestrian, vehicle, driver, crowd, sports, industrial human-pose, autonomous-driving, evacuation, or scene-level future motion prediction, including generative or diversity-seeking multi-future trajectory models.","exclusion":"Aggregate road traffic flow, speed, demand, or service usage belongs to aggregate transportation forecasting; general video or PDE prediction without an explicit agent or scene-motion objective belongs to general architectures and infrastructure.","nearest_boundary":"Aggregate transportation forecasting","representative_topics":["multi-future pedestrian trajectory prediction via graph spatial transformer and memory replay smoothing","lane boundary-aware graph networks for motion forecasting with traffic rule constraints","spatio-temporal transformer networks for heterogeneous multi-agent trajectory prediction in autonomous driving"]},{"direction_index":2,"label":"D2","name":"Weather and precipitation forecasting","description":"Forecasting atmospheric, climate, earth-system, precipitation, radar, satellite, and meteorological fields from nowcasting through medium-range or broader weather-state prediction.","inclusion":"Include global or regional weather models, climate-variable forecasting, AI-NWP hybrids, precipitation-specific forecasting, radar or satellite nowcasting, weather benchmark datasets, and meteorological post-processing when the target is an atmospheric or climate state.","exclusion":"Air quality, wildfire, drought, hydrology, soil moisture, ecological, agricultural, or health outcomes belong to environmental and public-health forecasting; energy generation from wind or solar belongs to energy and engineered infrastructure forecasting when framed as system operation.","nearest_boundary":"Environmental and public-health forecasting","representative_topics":["deep learning bias correction for short-range global precipitation in hybrid nwp models","space-time transformer with cuboid attention for earth system forecasting","video transformer with shifted window cross-attention for spatiotemporal weather forecasting"]},{"direction_index":3,"label":"D3","name":"Environmental and public-health forecasting","description":"Forecasting non-weather environmental hazards, ecological or resource variables, and public-health or epidemic dynamics over space and time.","inclusion":"Include air pollution, drought, wildfire, soil moisture, hydrology, crop or ecological indicators, epidemic cases, COVID-19 dynamics, antibiotic resistance, metapopulation disease spread, and mobility-aware public-health forecasts.","exclusion":"Broad atmospheric weather and precipitation fields belong to weather and precipitation forecasting; engineered energy production or utility operation belongs to energy and engineered infrastructure forecasting; generic mobility demand without health outcome belongs to aggregate transportation forecasting.","nearest_boundary":"Weather and precipitation forecasting","representative_topics":["graph neural network epidemic forecasting with spatial transmission risk encoding and region-aware learning","spatio-temporal pm2.5 trend forecasting in african urban regions using neural network models","spatio-temporal neural network forecasting of regional drought severity with climate priors"]},{"direction_index":4,"label":"D4","name":"Energy and engineered infrastructure forecasting","description":"Forecasting operational quantities in energy systems and other specialized engineered or industrial spatiotemporal infrastructures outside ordinary road traffic and urban mobility services.","inclusion":"Include wind or solar power and irradiance for generation, power-grid states, energy load, flight delay propagation, railroad track geometry, reservoirs, telecommunications traffic, industrial infrastructure, and other specialized operational networks.","exclusion":"EV-charging, parking, taxi, transit, or road traffic service usage belongs to aggregate transportation forecasting when framed as urban mobility demand; environmental geophysical variables without infrastructure operation framing belong to environmental and public-health or weather forecasting.","nearest_boundary":"Aggregate transportation forecasting","representative_topics":["spatiotemporal embedding-based solar irradiance forecasting with robustness to missing data","spatiotemporal graph convolutional networks for network-wide flight delay propagation modeling","spatiotemporal wind power forecasting with dynamic turbine context and spatial distribution"]},{"direction_index":5,"label":"D5","name":"Dependency-structure learning","description":"Learning, adapting, discovering, sparsifying, or searching spatial, relational, causal, latent, hypergraph, or virtual dependency structures for spatiotemporal forecasting.","inclusion":"Include work whose primary scientific objective is the dependency structure itself, such as adaptive adjacency learning, graph search, dynamic graph construction, latent relation inference, causal graph learning, hypergraph learning, graph sparsification, or explicit-prior and learned-link fusion.","exclusion":"Using a learned or fixed graph mainly to improve a domain-specific target belongs to that target program; general efficient architectures without a graph-discovery objective belong to general architectures and infrastructure.","nearest_boundary":"General architectures and infrastructure","representative_topics":["regularized graph structure learning fusing explicit and implicit graphs for multivariate time-series forecasting","meta-graph structure learning for spatio-temporal graph convolutional recurrent forecasting","graph state-space models for joint latent relational structure and dynamics learning on time series"]},{"direction_index":6,"label":"D6","name":"Probabilistic and data-limited forecasting","description":"Forecasting under uncertainty, distributional objectives, sparse observations, missingness, irregular sampling, unobserved locations, or generative ensemble requirements for non-trajectory spatiotemporal targets.","inclusion":"Include calibrated predictive distributions, prediction intervals, uncertainty decomposition, Bayesian or variational models, diffusion or flow-based probabilistic forecasts, neural processes, sparse-sensor inference, kriging, imputation, matrix or tensor completion, and unobserved-node forecasting when uncertainty or incomplete observation is central.","exclusion":"Multimodal future paths for individual agents belong to agent and scene motion forecasting; ordinary point forecasting with incidental uncertainty or missing-data handling belongs to the relevant target-centered program.","nearest_boundary":"Agent and scene motion forecasting","representative_topics":["uncertainty quantification for spatio-temporal traffic forecasting via aleatoric and epistemic decomposition","uncertainty quantification for sparse od travel demand using zero-inflated negative binomial graph neural networks","non-consecutive spatiotemporal forecasting under missing observations via factor-decoupled hierarchical learning"]},{"direction_index":7,"label":"D7","name":"General architectures and infrastructure","description":"Developing broadly applicable spatiotemporal forecasting representations, architectures, training paradigms, transfer mechanisms, robustness methods, benchmarks, surveys, toolboxes, and evaluation infrastructure.","inclusion":"Include general-purpose transformers, recurrent-free or lightweight architectures, continuous-time or implicit representations, PDE dynamics models, self-supervised, contrastive, federated, meta-learning, cross-region or cross-task generalization, domain shift robustness, benchmarks, surveys, datasets, and systematic comparisons when the objective is not a single application target.","exclusion":"Adaptive graph discovery belongs to dependency-structure learning; probabilistic or sparse-observation objectives belong to probabilistic and data-limited forecasting; domain-specific architectures whose main aim is skill on one target belong to that target program.","nearest_boundary":"Dependency-structure learning","representative_topics":["multi-scale recurrent framework for memory-efficient spatiotemporal predictive learning","federated spatial-temporal graph learning with dynamic topology and relevance-driven aggregation","identity-aware baseline for multivariate time series forecasting via spatial-temporal feature tagging"]}],"split":"dev"}
{"episode_id":"F197_T2025-01","field_index":197,"field":"spatiotemporal forecasting","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Aggregate transportation forecasting","description":"Forecasting aggregate states, demand, availability, flows, or service metrics for road networks and urban mobility systems.","inclusion":"Include papers whose primary objective is predicting road-segment or network traffic variables, travel time, incidents, OD demand, transit flow or arrivals, parking, EV-charging availability, construction activity, or other aggregate urban mobility service quantities.","exclusion":"Individual agent paths, poses, occupancy fields, or scene states belong to agent and scene motion forecasting; graph discovery as the main scientific objective belongs to dependency-structure learning; uncertainty-first formulations belong to probabilistic and data-limited forecasting.","nearest_boundary":"Agent and scene motion forecasting","representative_topics":["spatio-temporal road graph incident prediction using periodicity-aware traffic forecasting","decoupled diffusion-inherent signal separation for dynamic spatial-temporal traffic forecasting","social-aware transformer for long-term network-wide traffic forecasting using natural language"]},{"direction_index":1,"label":"D1","name":"Agent and scene motion forecasting","description":"Forecasting future motion of individual agents or dynamic scenes as trajectories, poses, multimodal futures, occupancy grids, flow fields, semantic maps, or complete scene states.","inclusion":"Include work centered on pedestrian, vehicle, driver, crowd, sports, industrial human-pose, autonomous-driving, evacuation, or scene-level future motion prediction, including generative or diversity-seeking multi-future trajectory models.","exclusion":"Aggregate road traffic flow, speed, demand, or service usage belongs to aggregate transportation forecasting; general video or PDE prediction without an explicit agent or scene-motion objective belongs to general architectures and infrastructure.","nearest_boundary":"Aggregate transportation forecasting","representative_topics":["multi-future pedestrian trajectory prediction via graph spatial transformer and memory replay smoothing","lane boundary-aware graph networks for motion forecasting with traffic rule constraints","spatio-temporal transformer networks for heterogeneous multi-agent trajectory prediction in autonomous driving"]},{"direction_index":2,"label":"D2","name":"Weather and precipitation forecasting","description":"Forecasting atmospheric, climate, earth-system, precipitation, radar, satellite, and meteorological fields from nowcasting through medium-range or broader weather-state prediction.","inclusion":"Include global or regional weather models, climate-variable forecasting, AI-NWP hybrids, precipitation-specific forecasting, radar or satellite nowcasting, weather benchmark datasets, and meteorological post-processing when the target is an atmospheric or climate state.","exclusion":"Air quality, wildfire, drought, hydrology, soil moisture, ecological, agricultural, or health outcomes belong to environmental and public-health forecasting; energy generation from wind or solar belongs to energy and engineered infrastructure forecasting when framed as system operation.","nearest_boundary":"Environmental and public-health forecasting","representative_topics":["deep learning bias correction for short-range global precipitation in hybrid nwp models","space-time transformer with cuboid attention for earth system forecasting","video transformer with shifted window cross-attention for spatiotemporal weather forecasting"]},{"direction_index":3,"label":"D3","name":"Environmental and public-health forecasting","description":"Forecasting non-weather environmental hazards, ecological or resource variables, and public-health or epidemic dynamics over space and time.","inclusion":"Include air pollution, drought, wildfire, soil moisture, hydrology, crop or ecological indicators, epidemic cases, COVID-19 dynamics, antibiotic resistance, metapopulation disease spread, and mobility-aware public-health forecasts.","exclusion":"Broad atmospheric weather and precipitation fields belong to weather and precipitation forecasting; engineered energy production or utility operation belongs to energy and engineered infrastructure forecasting; generic mobility demand without health outcome belongs to aggregate transportation forecasting.","nearest_boundary":"Weather and precipitation forecasting","representative_topics":["graph neural network epidemic forecasting with spatial transmission risk encoding and region-aware learning","spatio-temporal pm2.5 trend forecasting in african urban regions using neural network models","spatio-temporal neural network forecasting of regional drought severity with climate priors"]},{"direction_index":4,"label":"D4","name":"Energy and engineered infrastructure forecasting","description":"Forecasting operational quantities in energy systems and other specialized engineered or industrial spatiotemporal infrastructures outside ordinary road traffic and urban mobility services.","inclusion":"Include wind or solar power and irradiance for generation, power-grid states, energy load, flight delay propagation, railroad track geometry, reservoirs, telecommunications traffic, industrial infrastructure, and other specialized operational networks.","exclusion":"EV-charging, parking, taxi, transit, or road traffic service usage belongs to aggregate transportation forecasting when framed as urban mobility demand; environmental geophysical variables without infrastructure operation framing belong to environmental and public-health or weather forecasting.","nearest_boundary":"Aggregate transportation forecasting","representative_topics":["spatiotemporal embedding-based solar irradiance forecasting with robustness to missing data","spatiotemporal graph convolutional networks for network-wide flight delay propagation modeling","spatiotemporal wind power forecasting with dynamic turbine context and spatial distribution"]},{"direction_index":5,"label":"D5","name":"Dependency-structure learning","description":"Learning, adapting, discovering, sparsifying, or searching spatial, relational, causal, latent, hypergraph, or virtual dependency structures for spatiotemporal forecasting.","inclusion":"Include work whose primary scientific objective is the dependency structure itself, such as adaptive adjacency learning, graph search, dynamic graph construction, latent relation inference, causal graph learning, hypergraph learning, graph sparsification, or explicit-prior and learned-link fusion.","exclusion":"Using a learned or fixed graph mainly to improve a domain-specific target belongs to that target program; general efficient architectures without a graph-discovery objective belong to general architectures and infrastructure.","nearest_boundary":"General architectures and infrastructure","representative_topics":["regularized graph structure learning fusing explicit and implicit graphs for multivariate time-series forecasting","meta-graph structure learning for spatio-temporal graph convolutional recurrent forecasting","graph state-space models for joint latent relational structure and dynamics learning on time series"]},{"direction_index":6,"label":"D6","name":"Probabilistic and data-limited forecasting","description":"Forecasting under uncertainty, distributional objectives, sparse observations, missingness, irregular sampling, unobserved locations, or generative ensemble requirements for non-trajectory spatiotemporal targets.","inclusion":"Include calibrated predictive distributions, prediction intervals, uncertainty decomposition, Bayesian or variational models, diffusion or flow-based probabilistic forecasts, neural processes, sparse-sensor inference, kriging, imputation, matrix or tensor completion, and unobserved-node forecasting when uncertainty or incomplete observation is central.","exclusion":"Multimodal future paths for individual agents belong to agent and scene motion forecasting; ordinary point forecasting with incidental uncertainty or missing-data handling belongs to the relevant target-centered program.","nearest_boundary":"Agent and scene motion forecasting","representative_topics":["uncertainty quantification for spatio-temporal traffic forecasting via aleatoric and epistemic decomposition","uncertainty quantification for sparse od travel demand using zero-inflated negative binomial graph neural networks","non-consecutive spatiotemporal forecasting under missing observations via factor-decoupled hierarchical learning"]},{"direction_index":7,"label":"D7","name":"General architectures and infrastructure","description":"Developing broadly applicable spatiotemporal forecasting representations, architectures, training paradigms, transfer mechanisms, robustness methods, benchmarks, surveys, toolboxes, and evaluation infrastructure.","inclusion":"Include general-purpose transformers, recurrent-free or lightweight architectures, continuous-time or implicit representations, PDE dynamics models, self-supervised, contrastive, federated, meta-learning, cross-region or cross-task generalization, domain shift robustness, benchmarks, surveys, datasets, and systematic comparisons when the objective is not a single application target.","exclusion":"Adaptive graph discovery belongs to dependency-structure learning; probabilistic or sparse-observation objectives belong to probabilistic and data-limited forecasting; domain-specific architectures whose main aim is skill on one target belong to that target program.","nearest_boundary":"Dependency-structure learning","representative_topics":["multi-scale recurrent framework for memory-efficient spatiotemporal predictive learning","federated spatial-temporal graph learning with dynamic topology and relevance-driven aggregation","identity-aware baseline for multivariate time series forecasting via spatial-temporal feature tagging"]}],"split":"dev"}
{"episode_id":"F197_T2025-07","field_index":197,"field":"spatiotemporal forecasting","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Aggregate transportation forecasting","description":"Forecasting aggregate states, demand, availability, flows, or service metrics for road networks and urban mobility systems.","inclusion":"Include papers whose primary objective is predicting road-segment or network traffic variables, travel time, incidents, OD demand, transit flow or arrivals, parking, EV-charging availability, construction activity, or other aggregate urban mobility service quantities.","exclusion":"Individual agent paths, poses, occupancy fields, or scene states belong to agent and scene motion forecasting; graph discovery as the main scientific objective belongs to dependency-structure learning; uncertainty-first formulations belong to probabilistic and data-limited forecasting.","nearest_boundary":"Agent and scene motion forecasting","representative_topics":["spatio-temporal road graph incident prediction using periodicity-aware traffic forecasting","decoupled diffusion-inherent signal separation for dynamic spatial-temporal traffic forecasting","social-aware transformer for long-term network-wide traffic forecasting using natural language"]},{"direction_index":1,"label":"D1","name":"Agent and scene motion forecasting","description":"Forecasting future motion of individual agents or dynamic scenes as trajectories, poses, multimodal futures, occupancy grids, flow fields, semantic maps, or complete scene states.","inclusion":"Include work centered on pedestrian, vehicle, driver, crowd, sports, industrial human-pose, autonomous-driving, evacuation, or scene-level future motion prediction, including generative or diversity-seeking multi-future trajectory models.","exclusion":"Aggregate road traffic flow, speed, demand, or service usage belongs to aggregate transportation forecasting; general video or PDE prediction without an explicit agent or scene-motion objective belongs to general architectures and infrastructure.","nearest_boundary":"Aggregate transportation forecasting","representative_topics":["multi-future pedestrian trajectory prediction via graph spatial transformer and memory replay smoothing","lane boundary-aware graph networks for motion forecasting with traffic rule constraints","spatio-temporal transformer networks for heterogeneous multi-agent trajectory prediction in autonomous driving"]},{"direction_index":2,"label":"D2","name":"Weather and precipitation forecasting","description":"Forecasting atmospheric, climate, earth-system, precipitation, radar, satellite, and meteorological fields from nowcasting through medium-range or broader weather-state prediction.","inclusion":"Include global or regional weather models, climate-variable forecasting, AI-NWP hybrids, precipitation-specific forecasting, radar or satellite nowcasting, weather benchmark datasets, and meteorological post-processing when the target is an atmospheric or climate state.","exclusion":"Air quality, wildfire, drought, hydrology, soil moisture, ecological, agricultural, or health outcomes belong to environmental and public-health forecasting; energy generation from wind or solar belongs to energy and engineered infrastructure forecasting when framed as system operation.","nearest_boundary":"Environmental and public-health forecasting","representative_topics":["deep learning bias correction for short-range global precipitation in hybrid nwp models","space-time transformer with cuboid attention for earth system forecasting","video transformer with shifted window cross-attention for spatiotemporal weather forecasting"]},{"direction_index":3,"label":"D3","name":"Environmental and public-health forecasting","description":"Forecasting non-weather environmental hazards, ecological or resource variables, and public-health or epidemic dynamics over space and time.","inclusion":"Include air pollution, drought, wildfire, soil moisture, hydrology, crop or ecological indicators, epidemic cases, COVID-19 dynamics, antibiotic resistance, metapopulation disease spread, and mobility-aware public-health forecasts.","exclusion":"Broad atmospheric weather and precipitation fields belong to weather and precipitation forecasting; engineered energy production or utility operation belongs to energy and engineered infrastructure forecasting; generic mobility demand without health outcome belongs to aggregate transportation forecasting.","nearest_boundary":"Weather and precipitation forecasting","representative_topics":["graph neural network epidemic forecasting with spatial transmission risk encoding and region-aware learning","spatio-temporal pm2.5 trend forecasting in african urban regions using neural network models","spatio-temporal neural network forecasting of regional drought severity with climate priors"]},{"direction_index":4,"label":"D4","name":"Energy and engineered infrastructure forecasting","description":"Forecasting operational quantities in energy systems and other specialized engineered or industrial spatiotemporal infrastructures outside ordinary road traffic and urban mobility services.","inclusion":"Include wind or solar power and irradiance for generation, power-grid states, energy load, flight delay propagation, railroad track geometry, reservoirs, telecommunications traffic, industrial infrastructure, and other specialized operational networks.","exclusion":"EV-charging, parking, taxi, transit, or road traffic service usage belongs to aggregate transportation forecasting when framed as urban mobility demand; environmental geophysical variables without infrastructure operation framing belong to environmental and public-health or weather forecasting.","nearest_boundary":"Aggregate transportation forecasting","representative_topics":["spatiotemporal embedding-based solar irradiance forecasting with robustness to missing data","spatiotemporal graph convolutional networks for network-wide flight delay propagation modeling","spatiotemporal wind power forecasting with dynamic turbine context and spatial distribution"]},{"direction_index":5,"label":"D5","name":"Dependency-structure learning","description":"Learning, adapting, discovering, sparsifying, or searching spatial, relational, causal, latent, hypergraph, or virtual dependency structures for spatiotemporal forecasting.","inclusion":"Include work whose primary scientific objective is the dependency structure itself, such as adaptive adjacency learning, graph search, dynamic graph construction, latent relation inference, causal graph learning, hypergraph learning, graph sparsification, or explicit-prior and learned-link fusion.","exclusion":"Using a learned or fixed graph mainly to improve a domain-specific target belongs to that target program; general efficient architectures without a graph-discovery objective belong to general architectures and infrastructure.","nearest_boundary":"General architectures and infrastructure","representative_topics":["regularized graph structure learning fusing explicit and implicit graphs for multivariate time-series forecasting","meta-graph structure learning for spatio-temporal graph convolutional recurrent forecasting","graph state-space models for joint latent relational structure and dynamics learning on time series"]},{"direction_index":6,"label":"D6","name":"Probabilistic and data-limited forecasting","description":"Forecasting under uncertainty, distributional objectives, sparse observations, missingness, irregular sampling, unobserved locations, or generative ensemble requirements for non-trajectory spatiotemporal targets.","inclusion":"Include calibrated predictive distributions, prediction intervals, uncertainty decomposition, Bayesian or variational models, diffusion or flow-based probabilistic forecasts, neural processes, sparse-sensor inference, kriging, imputation, matrix or tensor completion, and unobserved-node forecasting when uncertainty or incomplete observation is central.","exclusion":"Multimodal future paths for individual agents belong to agent and scene motion forecasting; ordinary point forecasting with incidental uncertainty or missing-data handling belongs to the relevant target-centered program.","nearest_boundary":"Agent and scene motion forecasting","representative_topics":["uncertainty quantification for spatio-temporal traffic forecasting via aleatoric and epistemic decomposition","uncertainty quantification for sparse od travel demand using zero-inflated negative binomial graph neural networks","non-consecutive spatiotemporal forecasting under missing observations via factor-decoupled hierarchical learning"]},{"direction_index":7,"label":"D7","name":"General architectures and infrastructure","description":"Developing broadly applicable spatiotemporal forecasting representations, architectures, training paradigms, transfer mechanisms, robustness methods, benchmarks, surveys, toolboxes, and evaluation infrastructure.","inclusion":"Include general-purpose transformers, recurrent-free or lightweight architectures, continuous-time or implicit representations, PDE dynamics models, self-supervised, contrastive, federated, meta-learning, cross-region or cross-task generalization, domain shift robustness, benchmarks, surveys, datasets, and systematic comparisons when the objective is not a single application target.","exclusion":"Adaptive graph discovery belongs to dependency-structure learning; probabilistic or sparse-observation objectives belong to probabilistic and data-limited forecasting; domain-specific architectures whose main aim is skill on one target belong to that target program.","nearest_boundary":"Dependency-structure learning","representative_topics":["multi-scale recurrent framework for memory-efficient spatiotemporal predictive learning","federated spatial-temporal graph learning with dynamic topology and relevance-driven aggregation","identity-aware baseline for multivariate time series forecasting via spatial-temporal feature tagging"]}],"split":"dev"}
{"episode_id":"F197_T2026-01","field_index":197,"field":"spatiotemporal forecasting","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Aggregate transportation forecasting","description":"Forecasting aggregate states, demand, availability, flows, or service metrics for road networks and urban mobility systems.","inclusion":"Include papers whose primary objective is predicting road-segment or network traffic variables, travel time, incidents, OD demand, transit flow or arrivals, parking, EV-charging availability, construction activity, or other aggregate urban mobility service quantities.","exclusion":"Individual agent paths, poses, occupancy fields, or scene states belong to agent and scene motion forecasting; graph discovery as the main scientific objective belongs to dependency-structure learning; uncertainty-first formulations belong to probabilistic and data-limited forecasting.","nearest_boundary":"Agent and scene motion forecasting","representative_topics":["spatio-temporal road graph incident prediction using periodicity-aware traffic forecasting","decoupled diffusion-inherent signal separation for dynamic spatial-temporal traffic forecasting","social-aware transformer for long-term network-wide traffic forecasting using natural language"]},{"direction_index":1,"label":"D1","name":"Agent and scene motion forecasting","description":"Forecasting future motion of individual agents or dynamic scenes as trajectories, poses, multimodal futures, occupancy grids, flow fields, semantic maps, or complete scene states.","inclusion":"Include work centered on pedestrian, vehicle, driver, crowd, sports, industrial human-pose, autonomous-driving, evacuation, or scene-level future motion prediction, including generative or diversity-seeking multi-future trajectory models.","exclusion":"Aggregate road traffic flow, speed, demand, or service usage belongs to aggregate transportation forecasting; general video or PDE prediction without an explicit agent or scene-motion objective belongs to general architectures and infrastructure.","nearest_boundary":"Aggregate transportation forecasting","representative_topics":["multi-future pedestrian trajectory prediction via graph spatial transformer and memory replay smoothing","lane boundary-aware graph networks for motion forecasting with traffic rule constraints","spatio-temporal transformer networks for heterogeneous multi-agent trajectory prediction in autonomous driving"]},{"direction_index":2,"label":"D2","name":"Weather and precipitation forecasting","description":"Forecasting atmospheric, climate, earth-system, precipitation, radar, satellite, and meteorological fields from nowcasting through medium-range or broader weather-state prediction.","inclusion":"Include global or regional weather models, climate-variable forecasting, AI-NWP hybrids, precipitation-specific forecasting, radar or satellite nowcasting, weather benchmark datasets, and meteorological post-processing when the target is an atmospheric or climate state.","exclusion":"Air quality, wildfire, drought, hydrology, soil moisture, ecological, agricultural, or health outcomes belong to environmental and public-health forecasting; energy generation from wind or solar belongs to energy and engineered infrastructure forecasting when framed as system operation.","nearest_boundary":"Environmental and public-health forecasting","representative_topics":["deep learning bias correction for short-range global precipitation in hybrid nwp models","space-time transformer with cuboid attention for earth system forecasting","video transformer with shifted window cross-attention for spatiotemporal weather forecasting"]},{"direction_index":3,"label":"D3","name":"Environmental and public-health forecasting","description":"Forecasting non-weather environmental hazards, ecological or resource variables, and public-health or epidemic dynamics over space and time.","inclusion":"Include air pollution, drought, wildfire, soil moisture, hydrology, crop or ecological indicators, epidemic cases, COVID-19 dynamics, antibiotic resistance, metapopulation disease spread, and mobility-aware public-health forecasts.","exclusion":"Broad atmospheric weather and precipitation fields belong to weather and precipitation forecasting; engineered energy production or utility operation belongs to energy and engineered infrastructure forecasting; generic mobility demand without health outcome belongs to aggregate transportation forecasting.","nearest_boundary":"Weather and precipitation forecasting","representative_topics":["graph neural network epidemic forecasting with spatial transmission risk encoding and region-aware learning","spatio-temporal pm2.5 trend forecasting in african urban regions using neural network models","spatio-temporal neural network forecasting of regional drought severity with climate priors"]},{"direction_index":4,"label":"D4","name":"Energy and engineered infrastructure forecasting","description":"Forecasting operational quantities in energy systems and other specialized engineered or industrial spatiotemporal infrastructures outside ordinary road traffic and urban mobility services.","inclusion":"Include wind or solar power and irradiance for generation, power-grid states, energy load, flight delay propagation, railroad track geometry, reservoirs, telecommunications traffic, industrial infrastructure, and other specialized operational networks.","exclusion":"EV-charging, parking, taxi, transit, or road traffic service usage belongs to aggregate transportation forecasting when framed as urban mobility demand; environmental geophysical variables without infrastructure operation framing belong to environmental and public-health or weather forecasting.","nearest_boundary":"Aggregate transportation forecasting","representative_topics":["spatiotemporal embedding-based solar irradiance forecasting with robustness to missing data","spatiotemporal graph convolutional networks for network-wide flight delay propagation modeling","spatiotemporal wind power forecasting with dynamic turbine context and spatial distribution"]},{"direction_index":5,"label":"D5","name":"Dependency-structure learning","description":"Learning, adapting, discovering, sparsifying, or searching spatial, relational, causal, latent, hypergraph, or virtual dependency structures for spatiotemporal forecasting.","inclusion":"Include work whose primary scientific objective is the dependency structure itself, such as adaptive adjacency learning, graph search, dynamic graph construction, latent relation inference, causal graph learning, hypergraph learning, graph sparsification, or explicit-prior and learned-link fusion.","exclusion":"Using a learned or fixed graph mainly to improve a domain-specific target belongs to that target program; general efficient architectures without a graph-discovery objective belong to general architectures and infrastructure.","nearest_boundary":"General architectures and infrastructure","representative_topics":["regularized graph structure learning fusing explicit and implicit graphs for multivariate time-series forecasting","meta-graph structure learning for spatio-temporal graph convolutional recurrent forecasting","graph state-space models for joint latent relational structure and dynamics learning on time series"]},{"direction_index":6,"label":"D6","name":"Probabilistic and data-limited forecasting","description":"Forecasting under uncertainty, distributional objectives, sparse observations, missingness, irregular sampling, unobserved locations, or generative ensemble requirements for non-trajectory spatiotemporal targets.","inclusion":"Include calibrated predictive distributions, prediction intervals, uncertainty decomposition, Bayesian or variational models, diffusion or flow-based probabilistic forecasts, neural processes, sparse-sensor inference, kriging, imputation, matrix or tensor completion, and unobserved-node forecasting when uncertainty or incomplete observation is central.","exclusion":"Multimodal future paths for individual agents belong to agent and scene motion forecasting; ordinary point forecasting with incidental uncertainty or missing-data handling belongs to the relevant target-centered program.","nearest_boundary":"Agent and scene motion forecasting","representative_topics":["uncertainty quantification for spatio-temporal traffic forecasting via aleatoric and epistemic decomposition","uncertainty quantification for sparse od travel demand using zero-inflated negative binomial graph neural networks","non-consecutive spatiotemporal forecasting under missing observations via factor-decoupled hierarchical learning"]},{"direction_index":7,"label":"D7","name":"General architectures and infrastructure","description":"Developing broadly applicable spatiotemporal forecasting representations, architectures, training paradigms, transfer mechanisms, robustness methods, benchmarks, surveys, toolboxes, and evaluation infrastructure.","inclusion":"Include general-purpose transformers, recurrent-free or lightweight architectures, continuous-time or implicit representations, PDE dynamics models, self-supervised, contrastive, federated, meta-learning, cross-region or cross-task generalization, domain shift robustness, benchmarks, surveys, datasets, and systematic comparisons when the objective is not a single application target.","exclusion":"Adaptive graph discovery belongs to dependency-structure learning; probabilistic or sparse-observation objectives belong to probabilistic and data-limited forecasting; domain-specific architectures whose main aim is skill on one target belong to that target program.","nearest_boundary":"Dependency-structure learning","representative_topics":["multi-scale recurrent framework for memory-efficient spatiotemporal predictive learning","federated spatial-temporal graph learning with dynamic topology and relevance-driven aggregation","identity-aware baseline for multivariate time series forecasting via spatial-temporal feature tagging"]}],"split":"dev"}
{"episode_id":"F199_T2024-01","field_index":199,"field":"human-robot interaction","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Language-grounded embodied interaction","description":"Robots and embodied agents whose main objective is to use natural language, dialogue, or language/vision-language grounding to understand instructions, resolve ambiguity, explain behavior, or connect conversational interaction to situated action.","inclusion":"Include work where the central contribution is language grounding, action-language translation, situated dialogue policy, clarification or repair, referent disambiguation, language-model planning for embodied action, or multimodal face-to-face interaction organized around language.","exclusion":"If language is only a cue inside a primarily navigation, accessibility, shared-control, physical-skill, or team-planning contribution, assign to that nearer objective.","nearest_boundary":"Human-aware team planning and coordination","representative_topics":["multi-task robot action-language translation via paired gated autoencoders with direction-signal conditioning","multi-view referred object disambiguation and query generation for mobile embodied agents","dialogue policy design for confusion mitigation in situated human-robot interaction"]},{"direction_index":1,"label":"D1","name":"Social navigation in human spaces","description":"Navigation, locomotion, trajectory prediction, and local planning methods that enable robots to move safely, legibly, and socially appropriately around people in crowds or shared public environments.","inclusion":"Include work where the primary objective is human-aware mobile navigation, crowd interaction, socially compliant trajectory generation, pedestrian-aware local planning, gesture-aware path choice, occlusion-aware crowd reasoning, or social-navigation evaluation.","exclusion":"If the primary goal is accessibility-specific guiding or mobility assistance, assign to user-adaptive assistive and social support; if language reasoning dominates the navigation contribution, assign to language-grounded embodied interaction.","nearest_boundary":"User-adaptive assistive and social support","representative_topics":["flow-field-informed discrete path planning for socially compliant robot navigation in crowds","socially-aware dynamic local planning via trajectory-ranked deep inverse reinforcement learning","gesture-aware robot navigation via imitation learning and model-predictive control"]},{"direction_index":2,"label":"D2","name":"Human-aware team planning","description":"Planning, scheduling, role allocation, workspace/task design, and learned coordination methods for human-robot or human-AI teams that optimize joint task performance under human variability, constraints, and interdependence.","inclusion":"Include work whose central objective is collaborative task planning, role allocation, communication-need reasoning, zero-shot or adaptive human-AI coordination, learned costs or synergies for teamwork, risk-aware cooperative planning, or system-level workspace and task sequencing for collaboration.","exclusion":"If the contribution primarily models trust, stress, beliefs, or workload without driving concrete coordination, assign to human-state, trust, and ethics modeling; if it primarily controls contact-rich physical execution, assign to physical collaboration control and safety.","nearest_boundary":"Human-state, trust, and ethics modeling","representative_topics":["natural language driven collaborative planning with concept learning and htn solvers","multi-modal danger estimation and risk-aware planning for cooperative human-robot escape in hazardous environments","human-regularized search and learning for ai-human coordination in partially observable cooperative settings"]},{"direction_index":3,"label":"D3","name":"Human-state, trust, and ethics modeling","description":"Research whose main objective is to model, measure, explain, or intervene on human cognitive, affective, evaluative, ethical, or social states that shape HRI and human-AI teaming.","inclusion":"Include work focused on trust, stress, workload, mental models, beliefs, intentions, situation awareness, theory of mind, latent traits, ethical behavior, meaningful human control, social influence, or human-factor measurement in HRI.","exclusion":"If human-state estimates are mainly embedded as components of a specific navigation, physical-control, shared-control, or team-planning system, assign to that primary robot capability.","nearest_boundary":"Multimodal social signal understanding","representative_topics":["latent driver trait learning for personalized intervention policies in cooperative human-ai driving","trust factors in language grounding for human-robot teaming","mutual theory of mind framework for iterative human-ai cognitive shaping in social interaction"]},{"direction_index":4,"label":"D4","name":"User-adaptive assistive and social support","description":"Robotic systems and learning methods whose primary objective is personalized, longitudinal, therapeutic, educational, household, accessibility, or population-specific assistance for human users.","inclusion":"Include work on user-adaptive social robots, long-term contextual memory, proactive household assistance, therapeutic or educational child-robot interaction, accessibility-oriented perception and navigation, wearable or guide-robot assistance, and robotic support for specific populations.","exclusion":"If the main contribution is a general navigation, language, shared-control, perception, or manipulation algorithm without personalization, accessibility, or specific support goals as the central objective, assign to that technical program.","nearest_boundary":"Social navigation in human spaces","representative_topics":["contextually aware item availability prediction for assistive household robots via personalized learning","emotion-adaptive social robot therapy for autistic children","vision-based semantic environment perception for mobile assistive robots without a-priori knowledge"]},{"direction_index":5,"label":"D5","name":"Multimodal social signal understanding","description":"Analysis, annotation, recognition, modeling, or generation of nonverbal and multimodal social-communication signals used for human-robot interaction.","inclusion":"Include work where the main contribution is a representation, detector, benchmark, annotation scheme, or generative model for gaze, gesture, facial expression, posture, affect, attention, engagement, confusion, EEG/BCI interaction signals, body language, or dyadic interaction dynamics.","exclusion":"If signals are used mainly to infer trust, stress, workload, or mental models, assign to human-state, trust, and ethics modeling; if they are used mainly for collision monitoring or industrial safety, assign to physical collaboration control and safety.","nearest_boundary":"Human-state, trust, and ethics modeling","representative_topics":["cross-embodiment transfer of user confusion detection in human-robot and human-avatar studies","body gesture recognition for natural language control of social robots","multimodal human belief state prediction using object-context relations in nonverbal social interaction"]},{"direction_index":6,"label":"D6","name":"Physical collaboration control and safety","description":"Robot control, perception, learning, monitoring, and assurance methods for physical human-robot collaboration, including contact-rich interaction, cooperative manipulation, assembly, handovers, grasping, motion prediction, and runtime safety.","inclusion":"Include work whose primary objective is learning or controlling physical robot behavior with or from humans, transferring physical skills, generating human-compatible grasps or trajectories, forecasting human motion for collision avoidance, classifying contact, monitoring proximity, or testing robustness and failures in collaborative workspaces.","exclusion":"If the primary objective is operator authority allocation or teleoperation rather than physical execution or safety, assign to shared autonomy and teleoperation; if the objective is high-level task coordination without physical-control focus, assign to human-aware team planning.","nearest_boundary":"Shared autonomy and teleoperation","representative_topics":["force-based perception-intention-action cycle modeling for human-robot collaborative task execution","vision-based speed and separation monitoring for barrierless human-robot collaborative safety","sparse graph convolutional pose forecasting for industrial human-cobot collision awareness"]},{"direction_index":7,"label":"D7","name":"Shared autonomy and teleoperation","description":"Control architectures, interfaces, and learning systems that dynamically divide authority, information, preference elicitation, or teleoperation responsibility between humans and robots.","inclusion":"Include work focused on shared control, mixed-initiative autonomy, variable autonomy, authority switching, uncertainty-aware human querying, online preference elicitation, VR/AR teleoperation or visualization interfaces, operator skill assessment, or multimodal intention inference inside an operator-mediated control loop.","exclusion":"If human input is mainly natural-language instruction, assign to language-grounded embodied interaction; if it is demonstration for autonomous physical skill learning, assign to physical collaboration control and safety.","nearest_boundary":"Physical collaboration control and safety","representative_topics":["gaze and motion feature fusion for intention estimation in bimanual shared control manipulation","demonstration-based learning of augmented reality visualization policies for human-robot collaboration","hierarchical mixed-initiative control switching for human-ai authority transfer in mobile robot navigation"]}],"split":"dev"}
{"episode_id":"F199_T2024-07","field_index":199,"field":"human-robot interaction","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Language-grounded embodied interaction","description":"Robots and embodied agents whose main objective is to use natural language, dialogue, or language/vision-language grounding to understand instructions, resolve ambiguity, explain behavior, or connect conversational interaction to situated action.","inclusion":"Include work where the central contribution is language grounding, action-language translation, situated dialogue policy, clarification or repair, referent disambiguation, language-model planning for embodied action, or multimodal face-to-face interaction organized around language.","exclusion":"If language is only a cue inside a primarily navigation, accessibility, shared-control, physical-skill, or team-planning contribution, assign to that nearer objective.","nearest_boundary":"Human-aware team planning and coordination","representative_topics":["multi-task robot action-language translation via paired gated autoencoders with direction-signal conditioning","multi-view referred object disambiguation and query generation for mobile embodied agents","dialogue policy design for confusion mitigation in situated human-robot interaction"]},{"direction_index":1,"label":"D1","name":"Social navigation in human spaces","description":"Navigation, locomotion, trajectory prediction, and local planning methods that enable robots to move safely, legibly, and socially appropriately around people in crowds or shared public environments.","inclusion":"Include work where the primary objective is human-aware mobile navigation, crowd interaction, socially compliant trajectory generation, pedestrian-aware local planning, gesture-aware path choice, occlusion-aware crowd reasoning, or social-navigation evaluation.","exclusion":"If the primary goal is accessibility-specific guiding or mobility assistance, assign to user-adaptive assistive and social support; if language reasoning dominates the navigation contribution, assign to language-grounded embodied interaction.","nearest_boundary":"User-adaptive assistive and social support","representative_topics":["flow-field-informed discrete path planning for socially compliant robot navigation in crowds","socially-aware dynamic local planning via trajectory-ranked deep inverse reinforcement learning","gesture-aware robot navigation via imitation learning and model-predictive control"]},{"direction_index":2,"label":"D2","name":"Human-aware team planning","description":"Planning, scheduling, role allocation, workspace/task design, and learned coordination methods for human-robot or human-AI teams that optimize joint task performance under human variability, constraints, and interdependence.","inclusion":"Include work whose central objective is collaborative task planning, role allocation, communication-need reasoning, zero-shot or adaptive human-AI coordination, learned costs or synergies for teamwork, risk-aware cooperative planning, or system-level workspace and task sequencing for collaboration.","exclusion":"If the contribution primarily models trust, stress, beliefs, or workload without driving concrete coordination, assign to human-state, trust, and ethics modeling; if it primarily controls contact-rich physical execution, assign to physical collaboration control and safety.","nearest_boundary":"Human-state, trust, and ethics modeling","representative_topics":["natural language driven collaborative planning with concept learning and htn solvers","multi-modal danger estimation and risk-aware planning for cooperative human-robot escape in hazardous environments","human-regularized search and learning for ai-human coordination in partially observable cooperative settings"]},{"direction_index":3,"label":"D3","name":"Human-state, trust, and ethics modeling","description":"Research whose main objective is to model, measure, explain, or intervene on human cognitive, affective, evaluative, ethical, or social states that shape HRI and human-AI teaming.","inclusion":"Include work focused on trust, stress, workload, mental models, beliefs, intentions, situation awareness, theory of mind, latent traits, ethical behavior, meaningful human control, social influence, or human-factor measurement in HRI.","exclusion":"If human-state estimates are mainly embedded as components of a specific navigation, physical-control, shared-control, or team-planning system, assign to that primary robot capability.","nearest_boundary":"Multimodal social signal understanding","representative_topics":["latent driver trait learning for personalized intervention policies in cooperative human-ai driving","trust factors in language grounding for human-robot teaming","mutual theory of mind framework for iterative human-ai cognitive shaping in social interaction"]},{"direction_index":4,"label":"D4","name":"User-adaptive assistive and social support","description":"Robotic systems and learning methods whose primary objective is personalized, longitudinal, therapeutic, educational, household, accessibility, or population-specific assistance for human users.","inclusion":"Include work on user-adaptive social robots, long-term contextual memory, proactive household assistance, therapeutic or educational child-robot interaction, accessibility-oriented perception and navigation, wearable or guide-robot assistance, and robotic support for specific populations.","exclusion":"If the main contribution is a general navigation, language, shared-control, perception, or manipulation algorithm without personalization, accessibility, or specific support goals as the central objective, assign to that technical program.","nearest_boundary":"Social navigation in human spaces","representative_topics":["contextually aware item availability prediction for assistive household robots via personalized learning","emotion-adaptive social robot therapy for autistic children","vision-based semantic environment perception for mobile assistive robots without a-priori knowledge"]},{"direction_index":5,"label":"D5","name":"Multimodal social signal understanding","description":"Analysis, annotation, recognition, modeling, or generation of nonverbal and multimodal social-communication signals used for human-robot interaction.","inclusion":"Include work where the main contribution is a representation, detector, benchmark, annotation scheme, or generative model for gaze, gesture, facial expression, posture, affect, attention, engagement, confusion, EEG/BCI interaction signals, body language, or dyadic interaction dynamics.","exclusion":"If signals are used mainly to infer trust, stress, workload, or mental models, assign to human-state, trust, and ethics modeling; if they are used mainly for collision monitoring or industrial safety, assign to physical collaboration control and safety.","nearest_boundary":"Human-state, trust, and ethics modeling","representative_topics":["cross-embodiment transfer of user confusion detection in human-robot and human-avatar studies","body gesture recognition for natural language control of social robots","multimodal human belief state prediction using object-context relations in nonverbal social interaction"]},{"direction_index":6,"label":"D6","name":"Physical collaboration control and safety","description":"Robot control, perception, learning, monitoring, and assurance methods for physical human-robot collaboration, including contact-rich interaction, cooperative manipulation, assembly, handovers, grasping, motion prediction, and runtime safety.","inclusion":"Include work whose primary objective is learning or controlling physical robot behavior with or from humans, transferring physical skills, generating human-compatible grasps or trajectories, forecasting human motion for collision avoidance, classifying contact, monitoring proximity, or testing robustness and failures in collaborative workspaces.","exclusion":"If the primary objective is operator authority allocation or teleoperation rather than physical execution or safety, assign to shared autonomy and teleoperation; if the objective is high-level task coordination without physical-control focus, assign to human-aware team planning.","nearest_boundary":"Shared autonomy and teleoperation","representative_topics":["force-based perception-intention-action cycle modeling for human-robot collaborative task execution","vision-based speed and separation monitoring for barrierless human-robot collaborative safety","sparse graph convolutional pose forecasting for industrial human-cobot collision awareness"]},{"direction_index":7,"label":"D7","name":"Shared autonomy and teleoperation","description":"Control architectures, interfaces, and learning systems that dynamically divide authority, information, preference elicitation, or teleoperation responsibility between humans and robots.","inclusion":"Include work focused on shared control, mixed-initiative autonomy, variable autonomy, authority switching, uncertainty-aware human querying, online preference elicitation, VR/AR teleoperation or visualization interfaces, operator skill assessment, or multimodal intention inference inside an operator-mediated control loop.","exclusion":"If human input is mainly natural-language instruction, assign to language-grounded embodied interaction; if it is demonstration for autonomous physical skill learning, assign to physical collaboration control and safety.","nearest_boundary":"Physical collaboration control and safety","representative_topics":["gaze and motion feature fusion for intention estimation in bimanual shared control manipulation","demonstration-based learning of augmented reality visualization policies for human-robot collaboration","hierarchical mixed-initiative control switching for human-ai authority transfer in mobile robot navigation"]}],"split":"dev"}
{"episode_id":"F199_T2025-01","field_index":199,"field":"human-robot interaction","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Language-grounded embodied interaction","description":"Robots and embodied agents whose main objective is to use natural language, dialogue, or language/vision-language grounding to understand instructions, resolve ambiguity, explain behavior, or connect conversational interaction to situated action.","inclusion":"Include work where the central contribution is language grounding, action-language translation, situated dialogue policy, clarification or repair, referent disambiguation, language-model planning for embodied action, or multimodal face-to-face interaction organized around language.","exclusion":"If language is only a cue inside a primarily navigation, accessibility, shared-control, physical-skill, or team-planning contribution, assign to that nearer objective.","nearest_boundary":"Human-aware team planning and coordination","representative_topics":["multi-task robot action-language translation via paired gated autoencoders with direction-signal conditioning","multi-view referred object disambiguation and query generation for mobile embodied agents","dialogue policy design for confusion mitigation in situated human-robot interaction"]},{"direction_index":1,"label":"D1","name":"Social navigation in human spaces","description":"Navigation, locomotion, trajectory prediction, and local planning methods that enable robots to move safely, legibly, and socially appropriately around people in crowds or shared public environments.","inclusion":"Include work where the primary objective is human-aware mobile navigation, crowd interaction, socially compliant trajectory generation, pedestrian-aware local planning, gesture-aware path choice, occlusion-aware crowd reasoning, or social-navigation evaluation.","exclusion":"If the primary goal is accessibility-specific guiding or mobility assistance, assign to user-adaptive assistive and social support; if language reasoning dominates the navigation contribution, assign to language-grounded embodied interaction.","nearest_boundary":"User-adaptive assistive and social support","representative_topics":["flow-field-informed discrete path planning for socially compliant robot navigation in crowds","socially-aware dynamic local planning via trajectory-ranked deep inverse reinforcement learning","gesture-aware robot navigation via imitation learning and model-predictive control"]},{"direction_index":2,"label":"D2","name":"Human-aware team planning","description":"Planning, scheduling, role allocation, workspace/task design, and learned coordination methods for human-robot or human-AI teams that optimize joint task performance under human variability, constraints, and interdependence.","inclusion":"Include work whose central objective is collaborative task planning, role allocation, communication-need reasoning, zero-shot or adaptive human-AI coordination, learned costs or synergies for teamwork, risk-aware cooperative planning, or system-level workspace and task sequencing for collaboration.","exclusion":"If the contribution primarily models trust, stress, beliefs, or workload without driving concrete coordination, assign to human-state, trust, and ethics modeling; if it primarily controls contact-rich physical execution, assign to physical collaboration control and safety.","nearest_boundary":"Human-state, trust, and ethics modeling","representative_topics":["natural language driven collaborative planning with concept learning and htn solvers","multi-modal danger estimation and risk-aware planning for cooperative human-robot escape in hazardous environments","human-regularized search and learning for ai-human coordination in partially observable cooperative settings"]},{"direction_index":3,"label":"D3","name":"Human-state, trust, and ethics modeling","description":"Research whose main objective is to model, measure, explain, or intervene on human cognitive, affective, evaluative, ethical, or social states that shape HRI and human-AI teaming.","inclusion":"Include work focused on trust, stress, workload, mental models, beliefs, intentions, situation awareness, theory of mind, latent traits, ethical behavior, meaningful human control, social influence, or human-factor measurement in HRI.","exclusion":"If human-state estimates are mainly embedded as components of a specific navigation, physical-control, shared-control, or team-planning system, assign to that primary robot capability.","nearest_boundary":"Multimodal social signal understanding","representative_topics":["latent driver trait learning for personalized intervention policies in cooperative human-ai driving","trust factors in language grounding for human-robot teaming","mutual theory of mind framework for iterative human-ai cognitive shaping in social interaction"]},{"direction_index":4,"label":"D4","name":"User-adaptive assistive and social support","description":"Robotic systems and learning methods whose primary objective is personalized, longitudinal, therapeutic, educational, household, accessibility, or population-specific assistance for human users.","inclusion":"Include work on user-adaptive social robots, long-term contextual memory, proactive household assistance, therapeutic or educational child-robot interaction, accessibility-oriented perception and navigation, wearable or guide-robot assistance, and robotic support for specific populations.","exclusion":"If the main contribution is a general navigation, language, shared-control, perception, or manipulation algorithm without personalization, accessibility, or specific support goals as the central objective, assign to that technical program.","nearest_boundary":"Social navigation in human spaces","representative_topics":["contextually aware item availability prediction for assistive household robots via personalized learning","emotion-adaptive social robot therapy for autistic children","vision-based semantic environment perception for mobile assistive robots without a-priori knowledge"]},{"direction_index":5,"label":"D5","name":"Multimodal social signal understanding","description":"Analysis, annotation, recognition, modeling, or generation of nonverbal and multimodal social-communication signals used for human-robot interaction.","inclusion":"Include work where the main contribution is a representation, detector, benchmark, annotation scheme, or generative model for gaze, gesture, facial expression, posture, affect, attention, engagement, confusion, EEG/BCI interaction signals, body language, or dyadic interaction dynamics.","exclusion":"If signals are used mainly to infer trust, stress, workload, or mental models, assign to human-state, trust, and ethics modeling; if they are used mainly for collision monitoring or industrial safety, assign to physical collaboration control and safety.","nearest_boundary":"Human-state, trust, and ethics modeling","representative_topics":["cross-embodiment transfer of user confusion detection in human-robot and human-avatar studies","body gesture recognition for natural language control of social robots","multimodal human belief state prediction using object-context relations in nonverbal social interaction"]},{"direction_index":6,"label":"D6","name":"Physical collaboration control and safety","description":"Robot control, perception, learning, monitoring, and assurance methods for physical human-robot collaboration, including contact-rich interaction, cooperative manipulation, assembly, handovers, grasping, motion prediction, and runtime safety.","inclusion":"Include work whose primary objective is learning or controlling physical robot behavior with or from humans, transferring physical skills, generating human-compatible grasps or trajectories, forecasting human motion for collision avoidance, classifying contact, monitoring proximity, or testing robustness and failures in collaborative workspaces.","exclusion":"If the primary objective is operator authority allocation or teleoperation rather than physical execution or safety, assign to shared autonomy and teleoperation; if the objective is high-level task coordination without physical-control focus, assign to human-aware team planning.","nearest_boundary":"Shared autonomy and teleoperation","representative_topics":["force-based perception-intention-action cycle modeling for human-robot collaborative task execution","vision-based speed and separation monitoring for barrierless human-robot collaborative safety","sparse graph convolutional pose forecasting for industrial human-cobot collision awareness"]},{"direction_index":7,"label":"D7","name":"Shared autonomy and teleoperation","description":"Control architectures, interfaces, and learning systems that dynamically divide authority, information, preference elicitation, or teleoperation responsibility between humans and robots.","inclusion":"Include work focused on shared control, mixed-initiative autonomy, variable autonomy, authority switching, uncertainty-aware human querying, online preference elicitation, VR/AR teleoperation or visualization interfaces, operator skill assessment, or multimodal intention inference inside an operator-mediated control loop.","exclusion":"If human input is mainly natural-language instruction, assign to language-grounded embodied interaction; if it is demonstration for autonomous physical skill learning, assign to physical collaboration control and safety.","nearest_boundary":"Physical collaboration control and safety","representative_topics":["gaze and motion feature fusion for intention estimation in bimanual shared control manipulation","demonstration-based learning of augmented reality visualization policies for human-robot collaboration","hierarchical mixed-initiative control switching for human-ai authority transfer in mobile robot navigation"]}],"split":"dev"}
{"episode_id":"F199_T2025-07","field_index":199,"field":"human-robot interaction","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Language-grounded embodied interaction","description":"Robots and embodied agents whose main objective is to use natural language, dialogue, or language/vision-language grounding to understand instructions, resolve ambiguity, explain behavior, or connect conversational interaction to situated action.","inclusion":"Include work where the central contribution is language grounding, action-language translation, situated dialogue policy, clarification or repair, referent disambiguation, language-model planning for embodied action, or multimodal face-to-face interaction organized around language.","exclusion":"If language is only a cue inside a primarily navigation, accessibility, shared-control, physical-skill, or team-planning contribution, assign to that nearer objective.","nearest_boundary":"Human-aware team planning and coordination","representative_topics":["multi-task robot action-language translation via paired gated autoencoders with direction-signal conditioning","multi-view referred object disambiguation and query generation for mobile embodied agents","dialogue policy design for confusion mitigation in situated human-robot interaction"]},{"direction_index":1,"label":"D1","name":"Social navigation in human spaces","description":"Navigation, locomotion, trajectory prediction, and local planning methods that enable robots to move safely, legibly, and socially appropriately around people in crowds or shared public environments.","inclusion":"Include work where the primary objective is human-aware mobile navigation, crowd interaction, socially compliant trajectory generation, pedestrian-aware local planning, gesture-aware path choice, occlusion-aware crowd reasoning, or social-navigation evaluation.","exclusion":"If the primary goal is accessibility-specific guiding or mobility assistance, assign to user-adaptive assistive and social support; if language reasoning dominates the navigation contribution, assign to language-grounded embodied interaction.","nearest_boundary":"User-adaptive assistive and social support","representative_topics":["flow-field-informed discrete path planning for socially compliant robot navigation in crowds","socially-aware dynamic local planning via trajectory-ranked deep inverse reinforcement learning","gesture-aware robot navigation via imitation learning and model-predictive control"]},{"direction_index":2,"label":"D2","name":"Human-aware team planning","description":"Planning, scheduling, role allocation, workspace/task design, and learned coordination methods for human-robot or human-AI teams that optimize joint task performance under human variability, constraints, and interdependence.","inclusion":"Include work whose central objective is collaborative task planning, role allocation, communication-need reasoning, zero-shot or adaptive human-AI coordination, learned costs or synergies for teamwork, risk-aware cooperative planning, or system-level workspace and task sequencing for collaboration.","exclusion":"If the contribution primarily models trust, stress, beliefs, or workload without driving concrete coordination, assign to human-state, trust, and ethics modeling; if it primarily controls contact-rich physical execution, assign to physical collaboration control and safety.","nearest_boundary":"Human-state, trust, and ethics modeling","representative_topics":["natural language driven collaborative planning with concept learning and htn solvers","multi-modal danger estimation and risk-aware planning for cooperative human-robot escape in hazardous environments","human-regularized search and learning for ai-human coordination in partially observable cooperative settings"]},{"direction_index":3,"label":"D3","name":"Human-state, trust, and ethics modeling","description":"Research whose main objective is to model, measure, explain, or intervene on human cognitive, affective, evaluative, ethical, or social states that shape HRI and human-AI teaming.","inclusion":"Include work focused on trust, stress, workload, mental models, beliefs, intentions, situation awareness, theory of mind, latent traits, ethical behavior, meaningful human control, social influence, or human-factor measurement in HRI.","exclusion":"If human-state estimates are mainly embedded as components of a specific navigation, physical-control, shared-control, or team-planning system, assign to that primary robot capability.","nearest_boundary":"Multimodal social signal understanding","representative_topics":["latent driver trait learning for personalized intervention policies in cooperative human-ai driving","trust factors in language grounding for human-robot teaming","mutual theory of mind framework for iterative human-ai cognitive shaping in social interaction"]},{"direction_index":4,"label":"D4","name":"User-adaptive assistive and social support","description":"Robotic systems and learning methods whose primary objective is personalized, longitudinal, therapeutic, educational, household, accessibility, or population-specific assistance for human users.","inclusion":"Include work on user-adaptive social robots, long-term contextual memory, proactive household assistance, therapeutic or educational child-robot interaction, accessibility-oriented perception and navigation, wearable or guide-robot assistance, and robotic support for specific populations.","exclusion":"If the main contribution is a general navigation, language, shared-control, perception, or manipulation algorithm without personalization, accessibility, or specific support goals as the central objective, assign to that technical program.","nearest_boundary":"Social navigation in human spaces","representative_topics":["contextually aware item availability prediction for assistive household robots via personalized learning","emotion-adaptive social robot therapy for autistic children","vision-based semantic environment perception for mobile assistive robots without a-priori knowledge"]},{"direction_index":5,"label":"D5","name":"Multimodal social signal understanding","description":"Analysis, annotation, recognition, modeling, or generation of nonverbal and multimodal social-communication signals used for human-robot interaction.","inclusion":"Include work where the main contribution is a representation, detector, benchmark, annotation scheme, or generative model for gaze, gesture, facial expression, posture, affect, attention, engagement, confusion, EEG/BCI interaction signals, body language, or dyadic interaction dynamics.","exclusion":"If signals are used mainly to infer trust, stress, workload, or mental models, assign to human-state, trust, and ethics modeling; if they are used mainly for collision monitoring or industrial safety, assign to physical collaboration control and safety.","nearest_boundary":"Human-state, trust, and ethics modeling","representative_topics":["cross-embodiment transfer of user confusion detection in human-robot and human-avatar studies","body gesture recognition for natural language control of social robots","multimodal human belief state prediction using object-context relations in nonverbal social interaction"]},{"direction_index":6,"label":"D6","name":"Physical collaboration control and safety","description":"Robot control, perception, learning, monitoring, and assurance methods for physical human-robot collaboration, including contact-rich interaction, cooperative manipulation, assembly, handovers, grasping, motion prediction, and runtime safety.","inclusion":"Include work whose primary objective is learning or controlling physical robot behavior with or from humans, transferring physical skills, generating human-compatible grasps or trajectories, forecasting human motion for collision avoidance, classifying contact, monitoring proximity, or testing robustness and failures in collaborative workspaces.","exclusion":"If the primary objective is operator authority allocation or teleoperation rather than physical execution or safety, assign to shared autonomy and teleoperation; if the objective is high-level task coordination without physical-control focus, assign to human-aware team planning.","nearest_boundary":"Shared autonomy and teleoperation","representative_topics":["force-based perception-intention-action cycle modeling for human-robot collaborative task execution","vision-based speed and separation monitoring for barrierless human-robot collaborative safety","sparse graph convolutional pose forecasting for industrial human-cobot collision awareness"]},{"direction_index":7,"label":"D7","name":"Shared autonomy and teleoperation","description":"Control architectures, interfaces, and learning systems that dynamically divide authority, information, preference elicitation, or teleoperation responsibility between humans and robots.","inclusion":"Include work focused on shared control, mixed-initiative autonomy, variable autonomy, authority switching, uncertainty-aware human querying, online preference elicitation, VR/AR teleoperation or visualization interfaces, operator skill assessment, or multimodal intention inference inside an operator-mediated control loop.","exclusion":"If human input is mainly natural-language instruction, assign to language-grounded embodied interaction; if it is demonstration for autonomous physical skill learning, assign to physical collaboration control and safety.","nearest_boundary":"Physical collaboration control and safety","representative_topics":["gaze and motion feature fusion for intention estimation in bimanual shared control manipulation","demonstration-based learning of augmented reality visualization policies for human-robot collaboration","hierarchical mixed-initiative control switching for human-ai authority transfer in mobile robot navigation"]}],"split":"dev"}
{"episode_id":"F199_T2026-01","field_index":199,"field":"human-robot interaction","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Language-grounded embodied interaction","description":"Robots and embodied agents whose main objective is to use natural language, dialogue, or language/vision-language grounding to understand instructions, resolve ambiguity, explain behavior, or connect conversational interaction to situated action.","inclusion":"Include work where the central contribution is language grounding, action-language translation, situated dialogue policy, clarification or repair, referent disambiguation, language-model planning for embodied action, or multimodal face-to-face interaction organized around language.","exclusion":"If language is only a cue inside a primarily navigation, accessibility, shared-control, physical-skill, or team-planning contribution, assign to that nearer objective.","nearest_boundary":"Human-aware team planning and coordination","representative_topics":["multi-task robot action-language translation via paired gated autoencoders with direction-signal conditioning","multi-view referred object disambiguation and query generation for mobile embodied agents","dialogue policy design for confusion mitigation in situated human-robot interaction"]},{"direction_index":1,"label":"D1","name":"Social navigation in human spaces","description":"Navigation, locomotion, trajectory prediction, and local planning methods that enable robots to move safely, legibly, and socially appropriately around people in crowds or shared public environments.","inclusion":"Include work where the primary objective is human-aware mobile navigation, crowd interaction, socially compliant trajectory generation, pedestrian-aware local planning, gesture-aware path choice, occlusion-aware crowd reasoning, or social-navigation evaluation.","exclusion":"If the primary goal is accessibility-specific guiding or mobility assistance, assign to user-adaptive assistive and social support; if language reasoning dominates the navigation contribution, assign to language-grounded embodied interaction.","nearest_boundary":"User-adaptive assistive and social support","representative_topics":["flow-field-informed discrete path planning for socially compliant robot navigation in crowds","socially-aware dynamic local planning via trajectory-ranked deep inverse reinforcement learning","gesture-aware robot navigation via imitation learning and model-predictive control"]},{"direction_index":2,"label":"D2","name":"Human-aware team planning","description":"Planning, scheduling, role allocation, workspace/task design, and learned coordination methods for human-robot or human-AI teams that optimize joint task performance under human variability, constraints, and interdependence.","inclusion":"Include work whose central objective is collaborative task planning, role allocation, communication-need reasoning, zero-shot or adaptive human-AI coordination, learned costs or synergies for teamwork, risk-aware cooperative planning, or system-level workspace and task sequencing for collaboration.","exclusion":"If the contribution primarily models trust, stress, beliefs, or workload without driving concrete coordination, assign to human-state, trust, and ethics modeling; if it primarily controls contact-rich physical execution, assign to physical collaboration control and safety.","nearest_boundary":"Human-state, trust, and ethics modeling","representative_topics":["natural language driven collaborative planning with concept learning and htn solvers","multi-modal danger estimation and risk-aware planning for cooperative human-robot escape in hazardous environments","human-regularized search and learning for ai-human coordination in partially observable cooperative settings"]},{"direction_index":3,"label":"D3","name":"Human-state, trust, and ethics modeling","description":"Research whose main objective is to model, measure, explain, or intervene on human cognitive, affective, evaluative, ethical, or social states that shape HRI and human-AI teaming.","inclusion":"Include work focused on trust, stress, workload, mental models, beliefs, intentions, situation awareness, theory of mind, latent traits, ethical behavior, meaningful human control, social influence, or human-factor measurement in HRI.","exclusion":"If human-state estimates are mainly embedded as components of a specific navigation, physical-control, shared-control, or team-planning system, assign to that primary robot capability.","nearest_boundary":"Multimodal social signal understanding","representative_topics":["latent driver trait learning for personalized intervention policies in cooperative human-ai driving","trust factors in language grounding for human-robot teaming","mutual theory of mind framework for iterative human-ai cognitive shaping in social interaction"]},{"direction_index":4,"label":"D4","name":"User-adaptive assistive and social support","description":"Robotic systems and learning methods whose primary objective is personalized, longitudinal, therapeutic, educational, household, accessibility, or population-specific assistance for human users.","inclusion":"Include work on user-adaptive social robots, long-term contextual memory, proactive household assistance, therapeutic or educational child-robot interaction, accessibility-oriented perception and navigation, wearable or guide-robot assistance, and robotic support for specific populations.","exclusion":"If the main contribution is a general navigation, language, shared-control, perception, or manipulation algorithm without personalization, accessibility, or specific support goals as the central objective, assign to that technical program.","nearest_boundary":"Social navigation in human spaces","representative_topics":["contextually aware item availability prediction for assistive household robots via personalized learning","emotion-adaptive social robot therapy for autistic children","vision-based semantic environment perception for mobile assistive robots without a-priori knowledge"]},{"direction_index":5,"label":"D5","name":"Multimodal social signal understanding","description":"Analysis, annotation, recognition, modeling, or generation of nonverbal and multimodal social-communication signals used for human-robot interaction.","inclusion":"Include work where the main contribution is a representation, detector, benchmark, annotation scheme, or generative model for gaze, gesture, facial expression, posture, affect, attention, engagement, confusion, EEG/BCI interaction signals, body language, or dyadic interaction dynamics.","exclusion":"If signals are used mainly to infer trust, stress, workload, or mental models, assign to human-state, trust, and ethics modeling; if they are used mainly for collision monitoring or industrial safety, assign to physical collaboration control and safety.","nearest_boundary":"Human-state, trust, and ethics modeling","representative_topics":["cross-embodiment transfer of user confusion detection in human-robot and human-avatar studies","body gesture recognition for natural language control of social robots","multimodal human belief state prediction using object-context relations in nonverbal social interaction"]},{"direction_index":6,"label":"D6","name":"Physical collaboration control and safety","description":"Robot control, perception, learning, monitoring, and assurance methods for physical human-robot collaboration, including contact-rich interaction, cooperative manipulation, assembly, handovers, grasping, motion prediction, and runtime safety.","inclusion":"Include work whose primary objective is learning or controlling physical robot behavior with or from humans, transferring physical skills, generating human-compatible grasps or trajectories, forecasting human motion for collision avoidance, classifying contact, monitoring proximity, or testing robustness and failures in collaborative workspaces.","exclusion":"If the primary objective is operator authority allocation or teleoperation rather than physical execution or safety, assign to shared autonomy and teleoperation; if the objective is high-level task coordination without physical-control focus, assign to human-aware team planning.","nearest_boundary":"Shared autonomy and teleoperation","representative_topics":["force-based perception-intention-action cycle modeling for human-robot collaborative task execution","vision-based speed and separation monitoring for barrierless human-robot collaborative safety","sparse graph convolutional pose forecasting for industrial human-cobot collision awareness"]},{"direction_index":7,"label":"D7","name":"Shared autonomy and teleoperation","description":"Control architectures, interfaces, and learning systems that dynamically divide authority, information, preference elicitation, or teleoperation responsibility between humans and robots.","inclusion":"Include work focused on shared control, mixed-initiative autonomy, variable autonomy, authority switching, uncertainty-aware human querying, online preference elicitation, VR/AR teleoperation or visualization interfaces, operator skill assessment, or multimodal intention inference inside an operator-mediated control loop.","exclusion":"If human input is mainly natural-language instruction, assign to language-grounded embodied interaction; if it is demonstration for autonomous physical skill learning, assign to physical collaboration control and safety.","nearest_boundary":"Physical collaboration control and safety","representative_topics":["gaze and motion feature fusion for intention estimation in bimanual shared control manipulation","demonstration-based learning of augmented reality visualization policies for human-robot collaboration","hierarchical mixed-initiative control switching for human-ai authority transfer in mobile robot navigation"]}],"split":"dev"}
{"episode_id":"F205_T2024-01","field_index":205,"field":"fake news detection","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Social, propagation, and source-context detection","description":"Detect or characterize fake news using social diffusion, user/source relations, propagation graphs, website/source context, narrative spread, or network-level mitigation as the primary evidence.","inclusion":"Include work whose main objective is veracity, source reliability, or misinformation-spread assessment through graph, propagation, user-credibility, source-network, website-structural, or narrative-tracking signals, even when text or images are auxiliary.","exclusion":"If the main objective is domain-shift robustness rather than social/source modeling, assign to D2; if the main objective is human-readable explanation of a detector, assign to D7.","nearest_boundary":"D2","representative_topics":["graph neural network fake news detection via social propagation context and transformer content features","graph-aware multi-view fusion for rumor representation learning on social media","content-agnostic fake news website detection using network and structural domain features"]},{"direction_index":1,"label":"D1","name":"Text and domain-focused content classification","description":"Detect fake news, rumors, false claims, or health/COVID misinformation primarily from linguistic content, document structure, emotional or stylistic signals, or text-centered classifiers.","inclusion":"Include general or domain-specific text-primary detection where the main objective is content-level classification, including transformer, RNN/CNN, ensemble, weak-supervision, emotion/sentiment, and health/COVID-focused text detection.","exclusion":"If the primary objective is language-resource coverage, assign to D6; if the primary objective is LLM-as-fact-checker behavior, assign to D4; if multimodal fusion is central, assign to D3.","nearest_boundary":"D6","representative_topics":["fake news classification under class imbalance using pre-trained encoders and negation-based data augmentation","multi-policy deep learning framework for joint statement and article fake news detection","label noise-resistant weak supervision for fake news detection via mean teacher propagation"]},{"direction_index":2,"label":"D2","name":"Robust transfer and bias-resistant generalization","description":"Improve or evaluate fake-news detection reliability under domain shift, event/topic shift, missing evidence, adversarial perturbations, dataset bias, or out-of-distribution conditions.","inclusion":"Include work whose primary objective is transfer learning, domain adaptation, debiasing, adversarial robustness, uncertainty, missing-evidence realism, or generalization evaluation across distributional stressors.","exclusion":"If the primary transfer problem is low-resource or cross-lingual language coverage, assign to D6; if the contribution is mainly a dataset or reproducibility resource, assign to D5.","nearest_boundary":"D6","representative_topics":["multi-domain fake news detection via memory-guided multi-view domain adaptation","domain- and instance-level transfer for targeted fake news detection in influential domains","realistic nlp fact-checking under missing counter-evidence conditions for misinformation"]},{"direction_index":3,"label":"D3","name":"Multimodal and cross-modal verification","description":"Detect or verify fake news by jointly reasoning over text, images, video, or other modalities through fusion, alignment, cross-modal consistency, or media-authenticity cues.","inclusion":"Include work whose primary objective is multimodal fake-news detection, multimodal fact verification, image-text rumor classification, out-of-context media detection, or cross-modal representation learning for veracity.","exclusion":"If the main contribution is the dataset or benchmark rather than the multimodal method, assign to D5; if explanation is the central objective, assign to D7.","nearest_boundary":"D5","representative_topics":["bootstrapping multi-view representations for cross-modal consistency-aware fake news detection","dual-tag image tagging for authenticable and traceable fake news detection","multimodal fusion and ensemble learning for social media rumor veracity classification"]},{"direction_index":4,"label":"D4","name":"LLM-era disinformation detection and fact-checking","description":"Study misinformation involving large language models, either by detecting LLM-generated disinformation or by using/evaluating LLMs for verification, debunking, truthfulness judgment, or controlled correction.","inclusion":"Include work where LLMs are the object of study, the generator of deceptive content to be detected, or the primary instrument for fact-checking, debunking, claim editing, uncertainty-aware judgment, or misinformation mitigation.","exclusion":"If the work only fine-tunes ordinary pretrained encoders for text classification without an LLM-centric objective, assign to D1; if explanation of detector decisions is the primary objective, assign to D7.","nearest_boundary":"D7","representative_topics":["generalizable misinformation veracity evaluation with uncertainty and possibility-aware language models","zero-shot large language model prompting for political fake statement detection","comparative black-box evaluation of llms for news fact-checking and truthfulness classification"]},{"direction_index":5,"label":"D5","name":"Datasets, benchmarks, and audits","description":"Create, annotate, benchmark, audit, or analyze fake-news and misinformation datasets, shared tasks, evaluation resources, reproducibility, or dataset quality as the primary contribution.","inclusion":"Include work centered on resource construction, annotation schemes, benchmark design, shared-task resources, time-aware datasets, synthetic/evaluation resources, reproducibility audits, relabeling, or dataset quality and bias analysis rather than a new detector.","exclusion":"If a dataset is secondary to a new modeling method, assign to the relevant modeling direction; if the main goal is adversarial or transfer robustness rather than resource design, assign to D2.","nearest_boundary":"D2","representative_topics":["multi-event annotated tweet dataset for benchmarking misinformation detection across languages","time-aware dataset construction for detecting evolving misinformation in language models","multimodal claim verification dataset with satire and visual-textual entailment labels"]},{"direction_index":6,"label":"D6","name":"Multilingual and low-resource detection","description":"Develop fake-news, disinformation, or fact-checking systems for non-English, multilingual, cross-lingual, or low-resource language settings.","inclusion":"Include work where language coverage, low-resource modeling, multilingual representation, cross-lingual retrieval, cross-lingual transfer, or language-specific shared-task participation is the primary scientific objective.","exclusion":"If the central problem is topic/event/domain shift without language coverage as the main challenge, assign to D2; if COVID/health misinformation is dominant and the language setting is secondary, assign to D1.","nearest_boundary":"D2","representative_topics":["low-resource urdu fake news detection using transformer and traditional feature representations","low-resource urdu fake news detection using count-based features and neural classifiers","multilingual text embedding models for disinformation publisher detection in digital advertising"]},{"direction_index":7,"label":"D7","name":"Explainable and evidence-grounded detection","description":"Make fake-news, rumor, or misinformation judgments understandable or evidence-grounded through rationales, counterfactuals, evidence retrieval, evidence distillation, interpretable latent variables, or human-readable reasoning.","inclusion":"Include work where interpretability, explanation, rationale generation, evidence grounding, rebuttal support, or counterfactual evidence is the main scientific objective, regardless of whether the underlying evidence is textual, social, multimodal, or health-related.","exclusion":"If explanations are only auxiliary visualizations for a detector whose main contribution is graph, text, or multimodal modeling, assign to the relevant modeling direction; if the main focus is evaluating LLM fact-checking capability, assign to D4.","nearest_boundary":"D4","representative_topics":["counterfactual explanation generation for fake news using contradicted entailment reasoning","multi-modal explainable ai for interpretable fake news detection on social media","interpretable fake news classification using bert embeddings and word-token activation mapping"]}],"split":"dev"}
{"episode_id":"F205_T2024-07","field_index":205,"field":"fake news detection","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Social, propagation, and source-context detection","description":"Detect or characterize fake news using social diffusion, user/source relations, propagation graphs, website/source context, narrative spread, or network-level mitigation as the primary evidence.","inclusion":"Include work whose main objective is veracity, source reliability, or misinformation-spread assessment through graph, propagation, user-credibility, source-network, website-structural, or narrative-tracking signals, even when text or images are auxiliary.","exclusion":"If the main objective is domain-shift robustness rather than social/source modeling, assign to D2; if the main objective is human-readable explanation of a detector, assign to D7.","nearest_boundary":"D2","representative_topics":["graph neural network fake news detection via social propagation context and transformer content features","graph-aware multi-view fusion for rumor representation learning on social media","content-agnostic fake news website detection using network and structural domain features"]},{"direction_index":1,"label":"D1","name":"Text and domain-focused content classification","description":"Detect fake news, rumors, false claims, or health/COVID misinformation primarily from linguistic content, document structure, emotional or stylistic signals, or text-centered classifiers.","inclusion":"Include general or domain-specific text-primary detection where the main objective is content-level classification, including transformer, RNN/CNN, ensemble, weak-supervision, emotion/sentiment, and health/COVID-focused text detection.","exclusion":"If the primary objective is language-resource coverage, assign to D6; if the primary objective is LLM-as-fact-checker behavior, assign to D4; if multimodal fusion is central, assign to D3.","nearest_boundary":"D6","representative_topics":["fake news classification under class imbalance using pre-trained encoders and negation-based data augmentation","multi-policy deep learning framework for joint statement and article fake news detection","label noise-resistant weak supervision for fake news detection via mean teacher propagation"]},{"direction_index":2,"label":"D2","name":"Robust transfer and bias-resistant generalization","description":"Improve or evaluate fake-news detection reliability under domain shift, event/topic shift, missing evidence, adversarial perturbations, dataset bias, or out-of-distribution conditions.","inclusion":"Include work whose primary objective is transfer learning, domain adaptation, debiasing, adversarial robustness, uncertainty, missing-evidence realism, or generalization evaluation across distributional stressors.","exclusion":"If the primary transfer problem is low-resource or cross-lingual language coverage, assign to D6; if the contribution is mainly a dataset or reproducibility resource, assign to D5.","nearest_boundary":"D6","representative_topics":["multi-domain fake news detection via memory-guided multi-view domain adaptation","domain- and instance-level transfer for targeted fake news detection in influential domains","realistic nlp fact-checking under missing counter-evidence conditions for misinformation"]},{"direction_index":3,"label":"D3","name":"Multimodal and cross-modal verification","description":"Detect or verify fake news by jointly reasoning over text, images, video, or other modalities through fusion, alignment, cross-modal consistency, or media-authenticity cues.","inclusion":"Include work whose primary objective is multimodal fake-news detection, multimodal fact verification, image-text rumor classification, out-of-context media detection, or cross-modal representation learning for veracity.","exclusion":"If the main contribution is the dataset or benchmark rather than the multimodal method, assign to D5; if explanation is the central objective, assign to D7.","nearest_boundary":"D5","representative_topics":["bootstrapping multi-view representations for cross-modal consistency-aware fake news detection","dual-tag image tagging for authenticable and traceable fake news detection","multimodal fusion and ensemble learning for social media rumor veracity classification"]},{"direction_index":4,"label":"D4","name":"LLM-era disinformation detection and fact-checking","description":"Study misinformation involving large language models, either by detecting LLM-generated disinformation or by using/evaluating LLMs for verification, debunking, truthfulness judgment, or controlled correction.","inclusion":"Include work where LLMs are the object of study, the generator of deceptive content to be detected, or the primary instrument for fact-checking, debunking, claim editing, uncertainty-aware judgment, or misinformation mitigation.","exclusion":"If the work only fine-tunes ordinary pretrained encoders for text classification without an LLM-centric objective, assign to D1; if explanation of detector decisions is the primary objective, assign to D7.","nearest_boundary":"D7","representative_topics":["generalizable misinformation veracity evaluation with uncertainty and possibility-aware language models","zero-shot large language model prompting for political fake statement detection","comparative black-box evaluation of llms for news fact-checking and truthfulness classification"]},{"direction_index":5,"label":"D5","name":"Datasets, benchmarks, and audits","description":"Create, annotate, benchmark, audit, or analyze fake-news and misinformation datasets, shared tasks, evaluation resources, reproducibility, or dataset quality as the primary contribution.","inclusion":"Include work centered on resource construction, annotation schemes, benchmark design, shared-task resources, time-aware datasets, synthetic/evaluation resources, reproducibility audits, relabeling, or dataset quality and bias analysis rather than a new detector.","exclusion":"If a dataset is secondary to a new modeling method, assign to the relevant modeling direction; if the main goal is adversarial or transfer robustness rather than resource design, assign to D2.","nearest_boundary":"D2","representative_topics":["multi-event annotated tweet dataset for benchmarking misinformation detection across languages","time-aware dataset construction for detecting evolving misinformation in language models","multimodal claim verification dataset with satire and visual-textual entailment labels"]},{"direction_index":6,"label":"D6","name":"Multilingual and low-resource detection","description":"Develop fake-news, disinformation, or fact-checking systems for non-English, multilingual, cross-lingual, or low-resource language settings.","inclusion":"Include work where language coverage, low-resource modeling, multilingual representation, cross-lingual retrieval, cross-lingual transfer, or language-specific shared-task participation is the primary scientific objective.","exclusion":"If the central problem is topic/event/domain shift without language coverage as the main challenge, assign to D2; if COVID/health misinformation is dominant and the language setting is secondary, assign to D1.","nearest_boundary":"D2","representative_topics":["low-resource urdu fake news detection using transformer and traditional feature representations","low-resource urdu fake news detection using count-based features and neural classifiers","multilingual text embedding models for disinformation publisher detection in digital advertising"]},{"direction_index":7,"label":"D7","name":"Explainable and evidence-grounded detection","description":"Make fake-news, rumor, or misinformation judgments understandable or evidence-grounded through rationales, counterfactuals, evidence retrieval, evidence distillation, interpretable latent variables, or human-readable reasoning.","inclusion":"Include work where interpretability, explanation, rationale generation, evidence grounding, rebuttal support, or counterfactual evidence is the main scientific objective, regardless of whether the underlying evidence is textual, social, multimodal, or health-related.","exclusion":"If explanations are only auxiliary visualizations for a detector whose main contribution is graph, text, or multimodal modeling, assign to the relevant modeling direction; if the main focus is evaluating LLM fact-checking capability, assign to D4.","nearest_boundary":"D4","representative_topics":["counterfactual explanation generation for fake news using contradicted entailment reasoning","multi-modal explainable ai for interpretable fake news detection on social media","interpretable fake news classification using bert embeddings and word-token activation mapping"]}],"split":"dev"}
{"episode_id":"F205_T2025-01","field_index":205,"field":"fake news detection","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Social, propagation, and source-context detection","description":"Detect or characterize fake news using social diffusion, user/source relations, propagation graphs, website/source context, narrative spread, or network-level mitigation as the primary evidence.","inclusion":"Include work whose main objective is veracity, source reliability, or misinformation-spread assessment through graph, propagation, user-credibility, source-network, website-structural, or narrative-tracking signals, even when text or images are auxiliary.","exclusion":"If the main objective is domain-shift robustness rather than social/source modeling, assign to D2; if the main objective is human-readable explanation of a detector, assign to D7.","nearest_boundary":"D2","representative_topics":["graph neural network fake news detection via social propagation context and transformer content features","graph-aware multi-view fusion for rumor representation learning on social media","content-agnostic fake news website detection using network and structural domain features"]},{"direction_index":1,"label":"D1","name":"Text and domain-focused content classification","description":"Detect fake news, rumors, false claims, or health/COVID misinformation primarily from linguistic content, document structure, emotional or stylistic signals, or text-centered classifiers.","inclusion":"Include general or domain-specific text-primary detection where the main objective is content-level classification, including transformer, RNN/CNN, ensemble, weak-supervision, emotion/sentiment, and health/COVID-focused text detection.","exclusion":"If the primary objective is language-resource coverage, assign to D6; if the primary objective is LLM-as-fact-checker behavior, assign to D4; if multimodal fusion is central, assign to D3.","nearest_boundary":"D6","representative_topics":["fake news classification under class imbalance using pre-trained encoders and negation-based data augmentation","multi-policy deep learning framework for joint statement and article fake news detection","label noise-resistant weak supervision for fake news detection via mean teacher propagation"]},{"direction_index":2,"label":"D2","name":"Robust transfer and bias-resistant generalization","description":"Improve or evaluate fake-news detection reliability under domain shift, event/topic shift, missing evidence, adversarial perturbations, dataset bias, or out-of-distribution conditions.","inclusion":"Include work whose primary objective is transfer learning, domain adaptation, debiasing, adversarial robustness, uncertainty, missing-evidence realism, or generalization evaluation across distributional stressors.","exclusion":"If the primary transfer problem is low-resource or cross-lingual language coverage, assign to D6; if the contribution is mainly a dataset or reproducibility resource, assign to D5.","nearest_boundary":"D6","representative_topics":["multi-domain fake news detection via memory-guided multi-view domain adaptation","domain- and instance-level transfer for targeted fake news detection in influential domains","realistic nlp fact-checking under missing counter-evidence conditions for misinformation"]},{"direction_index":3,"label":"D3","name":"Multimodal and cross-modal verification","description":"Detect or verify fake news by jointly reasoning over text, images, video, or other modalities through fusion, alignment, cross-modal consistency, or media-authenticity cues.","inclusion":"Include work whose primary objective is multimodal fake-news detection, multimodal fact verification, image-text rumor classification, out-of-context media detection, or cross-modal representation learning for veracity.","exclusion":"If the main contribution is the dataset or benchmark rather than the multimodal method, assign to D5; if explanation is the central objective, assign to D7.","nearest_boundary":"D5","representative_topics":["bootstrapping multi-view representations for cross-modal consistency-aware fake news detection","dual-tag image tagging for authenticable and traceable fake news detection","multimodal fusion and ensemble learning for social media rumor veracity classification"]},{"direction_index":4,"label":"D4","name":"LLM-era disinformation detection and fact-checking","description":"Study misinformation involving large language models, either by detecting LLM-generated disinformation or by using/evaluating LLMs for verification, debunking, truthfulness judgment, or controlled correction.","inclusion":"Include work where LLMs are the object of study, the generator of deceptive content to be detected, or the primary instrument for fact-checking, debunking, claim editing, uncertainty-aware judgment, or misinformation mitigation.","exclusion":"If the work only fine-tunes ordinary pretrained encoders for text classification without an LLM-centric objective, assign to D1; if explanation of detector decisions is the primary objective, assign to D7.","nearest_boundary":"D7","representative_topics":["generalizable misinformation veracity evaluation with uncertainty and possibility-aware language models","zero-shot large language model prompting for political fake statement detection","comparative black-box evaluation of llms for news fact-checking and truthfulness classification"]},{"direction_index":5,"label":"D5","name":"Datasets, benchmarks, and audits","description":"Create, annotate, benchmark, audit, or analyze fake-news and misinformation datasets, shared tasks, evaluation resources, reproducibility, or dataset quality as the primary contribution.","inclusion":"Include work centered on resource construction, annotation schemes, benchmark design, shared-task resources, time-aware datasets, synthetic/evaluation resources, reproducibility audits, relabeling, or dataset quality and bias analysis rather than a new detector.","exclusion":"If a dataset is secondary to a new modeling method, assign to the relevant modeling direction; if the main goal is adversarial or transfer robustness rather than resource design, assign to D2.","nearest_boundary":"D2","representative_topics":["multi-event annotated tweet dataset for benchmarking misinformation detection across languages","time-aware dataset construction for detecting evolving misinformation in language models","multimodal claim verification dataset with satire and visual-textual entailment labels"]},{"direction_index":6,"label":"D6","name":"Multilingual and low-resource detection","description":"Develop fake-news, disinformation, or fact-checking systems for non-English, multilingual, cross-lingual, or low-resource language settings.","inclusion":"Include work where language coverage, low-resource modeling, multilingual representation, cross-lingual retrieval, cross-lingual transfer, or language-specific shared-task participation is the primary scientific objective.","exclusion":"If the central problem is topic/event/domain shift without language coverage as the main challenge, assign to D2; if COVID/health misinformation is dominant and the language setting is secondary, assign to D1.","nearest_boundary":"D2","representative_topics":["low-resource urdu fake news detection using transformer and traditional feature representations","low-resource urdu fake news detection using count-based features and neural classifiers","multilingual text embedding models for disinformation publisher detection in digital advertising"]},{"direction_index":7,"label":"D7","name":"Explainable and evidence-grounded detection","description":"Make fake-news, rumor, or misinformation judgments understandable or evidence-grounded through rationales, counterfactuals, evidence retrieval, evidence distillation, interpretable latent variables, or human-readable reasoning.","inclusion":"Include work where interpretability, explanation, rationale generation, evidence grounding, rebuttal support, or counterfactual evidence is the main scientific objective, regardless of whether the underlying evidence is textual, social, multimodal, or health-related.","exclusion":"If explanations are only auxiliary visualizations for a detector whose main contribution is graph, text, or multimodal modeling, assign to the relevant modeling direction; if the main focus is evaluating LLM fact-checking capability, assign to D4.","nearest_boundary":"D4","representative_topics":["counterfactual explanation generation for fake news using contradicted entailment reasoning","multi-modal explainable ai for interpretable fake news detection on social media","interpretable fake news classification using bert embeddings and word-token activation mapping"]}],"split":"dev"}
{"episode_id":"F205_T2025-07","field_index":205,"field":"fake news detection","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Social, propagation, and source-context detection","description":"Detect or characterize fake news using social diffusion, user/source relations, propagation graphs, website/source context, narrative spread, or network-level mitigation as the primary evidence.","inclusion":"Include work whose main objective is veracity, source reliability, or misinformation-spread assessment through graph, propagation, user-credibility, source-network, website-structural, or narrative-tracking signals, even when text or images are auxiliary.","exclusion":"If the main objective is domain-shift robustness rather than social/source modeling, assign to D2; if the main objective is human-readable explanation of a detector, assign to D7.","nearest_boundary":"D2","representative_topics":["graph neural network fake news detection via social propagation context and transformer content features","graph-aware multi-view fusion for rumor representation learning on social media","content-agnostic fake news website detection using network and structural domain features"]},{"direction_index":1,"label":"D1","name":"Text and domain-focused content classification","description":"Detect fake news, rumors, false claims, or health/COVID misinformation primarily from linguistic content, document structure, emotional or stylistic signals, or text-centered classifiers.","inclusion":"Include general or domain-specific text-primary detection where the main objective is content-level classification, including transformer, RNN/CNN, ensemble, weak-supervision, emotion/sentiment, and health/COVID-focused text detection.","exclusion":"If the primary objective is language-resource coverage, assign to D6; if the primary objective is LLM-as-fact-checker behavior, assign to D4; if multimodal fusion is central, assign to D3.","nearest_boundary":"D6","representative_topics":["fake news classification under class imbalance using pre-trained encoders and negation-based data augmentation","multi-policy deep learning framework for joint statement and article fake news detection","label noise-resistant weak supervision for fake news detection via mean teacher propagation"]},{"direction_index":2,"label":"D2","name":"Robust transfer and bias-resistant generalization","description":"Improve or evaluate fake-news detection reliability under domain shift, event/topic shift, missing evidence, adversarial perturbations, dataset bias, or out-of-distribution conditions.","inclusion":"Include work whose primary objective is transfer learning, domain adaptation, debiasing, adversarial robustness, uncertainty, missing-evidence realism, or generalization evaluation across distributional stressors.","exclusion":"If the primary transfer problem is low-resource or cross-lingual language coverage, assign to D6; if the contribution is mainly a dataset or reproducibility resource, assign to D5.","nearest_boundary":"D6","representative_topics":["multi-domain fake news detection via memory-guided multi-view domain adaptation","domain- and instance-level transfer for targeted fake news detection in influential domains","realistic nlp fact-checking under missing counter-evidence conditions for misinformation"]},{"direction_index":3,"label":"D3","name":"Multimodal and cross-modal verification","description":"Detect or verify fake news by jointly reasoning over text, images, video, or other modalities through fusion, alignment, cross-modal consistency, or media-authenticity cues.","inclusion":"Include work whose primary objective is multimodal fake-news detection, multimodal fact verification, image-text rumor classification, out-of-context media detection, or cross-modal representation learning for veracity.","exclusion":"If the main contribution is the dataset or benchmark rather than the multimodal method, assign to D5; if explanation is the central objective, assign to D7.","nearest_boundary":"D5","representative_topics":["bootstrapping multi-view representations for cross-modal consistency-aware fake news detection","dual-tag image tagging for authenticable and traceable fake news detection","multimodal fusion and ensemble learning for social media rumor veracity classification"]},{"direction_index":4,"label":"D4","name":"LLM-era disinformation detection and fact-checking","description":"Study misinformation involving large language models, either by detecting LLM-generated disinformation or by using/evaluating LLMs for verification, debunking, truthfulness judgment, or controlled correction.","inclusion":"Include work where LLMs are the object of study, the generator of deceptive content to be detected, or the primary instrument for fact-checking, debunking, claim editing, uncertainty-aware judgment, or misinformation mitigation.","exclusion":"If the work only fine-tunes ordinary pretrained encoders for text classification without an LLM-centric objective, assign to D1; if explanation of detector decisions is the primary objective, assign to D7.","nearest_boundary":"D7","representative_topics":["generalizable misinformation veracity evaluation with uncertainty and possibility-aware language models","zero-shot large language model prompting for political fake statement detection","comparative black-box evaluation of llms for news fact-checking and truthfulness classification"]},{"direction_index":5,"label":"D5","name":"Datasets, benchmarks, and audits","description":"Create, annotate, benchmark, audit, or analyze fake-news and misinformation datasets, shared tasks, evaluation resources, reproducibility, or dataset quality as the primary contribution.","inclusion":"Include work centered on resource construction, annotation schemes, benchmark design, shared-task resources, time-aware datasets, synthetic/evaluation resources, reproducibility audits, relabeling, or dataset quality and bias analysis rather than a new detector.","exclusion":"If a dataset is secondary to a new modeling method, assign to the relevant modeling direction; if the main goal is adversarial or transfer robustness rather than resource design, assign to D2.","nearest_boundary":"D2","representative_topics":["multi-event annotated tweet dataset for benchmarking misinformation detection across languages","time-aware dataset construction for detecting evolving misinformation in language models","multimodal claim verification dataset with satire and visual-textual entailment labels"]},{"direction_index":6,"label":"D6","name":"Multilingual and low-resource detection","description":"Develop fake-news, disinformation, or fact-checking systems for non-English, multilingual, cross-lingual, or low-resource language settings.","inclusion":"Include work where language coverage, low-resource modeling, multilingual representation, cross-lingual retrieval, cross-lingual transfer, or language-specific shared-task participation is the primary scientific objective.","exclusion":"If the central problem is topic/event/domain shift without language coverage as the main challenge, assign to D2; if COVID/health misinformation is dominant and the language setting is secondary, assign to D1.","nearest_boundary":"D2","representative_topics":["low-resource urdu fake news detection using transformer and traditional feature representations","low-resource urdu fake news detection using count-based features and neural classifiers","multilingual text embedding models for disinformation publisher detection in digital advertising"]},{"direction_index":7,"label":"D7","name":"Explainable and evidence-grounded detection","description":"Make fake-news, rumor, or misinformation judgments understandable or evidence-grounded through rationales, counterfactuals, evidence retrieval, evidence distillation, interpretable latent variables, or human-readable reasoning.","inclusion":"Include work where interpretability, explanation, rationale generation, evidence grounding, rebuttal support, or counterfactual evidence is the main scientific objective, regardless of whether the underlying evidence is textual, social, multimodal, or health-related.","exclusion":"If explanations are only auxiliary visualizations for a detector whose main contribution is graph, text, or multimodal modeling, assign to the relevant modeling direction; if the main focus is evaluating LLM fact-checking capability, assign to D4.","nearest_boundary":"D4","representative_topics":["counterfactual explanation generation for fake news using contradicted entailment reasoning","multi-modal explainable ai for interpretable fake news detection on social media","interpretable fake news classification using bert embeddings and word-token activation mapping"]}],"split":"dev"}
{"episode_id":"F205_T2026-01","field_index":205,"field":"fake news detection","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Social, propagation, and source-context detection","description":"Detect or characterize fake news using social diffusion, user/source relations, propagation graphs, website/source context, narrative spread, or network-level mitigation as the primary evidence.","inclusion":"Include work whose main objective is veracity, source reliability, or misinformation-spread assessment through graph, propagation, user-credibility, source-network, website-structural, or narrative-tracking signals, even when text or images are auxiliary.","exclusion":"If the main objective is domain-shift robustness rather than social/source modeling, assign to D2; if the main objective is human-readable explanation of a detector, assign to D7.","nearest_boundary":"D2","representative_topics":["graph neural network fake news detection via social propagation context and transformer content features","graph-aware multi-view fusion for rumor representation learning on social media","content-agnostic fake news website detection using network and structural domain features"]},{"direction_index":1,"label":"D1","name":"Text and domain-focused content classification","description":"Detect fake news, rumors, false claims, or health/COVID misinformation primarily from linguistic content, document structure, emotional or stylistic signals, or text-centered classifiers.","inclusion":"Include general or domain-specific text-primary detection where the main objective is content-level classification, including transformer, RNN/CNN, ensemble, weak-supervision, emotion/sentiment, and health/COVID-focused text detection.","exclusion":"If the primary objective is language-resource coverage, assign to D6; if the primary objective is LLM-as-fact-checker behavior, assign to D4; if multimodal fusion is central, assign to D3.","nearest_boundary":"D6","representative_topics":["fake news classification under class imbalance using pre-trained encoders and negation-based data augmentation","multi-policy deep learning framework for joint statement and article fake news detection","label noise-resistant weak supervision for fake news detection via mean teacher propagation"]},{"direction_index":2,"label":"D2","name":"Robust transfer and bias-resistant generalization","description":"Improve or evaluate fake-news detection reliability under domain shift, event/topic shift, missing evidence, adversarial perturbations, dataset bias, or out-of-distribution conditions.","inclusion":"Include work whose primary objective is transfer learning, domain adaptation, debiasing, adversarial robustness, uncertainty, missing-evidence realism, or generalization evaluation across distributional stressors.","exclusion":"If the primary transfer problem is low-resource or cross-lingual language coverage, assign to D6; if the contribution is mainly a dataset or reproducibility resource, assign to D5.","nearest_boundary":"D6","representative_topics":["multi-domain fake news detection via memory-guided multi-view domain adaptation","domain- and instance-level transfer for targeted fake news detection in influential domains","realistic nlp fact-checking under missing counter-evidence conditions for misinformation"]},{"direction_index":3,"label":"D3","name":"Multimodal and cross-modal verification","description":"Detect or verify fake news by jointly reasoning over text, images, video, or other modalities through fusion, alignment, cross-modal consistency, or media-authenticity cues.","inclusion":"Include work whose primary objective is multimodal fake-news detection, multimodal fact verification, image-text rumor classification, out-of-context media detection, or cross-modal representation learning for veracity.","exclusion":"If the main contribution is the dataset or benchmark rather than the multimodal method, assign to D5; if explanation is the central objective, assign to D7.","nearest_boundary":"D5","representative_topics":["bootstrapping multi-view representations for cross-modal consistency-aware fake news detection","dual-tag image tagging for authenticable and traceable fake news detection","multimodal fusion and ensemble learning for social media rumor veracity classification"]},{"direction_index":4,"label":"D4","name":"LLM-era disinformation detection and fact-checking","description":"Study misinformation involving large language models, either by detecting LLM-generated disinformation or by using/evaluating LLMs for verification, debunking, truthfulness judgment, or controlled correction.","inclusion":"Include work where LLMs are the object of study, the generator of deceptive content to be detected, or the primary instrument for fact-checking, debunking, claim editing, uncertainty-aware judgment, or misinformation mitigation.","exclusion":"If the work only fine-tunes ordinary pretrained encoders for text classification without an LLM-centric objective, assign to D1; if explanation of detector decisions is the primary objective, assign to D7.","nearest_boundary":"D7","representative_topics":["generalizable misinformation veracity evaluation with uncertainty and possibility-aware language models","zero-shot large language model prompting for political fake statement detection","comparative black-box evaluation of llms for news fact-checking and truthfulness classification"]},{"direction_index":5,"label":"D5","name":"Datasets, benchmarks, and audits","description":"Create, annotate, benchmark, audit, or analyze fake-news and misinformation datasets, shared tasks, evaluation resources, reproducibility, or dataset quality as the primary contribution.","inclusion":"Include work centered on resource construction, annotation schemes, benchmark design, shared-task resources, time-aware datasets, synthetic/evaluation resources, reproducibility audits, relabeling, or dataset quality and bias analysis rather than a new detector.","exclusion":"If a dataset is secondary to a new modeling method, assign to the relevant modeling direction; if the main goal is adversarial or transfer robustness rather than resource design, assign to D2.","nearest_boundary":"D2","representative_topics":["multi-event annotated tweet dataset for benchmarking misinformation detection across languages","time-aware dataset construction for detecting evolving misinformation in language models","multimodal claim verification dataset with satire and visual-textual entailment labels"]},{"direction_index":6,"label":"D6","name":"Multilingual and low-resource detection","description":"Develop fake-news, disinformation, or fact-checking systems for non-English, multilingual, cross-lingual, or low-resource language settings.","inclusion":"Include work where language coverage, low-resource modeling, multilingual representation, cross-lingual retrieval, cross-lingual transfer, or language-specific shared-task participation is the primary scientific objective.","exclusion":"If the central problem is topic/event/domain shift without language coverage as the main challenge, assign to D2; if COVID/health misinformation is dominant and the language setting is secondary, assign to D1.","nearest_boundary":"D2","representative_topics":["low-resource urdu fake news detection using transformer and traditional feature representations","low-resource urdu fake news detection using count-based features and neural classifiers","multilingual text embedding models for disinformation publisher detection in digital advertising"]},{"direction_index":7,"label":"D7","name":"Explainable and evidence-grounded detection","description":"Make fake-news, rumor, or misinformation judgments understandable or evidence-grounded through rationales, counterfactuals, evidence retrieval, evidence distillation, interpretable latent variables, or human-readable reasoning.","inclusion":"Include work where interpretability, explanation, rationale generation, evidence grounding, rebuttal support, or counterfactual evidence is the main scientific objective, regardless of whether the underlying evidence is textual, social, multimodal, or health-related.","exclusion":"If explanations are only auxiliary visualizations for a detector whose main contribution is graph, text, or multimodal modeling, assign to the relevant modeling direction; if the main focus is evaluating LLM fact-checking capability, assign to D4.","nearest_boundary":"D4","representative_topics":["counterfactual explanation generation for fake news using contradicted entailment reasoning","multi-modal explainable ai for interpretable fake news detection on social media","interpretable fake news classification using bert embeddings and word-token activation mapping"]}],"split":"dev"}
{"episode_id":"F209_T2024-01","field_index":209,"field":"bayesian deep learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Variational posterior approximation","description":"Formulate, improve, or analyze variational objectives and posterior families for Bayesian neural networks and Bayesian deep models, including mean-field, structured, implicit, amortized, flow-based, function-space, and symmetry-aware approximations.","inclusion":"Include work whose primary research objective is a new variational family, variational learning objective, deterministic learned posterior approximation, or theory specific to variational inference quality in Bayesian deep learning.","exclusion":"If the main objective is explicit Monte Carlo posterior simulation, assign to sampling-based inference; if it is local curvature Gaussianization, assign to Laplace and local geometry; if it is primarily coresets, subspaces, systems, or compression for tractability, assign to scalable efficient inference.","nearest_boundary":"D1","representative_topics":["deep variational implicit processes for scalable non-gaussian function-space inference","variational uncertainty estimation via layer-wise gaussian output distributions in neural networks","relu activation functions causing poor mean-field variational inference posteriors in bayesian neural networks"]},{"direction_index":1,"label":"D1","name":"Sampling-based posterior inference","description":"Develop and analyze explicit sample-based posterior inference for Bayesian deep learning, including MCMC, Langevin and Hamiltonian dynamics, Gibbs sampling, particle methods, importance sampling, SMC, and related stochastic simulation methods.","inclusion":"Include work whose primary research objective is approximating Bayesian neural network or Bayesian deep model posteriors through sampler design, particle dynamics, importance weighting, convergence analysis, or stochastic posterior simulation.","exclusion":"If sampling is incidental inside a broader transfer, federated, or application pipeline, assign to that broader objective; if the main approximation is variational or local Gaussian/Laplace rather than posterior simulation, assign to the corresponding inference direction.","nearest_boundary":"D0","representative_topics":["feature-space particle inference for diverse neural network ensembles via bayesian posterior sampling","blocked gibbs sampling for minibatch mcmc in deep bayesian feedforward neural networks","adaptive importance sampling for scalable bayesian neural network posterior inference"]},{"direction_index":2,"label":"D2","name":"Scalable efficient representations","description":"Make Bayesian deep learning computationally feasible through scalable inference reductions, reusable systems, coresets, low-dimensional subspaces, distillation, post-hoc Bayesian wrappers, sparse subnetworks, pruning, masks, compression, or hardware-efficient Bayesian representations.","inclusion":"Include work whose primary research objective is tractability, reduced cost, model or data compression, sparse or structured Bayesian representation, scalable infrastructure, or efficient deployment while preserving Bayesian uncertainty semantics.","exclusion":"If the central novelty is the posterior approximation family itself, assign to variational, sampling, or Laplace inference; if the main objective is adaptation across tasks or clients rather than efficiency alone, assign to Bayesian transfer and federated learning.","nearest_boundary":"D0","representative_topics":["sequential bayesian sparse subnetwork ensembles for robust prediction and uncertainty estimation","sparse bayesian neural networks with node-wise masking and minimax-optimal posterior structure","functional ensemble distillation for covariance-aware uncertainty estimation under limited data"]},{"direction_index":3,"label":"D3","name":"Laplace and local geometry","description":"Construct posterior approximations, evidence estimates, and uncertainty mechanisms using local curvature, Laplace approximations, linearized neural networks, neural tangent kernels, Riemannian geometry, or related local Gaussian methods.","inclusion":"Include work whose primary research objective is a Laplace-style, local Gaussian, linearized, curvature-based, Riemannian, or NTK-connected Bayesian approximation for deep networks.","exclusion":"If the approximation is a learned variational posterior not centered on local curvature, assign to variational posterior approximation; if general posterior samples are the central object rather than a local Gaussian or linear surrogate, assign to sampling-based posterior inference.","nearest_boundary":"D0","representative_topics":["adapting linearised laplace model evidence estimation for bayesian deep learning with normalisation and stochastic approximation","scalable sample-based bayesian inference for large linearised laplace surrogate models","accelerating linearized laplace approximation via ntk nyström for bayesian pretrained network adaptation"]},{"direction_index":4,"label":"D4","name":"Uncertainty calibration and robustness","description":"Assess, calibrate, interpret, improve, or certify predictive uncertainty and robustness of Bayesian deep models under ordinary prediction, covariate shift, corruption, out-of-distribution stress, or adversarial perturbation.","inclusion":"Include work whose primary research objective is uncertainty quality, calibration, benchmarking, decomposition, coverage, OOD behavior, corruption robustness, adversarial robustness, or certification rather than posterior inference as an end in itself.","exclusion":"If uncertainty or robustness is only an evaluation endpoint for a new inference algorithm, assign to that inference direction; if the main contribution is a domain-specific mechanistic model, assign to domain- and knowledge-constrained modeling; if it is primarily proof-driven general theory, assign to theoretical foundations.","nearest_boundary":"D6","representative_topics":["node-based bayesian neural networks for covariate shift via implicit input corruptions","benchmarking bayesian neural network posterior approximation methods for regression uncertainty quantification","corruption-robust bayesian neural networks via training-data-dependent likelihood matching"]},{"direction_index":5,"label":"D5","name":"Domain and knowledge modeling","description":"Adapt Bayesian deep learning to domain-structured, scientific, mechanistic, or application-shaped problems through domain-aware architectures, physical constraints, knowledge constraints, informative priors, posterior regularization, specialized likelihoods, neural operators, kernels, or task-specific Bayesian model design.","inclusion":"Include work whose primary research objective is incorporating external knowledge, domain structure, physical or scientific constraints, learned or informative priors, posterior tempering, or specialized Bayesian architectures for a particular structured problem setting.","exclusion":"If the contribution is a general-purpose inference algorithm only demonstrated on a domain dataset, assign to the relevant inference direction; if it is an uncertainty benchmark without domain-shaped modeling, assign to uncertainty calibration and robustness; if the main objective is cross-task or client adaptation, assign to Bayesian transfer and federated learning.","nearest_boundary":"D4","representative_topics":["variational bayesian deep operator networks for uncertainty-aware parametric differential equation solving","posterior-regularized bayesian neural networks with soft and hard knowledge constraints","hierarchical bayesian graph neural networks for weak lensing convergence estimation in strong time delay cosmography"]},{"direction_index":6,"label":"D6","name":"Theoretical foundations","description":"Develop formal mathematical understanding of Bayesian deep learning, including posterior contraction, asymptotics, width limits, generalization, PAC-Bayes risk, posterior geometry, tempered behavior, function-space versus parameter-space properties, and expressivity of stochastic parameters.","inclusion":"Include work whose primary research objective is a theorem, bound, limit, equivalence, convergence result, or formal characterization of Bayesian neural networks or Bayesian deep learning behavior.","exclusion":"If theory mainly supports a specific new inference algorithm, robustness method, or application system, assign to that method's primary objective; if the result is specifically about a variational objective rather than broader BDL behavior, assign to variational posterior approximation.","nearest_boundary":"D0","representative_topics":["bayesian neural network posterior convergence rates in besov space with adaptive smoothness","pac-bayesian excess risk analysis of epistemic uncertainty in variational inference","theoretical analysis of bayesian neural network robustness to adversarial attacks via posterior geometry"]},{"direction_index":7,"label":"D7","name":"Bayesian transfer and federated learning","description":"Use Bayesian deep learning for learning across heterogeneous tasks, clients, populations, or time-varying data sources, including meta-learning, few-shot learning, federated personalization, hierarchical Bayesian federation, continual updating, and cross-task prior or posterior transfer.","inclusion":"Include work whose primary research objective is transferable probabilistic representation, meta-learned prior or posterior, client personalization, federated Bayesian aggregation, task heterogeneity, or adaptation across data partitions or time.","exclusion":"If federation is only a distributed implementation of a sampler, assign to sampling-based posterior inference; if prior learning is single-task rather than cross-task or cross-client, assign to domain and knowledge modeling; if the central goal is generic computational scalability, assign to scalable efficient representations.","nearest_boundary":"D2","representative_topics":["functional bayesian meta-learning via score matching over data-generating process marginals","federated langevin dynamics with heterogeneity-aware convergence guarantees and drift-correcting control variates","variational bayesian personalized federated learning with sparsity and clustering for non-i.i.d. data"]}],"split":"dev"}
{"episode_id":"F209_T2024-07","field_index":209,"field":"bayesian deep learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Variational posterior approximation","description":"Formulate, improve, or analyze variational objectives and posterior families for Bayesian neural networks and Bayesian deep models, including mean-field, structured, implicit, amortized, flow-based, function-space, and symmetry-aware approximations.","inclusion":"Include work whose primary research objective is a new variational family, variational learning objective, deterministic learned posterior approximation, or theory specific to variational inference quality in Bayesian deep learning.","exclusion":"If the main objective is explicit Monte Carlo posterior simulation, assign to sampling-based inference; if it is local curvature Gaussianization, assign to Laplace and local geometry; if it is primarily coresets, subspaces, systems, or compression for tractability, assign to scalable efficient inference.","nearest_boundary":"D1","representative_topics":["deep variational implicit processes for scalable non-gaussian function-space inference","variational uncertainty estimation via layer-wise gaussian output distributions in neural networks","relu activation functions causing poor mean-field variational inference posteriors in bayesian neural networks"]},{"direction_index":1,"label":"D1","name":"Sampling-based posterior inference","description":"Develop and analyze explicit sample-based posterior inference for Bayesian deep learning, including MCMC, Langevin and Hamiltonian dynamics, Gibbs sampling, particle methods, importance sampling, SMC, and related stochastic simulation methods.","inclusion":"Include work whose primary research objective is approximating Bayesian neural network or Bayesian deep model posteriors through sampler design, particle dynamics, importance weighting, convergence analysis, or stochastic posterior simulation.","exclusion":"If sampling is incidental inside a broader transfer, federated, or application pipeline, assign to that broader objective; if the main approximation is variational or local Gaussian/Laplace rather than posterior simulation, assign to the corresponding inference direction.","nearest_boundary":"D0","representative_topics":["feature-space particle inference for diverse neural network ensembles via bayesian posterior sampling","blocked gibbs sampling for minibatch mcmc in deep bayesian feedforward neural networks","adaptive importance sampling for scalable bayesian neural network posterior inference"]},{"direction_index":2,"label":"D2","name":"Scalable efficient representations","description":"Make Bayesian deep learning computationally feasible through scalable inference reductions, reusable systems, coresets, low-dimensional subspaces, distillation, post-hoc Bayesian wrappers, sparse subnetworks, pruning, masks, compression, or hardware-efficient Bayesian representations.","inclusion":"Include work whose primary research objective is tractability, reduced cost, model or data compression, sparse or structured Bayesian representation, scalable infrastructure, or efficient deployment while preserving Bayesian uncertainty semantics.","exclusion":"If the central novelty is the posterior approximation family itself, assign to variational, sampling, or Laplace inference; if the main objective is adaptation across tasks or clients rather than efficiency alone, assign to Bayesian transfer and federated learning.","nearest_boundary":"D0","representative_topics":["sequential bayesian sparse subnetwork ensembles for robust prediction and uncertainty estimation","sparse bayesian neural networks with node-wise masking and minimax-optimal posterior structure","functional ensemble distillation for covariance-aware uncertainty estimation under limited data"]},{"direction_index":3,"label":"D3","name":"Laplace and local geometry","description":"Construct posterior approximations, evidence estimates, and uncertainty mechanisms using local curvature, Laplace approximations, linearized neural networks, neural tangent kernels, Riemannian geometry, or related local Gaussian methods.","inclusion":"Include work whose primary research objective is a Laplace-style, local Gaussian, linearized, curvature-based, Riemannian, or NTK-connected Bayesian approximation for deep networks.","exclusion":"If the approximation is a learned variational posterior not centered on local curvature, assign to variational posterior approximation; if general posterior samples are the central object rather than a local Gaussian or linear surrogate, assign to sampling-based posterior inference.","nearest_boundary":"D0","representative_topics":["adapting linearised laplace model evidence estimation for bayesian deep learning with normalisation and stochastic approximation","scalable sample-based bayesian inference for large linearised laplace surrogate models","accelerating linearized laplace approximation via ntk nyström for bayesian pretrained network adaptation"]},{"direction_index":4,"label":"D4","name":"Uncertainty calibration and robustness","description":"Assess, calibrate, interpret, improve, or certify predictive uncertainty and robustness of Bayesian deep models under ordinary prediction, covariate shift, corruption, out-of-distribution stress, or adversarial perturbation.","inclusion":"Include work whose primary research objective is uncertainty quality, calibration, benchmarking, decomposition, coverage, OOD behavior, corruption robustness, adversarial robustness, or certification rather than posterior inference as an end in itself.","exclusion":"If uncertainty or robustness is only an evaluation endpoint for a new inference algorithm, assign to that inference direction; if the main contribution is a domain-specific mechanistic model, assign to domain- and knowledge-constrained modeling; if it is primarily proof-driven general theory, assign to theoretical foundations.","nearest_boundary":"D6","representative_topics":["node-based bayesian neural networks for covariate shift via implicit input corruptions","benchmarking bayesian neural network posterior approximation methods for regression uncertainty quantification","corruption-robust bayesian neural networks via training-data-dependent likelihood matching"]},{"direction_index":5,"label":"D5","name":"Domain and knowledge modeling","description":"Adapt Bayesian deep learning to domain-structured, scientific, mechanistic, or application-shaped problems through domain-aware architectures, physical constraints, knowledge constraints, informative priors, posterior regularization, specialized likelihoods, neural operators, kernels, or task-specific Bayesian model design.","inclusion":"Include work whose primary research objective is incorporating external knowledge, domain structure, physical or scientific constraints, learned or informative priors, posterior tempering, or specialized Bayesian architectures for a particular structured problem setting.","exclusion":"If the contribution is a general-purpose inference algorithm only demonstrated on a domain dataset, assign to the relevant inference direction; if it is an uncertainty benchmark without domain-shaped modeling, assign to uncertainty calibration and robustness; if the main objective is cross-task or client adaptation, assign to Bayesian transfer and federated learning.","nearest_boundary":"D4","representative_topics":["variational bayesian deep operator networks for uncertainty-aware parametric differential equation solving","posterior-regularized bayesian neural networks with soft and hard knowledge constraints","hierarchical bayesian graph neural networks for weak lensing convergence estimation in strong time delay cosmography"]},{"direction_index":6,"label":"D6","name":"Theoretical foundations","description":"Develop formal mathematical understanding of Bayesian deep learning, including posterior contraction, asymptotics, width limits, generalization, PAC-Bayes risk, posterior geometry, tempered behavior, function-space versus parameter-space properties, and expressivity of stochastic parameters.","inclusion":"Include work whose primary research objective is a theorem, bound, limit, equivalence, convergence result, or formal characterization of Bayesian neural networks or Bayesian deep learning behavior.","exclusion":"If theory mainly supports a specific new inference algorithm, robustness method, or application system, assign to that method's primary objective; if the result is specifically about a variational objective rather than broader BDL behavior, assign to variational posterior approximation.","nearest_boundary":"D0","representative_topics":["bayesian neural network posterior convergence rates in besov space with adaptive smoothness","pac-bayesian excess risk analysis of epistemic uncertainty in variational inference","theoretical analysis of bayesian neural network robustness to adversarial attacks via posterior geometry"]},{"direction_index":7,"label":"D7","name":"Bayesian transfer and federated learning","description":"Use Bayesian deep learning for learning across heterogeneous tasks, clients, populations, or time-varying data sources, including meta-learning, few-shot learning, federated personalization, hierarchical Bayesian federation, continual updating, and cross-task prior or posterior transfer.","inclusion":"Include work whose primary research objective is transferable probabilistic representation, meta-learned prior or posterior, client personalization, federated Bayesian aggregation, task heterogeneity, or adaptation across data partitions or time.","exclusion":"If federation is only a distributed implementation of a sampler, assign to sampling-based posterior inference; if prior learning is single-task rather than cross-task or cross-client, assign to domain and knowledge modeling; if the central goal is generic computational scalability, assign to scalable efficient representations.","nearest_boundary":"D2","representative_topics":["functional bayesian meta-learning via score matching over data-generating process marginals","federated langevin dynamics with heterogeneity-aware convergence guarantees and drift-correcting control variates","variational bayesian personalized federated learning with sparsity and clustering for non-i.i.d. data"]}],"split":"dev"}
{"episode_id":"F209_T2025-01","field_index":209,"field":"bayesian deep learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Variational posterior approximation","description":"Formulate, improve, or analyze variational objectives and posterior families for Bayesian neural networks and Bayesian deep models, including mean-field, structured, implicit, amortized, flow-based, function-space, and symmetry-aware approximations.","inclusion":"Include work whose primary research objective is a new variational family, variational learning objective, deterministic learned posterior approximation, or theory specific to variational inference quality in Bayesian deep learning.","exclusion":"If the main objective is explicit Monte Carlo posterior simulation, assign to sampling-based inference; if it is local curvature Gaussianization, assign to Laplace and local geometry; if it is primarily coresets, subspaces, systems, or compression for tractability, assign to scalable efficient inference.","nearest_boundary":"D1","representative_topics":["deep variational implicit processes for scalable non-gaussian function-space inference","variational uncertainty estimation via layer-wise gaussian output distributions in neural networks","relu activation functions causing poor mean-field variational inference posteriors in bayesian neural networks"]},{"direction_index":1,"label":"D1","name":"Sampling-based posterior inference","description":"Develop and analyze explicit sample-based posterior inference for Bayesian deep learning, including MCMC, Langevin and Hamiltonian dynamics, Gibbs sampling, particle methods, importance sampling, SMC, and related stochastic simulation methods.","inclusion":"Include work whose primary research objective is approximating Bayesian neural network or Bayesian deep model posteriors through sampler design, particle dynamics, importance weighting, convergence analysis, or stochastic posterior simulation.","exclusion":"If sampling is incidental inside a broader transfer, federated, or application pipeline, assign to that broader objective; if the main approximation is variational or local Gaussian/Laplace rather than posterior simulation, assign to the corresponding inference direction.","nearest_boundary":"D0","representative_topics":["feature-space particle inference for diverse neural network ensembles via bayesian posterior sampling","blocked gibbs sampling for minibatch mcmc in deep bayesian feedforward neural networks","adaptive importance sampling for scalable bayesian neural network posterior inference"]},{"direction_index":2,"label":"D2","name":"Scalable efficient representations","description":"Make Bayesian deep learning computationally feasible through scalable inference reductions, reusable systems, coresets, low-dimensional subspaces, distillation, post-hoc Bayesian wrappers, sparse subnetworks, pruning, masks, compression, or hardware-efficient Bayesian representations.","inclusion":"Include work whose primary research objective is tractability, reduced cost, model or data compression, sparse or structured Bayesian representation, scalable infrastructure, or efficient deployment while preserving Bayesian uncertainty semantics.","exclusion":"If the central novelty is the posterior approximation family itself, assign to variational, sampling, or Laplace inference; if the main objective is adaptation across tasks or clients rather than efficiency alone, assign to Bayesian transfer and federated learning.","nearest_boundary":"D0","representative_topics":["sequential bayesian sparse subnetwork ensembles for robust prediction and uncertainty estimation","sparse bayesian neural networks with node-wise masking and minimax-optimal posterior structure","functional ensemble distillation for covariance-aware uncertainty estimation under limited data"]},{"direction_index":3,"label":"D3","name":"Laplace and local geometry","description":"Construct posterior approximations, evidence estimates, and uncertainty mechanisms using local curvature, Laplace approximations, linearized neural networks, neural tangent kernels, Riemannian geometry, or related local Gaussian methods.","inclusion":"Include work whose primary research objective is a Laplace-style, local Gaussian, linearized, curvature-based, Riemannian, or NTK-connected Bayesian approximation for deep networks.","exclusion":"If the approximation is a learned variational posterior not centered on local curvature, assign to variational posterior approximation; if general posterior samples are the central object rather than a local Gaussian or linear surrogate, assign to sampling-based posterior inference.","nearest_boundary":"D0","representative_topics":["adapting linearised laplace model evidence estimation for bayesian deep learning with normalisation and stochastic approximation","scalable sample-based bayesian inference for large linearised laplace surrogate models","accelerating linearized laplace approximation via ntk nyström for bayesian pretrained network adaptation"]},{"direction_index":4,"label":"D4","name":"Uncertainty calibration and robustness","description":"Assess, calibrate, interpret, improve, or certify predictive uncertainty and robustness of Bayesian deep models under ordinary prediction, covariate shift, corruption, out-of-distribution stress, or adversarial perturbation.","inclusion":"Include work whose primary research objective is uncertainty quality, calibration, benchmarking, decomposition, coverage, OOD behavior, corruption robustness, adversarial robustness, or certification rather than posterior inference as an end in itself.","exclusion":"If uncertainty or robustness is only an evaluation endpoint for a new inference algorithm, assign to that inference direction; if the main contribution is a domain-specific mechanistic model, assign to domain- and knowledge-constrained modeling; if it is primarily proof-driven general theory, assign to theoretical foundations.","nearest_boundary":"D6","representative_topics":["node-based bayesian neural networks for covariate shift via implicit input corruptions","benchmarking bayesian neural network posterior approximation methods for regression uncertainty quantification","corruption-robust bayesian neural networks via training-data-dependent likelihood matching"]},{"direction_index":5,"label":"D5","name":"Domain and knowledge modeling","description":"Adapt Bayesian deep learning to domain-structured, scientific, mechanistic, or application-shaped problems through domain-aware architectures, physical constraints, knowledge constraints, informative priors, posterior regularization, specialized likelihoods, neural operators, kernels, or task-specific Bayesian model design.","inclusion":"Include work whose primary research objective is incorporating external knowledge, domain structure, physical or scientific constraints, learned or informative priors, posterior tempering, or specialized Bayesian architectures for a particular structured problem setting.","exclusion":"If the contribution is a general-purpose inference algorithm only demonstrated on a domain dataset, assign to the relevant inference direction; if it is an uncertainty benchmark without domain-shaped modeling, assign to uncertainty calibration and robustness; if the main objective is cross-task or client adaptation, assign to Bayesian transfer and federated learning.","nearest_boundary":"D4","representative_topics":["variational bayesian deep operator networks for uncertainty-aware parametric differential equation solving","posterior-regularized bayesian neural networks with soft and hard knowledge constraints","hierarchical bayesian graph neural networks for weak lensing convergence estimation in strong time delay cosmography"]},{"direction_index":6,"label":"D6","name":"Theoretical foundations","description":"Develop formal mathematical understanding of Bayesian deep learning, including posterior contraction, asymptotics, width limits, generalization, PAC-Bayes risk, posterior geometry, tempered behavior, function-space versus parameter-space properties, and expressivity of stochastic parameters.","inclusion":"Include work whose primary research objective is a theorem, bound, limit, equivalence, convergence result, or formal characterization of Bayesian neural networks or Bayesian deep learning behavior.","exclusion":"If theory mainly supports a specific new inference algorithm, robustness method, or application system, assign to that method's primary objective; if the result is specifically about a variational objective rather than broader BDL behavior, assign to variational posterior approximation.","nearest_boundary":"D0","representative_topics":["bayesian neural network posterior convergence rates in besov space with adaptive smoothness","pac-bayesian excess risk analysis of epistemic uncertainty in variational inference","theoretical analysis of bayesian neural network robustness to adversarial attacks via posterior geometry"]},{"direction_index":7,"label":"D7","name":"Bayesian transfer and federated learning","description":"Use Bayesian deep learning for learning across heterogeneous tasks, clients, populations, or time-varying data sources, including meta-learning, few-shot learning, federated personalization, hierarchical Bayesian federation, continual updating, and cross-task prior or posterior transfer.","inclusion":"Include work whose primary research objective is transferable probabilistic representation, meta-learned prior or posterior, client personalization, federated Bayesian aggregation, task heterogeneity, or adaptation across data partitions or time.","exclusion":"If federation is only a distributed implementation of a sampler, assign to sampling-based posterior inference; if prior learning is single-task rather than cross-task or cross-client, assign to domain and knowledge modeling; if the central goal is generic computational scalability, assign to scalable efficient representations.","nearest_boundary":"D2","representative_topics":["functional bayesian meta-learning via score matching over data-generating process marginals","federated langevin dynamics with heterogeneity-aware convergence guarantees and drift-correcting control variates","variational bayesian personalized federated learning with sparsity and clustering for non-i.i.d. data"]}],"split":"dev"}
{"episode_id":"F209_T2025-07","field_index":209,"field":"bayesian deep learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Variational posterior approximation","description":"Formulate, improve, or analyze variational objectives and posterior families for Bayesian neural networks and Bayesian deep models, including mean-field, structured, implicit, amortized, flow-based, function-space, and symmetry-aware approximations.","inclusion":"Include work whose primary research objective is a new variational family, variational learning objective, deterministic learned posterior approximation, or theory specific to variational inference quality in Bayesian deep learning.","exclusion":"If the main objective is explicit Monte Carlo posterior simulation, assign to sampling-based inference; if it is local curvature Gaussianization, assign to Laplace and local geometry; if it is primarily coresets, subspaces, systems, or compression for tractability, assign to scalable efficient inference.","nearest_boundary":"D1","representative_topics":["deep variational implicit processes for scalable non-gaussian function-space inference","variational uncertainty estimation via layer-wise gaussian output distributions in neural networks","relu activation functions causing poor mean-field variational inference posteriors in bayesian neural networks"]},{"direction_index":1,"label":"D1","name":"Sampling-based posterior inference","description":"Develop and analyze explicit sample-based posterior inference for Bayesian deep learning, including MCMC, Langevin and Hamiltonian dynamics, Gibbs sampling, particle methods, importance sampling, SMC, and related stochastic simulation methods.","inclusion":"Include work whose primary research objective is approximating Bayesian neural network or Bayesian deep model posteriors through sampler design, particle dynamics, importance weighting, convergence analysis, or stochastic posterior simulation.","exclusion":"If sampling is incidental inside a broader transfer, federated, or application pipeline, assign to that broader objective; if the main approximation is variational or local Gaussian/Laplace rather than posterior simulation, assign to the corresponding inference direction.","nearest_boundary":"D0","representative_topics":["feature-space particle inference for diverse neural network ensembles via bayesian posterior sampling","blocked gibbs sampling for minibatch mcmc in deep bayesian feedforward neural networks","adaptive importance sampling for scalable bayesian neural network posterior inference"]},{"direction_index":2,"label":"D2","name":"Scalable efficient representations","description":"Make Bayesian deep learning computationally feasible through scalable inference reductions, reusable systems, coresets, low-dimensional subspaces, distillation, post-hoc Bayesian wrappers, sparse subnetworks, pruning, masks, compression, or hardware-efficient Bayesian representations.","inclusion":"Include work whose primary research objective is tractability, reduced cost, model or data compression, sparse or structured Bayesian representation, scalable infrastructure, or efficient deployment while preserving Bayesian uncertainty semantics.","exclusion":"If the central novelty is the posterior approximation family itself, assign to variational, sampling, or Laplace inference; if the main objective is adaptation across tasks or clients rather than efficiency alone, assign to Bayesian transfer and federated learning.","nearest_boundary":"D0","representative_topics":["sequential bayesian sparse subnetwork ensembles for robust prediction and uncertainty estimation","sparse bayesian neural networks with node-wise masking and minimax-optimal posterior structure","functional ensemble distillation for covariance-aware uncertainty estimation under limited data"]},{"direction_index":3,"label":"D3","name":"Laplace and local geometry","description":"Construct posterior approximations, evidence estimates, and uncertainty mechanisms using local curvature, Laplace approximations, linearized neural networks, neural tangent kernels, Riemannian geometry, or related local Gaussian methods.","inclusion":"Include work whose primary research objective is a Laplace-style, local Gaussian, linearized, curvature-based, Riemannian, or NTK-connected Bayesian approximation for deep networks.","exclusion":"If the approximation is a learned variational posterior not centered on local curvature, assign to variational posterior approximation; if general posterior samples are the central object rather than a local Gaussian or linear surrogate, assign to sampling-based posterior inference.","nearest_boundary":"D0","representative_topics":["adapting linearised laplace model evidence estimation for bayesian deep learning with normalisation and stochastic approximation","scalable sample-based bayesian inference for large linearised laplace surrogate models","accelerating linearized laplace approximation via ntk nyström for bayesian pretrained network adaptation"]},{"direction_index":4,"label":"D4","name":"Uncertainty calibration and robustness","description":"Assess, calibrate, interpret, improve, or certify predictive uncertainty and robustness of Bayesian deep models under ordinary prediction, covariate shift, corruption, out-of-distribution stress, or adversarial perturbation.","inclusion":"Include work whose primary research objective is uncertainty quality, calibration, benchmarking, decomposition, coverage, OOD behavior, corruption robustness, adversarial robustness, or certification rather than posterior inference as an end in itself.","exclusion":"If uncertainty or robustness is only an evaluation endpoint for a new inference algorithm, assign to that inference direction; if the main contribution is a domain-specific mechanistic model, assign to domain- and knowledge-constrained modeling; if it is primarily proof-driven general theory, assign to theoretical foundations.","nearest_boundary":"D6","representative_topics":["node-based bayesian neural networks for covariate shift via implicit input corruptions","benchmarking bayesian neural network posterior approximation methods for regression uncertainty quantification","corruption-robust bayesian neural networks via training-data-dependent likelihood matching"]},{"direction_index":5,"label":"D5","name":"Domain and knowledge modeling","description":"Adapt Bayesian deep learning to domain-structured, scientific, mechanistic, or application-shaped problems through domain-aware architectures, physical constraints, knowledge constraints, informative priors, posterior regularization, specialized likelihoods, neural operators, kernels, or task-specific Bayesian model design.","inclusion":"Include work whose primary research objective is incorporating external knowledge, domain structure, physical or scientific constraints, learned or informative priors, posterior tempering, or specialized Bayesian architectures for a particular structured problem setting.","exclusion":"If the contribution is a general-purpose inference algorithm only demonstrated on a domain dataset, assign to the relevant inference direction; if it is an uncertainty benchmark without domain-shaped modeling, assign to uncertainty calibration and robustness; if the main objective is cross-task or client adaptation, assign to Bayesian transfer and federated learning.","nearest_boundary":"D4","representative_topics":["variational bayesian deep operator networks for uncertainty-aware parametric differential equation solving","posterior-regularized bayesian neural networks with soft and hard knowledge constraints","hierarchical bayesian graph neural networks for weak lensing convergence estimation in strong time delay cosmography"]},{"direction_index":6,"label":"D6","name":"Theoretical foundations","description":"Develop formal mathematical understanding of Bayesian deep learning, including posterior contraction, asymptotics, width limits, generalization, PAC-Bayes risk, posterior geometry, tempered behavior, function-space versus parameter-space properties, and expressivity of stochastic parameters.","inclusion":"Include work whose primary research objective is a theorem, bound, limit, equivalence, convergence result, or formal characterization of Bayesian neural networks or Bayesian deep learning behavior.","exclusion":"If theory mainly supports a specific new inference algorithm, robustness method, or application system, assign to that method's primary objective; if the result is specifically about a variational objective rather than broader BDL behavior, assign to variational posterior approximation.","nearest_boundary":"D0","representative_topics":["bayesian neural network posterior convergence rates in besov space with adaptive smoothness","pac-bayesian excess risk analysis of epistemic uncertainty in variational inference","theoretical analysis of bayesian neural network robustness to adversarial attacks via posterior geometry"]},{"direction_index":7,"label":"D7","name":"Bayesian transfer and federated learning","description":"Use Bayesian deep learning for learning across heterogeneous tasks, clients, populations, or time-varying data sources, including meta-learning, few-shot learning, federated personalization, hierarchical Bayesian federation, continual updating, and cross-task prior or posterior transfer.","inclusion":"Include work whose primary research objective is transferable probabilistic representation, meta-learned prior or posterior, client personalization, federated Bayesian aggregation, task heterogeneity, or adaptation across data partitions or time.","exclusion":"If federation is only a distributed implementation of a sampler, assign to sampling-based posterior inference; if prior learning is single-task rather than cross-task or cross-client, assign to domain and knowledge modeling; if the central goal is generic computational scalability, assign to scalable efficient representations.","nearest_boundary":"D2","representative_topics":["functional bayesian meta-learning via score matching over data-generating process marginals","federated langevin dynamics with heterogeneity-aware convergence guarantees and drift-correcting control variates","variational bayesian personalized federated learning with sparsity and clustering for non-i.i.d. data"]}],"split":"dev"}
{"episode_id":"F209_T2026-01","field_index":209,"field":"bayesian deep learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Variational posterior approximation","description":"Formulate, improve, or analyze variational objectives and posterior families for Bayesian neural networks and Bayesian deep models, including mean-field, structured, implicit, amortized, flow-based, function-space, and symmetry-aware approximations.","inclusion":"Include work whose primary research objective is a new variational family, variational learning objective, deterministic learned posterior approximation, or theory specific to variational inference quality in Bayesian deep learning.","exclusion":"If the main objective is explicit Monte Carlo posterior simulation, assign to sampling-based inference; if it is local curvature Gaussianization, assign to Laplace and local geometry; if it is primarily coresets, subspaces, systems, or compression for tractability, assign to scalable efficient inference.","nearest_boundary":"D1","representative_topics":["deep variational implicit processes for scalable non-gaussian function-space inference","variational uncertainty estimation via layer-wise gaussian output distributions in neural networks","relu activation functions causing poor mean-field variational inference posteriors in bayesian neural networks"]},{"direction_index":1,"label":"D1","name":"Sampling-based posterior inference","description":"Develop and analyze explicit sample-based posterior inference for Bayesian deep learning, including MCMC, Langevin and Hamiltonian dynamics, Gibbs sampling, particle methods, importance sampling, SMC, and related stochastic simulation methods.","inclusion":"Include work whose primary research objective is approximating Bayesian neural network or Bayesian deep model posteriors through sampler design, particle dynamics, importance weighting, convergence analysis, or stochastic posterior simulation.","exclusion":"If sampling is incidental inside a broader transfer, federated, or application pipeline, assign to that broader objective; if the main approximation is variational or local Gaussian/Laplace rather than posterior simulation, assign to the corresponding inference direction.","nearest_boundary":"D0","representative_topics":["feature-space particle inference for diverse neural network ensembles via bayesian posterior sampling","blocked gibbs sampling for minibatch mcmc in deep bayesian feedforward neural networks","adaptive importance sampling for scalable bayesian neural network posterior inference"]},{"direction_index":2,"label":"D2","name":"Scalable efficient representations","description":"Make Bayesian deep learning computationally feasible through scalable inference reductions, reusable systems, coresets, low-dimensional subspaces, distillation, post-hoc Bayesian wrappers, sparse subnetworks, pruning, masks, compression, or hardware-efficient Bayesian representations.","inclusion":"Include work whose primary research objective is tractability, reduced cost, model or data compression, sparse or structured Bayesian representation, scalable infrastructure, or efficient deployment while preserving Bayesian uncertainty semantics.","exclusion":"If the central novelty is the posterior approximation family itself, assign to variational, sampling, or Laplace inference; if the main objective is adaptation across tasks or clients rather than efficiency alone, assign to Bayesian transfer and federated learning.","nearest_boundary":"D0","representative_topics":["sequential bayesian sparse subnetwork ensembles for robust prediction and uncertainty estimation","sparse bayesian neural networks with node-wise masking and minimax-optimal posterior structure","functional ensemble distillation for covariance-aware uncertainty estimation under limited data"]},{"direction_index":3,"label":"D3","name":"Laplace and local geometry","description":"Construct posterior approximations, evidence estimates, and uncertainty mechanisms using local curvature, Laplace approximations, linearized neural networks, neural tangent kernels, Riemannian geometry, or related local Gaussian methods.","inclusion":"Include work whose primary research objective is a Laplace-style, local Gaussian, linearized, curvature-based, Riemannian, or NTK-connected Bayesian approximation for deep networks.","exclusion":"If the approximation is a learned variational posterior not centered on local curvature, assign to variational posterior approximation; if general posterior samples are the central object rather than a local Gaussian or linear surrogate, assign to sampling-based posterior inference.","nearest_boundary":"D0","representative_topics":["adapting linearised laplace model evidence estimation for bayesian deep learning with normalisation and stochastic approximation","scalable sample-based bayesian inference for large linearised laplace surrogate models","accelerating linearized laplace approximation via ntk nyström for bayesian pretrained network adaptation"]},{"direction_index":4,"label":"D4","name":"Uncertainty calibration and robustness","description":"Assess, calibrate, interpret, improve, or certify predictive uncertainty and robustness of Bayesian deep models under ordinary prediction, covariate shift, corruption, out-of-distribution stress, or adversarial perturbation.","inclusion":"Include work whose primary research objective is uncertainty quality, calibration, benchmarking, decomposition, coverage, OOD behavior, corruption robustness, adversarial robustness, or certification rather than posterior inference as an end in itself.","exclusion":"If uncertainty or robustness is only an evaluation endpoint for a new inference algorithm, assign to that inference direction; if the main contribution is a domain-specific mechanistic model, assign to domain- and knowledge-constrained modeling; if it is primarily proof-driven general theory, assign to theoretical foundations.","nearest_boundary":"D6","representative_topics":["node-based bayesian neural networks for covariate shift via implicit input corruptions","benchmarking bayesian neural network posterior approximation methods for regression uncertainty quantification","corruption-robust bayesian neural networks via training-data-dependent likelihood matching"]},{"direction_index":5,"label":"D5","name":"Domain and knowledge modeling","description":"Adapt Bayesian deep learning to domain-structured, scientific, mechanistic, or application-shaped problems through domain-aware architectures, physical constraints, knowledge constraints, informative priors, posterior regularization, specialized likelihoods, neural operators, kernels, or task-specific Bayesian model design.","inclusion":"Include work whose primary research objective is incorporating external knowledge, domain structure, physical or scientific constraints, learned or informative priors, posterior tempering, or specialized Bayesian architectures for a particular structured problem setting.","exclusion":"If the contribution is a general-purpose inference algorithm only demonstrated on a domain dataset, assign to the relevant inference direction; if it is an uncertainty benchmark without domain-shaped modeling, assign to uncertainty calibration and robustness; if the main objective is cross-task or client adaptation, assign to Bayesian transfer and federated learning.","nearest_boundary":"D4","representative_topics":["variational bayesian deep operator networks for uncertainty-aware parametric differential equation solving","posterior-regularized bayesian neural networks with soft and hard knowledge constraints","hierarchical bayesian graph neural networks for weak lensing convergence estimation in strong time delay cosmography"]},{"direction_index":6,"label":"D6","name":"Theoretical foundations","description":"Develop formal mathematical understanding of Bayesian deep learning, including posterior contraction, asymptotics, width limits, generalization, PAC-Bayes risk, posterior geometry, tempered behavior, function-space versus parameter-space properties, and expressivity of stochastic parameters.","inclusion":"Include work whose primary research objective is a theorem, bound, limit, equivalence, convergence result, or formal characterization of Bayesian neural networks or Bayesian deep learning behavior.","exclusion":"If theory mainly supports a specific new inference algorithm, robustness method, or application system, assign to that method's primary objective; if the result is specifically about a variational objective rather than broader BDL behavior, assign to variational posterior approximation.","nearest_boundary":"D0","representative_topics":["bayesian neural network posterior convergence rates in besov space with adaptive smoothness","pac-bayesian excess risk analysis of epistemic uncertainty in variational inference","theoretical analysis of bayesian neural network robustness to adversarial attacks via posterior geometry"]},{"direction_index":7,"label":"D7","name":"Bayesian transfer and federated learning","description":"Use Bayesian deep learning for learning across heterogeneous tasks, clients, populations, or time-varying data sources, including meta-learning, few-shot learning, federated personalization, hierarchical Bayesian federation, continual updating, and cross-task prior or posterior transfer.","inclusion":"Include work whose primary research objective is transferable probabilistic representation, meta-learned prior or posterior, client personalization, federated Bayesian aggregation, task heterogeneity, or adaptation across data partitions or time.","exclusion":"If federation is only a distributed implementation of a sampler, assign to sampling-based posterior inference; if prior learning is single-task rather than cross-task or cross-client, assign to domain and knowledge modeling; if the central goal is generic computational scalability, assign to scalable efficient representations.","nearest_boundary":"D2","representative_topics":["functional bayesian meta-learning via score matching over data-generating process marginals","federated langevin dynamics with heterogeneity-aware convergence guarantees and drift-correcting control variates","variational bayesian personalized federated learning with sparsity and clustering for non-i.i.d. data"]}],"split":"dev"}
{"episode_id":"F210_T2024-01","field_index":210,"field":"molecular generation","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Geometry-first structure generation","description":"Generate molecular, reaction-state, conformational, shape-matched, or molecular-material structures by directly modeling 3D geometry, coordinates, equivariance, physical validity, or structural arrangements without a specified biomolecular target as the primary conditioning object.","inclusion":"Include work whose main objective is de novo 3D molecular or molecular-material structure generation, coordinate diffusion or flow, conformer or transition-state generation, energy/force-guided geometry, shape-conditioned ligand-based generation, or MOF/framework generation.","exclusion":"If the primary objective is designing a ligand for a specified protein, pocket, sequence, or biological target, assign to target-conditioned ligand design; if the main objective is generic property/Pareto search, assign to property-directed optimization.","nearest_boundary":"D1","representative_topics":["diffusion bridge priors for physically informed 3d molecule generation","diffusion-based 3d de novo molecule generation with interatomic force encoding and diversity control","equivariant energy-guided sde diffusion models for controllable 3d inverse molecular design"]},{"direction_index":1,"label":"D1","name":"Target-conditioned ligand design","description":"Generate candidate ligands, linkers, PROTACs, inhibitors, or drug-like molecules conditioned on a specified biological target, including target identity, sequence, biochemical context, binding affinity, protein structure, or 3D binding pocket.","inclusion":"Include work whose main objective is target-specific molecular design, whether conditioned on explicit pocket geometry, protein structure, protein sequence, target identifiers, docking or affinity signals, or pretrained biochemical representations.","exclusion":"If no specific biological target is central and the contribution is general 3D structure generation, assign to geometry-first structure generation; if target affinity is only a reward example in an RL-method paper, assign to reinforcement-learning molecular design.","nearest_boundary":"D0","representative_topics":["latent space reverse optimization for high-affinity drug-like molecule generation","ligand-and-structure dual-driven molecular generation for target-specific antagonist discovery","warm-starting targeted molecular generation using pretrained biochemical language models"]},{"direction_index":2,"label":"D2","name":"Property-directed optimization","description":"Search, rank, sample, and generate molecules to satisfy molecular properties, desirability functions, novelty, synthesizability, ADMET, scaffold constraints, or multiple competing objectives in chemical space.","inclusion":"Include work whose main objective is molecular optimization using Pareto search, Bayesian optimization, active learning, GFlowNets, evolutionary search, integer programming, MCMC, retrieval guidance, scaffold-constrained search, or expensive-oracle sample-efficiency analysis.","exclusion":"If reinforcement learning policy design is the primary research object, assign to reinforcement-learning molecular design; if target-conditioned binding design is the primary objective, assign to target-conditioned ligand design.","nearest_boundary":"D3","representative_topics":["out-of-distribution score-based diffusion for novel molecule generation guided by property predictors","retrieval-based controllable molecule generation using exemplar fusion and iterative refinement","pareto-based multi-objective optimization for small molecule discovery via generative and bayesian methods"]},{"direction_index":3,"label":"D3","name":"Reinforcement-learning molecular design","description":"Develop, analyze, or improve reinforcement-learning policies for sequential molecular generation or optimization under explicit reward functions.","inclusion":"Include work where RL is the central optimization mechanism or scientific contribution, including policy-gradient, actor-critic, double-loop, multi-agent, graph-based, SMILES-based, augmented-SMILES, target-aware, or property-rewarded molecular generation.","exclusion":"If RL is only an implementation detail inside a target-specific ligand design study, assign to target-conditioned ligand design; if multi-objective or Bayesian search rather than RL is foregrounded, assign to property-directed optimization.","nearest_boundary":"D2","representative_topics":["reinforcement learning for natural flavor molecule discovery balancing synthetic accessibility and structural similarity","sequence-based reinforcement learning for structure-free de novo drug design","double-loop reinforcement learning with smiles augmentation for efficient diverse de novo molecular optimization"]},{"direction_index":4,"label":"D4","name":"Chemical language generation","description":"Generate molecules, reactions, molecular structures, or coordinate-file structures by treating chemistry as strings, tokens, file-format sequences, or autoregressive language-model distributions.","inclusion":"Include work whose main objective is SMILES, SELFIES, XYZ, CIF, PDB, reaction-token, transformer, probabilistic language-model, retrieval-augmented sequence, pretraining, prompting, or multi-conditional sequence generation for molecular design.","exclusion":"If the sequence generator is primarily optimized through RL rewards, assign to reinforcement-learning molecular design; if the central contribution is explicit fragment vocabulary construction or subgraph assembly, assign to fragment-compositional graph generation.","nearest_boundary":"D6","representative_topics":["probabilistic transformer language models for interpretable generative molecular design","unified pretraining framework for reaction representation learning and conditional molecule generation","3d molecular and biomolecular structure generation via language models trained on coordinate file sequences"]},{"direction_index":5,"label":"D5","name":"Latent-variable generative design","description":"Learn and exploit molecular latent spaces for reconstruction, interpolation, controllable sampling, conditional generation, property-guided navigation, or latent-space optimization.","inclusion":"Include work whose main objective is a VAE, autoencoder, fingerprint latent model, mutual-information latent model, hierarchical latent generator, adversarial latent generator, energy-based latent model, quantum-inspired latent generator, or latent-inception-style molecular design framework.","exclusion":"If the main contribution is an external property-optimization loop rather than the latent generator, assign to property-directed optimization; if the model is a flat chemical language model without a latent bottleneck, assign to chemical language generation.","nearest_boundary":"D4","representative_topics":["latent space reverse optimization for high-affinity drug-like molecule generation","pharmacophore-guided variational autoencoder for bioactive molecule generation and lead optimization","mutual information machine autoencoder for property-guided small molecule generation and optimization"]},{"direction_index":6,"label":"D6","name":"Fragment-compositional graph generation","description":"Generate molecules by decomposing, representing, selecting, connecting, diffusing over, or reassembling chemically meaningful fragments, motifs, subgraphs, scaffolds, or learned vocabularies.","inclusion":"Include work whose main objective is fragment extraction, motif mining, subgraph vocabulary learning, decomposition-and-reassembly, connection-aware assembly, dynamic fragment vocabularies, motif inductive bias, fragment-level graph generation, or fragment-based analog design.","exclusion":"If fragments merely serve as tokens in a flat language-model sequence, assign to chemical language generation; if motifs are used mainly to model protein-pocket interactions for target-specific binders, assign to target-conditioned ligand design.","nearest_boundary":"D4","representative_topics":["fragment-based multiscale molecular representations for scalable de novo molecule generation","conditional diffusion models for molecular graph generation using discrete graph structure conditioning","decomposition-and-reassembling molecular graph generation via subgraph mining and reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Generator evaluation and diagnostics","description":"Assess, benchmark, interpret, explain, diagnose, or stress-test molecular generative models rather than primarily proposing a new generator for a design objective.","inclusion":"Include work whose main objective is a benchmark, biological relevance assessment, metric suite, sample-efficiency evaluation, physical or random baseline comparison, reconstruction-failure diagnosis, GNN expressiveness analysis, explainability study, or dataset curation for evaluating generative models.","exclusion":"If evaluation is secondary to a new generator, assign by that generator's primary objective; if oracle-efficiency analysis is embedded in a new optimization method, assign to property-directed optimization.","nearest_boundary":"D2","representative_topics":["biologically-inspired multi-modal benchmarking of molecular generative models for drug discovery","sample-efficient de novo molecule generation benchmarking with chemical desirability constraints","patent-driven generative modeling for property optimization in focused chemical spaces"]}],"split":"dev"}
{"episode_id":"F210_T2024-07","field_index":210,"field":"molecular generation","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Geometry-first structure generation","description":"Generate molecular, reaction-state, conformational, shape-matched, or molecular-material structures by directly modeling 3D geometry, coordinates, equivariance, physical validity, or structural arrangements without a specified biomolecular target as the primary conditioning object.","inclusion":"Include work whose main objective is de novo 3D molecular or molecular-material structure generation, coordinate diffusion or flow, conformer or transition-state generation, energy/force-guided geometry, shape-conditioned ligand-based generation, or MOF/framework generation.","exclusion":"If the primary objective is designing a ligand for a specified protein, pocket, sequence, or biological target, assign to target-conditioned ligand design; if the main objective is generic property/Pareto search, assign to property-directed optimization.","nearest_boundary":"D1","representative_topics":["diffusion bridge priors for physically informed 3d molecule generation","diffusion-based 3d de novo molecule generation with interatomic force encoding and diversity control","equivariant energy-guided sde diffusion models for controllable 3d inverse molecular design"]},{"direction_index":1,"label":"D1","name":"Target-conditioned ligand design","description":"Generate candidate ligands, linkers, PROTACs, inhibitors, or drug-like molecules conditioned on a specified biological target, including target identity, sequence, biochemical context, binding affinity, protein structure, or 3D binding pocket.","inclusion":"Include work whose main objective is target-specific molecular design, whether conditioned on explicit pocket geometry, protein structure, protein sequence, target identifiers, docking or affinity signals, or pretrained biochemical representations.","exclusion":"If no specific biological target is central and the contribution is general 3D structure generation, assign to geometry-first structure generation; if target affinity is only a reward example in an RL-method paper, assign to reinforcement-learning molecular design.","nearest_boundary":"D0","representative_topics":["latent space reverse optimization for high-affinity drug-like molecule generation","ligand-and-structure dual-driven molecular generation for target-specific antagonist discovery","warm-starting targeted molecular generation using pretrained biochemical language models"]},{"direction_index":2,"label":"D2","name":"Property-directed optimization","description":"Search, rank, sample, and generate molecules to satisfy molecular properties, desirability functions, novelty, synthesizability, ADMET, scaffold constraints, or multiple competing objectives in chemical space.","inclusion":"Include work whose main objective is molecular optimization using Pareto search, Bayesian optimization, active learning, GFlowNets, evolutionary search, integer programming, MCMC, retrieval guidance, scaffold-constrained search, or expensive-oracle sample-efficiency analysis.","exclusion":"If reinforcement learning policy design is the primary research object, assign to reinforcement-learning molecular design; if target-conditioned binding design is the primary objective, assign to target-conditioned ligand design.","nearest_boundary":"D3","representative_topics":["out-of-distribution score-based diffusion for novel molecule generation guided by property predictors","retrieval-based controllable molecule generation using exemplar fusion and iterative refinement","pareto-based multi-objective optimization for small molecule discovery via generative and bayesian methods"]},{"direction_index":3,"label":"D3","name":"Reinforcement-learning molecular design","description":"Develop, analyze, or improve reinforcement-learning policies for sequential molecular generation or optimization under explicit reward functions.","inclusion":"Include work where RL is the central optimization mechanism or scientific contribution, including policy-gradient, actor-critic, double-loop, multi-agent, graph-based, SMILES-based, augmented-SMILES, target-aware, or property-rewarded molecular generation.","exclusion":"If RL is only an implementation detail inside a target-specific ligand design study, assign to target-conditioned ligand design; if multi-objective or Bayesian search rather than RL is foregrounded, assign to property-directed optimization.","nearest_boundary":"D2","representative_topics":["reinforcement learning for natural flavor molecule discovery balancing synthetic accessibility and structural similarity","sequence-based reinforcement learning for structure-free de novo drug design","double-loop reinforcement learning with smiles augmentation for efficient diverse de novo molecular optimization"]},{"direction_index":4,"label":"D4","name":"Chemical language generation","description":"Generate molecules, reactions, molecular structures, or coordinate-file structures by treating chemistry as strings, tokens, file-format sequences, or autoregressive language-model distributions.","inclusion":"Include work whose main objective is SMILES, SELFIES, XYZ, CIF, PDB, reaction-token, transformer, probabilistic language-model, retrieval-augmented sequence, pretraining, prompting, or multi-conditional sequence generation for molecular design.","exclusion":"If the sequence generator is primarily optimized through RL rewards, assign to reinforcement-learning molecular design; if the central contribution is explicit fragment vocabulary construction or subgraph assembly, assign to fragment-compositional graph generation.","nearest_boundary":"D6","representative_topics":["probabilistic transformer language models for interpretable generative molecular design","unified pretraining framework for reaction representation learning and conditional molecule generation","3d molecular and biomolecular structure generation via language models trained on coordinate file sequences"]},{"direction_index":5,"label":"D5","name":"Latent-variable generative design","description":"Learn and exploit molecular latent spaces for reconstruction, interpolation, controllable sampling, conditional generation, property-guided navigation, or latent-space optimization.","inclusion":"Include work whose main objective is a VAE, autoencoder, fingerprint latent model, mutual-information latent model, hierarchical latent generator, adversarial latent generator, energy-based latent model, quantum-inspired latent generator, or latent-inception-style molecular design framework.","exclusion":"If the main contribution is an external property-optimization loop rather than the latent generator, assign to property-directed optimization; if the model is a flat chemical language model without a latent bottleneck, assign to chemical language generation.","nearest_boundary":"D4","representative_topics":["latent space reverse optimization for high-affinity drug-like molecule generation","pharmacophore-guided variational autoencoder for bioactive molecule generation and lead optimization","mutual information machine autoencoder for property-guided small molecule generation and optimization"]},{"direction_index":6,"label":"D6","name":"Fragment-compositional graph generation","description":"Generate molecules by decomposing, representing, selecting, connecting, diffusing over, or reassembling chemically meaningful fragments, motifs, subgraphs, scaffolds, or learned vocabularies.","inclusion":"Include work whose main objective is fragment extraction, motif mining, subgraph vocabulary learning, decomposition-and-reassembly, connection-aware assembly, dynamic fragment vocabularies, motif inductive bias, fragment-level graph generation, or fragment-based analog design.","exclusion":"If fragments merely serve as tokens in a flat language-model sequence, assign to chemical language generation; if motifs are used mainly to model protein-pocket interactions for target-specific binders, assign to target-conditioned ligand design.","nearest_boundary":"D4","representative_topics":["fragment-based multiscale molecular representations for scalable de novo molecule generation","conditional diffusion models for molecular graph generation using discrete graph structure conditioning","decomposition-and-reassembling molecular graph generation via subgraph mining and reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Generator evaluation and diagnostics","description":"Assess, benchmark, interpret, explain, diagnose, or stress-test molecular generative models rather than primarily proposing a new generator for a design objective.","inclusion":"Include work whose main objective is a benchmark, biological relevance assessment, metric suite, sample-efficiency evaluation, physical or random baseline comparison, reconstruction-failure diagnosis, GNN expressiveness analysis, explainability study, or dataset curation for evaluating generative models.","exclusion":"If evaluation is secondary to a new generator, assign by that generator's primary objective; if oracle-efficiency analysis is embedded in a new optimization method, assign to property-directed optimization.","nearest_boundary":"D2","representative_topics":["biologically-inspired multi-modal benchmarking of molecular generative models for drug discovery","sample-efficient de novo molecule generation benchmarking with chemical desirability constraints","patent-driven generative modeling for property optimization in focused chemical spaces"]}],"split":"dev"}
{"episode_id":"F210_T2025-01","field_index":210,"field":"molecular generation","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Geometry-first structure generation","description":"Generate molecular, reaction-state, conformational, shape-matched, or molecular-material structures by directly modeling 3D geometry, coordinates, equivariance, physical validity, or structural arrangements without a specified biomolecular target as the primary conditioning object.","inclusion":"Include work whose main objective is de novo 3D molecular or molecular-material structure generation, coordinate diffusion or flow, conformer or transition-state generation, energy/force-guided geometry, shape-conditioned ligand-based generation, or MOF/framework generation.","exclusion":"If the primary objective is designing a ligand for a specified protein, pocket, sequence, or biological target, assign to target-conditioned ligand design; if the main objective is generic property/Pareto search, assign to property-directed optimization.","nearest_boundary":"D1","representative_topics":["diffusion bridge priors for physically informed 3d molecule generation","diffusion-based 3d de novo molecule generation with interatomic force encoding and diversity control","equivariant energy-guided sde diffusion models for controllable 3d inverse molecular design"]},{"direction_index":1,"label":"D1","name":"Target-conditioned ligand design","description":"Generate candidate ligands, linkers, PROTACs, inhibitors, or drug-like molecules conditioned on a specified biological target, including target identity, sequence, biochemical context, binding affinity, protein structure, or 3D binding pocket.","inclusion":"Include work whose main objective is target-specific molecular design, whether conditioned on explicit pocket geometry, protein structure, protein sequence, target identifiers, docking or affinity signals, or pretrained biochemical representations.","exclusion":"If no specific biological target is central and the contribution is general 3D structure generation, assign to geometry-first structure generation; if target affinity is only a reward example in an RL-method paper, assign to reinforcement-learning molecular design.","nearest_boundary":"D0","representative_topics":["latent space reverse optimization for high-affinity drug-like molecule generation","ligand-and-structure dual-driven molecular generation for target-specific antagonist discovery","warm-starting targeted molecular generation using pretrained biochemical language models"]},{"direction_index":2,"label":"D2","name":"Property-directed optimization","description":"Search, rank, sample, and generate molecules to satisfy molecular properties, desirability functions, novelty, synthesizability, ADMET, scaffold constraints, or multiple competing objectives in chemical space.","inclusion":"Include work whose main objective is molecular optimization using Pareto search, Bayesian optimization, active learning, GFlowNets, evolutionary search, integer programming, MCMC, retrieval guidance, scaffold-constrained search, or expensive-oracle sample-efficiency analysis.","exclusion":"If reinforcement learning policy design is the primary research object, assign to reinforcement-learning molecular design; if target-conditioned binding design is the primary objective, assign to target-conditioned ligand design.","nearest_boundary":"D3","representative_topics":["out-of-distribution score-based diffusion for novel molecule generation guided by property predictors","retrieval-based controllable molecule generation using exemplar fusion and iterative refinement","pareto-based multi-objective optimization for small molecule discovery via generative and bayesian methods"]},{"direction_index":3,"label":"D3","name":"Reinforcement-learning molecular design","description":"Develop, analyze, or improve reinforcement-learning policies for sequential molecular generation or optimization under explicit reward functions.","inclusion":"Include work where RL is the central optimization mechanism or scientific contribution, including policy-gradient, actor-critic, double-loop, multi-agent, graph-based, SMILES-based, augmented-SMILES, target-aware, or property-rewarded molecular generation.","exclusion":"If RL is only an implementation detail inside a target-specific ligand design study, assign to target-conditioned ligand design; if multi-objective or Bayesian search rather than RL is foregrounded, assign to property-directed optimization.","nearest_boundary":"D2","representative_topics":["reinforcement learning for natural flavor molecule discovery balancing synthetic accessibility and structural similarity","sequence-based reinforcement learning for structure-free de novo drug design","double-loop reinforcement learning with smiles augmentation for efficient diverse de novo molecular optimization"]},{"direction_index":4,"label":"D4","name":"Chemical language generation","description":"Generate molecules, reactions, molecular structures, or coordinate-file structures by treating chemistry as strings, tokens, file-format sequences, or autoregressive language-model distributions.","inclusion":"Include work whose main objective is SMILES, SELFIES, XYZ, CIF, PDB, reaction-token, transformer, probabilistic language-model, retrieval-augmented sequence, pretraining, prompting, or multi-conditional sequence generation for molecular design.","exclusion":"If the sequence generator is primarily optimized through RL rewards, assign to reinforcement-learning molecular design; if the central contribution is explicit fragment vocabulary construction or subgraph assembly, assign to fragment-compositional graph generation.","nearest_boundary":"D6","representative_topics":["probabilistic transformer language models for interpretable generative molecular design","unified pretraining framework for reaction representation learning and conditional molecule generation","3d molecular and biomolecular structure generation via language models trained on coordinate file sequences"]},{"direction_index":5,"label":"D5","name":"Latent-variable generative design","description":"Learn and exploit molecular latent spaces for reconstruction, interpolation, controllable sampling, conditional generation, property-guided navigation, or latent-space optimization.","inclusion":"Include work whose main objective is a VAE, autoencoder, fingerprint latent model, mutual-information latent model, hierarchical latent generator, adversarial latent generator, energy-based latent model, quantum-inspired latent generator, or latent-inception-style molecular design framework.","exclusion":"If the main contribution is an external property-optimization loop rather than the latent generator, assign to property-directed optimization; if the model is a flat chemical language model without a latent bottleneck, assign to chemical language generation.","nearest_boundary":"D4","representative_topics":["latent space reverse optimization for high-affinity drug-like molecule generation","pharmacophore-guided variational autoencoder for bioactive molecule generation and lead optimization","mutual information machine autoencoder for property-guided small molecule generation and optimization"]},{"direction_index":6,"label":"D6","name":"Fragment-compositional graph generation","description":"Generate molecules by decomposing, representing, selecting, connecting, diffusing over, or reassembling chemically meaningful fragments, motifs, subgraphs, scaffolds, or learned vocabularies.","inclusion":"Include work whose main objective is fragment extraction, motif mining, subgraph vocabulary learning, decomposition-and-reassembly, connection-aware assembly, dynamic fragment vocabularies, motif inductive bias, fragment-level graph generation, or fragment-based analog design.","exclusion":"If fragments merely serve as tokens in a flat language-model sequence, assign to chemical language generation; if motifs are used mainly to model protein-pocket interactions for target-specific binders, assign to target-conditioned ligand design.","nearest_boundary":"D4","representative_topics":["fragment-based multiscale molecular representations for scalable de novo molecule generation","conditional diffusion models for molecular graph generation using discrete graph structure conditioning","decomposition-and-reassembling molecular graph generation via subgraph mining and reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Generator evaluation and diagnostics","description":"Assess, benchmark, interpret, explain, diagnose, or stress-test molecular generative models rather than primarily proposing a new generator for a design objective.","inclusion":"Include work whose main objective is a benchmark, biological relevance assessment, metric suite, sample-efficiency evaluation, physical or random baseline comparison, reconstruction-failure diagnosis, GNN expressiveness analysis, explainability study, or dataset curation for evaluating generative models.","exclusion":"If evaluation is secondary to a new generator, assign by that generator's primary objective; if oracle-efficiency analysis is embedded in a new optimization method, assign to property-directed optimization.","nearest_boundary":"D2","representative_topics":["biologically-inspired multi-modal benchmarking of molecular generative models for drug discovery","sample-efficient de novo molecule generation benchmarking with chemical desirability constraints","patent-driven generative modeling for property optimization in focused chemical spaces"]}],"split":"dev"}
{"episode_id":"F210_T2025-07","field_index":210,"field":"molecular generation","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Geometry-first structure generation","description":"Generate molecular, reaction-state, conformational, shape-matched, or molecular-material structures by directly modeling 3D geometry, coordinates, equivariance, physical validity, or structural arrangements without a specified biomolecular target as the primary conditioning object.","inclusion":"Include work whose main objective is de novo 3D molecular or molecular-material structure generation, coordinate diffusion or flow, conformer or transition-state generation, energy/force-guided geometry, shape-conditioned ligand-based generation, or MOF/framework generation.","exclusion":"If the primary objective is designing a ligand for a specified protein, pocket, sequence, or biological target, assign to target-conditioned ligand design; if the main objective is generic property/Pareto search, assign to property-directed optimization.","nearest_boundary":"D1","representative_topics":["diffusion bridge priors for physically informed 3d molecule generation","diffusion-based 3d de novo molecule generation with interatomic force encoding and diversity control","equivariant energy-guided sde diffusion models for controllable 3d inverse molecular design"]},{"direction_index":1,"label":"D1","name":"Target-conditioned ligand design","description":"Generate candidate ligands, linkers, PROTACs, inhibitors, or drug-like molecules conditioned on a specified biological target, including target identity, sequence, biochemical context, binding affinity, protein structure, or 3D binding pocket.","inclusion":"Include work whose main objective is target-specific molecular design, whether conditioned on explicit pocket geometry, protein structure, protein sequence, target identifiers, docking or affinity signals, or pretrained biochemical representations.","exclusion":"If no specific biological target is central and the contribution is general 3D structure generation, assign to geometry-first structure generation; if target affinity is only a reward example in an RL-method paper, assign to reinforcement-learning molecular design.","nearest_boundary":"D0","representative_topics":["latent space reverse optimization for high-affinity drug-like molecule generation","ligand-and-structure dual-driven molecular generation for target-specific antagonist discovery","warm-starting targeted molecular generation using pretrained biochemical language models"]},{"direction_index":2,"label":"D2","name":"Property-directed optimization","description":"Search, rank, sample, and generate molecules to satisfy molecular properties, desirability functions, novelty, synthesizability, ADMET, scaffold constraints, or multiple competing objectives in chemical space.","inclusion":"Include work whose main objective is molecular optimization using Pareto search, Bayesian optimization, active learning, GFlowNets, evolutionary search, integer programming, MCMC, retrieval guidance, scaffold-constrained search, or expensive-oracle sample-efficiency analysis.","exclusion":"If reinforcement learning policy design is the primary research object, assign to reinforcement-learning molecular design; if target-conditioned binding design is the primary objective, assign to target-conditioned ligand design.","nearest_boundary":"D3","representative_topics":["out-of-distribution score-based diffusion for novel molecule generation guided by property predictors","retrieval-based controllable molecule generation using exemplar fusion and iterative refinement","pareto-based multi-objective optimization for small molecule discovery via generative and bayesian methods"]},{"direction_index":3,"label":"D3","name":"Reinforcement-learning molecular design","description":"Develop, analyze, or improve reinforcement-learning policies for sequential molecular generation or optimization under explicit reward functions.","inclusion":"Include work where RL is the central optimization mechanism or scientific contribution, including policy-gradient, actor-critic, double-loop, multi-agent, graph-based, SMILES-based, augmented-SMILES, target-aware, or property-rewarded molecular generation.","exclusion":"If RL is only an implementation detail inside a target-specific ligand design study, assign to target-conditioned ligand design; if multi-objective or Bayesian search rather than RL is foregrounded, assign to property-directed optimization.","nearest_boundary":"D2","representative_topics":["reinforcement learning for natural flavor molecule discovery balancing synthetic accessibility and structural similarity","sequence-based reinforcement learning for structure-free de novo drug design","double-loop reinforcement learning with smiles augmentation for efficient diverse de novo molecular optimization"]},{"direction_index":4,"label":"D4","name":"Chemical language generation","description":"Generate molecules, reactions, molecular structures, or coordinate-file structures by treating chemistry as strings, tokens, file-format sequences, or autoregressive language-model distributions.","inclusion":"Include work whose main objective is SMILES, SELFIES, XYZ, CIF, PDB, reaction-token, transformer, probabilistic language-model, retrieval-augmented sequence, pretraining, prompting, or multi-conditional sequence generation for molecular design.","exclusion":"If the sequence generator is primarily optimized through RL rewards, assign to reinforcement-learning molecular design; if the central contribution is explicit fragment vocabulary construction or subgraph assembly, assign to fragment-compositional graph generation.","nearest_boundary":"D6","representative_topics":["probabilistic transformer language models for interpretable generative molecular design","unified pretraining framework for reaction representation learning and conditional molecule generation","3d molecular and biomolecular structure generation via language models trained on coordinate file sequences"]},{"direction_index":5,"label":"D5","name":"Latent-variable generative design","description":"Learn and exploit molecular latent spaces for reconstruction, interpolation, controllable sampling, conditional generation, property-guided navigation, or latent-space optimization.","inclusion":"Include work whose main objective is a VAE, autoencoder, fingerprint latent model, mutual-information latent model, hierarchical latent generator, adversarial latent generator, energy-based latent model, quantum-inspired latent generator, or latent-inception-style molecular design framework.","exclusion":"If the main contribution is an external property-optimization loop rather than the latent generator, assign to property-directed optimization; if the model is a flat chemical language model without a latent bottleneck, assign to chemical language generation.","nearest_boundary":"D4","representative_topics":["latent space reverse optimization for high-affinity drug-like molecule generation","pharmacophore-guided variational autoencoder for bioactive molecule generation and lead optimization","mutual information machine autoencoder for property-guided small molecule generation and optimization"]},{"direction_index":6,"label":"D6","name":"Fragment-compositional graph generation","description":"Generate molecules by decomposing, representing, selecting, connecting, diffusing over, or reassembling chemically meaningful fragments, motifs, subgraphs, scaffolds, or learned vocabularies.","inclusion":"Include work whose main objective is fragment extraction, motif mining, subgraph vocabulary learning, decomposition-and-reassembly, connection-aware assembly, dynamic fragment vocabularies, motif inductive bias, fragment-level graph generation, or fragment-based analog design.","exclusion":"If fragments merely serve as tokens in a flat language-model sequence, assign to chemical language generation; if motifs are used mainly to model protein-pocket interactions for target-specific binders, assign to target-conditioned ligand design.","nearest_boundary":"D4","representative_topics":["fragment-based multiscale molecular representations for scalable de novo molecule generation","conditional diffusion models for molecular graph generation using discrete graph structure conditioning","decomposition-and-reassembling molecular graph generation via subgraph mining and reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Generator evaluation and diagnostics","description":"Assess, benchmark, interpret, explain, diagnose, or stress-test molecular generative models rather than primarily proposing a new generator for a design objective.","inclusion":"Include work whose main objective is a benchmark, biological relevance assessment, metric suite, sample-efficiency evaluation, physical or random baseline comparison, reconstruction-failure diagnosis, GNN expressiveness analysis, explainability study, or dataset curation for evaluating generative models.","exclusion":"If evaluation is secondary to a new generator, assign by that generator's primary objective; if oracle-efficiency analysis is embedded in a new optimization method, assign to property-directed optimization.","nearest_boundary":"D2","representative_topics":["biologically-inspired multi-modal benchmarking of molecular generative models for drug discovery","sample-efficient de novo molecule generation benchmarking with chemical desirability constraints","patent-driven generative modeling for property optimization in focused chemical spaces"]}],"split":"dev"}
{"episode_id":"F210_T2026-01","field_index":210,"field":"molecular generation","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Geometry-first structure generation","description":"Generate molecular, reaction-state, conformational, shape-matched, or molecular-material structures by directly modeling 3D geometry, coordinates, equivariance, physical validity, or structural arrangements without a specified biomolecular target as the primary conditioning object.","inclusion":"Include work whose main objective is de novo 3D molecular or molecular-material structure generation, coordinate diffusion or flow, conformer or transition-state generation, energy/force-guided geometry, shape-conditioned ligand-based generation, or MOF/framework generation.","exclusion":"If the primary objective is designing a ligand for a specified protein, pocket, sequence, or biological target, assign to target-conditioned ligand design; if the main objective is generic property/Pareto search, assign to property-directed optimization.","nearest_boundary":"D1","representative_topics":["diffusion bridge priors for physically informed 3d molecule generation","diffusion-based 3d de novo molecule generation with interatomic force encoding and diversity control","equivariant energy-guided sde diffusion models for controllable 3d inverse molecular design"]},{"direction_index":1,"label":"D1","name":"Target-conditioned ligand design","description":"Generate candidate ligands, linkers, PROTACs, inhibitors, or drug-like molecules conditioned on a specified biological target, including target identity, sequence, biochemical context, binding affinity, protein structure, or 3D binding pocket.","inclusion":"Include work whose main objective is target-specific molecular design, whether conditioned on explicit pocket geometry, protein structure, protein sequence, target identifiers, docking or affinity signals, or pretrained biochemical representations.","exclusion":"If no specific biological target is central and the contribution is general 3D structure generation, assign to geometry-first structure generation; if target affinity is only a reward example in an RL-method paper, assign to reinforcement-learning molecular design.","nearest_boundary":"D0","representative_topics":["latent space reverse optimization for high-affinity drug-like molecule generation","ligand-and-structure dual-driven molecular generation for target-specific antagonist discovery","warm-starting targeted molecular generation using pretrained biochemical language models"]},{"direction_index":2,"label":"D2","name":"Property-directed optimization","description":"Search, rank, sample, and generate molecules to satisfy molecular properties, desirability functions, novelty, synthesizability, ADMET, scaffold constraints, or multiple competing objectives in chemical space.","inclusion":"Include work whose main objective is molecular optimization using Pareto search, Bayesian optimization, active learning, GFlowNets, evolutionary search, integer programming, MCMC, retrieval guidance, scaffold-constrained search, or expensive-oracle sample-efficiency analysis.","exclusion":"If reinforcement learning policy design is the primary research object, assign to reinforcement-learning molecular design; if target-conditioned binding design is the primary objective, assign to target-conditioned ligand design.","nearest_boundary":"D3","representative_topics":["out-of-distribution score-based diffusion for novel molecule generation guided by property predictors","retrieval-based controllable molecule generation using exemplar fusion and iterative refinement","pareto-based multi-objective optimization for small molecule discovery via generative and bayesian methods"]},{"direction_index":3,"label":"D3","name":"Reinforcement-learning molecular design","description":"Develop, analyze, or improve reinforcement-learning policies for sequential molecular generation or optimization under explicit reward functions.","inclusion":"Include work where RL is the central optimization mechanism or scientific contribution, including policy-gradient, actor-critic, double-loop, multi-agent, graph-based, SMILES-based, augmented-SMILES, target-aware, or property-rewarded molecular generation.","exclusion":"If RL is only an implementation detail inside a target-specific ligand design study, assign to target-conditioned ligand design; if multi-objective or Bayesian search rather than RL is foregrounded, assign to property-directed optimization.","nearest_boundary":"D2","representative_topics":["reinforcement learning for natural flavor molecule discovery balancing synthetic accessibility and structural similarity","sequence-based reinforcement learning for structure-free de novo drug design","double-loop reinforcement learning with smiles augmentation for efficient diverse de novo molecular optimization"]},{"direction_index":4,"label":"D4","name":"Chemical language generation","description":"Generate molecules, reactions, molecular structures, or coordinate-file structures by treating chemistry as strings, tokens, file-format sequences, or autoregressive language-model distributions.","inclusion":"Include work whose main objective is SMILES, SELFIES, XYZ, CIF, PDB, reaction-token, transformer, probabilistic language-model, retrieval-augmented sequence, pretraining, prompting, or multi-conditional sequence generation for molecular design.","exclusion":"If the sequence generator is primarily optimized through RL rewards, assign to reinforcement-learning molecular design; if the central contribution is explicit fragment vocabulary construction or subgraph assembly, assign to fragment-compositional graph generation.","nearest_boundary":"D6","representative_topics":["probabilistic transformer language models for interpretable generative molecular design","unified pretraining framework for reaction representation learning and conditional molecule generation","3d molecular and biomolecular structure generation via language models trained on coordinate file sequences"]},{"direction_index":5,"label":"D5","name":"Latent-variable generative design","description":"Learn and exploit molecular latent spaces for reconstruction, interpolation, controllable sampling, conditional generation, property-guided navigation, or latent-space optimization.","inclusion":"Include work whose main objective is a VAE, autoencoder, fingerprint latent model, mutual-information latent model, hierarchical latent generator, adversarial latent generator, energy-based latent model, quantum-inspired latent generator, or latent-inception-style molecular design framework.","exclusion":"If the main contribution is an external property-optimization loop rather than the latent generator, assign to property-directed optimization; if the model is a flat chemical language model without a latent bottleneck, assign to chemical language generation.","nearest_boundary":"D4","representative_topics":["latent space reverse optimization for high-affinity drug-like molecule generation","pharmacophore-guided variational autoencoder for bioactive molecule generation and lead optimization","mutual information machine autoencoder for property-guided small molecule generation and optimization"]},{"direction_index":6,"label":"D6","name":"Fragment-compositional graph generation","description":"Generate molecules by decomposing, representing, selecting, connecting, diffusing over, or reassembling chemically meaningful fragments, motifs, subgraphs, scaffolds, or learned vocabularies.","inclusion":"Include work whose main objective is fragment extraction, motif mining, subgraph vocabulary learning, decomposition-and-reassembly, connection-aware assembly, dynamic fragment vocabularies, motif inductive bias, fragment-level graph generation, or fragment-based analog design.","exclusion":"If fragments merely serve as tokens in a flat language-model sequence, assign to chemical language generation; if motifs are used mainly to model protein-pocket interactions for target-specific binders, assign to target-conditioned ligand design.","nearest_boundary":"D4","representative_topics":["fragment-based multiscale molecular representations for scalable de novo molecule generation","conditional diffusion models for molecular graph generation using discrete graph structure conditioning","decomposition-and-reassembling molecular graph generation via subgraph mining and reinforcement learning"]},{"direction_index":7,"label":"D7","name":"Generator evaluation and diagnostics","description":"Assess, benchmark, interpret, explain, diagnose, or stress-test molecular generative models rather than primarily proposing a new generator for a design objective.","inclusion":"Include work whose main objective is a benchmark, biological relevance assessment, metric suite, sample-efficiency evaluation, physical or random baseline comparison, reconstruction-failure diagnosis, GNN expressiveness analysis, explainability study, or dataset curation for evaluating generative models.","exclusion":"If evaluation is secondary to a new generator, assign by that generator's primary objective; if oracle-efficiency analysis is embedded in a new optimization method, assign to property-directed optimization.","nearest_boundary":"D2","representative_topics":["biologically-inspired multi-modal benchmarking of molecular generative models for drug discovery","sample-efficient de novo molecule generation benchmarking with chemical desirability constraints","patent-driven generative modeling for property optimization in focused chemical spaces"]}],"split":"dev"}
{"episode_id":"F226_T2024-01","field_index":226,"field":"neuroimaging analysis","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Subject-level clinical and normative phenotyping","description":"Develop neuroimaging biomarkers and predictive models for individual diagnosis, prognosis, demographic phenotyping, cognitive or clinical outcome prediction, disease staging, and deviation from healthy population norms.","inclusion":"Include work whose primary output is a subject-level label, risk score, stage, conversion prediction, survival or outcome estimate, cognitive score, brain-age or normative-deviation measure, sex or demographic phenotype, or clinically interpretable biomarker for AD, non-AD disorders, psychiatric conditions, tumors, TBI, addiction, ASD, schizophrenia, or aging-related variation.","exclusion":"If the main objective is voxelwise delineation of lesions or tumors, assign to Anatomical segmentation; if the main objective is discovering network organization rather than predicting a subject endpoint, assign to Connectome organization; if the main objective is reusable pretraining or harmonization independent of one endpoint, assign to Reusable methods and evaluation infrastructure.","nearest_boundary":"D4 Connectome organization and dynamics","representative_topics":["graph-based ensemble u-net framework for interpretable alzheimer's disease detection and prognosis from structural mri","multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","self-supervised graph state space models for single-subject psychiatric disorder diagnosis from resting-state fmri"]},{"direction_index":1,"label":"D1","name":"Fetal and neonatal neurodevelopmental imaging","description":"Analyze fetal, perinatal, and neonatal brain imaging with methods tailored to early-life anatomy, motion, low signal, thick slices, atlas construction, super-resolution, segmentation, surface development, quality control, and normative developmental baselines.","inclusion":"Include work where fetal, perinatal, or neonatal imaging is central to the scientific problem, including early-life MRI or fMRI segmentation, motion or quality control, super-resolution, atlas construction, neonatal cortical maturation, low-field or thick-slice adaptation, and perinatal structural or functional connectivity baselines.","exclusion":"Adult or general pediatric segmentation, reconstruction, registration, or brain aging studies belong to the corresponding non-developmental program unless fetal or neonatal acquisition and anatomy define the problem.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["surface-based generative modeling of neonatal cortical maturation across gestational age","continuous longitudinal fetal brain atlas via implicit neural representation for temporal consistency","self-supervised single-volume isotropic superresolution for fetal brain mri reconstruction"]},{"direction_index":2,"label":"D2","name":"Anatomical segmentation and lesion delineation","description":"Automate voxelwise, regionwise, or object-level delineation of brain tissues, anatomical structures, tumors, white matter hyperintensities, lesions, skull-stripping targets, and related volumetric targets.","inclusion":"Include work whose primary contribution is segmentation, detection, or delineation of brain anatomy, lesions, tumors, white matter abnormalities, hippocampus, fiber bundles as segmentation targets, or longitudinally consistent anatomical labels, including robustness to missing modalities or acquisition perturbations.","exclusion":"If the main target is cortical meshes, surfaces, tracts as geometric objects, registration, or morphometric shape analysis, assign to Geometric brain structure analysis; if segmentation is only a preprocessing step for clinical prediction, assign to Subject-level clinical and normative phenotyping.","nearest_boundary":"D3 Geometric brain structure analysis","representative_topics":["multimodal mri glioma survival prediction via cross-scale nonlocal feature fusion","longitudinal whole-brain and white matter lesion segmentation with temporal consistency constraints","3d flair-weighted white matter hyperintensity segmentation using deep learning on multi-site neuroimaging data"]},{"direction_index":3,"label":"D3","name":"Geometric brain structure analysis","description":"Reconstruct, register, parcellate, quantify, and model brain geometry, cortical surfaces, white matter tracts, diffusion structures, meshes, point clouds, deformation fields, longitudinal morphometry, and shape-based anatomical variation.","inclusion":"Include work operating primarily on anatomical objects or spatial transformations, including cortical surface segmentation or registration, mesh analysis, sulcal labeling, tractography and fiber-level representations, tract parcellation, diffusion signal geometry, morphometric population analysis, deformable registration, diffeomorphic mapping, and longitudinal deformation or volumetric change modeling.","exclusion":"Voxelwise tissue, tumor, or lesion segmentation belongs to Anatomical segmentation; MRI signal reconstruction or contrast synthesis belongs to MRI reconstruction and synthesis; connectome graphs whose nodes are ROIs belong to Connectome organization and dynamics.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["equivariant graph neural networks for robust cortical mesh segmentation under misalignment","fiber-level whole brain tractography classification using spectral embedding and vision transformer attention","point cloud geometric deep learning for neuropsychological score prediction and tract region localization"]},{"direction_index":4,"label":"D4","name":"Connectome organization and dynamics","description":"Infer, characterize, label, generate, or model functional, structural, temporal, effective, manifold-valued, or graph-based brain connectivity to understand network organization and dynamics.","inclusion":"Include work whose main objective is brain network structure or dynamics: graph neural network methodology for connectomes, dynamic graph structure learning, effective or temporal-lag connectivity, directed communication learning, network labeling, universal connectome template estimation, connectivity generation, high-order graph relationships, Riemannian or SPD-manifold connectivity modeling, and temporal models of fMRI dependencies when network dynamics are central.","exclusion":"Connectome-based studies primarily optimized for a disease, demographic, or cognitive subject-level endpoint belong to Subject-level clinical and normative phenotyping; general-purpose graph pretraining or benchmark construction belongs to Reusable methods and evaluation infrastructure; stimulus decoding from fMRI belongs to Brain encoding and decoding.","nearest_boundary":"D0 Subject-level clinical and normative phenotyping","representative_topics":["deep learning classification of resting state fmri brain networks using cloud-based reproducible pipelines","multimodal multiresolution brain graph integration for universal connectome template estimation","sparse directed dynamic communication learning in brain networks via coupled sequential vae"]},{"direction_index":5,"label":"D5","name":"Brain encoding and decoding","description":"Map between brain activity and external perceptual, semantic, linguistic, speech, visual, or cognitive representations, including encoding models, neural decoding, stimulus reconstruction, retrieval, and brain-AI representational alignment.","inclusion":"Include work whose primary objective is predicting neural responses from stimuli, reconstructing or retrieving images, text, speech, music, captions, or semantic content from fMRI or other brain recordings, aligning AI representations with brain activity, or pretraining explicitly for brain decoding and stimulus-related interpretation.","exclusion":"Disease or trait classification from fMRI belongs to Subject-level clinical and normative phenotyping; general-purpose neuroimaging representation learning not centered on stimulus content belongs to Reusable methods and evaluation infrastructure; MRI-to-image synthesis as an acquisition or reconstruction task belongs to MRI reconstruction and synthesis.","nearest_boundary":"D7 Reusable methods and evaluation infrastructure","representative_topics":["self-supervised speech representations aligned with cortical hierarchy and behavioral responses","cross-modal fmri surrogate marker prediction from fnirs using neural data augmentation","self-taught pre-training for brain decoding via convolutional autoencoder initialization"]},{"direction_index":6,"label":"D6","name":"MRI reconstruction and synthesis","description":"Recover, synthesize, complete, translate, or quantitatively estimate MRI-derived images, contrasts, signals, and parameter maps from undersampled, incomplete, noisy, cross-modal, longitudinal, or physics-constrained measurements.","inclusion":"Include work whose primary contribution is MRI inverse reconstruction, accelerated acquisition, quantitative susceptibility, perfusion or fingerprinting parameter mapping, modality synthesis, longitudinal image synthesis, healthy-anatomy reconstruction for anomaly detection, missing modality completion, or generative image translation.","exclusion":"Downstream segmentation of reconstructed images belongs to Anatomical segmentation; clinical classification using synthesized or reconstructed images belongs to Subject-level clinical and normative phenotyping unless reconstruction or synthesis is the main contribution; fMRI-to-natural-image stimulus reconstruction belongs to Brain encoding and decoding.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["deep learning-based reduction of post-labeling delay acquisitions for cerebral blood flow and arterial transit time map estimation","generative adversarial networks for disease-specific brain imaging synthesis and analysis","temporally adjustable longitudinal flair mri synthesis for multiple sclerosis disease progression monitoring"]},{"direction_index":7,"label":"D7","name":"Reusable methods and evaluation infrastructure","description":"Build general-purpose neuroimaging representations, multimodal fusion methods, domain adaptation and harmonization approaches, benchmarking resources, quality and reliability infrastructure, and methodological foundations for interpretable and generalizable neuroimaging machine learning.","inclusion":"Include work whose primary objective is reusable representation learning, self-supervised or contrastive pretraining across tasks, multimodal fusion as a general method, cross-modal surrogate mapping, domain adaptation, scanner or site harmonization, acquisition-shift mitigation, quality control, benchmark construction, standardized evaluation, interpretability surveys, theoretical limits, or broad methodological characterization not tied to a single endpoint.","exclusion":"Disease-specific fusion, robustness, or interpretability belongs to the relevant clinical program when the endpoint is dominant; reconstruction or synthesis belongs to MRI reconstruction and synthesis; stimulus-focused pretraining belongs to Brain encoding and decoding.","nearest_boundary":"D5 Brain encoding and decoding","representative_topics":["benchmarking brain network graph datasets for neurodegenerative and developmental condition prediction","multi-task benchmark for multi-scale brain structure representation learning from microtomography","source-free domain adaptation for scanner harmonization in distributed neuroimaging"]}],"split":"dev"}
{"episode_id":"F226_T2024-07","field_index":226,"field":"neuroimaging analysis","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Subject-level clinical and normative phenotyping","description":"Develop neuroimaging biomarkers and predictive models for individual diagnosis, prognosis, demographic phenotyping, cognitive or clinical outcome prediction, disease staging, and deviation from healthy population norms.","inclusion":"Include work whose primary output is a subject-level label, risk score, stage, conversion prediction, survival or outcome estimate, cognitive score, brain-age or normative-deviation measure, sex or demographic phenotype, or clinically interpretable biomarker for AD, non-AD disorders, psychiatric conditions, tumors, TBI, addiction, ASD, schizophrenia, or aging-related variation.","exclusion":"If the main objective is voxelwise delineation of lesions or tumors, assign to Anatomical segmentation; if the main objective is discovering network organization rather than predicting a subject endpoint, assign to Connectome organization; if the main objective is reusable pretraining or harmonization independent of one endpoint, assign to Reusable methods and evaluation infrastructure.","nearest_boundary":"D4 Connectome organization and dynamics","representative_topics":["graph-based ensemble u-net framework for interpretable alzheimer's disease detection and prognosis from structural mri","multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","self-supervised graph state space models for single-subject psychiatric disorder diagnosis from resting-state fmri"]},{"direction_index":1,"label":"D1","name":"Fetal and neonatal neurodevelopmental imaging","description":"Analyze fetal, perinatal, and neonatal brain imaging with methods tailored to early-life anatomy, motion, low signal, thick slices, atlas construction, super-resolution, segmentation, surface development, quality control, and normative developmental baselines.","inclusion":"Include work where fetal, perinatal, or neonatal imaging is central to the scientific problem, including early-life MRI or fMRI segmentation, motion or quality control, super-resolution, atlas construction, neonatal cortical maturation, low-field or thick-slice adaptation, and perinatal structural or functional connectivity baselines.","exclusion":"Adult or general pediatric segmentation, reconstruction, registration, or brain aging studies belong to the corresponding non-developmental program unless fetal or neonatal acquisition and anatomy define the problem.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["surface-based generative modeling of neonatal cortical maturation across gestational age","continuous longitudinal fetal brain atlas via implicit neural representation for temporal consistency","self-supervised single-volume isotropic superresolution for fetal brain mri reconstruction"]},{"direction_index":2,"label":"D2","name":"Anatomical segmentation and lesion delineation","description":"Automate voxelwise, regionwise, or object-level delineation of brain tissues, anatomical structures, tumors, white matter hyperintensities, lesions, skull-stripping targets, and related volumetric targets.","inclusion":"Include work whose primary contribution is segmentation, detection, or delineation of brain anatomy, lesions, tumors, white matter abnormalities, hippocampus, fiber bundles as segmentation targets, or longitudinally consistent anatomical labels, including robustness to missing modalities or acquisition perturbations.","exclusion":"If the main target is cortical meshes, surfaces, tracts as geometric objects, registration, or morphometric shape analysis, assign to Geometric brain structure analysis; if segmentation is only a preprocessing step for clinical prediction, assign to Subject-level clinical and normative phenotyping.","nearest_boundary":"D3 Geometric brain structure analysis","representative_topics":["multimodal mri glioma survival prediction via cross-scale nonlocal feature fusion","longitudinal whole-brain and white matter lesion segmentation with temporal consistency constraints","3d flair-weighted white matter hyperintensity segmentation using deep learning on multi-site neuroimaging data"]},{"direction_index":3,"label":"D3","name":"Geometric brain structure analysis","description":"Reconstruct, register, parcellate, quantify, and model brain geometry, cortical surfaces, white matter tracts, diffusion structures, meshes, point clouds, deformation fields, longitudinal morphometry, and shape-based anatomical variation.","inclusion":"Include work operating primarily on anatomical objects or spatial transformations, including cortical surface segmentation or registration, mesh analysis, sulcal labeling, tractography and fiber-level representations, tract parcellation, diffusion signal geometry, morphometric population analysis, deformable registration, diffeomorphic mapping, and longitudinal deformation or volumetric change modeling.","exclusion":"Voxelwise tissue, tumor, or lesion segmentation belongs to Anatomical segmentation; MRI signal reconstruction or contrast synthesis belongs to MRI reconstruction and synthesis; connectome graphs whose nodes are ROIs belong to Connectome organization and dynamics.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["equivariant graph neural networks for robust cortical mesh segmentation under misalignment","fiber-level whole brain tractography classification using spectral embedding and vision transformer attention","point cloud geometric deep learning for neuropsychological score prediction and tract region localization"]},{"direction_index":4,"label":"D4","name":"Connectome organization and dynamics","description":"Infer, characterize, label, generate, or model functional, structural, temporal, effective, manifold-valued, or graph-based brain connectivity to understand network organization and dynamics.","inclusion":"Include work whose main objective is brain network structure or dynamics: graph neural network methodology for connectomes, dynamic graph structure learning, effective or temporal-lag connectivity, directed communication learning, network labeling, universal connectome template estimation, connectivity generation, high-order graph relationships, Riemannian or SPD-manifold connectivity modeling, and temporal models of fMRI dependencies when network dynamics are central.","exclusion":"Connectome-based studies primarily optimized for a disease, demographic, or cognitive subject-level endpoint belong to Subject-level clinical and normative phenotyping; general-purpose graph pretraining or benchmark construction belongs to Reusable methods and evaluation infrastructure; stimulus decoding from fMRI belongs to Brain encoding and decoding.","nearest_boundary":"D0 Subject-level clinical and normative phenotyping","representative_topics":["deep learning classification of resting state fmri brain networks using cloud-based reproducible pipelines","multimodal multiresolution brain graph integration for universal connectome template estimation","sparse directed dynamic communication learning in brain networks via coupled sequential vae"]},{"direction_index":5,"label":"D5","name":"Brain encoding and decoding","description":"Map between brain activity and external perceptual, semantic, linguistic, speech, visual, or cognitive representations, including encoding models, neural decoding, stimulus reconstruction, retrieval, and brain-AI representational alignment.","inclusion":"Include work whose primary objective is predicting neural responses from stimuli, reconstructing or retrieving images, text, speech, music, captions, or semantic content from fMRI or other brain recordings, aligning AI representations with brain activity, or pretraining explicitly for brain decoding and stimulus-related interpretation.","exclusion":"Disease or trait classification from fMRI belongs to Subject-level clinical and normative phenotyping; general-purpose neuroimaging representation learning not centered on stimulus content belongs to Reusable methods and evaluation infrastructure; MRI-to-image synthesis as an acquisition or reconstruction task belongs to MRI reconstruction and synthesis.","nearest_boundary":"D7 Reusable methods and evaluation infrastructure","representative_topics":["self-supervised speech representations aligned with cortical hierarchy and behavioral responses","cross-modal fmri surrogate marker prediction from fnirs using neural data augmentation","self-taught pre-training for brain decoding via convolutional autoencoder initialization"]},{"direction_index":6,"label":"D6","name":"MRI reconstruction and synthesis","description":"Recover, synthesize, complete, translate, or quantitatively estimate MRI-derived images, contrasts, signals, and parameter maps from undersampled, incomplete, noisy, cross-modal, longitudinal, or physics-constrained measurements.","inclusion":"Include work whose primary contribution is MRI inverse reconstruction, accelerated acquisition, quantitative susceptibility, perfusion or fingerprinting parameter mapping, modality synthesis, longitudinal image synthesis, healthy-anatomy reconstruction for anomaly detection, missing modality completion, or generative image translation.","exclusion":"Downstream segmentation of reconstructed images belongs to Anatomical segmentation; clinical classification using synthesized or reconstructed images belongs to Subject-level clinical and normative phenotyping unless reconstruction or synthesis is the main contribution; fMRI-to-natural-image stimulus reconstruction belongs to Brain encoding and decoding.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["deep learning-based reduction of post-labeling delay acquisitions for cerebral blood flow and arterial transit time map estimation","generative adversarial networks for disease-specific brain imaging synthesis and analysis","temporally adjustable longitudinal flair mri synthesis for multiple sclerosis disease progression monitoring"]},{"direction_index":7,"label":"D7","name":"Reusable methods and evaluation infrastructure","description":"Build general-purpose neuroimaging representations, multimodal fusion methods, domain adaptation and harmonization approaches, benchmarking resources, quality and reliability infrastructure, and methodological foundations for interpretable and generalizable neuroimaging machine learning.","inclusion":"Include work whose primary objective is reusable representation learning, self-supervised or contrastive pretraining across tasks, multimodal fusion as a general method, cross-modal surrogate mapping, domain adaptation, scanner or site harmonization, acquisition-shift mitigation, quality control, benchmark construction, standardized evaluation, interpretability surveys, theoretical limits, or broad methodological characterization not tied to a single endpoint.","exclusion":"Disease-specific fusion, robustness, or interpretability belongs to the relevant clinical program when the endpoint is dominant; reconstruction or synthesis belongs to MRI reconstruction and synthesis; stimulus-focused pretraining belongs to Brain encoding and decoding.","nearest_boundary":"D5 Brain encoding and decoding","representative_topics":["benchmarking brain network graph datasets for neurodegenerative and developmental condition prediction","multi-task benchmark for multi-scale brain structure representation learning from microtomography","source-free domain adaptation for scanner harmonization in distributed neuroimaging"]}],"split":"dev"}
{"episode_id":"F226_T2025-01","field_index":226,"field":"neuroimaging analysis","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Subject-level clinical and normative phenotyping","description":"Develop neuroimaging biomarkers and predictive models for individual diagnosis, prognosis, demographic phenotyping, cognitive or clinical outcome prediction, disease staging, and deviation from healthy population norms.","inclusion":"Include work whose primary output is a subject-level label, risk score, stage, conversion prediction, survival or outcome estimate, cognitive score, brain-age or normative-deviation measure, sex or demographic phenotype, or clinically interpretable biomarker for AD, non-AD disorders, psychiatric conditions, tumors, TBI, addiction, ASD, schizophrenia, or aging-related variation.","exclusion":"If the main objective is voxelwise delineation of lesions or tumors, assign to Anatomical segmentation; if the main objective is discovering network organization rather than predicting a subject endpoint, assign to Connectome organization; if the main objective is reusable pretraining or harmonization independent of one endpoint, assign to Reusable methods and evaluation infrastructure.","nearest_boundary":"D4 Connectome organization and dynamics","representative_topics":["graph-based ensemble u-net framework for interpretable alzheimer's disease detection and prognosis from structural mri","multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","self-supervised graph state space models for single-subject psychiatric disorder diagnosis from resting-state fmri"]},{"direction_index":1,"label":"D1","name":"Fetal and neonatal neurodevelopmental imaging","description":"Analyze fetal, perinatal, and neonatal brain imaging with methods tailored to early-life anatomy, motion, low signal, thick slices, atlas construction, super-resolution, segmentation, surface development, quality control, and normative developmental baselines.","inclusion":"Include work where fetal, perinatal, or neonatal imaging is central to the scientific problem, including early-life MRI or fMRI segmentation, motion or quality control, super-resolution, atlas construction, neonatal cortical maturation, low-field or thick-slice adaptation, and perinatal structural or functional connectivity baselines.","exclusion":"Adult or general pediatric segmentation, reconstruction, registration, or brain aging studies belong to the corresponding non-developmental program unless fetal or neonatal acquisition and anatomy define the problem.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["surface-based generative modeling of neonatal cortical maturation across gestational age","continuous longitudinal fetal brain atlas via implicit neural representation for temporal consistency","self-supervised single-volume isotropic superresolution for fetal brain mri reconstruction"]},{"direction_index":2,"label":"D2","name":"Anatomical segmentation and lesion delineation","description":"Automate voxelwise, regionwise, or object-level delineation of brain tissues, anatomical structures, tumors, white matter hyperintensities, lesions, skull-stripping targets, and related volumetric targets.","inclusion":"Include work whose primary contribution is segmentation, detection, or delineation of brain anatomy, lesions, tumors, white matter abnormalities, hippocampus, fiber bundles as segmentation targets, or longitudinally consistent anatomical labels, including robustness to missing modalities or acquisition perturbations.","exclusion":"If the main target is cortical meshes, surfaces, tracts as geometric objects, registration, or morphometric shape analysis, assign to Geometric brain structure analysis; if segmentation is only a preprocessing step for clinical prediction, assign to Subject-level clinical and normative phenotyping.","nearest_boundary":"D3 Geometric brain structure analysis","representative_topics":["multimodal mri glioma survival prediction via cross-scale nonlocal feature fusion","longitudinal whole-brain and white matter lesion segmentation with temporal consistency constraints","3d flair-weighted white matter hyperintensity segmentation using deep learning on multi-site neuroimaging data"]},{"direction_index":3,"label":"D3","name":"Geometric brain structure analysis","description":"Reconstruct, register, parcellate, quantify, and model brain geometry, cortical surfaces, white matter tracts, diffusion structures, meshes, point clouds, deformation fields, longitudinal morphometry, and shape-based anatomical variation.","inclusion":"Include work operating primarily on anatomical objects or spatial transformations, including cortical surface segmentation or registration, mesh analysis, sulcal labeling, tractography and fiber-level representations, tract parcellation, diffusion signal geometry, morphometric population analysis, deformable registration, diffeomorphic mapping, and longitudinal deformation or volumetric change modeling.","exclusion":"Voxelwise tissue, tumor, or lesion segmentation belongs to Anatomical segmentation; MRI signal reconstruction or contrast synthesis belongs to MRI reconstruction and synthesis; connectome graphs whose nodes are ROIs belong to Connectome organization and dynamics.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["equivariant graph neural networks for robust cortical mesh segmentation under misalignment","fiber-level whole brain tractography classification using spectral embedding and vision transformer attention","point cloud geometric deep learning for neuropsychological score prediction and tract region localization"]},{"direction_index":4,"label":"D4","name":"Connectome organization and dynamics","description":"Infer, characterize, label, generate, or model functional, structural, temporal, effective, manifold-valued, or graph-based brain connectivity to understand network organization and dynamics.","inclusion":"Include work whose main objective is brain network structure or dynamics: graph neural network methodology for connectomes, dynamic graph structure learning, effective or temporal-lag connectivity, directed communication learning, network labeling, universal connectome template estimation, connectivity generation, high-order graph relationships, Riemannian or SPD-manifold connectivity modeling, and temporal models of fMRI dependencies when network dynamics are central.","exclusion":"Connectome-based studies primarily optimized for a disease, demographic, or cognitive subject-level endpoint belong to Subject-level clinical and normative phenotyping; general-purpose graph pretraining or benchmark construction belongs to Reusable methods and evaluation infrastructure; stimulus decoding from fMRI belongs to Brain encoding and decoding.","nearest_boundary":"D0 Subject-level clinical and normative phenotyping","representative_topics":["deep learning classification of resting state fmri brain networks using cloud-based reproducible pipelines","multimodal multiresolution brain graph integration for universal connectome template estimation","sparse directed dynamic communication learning in brain networks via coupled sequential vae"]},{"direction_index":5,"label":"D5","name":"Brain encoding and decoding","description":"Map between brain activity and external perceptual, semantic, linguistic, speech, visual, or cognitive representations, including encoding models, neural decoding, stimulus reconstruction, retrieval, and brain-AI representational alignment.","inclusion":"Include work whose primary objective is predicting neural responses from stimuli, reconstructing or retrieving images, text, speech, music, captions, or semantic content from fMRI or other brain recordings, aligning AI representations with brain activity, or pretraining explicitly for brain decoding and stimulus-related interpretation.","exclusion":"Disease or trait classification from fMRI belongs to Subject-level clinical and normative phenotyping; general-purpose neuroimaging representation learning not centered on stimulus content belongs to Reusable methods and evaluation infrastructure; MRI-to-image synthesis as an acquisition or reconstruction task belongs to MRI reconstruction and synthesis.","nearest_boundary":"D7 Reusable methods and evaluation infrastructure","representative_topics":["self-supervised speech representations aligned with cortical hierarchy and behavioral responses","cross-modal fmri surrogate marker prediction from fnirs using neural data augmentation","self-taught pre-training for brain decoding via convolutional autoencoder initialization"]},{"direction_index":6,"label":"D6","name":"MRI reconstruction and synthesis","description":"Recover, synthesize, complete, translate, or quantitatively estimate MRI-derived images, contrasts, signals, and parameter maps from undersampled, incomplete, noisy, cross-modal, longitudinal, or physics-constrained measurements.","inclusion":"Include work whose primary contribution is MRI inverse reconstruction, accelerated acquisition, quantitative susceptibility, perfusion or fingerprinting parameter mapping, modality synthesis, longitudinal image synthesis, healthy-anatomy reconstruction for anomaly detection, missing modality completion, or generative image translation.","exclusion":"Downstream segmentation of reconstructed images belongs to Anatomical segmentation; clinical classification using synthesized or reconstructed images belongs to Subject-level clinical and normative phenotyping unless reconstruction or synthesis is the main contribution; fMRI-to-natural-image stimulus reconstruction belongs to Brain encoding and decoding.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["deep learning-based reduction of post-labeling delay acquisitions for cerebral blood flow and arterial transit time map estimation","generative adversarial networks for disease-specific brain imaging synthesis and analysis","temporally adjustable longitudinal flair mri synthesis for multiple sclerosis disease progression monitoring"]},{"direction_index":7,"label":"D7","name":"Reusable methods and evaluation infrastructure","description":"Build general-purpose neuroimaging representations, multimodal fusion methods, domain adaptation and harmonization approaches, benchmarking resources, quality and reliability infrastructure, and methodological foundations for interpretable and generalizable neuroimaging machine learning.","inclusion":"Include work whose primary objective is reusable representation learning, self-supervised or contrastive pretraining across tasks, multimodal fusion as a general method, cross-modal surrogate mapping, domain adaptation, scanner or site harmonization, acquisition-shift mitigation, quality control, benchmark construction, standardized evaluation, interpretability surveys, theoretical limits, or broad methodological characterization not tied to a single endpoint.","exclusion":"Disease-specific fusion, robustness, or interpretability belongs to the relevant clinical program when the endpoint is dominant; reconstruction or synthesis belongs to MRI reconstruction and synthesis; stimulus-focused pretraining belongs to Brain encoding and decoding.","nearest_boundary":"D5 Brain encoding and decoding","representative_topics":["benchmarking brain network graph datasets for neurodegenerative and developmental condition prediction","multi-task benchmark for multi-scale brain structure representation learning from microtomography","source-free domain adaptation for scanner harmonization in distributed neuroimaging"]}],"split":"dev"}
{"episode_id":"F226_T2025-07","field_index":226,"field":"neuroimaging analysis","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Subject-level clinical and normative phenotyping","description":"Develop neuroimaging biomarkers and predictive models for individual diagnosis, prognosis, demographic phenotyping, cognitive or clinical outcome prediction, disease staging, and deviation from healthy population norms.","inclusion":"Include work whose primary output is a subject-level label, risk score, stage, conversion prediction, survival or outcome estimate, cognitive score, brain-age or normative-deviation measure, sex or demographic phenotype, or clinically interpretable biomarker for AD, non-AD disorders, psychiatric conditions, tumors, TBI, addiction, ASD, schizophrenia, or aging-related variation.","exclusion":"If the main objective is voxelwise delineation of lesions or tumors, assign to Anatomical segmentation; if the main objective is discovering network organization rather than predicting a subject endpoint, assign to Connectome organization; if the main objective is reusable pretraining or harmonization independent of one endpoint, assign to Reusable methods and evaluation infrastructure.","nearest_boundary":"D4 Connectome organization and dynamics","representative_topics":["graph-based ensemble u-net framework for interpretable alzheimer's disease detection and prognosis from structural mri","multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","self-supervised graph state space models for single-subject psychiatric disorder diagnosis from resting-state fmri"]},{"direction_index":1,"label":"D1","name":"Fetal and neonatal neurodevelopmental imaging","description":"Analyze fetal, perinatal, and neonatal brain imaging with methods tailored to early-life anatomy, motion, low signal, thick slices, atlas construction, super-resolution, segmentation, surface development, quality control, and normative developmental baselines.","inclusion":"Include work where fetal, perinatal, or neonatal imaging is central to the scientific problem, including early-life MRI or fMRI segmentation, motion or quality control, super-resolution, atlas construction, neonatal cortical maturation, low-field or thick-slice adaptation, and perinatal structural or functional connectivity baselines.","exclusion":"Adult or general pediatric segmentation, reconstruction, registration, or brain aging studies belong to the corresponding non-developmental program unless fetal or neonatal acquisition and anatomy define the problem.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["surface-based generative modeling of neonatal cortical maturation across gestational age","continuous longitudinal fetal brain atlas via implicit neural representation for temporal consistency","self-supervised single-volume isotropic superresolution for fetal brain mri reconstruction"]},{"direction_index":2,"label":"D2","name":"Anatomical segmentation and lesion delineation","description":"Automate voxelwise, regionwise, or object-level delineation of brain tissues, anatomical structures, tumors, white matter hyperintensities, lesions, skull-stripping targets, and related volumetric targets.","inclusion":"Include work whose primary contribution is segmentation, detection, or delineation of brain anatomy, lesions, tumors, white matter abnormalities, hippocampus, fiber bundles as segmentation targets, or longitudinally consistent anatomical labels, including robustness to missing modalities or acquisition perturbations.","exclusion":"If the main target is cortical meshes, surfaces, tracts as geometric objects, registration, or morphometric shape analysis, assign to Geometric brain structure analysis; if segmentation is only a preprocessing step for clinical prediction, assign to Subject-level clinical and normative phenotyping.","nearest_boundary":"D3 Geometric brain structure analysis","representative_topics":["multimodal mri glioma survival prediction via cross-scale nonlocal feature fusion","longitudinal whole-brain and white matter lesion segmentation with temporal consistency constraints","3d flair-weighted white matter hyperintensity segmentation using deep learning on multi-site neuroimaging data"]},{"direction_index":3,"label":"D3","name":"Geometric brain structure analysis","description":"Reconstruct, register, parcellate, quantify, and model brain geometry, cortical surfaces, white matter tracts, diffusion structures, meshes, point clouds, deformation fields, longitudinal morphometry, and shape-based anatomical variation.","inclusion":"Include work operating primarily on anatomical objects or spatial transformations, including cortical surface segmentation or registration, mesh analysis, sulcal labeling, tractography and fiber-level representations, tract parcellation, diffusion signal geometry, morphometric population analysis, deformable registration, diffeomorphic mapping, and longitudinal deformation or volumetric change modeling.","exclusion":"Voxelwise tissue, tumor, or lesion segmentation belongs to Anatomical segmentation; MRI signal reconstruction or contrast synthesis belongs to MRI reconstruction and synthesis; connectome graphs whose nodes are ROIs belong to Connectome organization and dynamics.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["equivariant graph neural networks for robust cortical mesh segmentation under misalignment","fiber-level whole brain tractography classification using spectral embedding and vision transformer attention","point cloud geometric deep learning for neuropsychological score prediction and tract region localization"]},{"direction_index":4,"label":"D4","name":"Connectome organization and dynamics","description":"Infer, characterize, label, generate, or model functional, structural, temporal, effective, manifold-valued, or graph-based brain connectivity to understand network organization and dynamics.","inclusion":"Include work whose main objective is brain network structure or dynamics: graph neural network methodology for connectomes, dynamic graph structure learning, effective or temporal-lag connectivity, directed communication learning, network labeling, universal connectome template estimation, connectivity generation, high-order graph relationships, Riemannian or SPD-manifold connectivity modeling, and temporal models of fMRI dependencies when network dynamics are central.","exclusion":"Connectome-based studies primarily optimized for a disease, demographic, or cognitive subject-level endpoint belong to Subject-level clinical and normative phenotyping; general-purpose graph pretraining or benchmark construction belongs to Reusable methods and evaluation infrastructure; stimulus decoding from fMRI belongs to Brain encoding and decoding.","nearest_boundary":"D0 Subject-level clinical and normative phenotyping","representative_topics":["deep learning classification of resting state fmri brain networks using cloud-based reproducible pipelines","multimodal multiresolution brain graph integration for universal connectome template estimation","sparse directed dynamic communication learning in brain networks via coupled sequential vae"]},{"direction_index":5,"label":"D5","name":"Brain encoding and decoding","description":"Map between brain activity and external perceptual, semantic, linguistic, speech, visual, or cognitive representations, including encoding models, neural decoding, stimulus reconstruction, retrieval, and brain-AI representational alignment.","inclusion":"Include work whose primary objective is predicting neural responses from stimuli, reconstructing or retrieving images, text, speech, music, captions, or semantic content from fMRI or other brain recordings, aligning AI representations with brain activity, or pretraining explicitly for brain decoding and stimulus-related interpretation.","exclusion":"Disease or trait classification from fMRI belongs to Subject-level clinical and normative phenotyping; general-purpose neuroimaging representation learning not centered on stimulus content belongs to Reusable methods and evaluation infrastructure; MRI-to-image synthesis as an acquisition or reconstruction task belongs to MRI reconstruction and synthesis.","nearest_boundary":"D7 Reusable methods and evaluation infrastructure","representative_topics":["self-supervised speech representations aligned with cortical hierarchy and behavioral responses","cross-modal fmri surrogate marker prediction from fnirs using neural data augmentation","self-taught pre-training for brain decoding via convolutional autoencoder initialization"]},{"direction_index":6,"label":"D6","name":"MRI reconstruction and synthesis","description":"Recover, synthesize, complete, translate, or quantitatively estimate MRI-derived images, contrasts, signals, and parameter maps from undersampled, incomplete, noisy, cross-modal, longitudinal, or physics-constrained measurements.","inclusion":"Include work whose primary contribution is MRI inverse reconstruction, accelerated acquisition, quantitative susceptibility, perfusion or fingerprinting parameter mapping, modality synthesis, longitudinal image synthesis, healthy-anatomy reconstruction for anomaly detection, missing modality completion, or generative image translation.","exclusion":"Downstream segmentation of reconstructed images belongs to Anatomical segmentation; clinical classification using synthesized or reconstructed images belongs to Subject-level clinical and normative phenotyping unless reconstruction or synthesis is the main contribution; fMRI-to-natural-image stimulus reconstruction belongs to Brain encoding and decoding.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["deep learning-based reduction of post-labeling delay acquisitions for cerebral blood flow and arterial transit time map estimation","generative adversarial networks for disease-specific brain imaging synthesis and analysis","temporally adjustable longitudinal flair mri synthesis for multiple sclerosis disease progression monitoring"]},{"direction_index":7,"label":"D7","name":"Reusable methods and evaluation infrastructure","description":"Build general-purpose neuroimaging representations, multimodal fusion methods, domain adaptation and harmonization approaches, benchmarking resources, quality and reliability infrastructure, and methodological foundations for interpretable and generalizable neuroimaging machine learning.","inclusion":"Include work whose primary objective is reusable representation learning, self-supervised or contrastive pretraining across tasks, multimodal fusion as a general method, cross-modal surrogate mapping, domain adaptation, scanner or site harmonization, acquisition-shift mitigation, quality control, benchmark construction, standardized evaluation, interpretability surveys, theoretical limits, or broad methodological characterization not tied to a single endpoint.","exclusion":"Disease-specific fusion, robustness, or interpretability belongs to the relevant clinical program when the endpoint is dominant; reconstruction or synthesis belongs to MRI reconstruction and synthesis; stimulus-focused pretraining belongs to Brain encoding and decoding.","nearest_boundary":"D5 Brain encoding and decoding","representative_topics":["benchmarking brain network graph datasets for neurodegenerative and developmental condition prediction","multi-task benchmark for multi-scale brain structure representation learning from microtomography","source-free domain adaptation for scanner harmonization in distributed neuroimaging"]}],"split":"dev"}
{"episode_id":"F226_T2026-01","field_index":226,"field":"neuroimaging analysis","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Subject-level clinical and normative phenotyping","description":"Develop neuroimaging biomarkers and predictive models for individual diagnosis, prognosis, demographic phenotyping, cognitive or clinical outcome prediction, disease staging, and deviation from healthy population norms.","inclusion":"Include work whose primary output is a subject-level label, risk score, stage, conversion prediction, survival or outcome estimate, cognitive score, brain-age or normative-deviation measure, sex or demographic phenotype, or clinically interpretable biomarker for AD, non-AD disorders, psychiatric conditions, tumors, TBI, addiction, ASD, schizophrenia, or aging-related variation.","exclusion":"If the main objective is voxelwise delineation of lesions or tumors, assign to Anatomical segmentation; if the main objective is discovering network organization rather than predicting a subject endpoint, assign to Connectome organization; if the main objective is reusable pretraining or harmonization independent of one endpoint, assign to Reusable methods and evaluation infrastructure.","nearest_boundary":"D4 Connectome organization and dynamics","representative_topics":["graph-based ensemble u-net framework for interpretable alzheimer's disease detection and prognosis from structural mri","multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","self-supervised graph state space models for single-subject psychiatric disorder diagnosis from resting-state fmri"]},{"direction_index":1,"label":"D1","name":"Fetal and neonatal neurodevelopmental imaging","description":"Analyze fetal, perinatal, and neonatal brain imaging with methods tailored to early-life anatomy, motion, low signal, thick slices, atlas construction, super-resolution, segmentation, surface development, quality control, and normative developmental baselines.","inclusion":"Include work where fetal, perinatal, or neonatal imaging is central to the scientific problem, including early-life MRI or fMRI segmentation, motion or quality control, super-resolution, atlas construction, neonatal cortical maturation, low-field or thick-slice adaptation, and perinatal structural or functional connectivity baselines.","exclusion":"Adult or general pediatric segmentation, reconstruction, registration, or brain aging studies belong to the corresponding non-developmental program unless fetal or neonatal acquisition and anatomy define the problem.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["surface-based generative modeling of neonatal cortical maturation across gestational age","continuous longitudinal fetal brain atlas via implicit neural representation for temporal consistency","self-supervised single-volume isotropic superresolution for fetal brain mri reconstruction"]},{"direction_index":2,"label":"D2","name":"Anatomical segmentation and lesion delineation","description":"Automate voxelwise, regionwise, or object-level delineation of brain tissues, anatomical structures, tumors, white matter hyperintensities, lesions, skull-stripping targets, and related volumetric targets.","inclusion":"Include work whose primary contribution is segmentation, detection, or delineation of brain anatomy, lesions, tumors, white matter abnormalities, hippocampus, fiber bundles as segmentation targets, or longitudinally consistent anatomical labels, including robustness to missing modalities or acquisition perturbations.","exclusion":"If the main target is cortical meshes, surfaces, tracts as geometric objects, registration, or morphometric shape analysis, assign to Geometric brain structure analysis; if segmentation is only a preprocessing step for clinical prediction, assign to Subject-level clinical and normative phenotyping.","nearest_boundary":"D3 Geometric brain structure analysis","representative_topics":["multimodal mri glioma survival prediction via cross-scale nonlocal feature fusion","longitudinal whole-brain and white matter lesion segmentation with temporal consistency constraints","3d flair-weighted white matter hyperintensity segmentation using deep learning on multi-site neuroimaging data"]},{"direction_index":3,"label":"D3","name":"Geometric brain structure analysis","description":"Reconstruct, register, parcellate, quantify, and model brain geometry, cortical surfaces, white matter tracts, diffusion structures, meshes, point clouds, deformation fields, longitudinal morphometry, and shape-based anatomical variation.","inclusion":"Include work operating primarily on anatomical objects or spatial transformations, including cortical surface segmentation or registration, mesh analysis, sulcal labeling, tractography and fiber-level representations, tract parcellation, diffusion signal geometry, morphometric population analysis, deformable registration, diffeomorphic mapping, and longitudinal deformation or volumetric change modeling.","exclusion":"Voxelwise tissue, tumor, or lesion segmentation belongs to Anatomical segmentation; MRI signal reconstruction or contrast synthesis belongs to MRI reconstruction and synthesis; connectome graphs whose nodes are ROIs belong to Connectome organization and dynamics.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["equivariant graph neural networks for robust cortical mesh segmentation under misalignment","fiber-level whole brain tractography classification using spectral embedding and vision transformer attention","point cloud geometric deep learning for neuropsychological score prediction and tract region localization"]},{"direction_index":4,"label":"D4","name":"Connectome organization and dynamics","description":"Infer, characterize, label, generate, or model functional, structural, temporal, effective, manifold-valued, or graph-based brain connectivity to understand network organization and dynamics.","inclusion":"Include work whose main objective is brain network structure or dynamics: graph neural network methodology for connectomes, dynamic graph structure learning, effective or temporal-lag connectivity, directed communication learning, network labeling, universal connectome template estimation, connectivity generation, high-order graph relationships, Riemannian or SPD-manifold connectivity modeling, and temporal models of fMRI dependencies when network dynamics are central.","exclusion":"Connectome-based studies primarily optimized for a disease, demographic, or cognitive subject-level endpoint belong to Subject-level clinical and normative phenotyping; general-purpose graph pretraining or benchmark construction belongs to Reusable methods and evaluation infrastructure; stimulus decoding from fMRI belongs to Brain encoding and decoding.","nearest_boundary":"D0 Subject-level clinical and normative phenotyping","representative_topics":["deep learning classification of resting state fmri brain networks using cloud-based reproducible pipelines","multimodal multiresolution brain graph integration for universal connectome template estimation","sparse directed dynamic communication learning in brain networks via coupled sequential vae"]},{"direction_index":5,"label":"D5","name":"Brain encoding and decoding","description":"Map between brain activity and external perceptual, semantic, linguistic, speech, visual, or cognitive representations, including encoding models, neural decoding, stimulus reconstruction, retrieval, and brain-AI representational alignment.","inclusion":"Include work whose primary objective is predicting neural responses from stimuli, reconstructing or retrieving images, text, speech, music, captions, or semantic content from fMRI or other brain recordings, aligning AI representations with brain activity, or pretraining explicitly for brain decoding and stimulus-related interpretation.","exclusion":"Disease or trait classification from fMRI belongs to Subject-level clinical and normative phenotyping; general-purpose neuroimaging representation learning not centered on stimulus content belongs to Reusable methods and evaluation infrastructure; MRI-to-image synthesis as an acquisition or reconstruction task belongs to MRI reconstruction and synthesis.","nearest_boundary":"D7 Reusable methods and evaluation infrastructure","representative_topics":["self-supervised speech representations aligned with cortical hierarchy and behavioral responses","cross-modal fmri surrogate marker prediction from fnirs using neural data augmentation","self-taught pre-training for brain decoding via convolutional autoencoder initialization"]},{"direction_index":6,"label":"D6","name":"MRI reconstruction and synthesis","description":"Recover, synthesize, complete, translate, or quantitatively estimate MRI-derived images, contrasts, signals, and parameter maps from undersampled, incomplete, noisy, cross-modal, longitudinal, or physics-constrained measurements.","inclusion":"Include work whose primary contribution is MRI inverse reconstruction, accelerated acquisition, quantitative susceptibility, perfusion or fingerprinting parameter mapping, modality synthesis, longitudinal image synthesis, healthy-anatomy reconstruction for anomaly detection, missing modality completion, or generative image translation.","exclusion":"Downstream segmentation of reconstructed images belongs to Anatomical segmentation; clinical classification using synthesized or reconstructed images belongs to Subject-level clinical and normative phenotyping unless reconstruction or synthesis is the main contribution; fMRI-to-natural-image stimulus reconstruction belongs to Brain encoding and decoding.","nearest_boundary":"D2 Anatomical segmentation and lesion delineation","representative_topics":["deep learning-based reduction of post-labeling delay acquisitions for cerebral blood flow and arterial transit time map estimation","generative adversarial networks for disease-specific brain imaging synthesis and analysis","temporally adjustable longitudinal flair mri synthesis for multiple sclerosis disease progression monitoring"]},{"direction_index":7,"label":"D7","name":"Reusable methods and evaluation infrastructure","description":"Build general-purpose neuroimaging representations, multimodal fusion methods, domain adaptation and harmonization approaches, benchmarking resources, quality and reliability infrastructure, and methodological foundations for interpretable and generalizable neuroimaging machine learning.","inclusion":"Include work whose primary objective is reusable representation learning, self-supervised or contrastive pretraining across tasks, multimodal fusion as a general method, cross-modal surrogate mapping, domain adaptation, scanner or site harmonization, acquisition-shift mitigation, quality control, benchmark construction, standardized evaluation, interpretability surveys, theoretical limits, or broad methodological characterization not tied to a single endpoint.","exclusion":"Disease-specific fusion, robustness, or interpretability belongs to the relevant clinical program when the endpoint is dominant; reconstruction or synthesis belongs to MRI reconstruction and synthesis; stimulus-focused pretraining belongs to Brain encoding and decoding.","nearest_boundary":"D5 Brain encoding and decoding","representative_topics":["benchmarking brain network graph datasets for neurodegenerative and developmental condition prediction","multi-task benchmark for multi-scale brain structure representation learning from microtomography","source-free domain adaptation for scanner harmonization in distributed neuroimaging"]}],"split":"dev"}
{"episode_id":"F229_T2024-01","field_index":229,"field":"multimodal reasoning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Diagnostic reasoning evaluation","description":"Create, validate, or analyze benchmarks, diagnostic datasets, probes, and evaluation protocols for measuring multimodal reasoning abilities, failures, consistency, robustness, and generalization.","inclusion":"Assign work whose primary objective is measurement of multimodal reasoning capability, including new benchmark suites, evaluation taxonomies, diagnostic tasks, or comparative capability analyses.","exclusion":"If the main contribution is a solver, training method, architecture, rationale method, tool system, or structured representation and evaluation is only validation, assign to that method-oriented direction.","nearest_boundary":"Rationale-based multimodal reasoning","representative_topics":["multimodal word problem solving requiring compositional language, visual, and mathematical reasoning","vision-language benchmarking for compositional generalization in embodied human activity forecasting","benchmarking deep model generalization on elementary visuo-linguistic reasoning tasks"]},{"direction_index":1,"label":"D1","name":"Rationale-based multimodal reasoning","description":"Develop methods that make multimodal inference proceed through explicit intermediate natural-language, quasi-symbolic, or explanation-like reasoning traces such as chains of thought, rationales, self-consistency signals, or visual infillings.","inclusion":"Assign work whose core mechanism is eliciting, supervising, generating, refining, distilling, or using intermediate rationale chains between multimodal inputs and final answers.","exclusion":"If decomposition produces executable programs, tool calls, or external subtask plans, assign to tool-augmented orchestration; if the intermediate structure is a persistent graph, module, or symbolic representation, assign to structured compositional reasoning.","nearest_boundary":"Tool-augmented orchestration","representative_topics":["multimodal chain-of-thought reasoning for interpretable science question answering","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","multimodal chain-of-thought augmentation via recursive visual infillings for temporal reasoning"]},{"direction_index":2,"label":"D2","name":"Abstract visual rule induction","description":"Solve abstract, formal, geometric, analogy, puzzle-like, or visual pattern reasoning tasks by inducing latent rules, transformations, programs, object relations, or algebraic/logical structures from visual stimuli.","inclusion":"Assign work whose primary objective is rule induction for abstract visual reasoning settings such as Raven-style matrices, ARC-like tasks, tangrams, visual analogies, constructible geometry, or abstract object-pattern completion.","exclusion":"If the same symbolic or structural machinery is applied mainly to natural-image vision-language QA, dialog, or grounded language tasks, assign to structured compositional reasoning; if the contribution is only a benchmark, assign to diagnostic evaluation.","nearest_boundary":"Structured compositional vision-language reasoning","representative_topics":["transformer-based spatial token reasoning for raven's progressive matrix rule induction","abstract visual reasoning over part-segmented shape stimuli with multimodal fine-tuning","visual analogy completion over complex scene understanding using situation recognition annotations"]},{"direction_index":3,"label":"D3","name":"Spatial and 3D grounding","description":"Improve multimodal understanding of spatial, regional, depth, orientation, positional, 3D, point-cloud, and spatio-temporal grounding relations.","inclusion":"Assign work whose primary objective is modeling, grounding, evaluating, or enhancing spatial or spatio-temporal relation comprehension, including 3D object grounding, pseudo-3D perception, region-specific prompts, spatial relation graphs, and target-aware audio-visual grounding.","exclusion":"If temporal causality, physical outcome prediction, or state-change dynamics dominate over spatial structure, assign to causal and dynamic reasoning; if spatial relations are merely one part of a general model architecture, assign to broad model integration.","nearest_boundary":"Causal and dynamic reasoning","representative_topics":["language-conditioned spatial relation reasoning for 3d object grounding in point clouds","depth-aware pseudo 3d perception framework for visual commonsense reasoning with multimodal large language models","target-aware spatio-temporal grounding for audio-visual question answering via joint temporal fusion"]},{"direction_index":4,"label":"D4","name":"Structured compositional vision-language reasoning","description":"Build multimodal reasoning systems that explicitly represent, compose, and manipulate objects, attributes, relations, scene graphs, neural modules, programs, queries, or symbolic structures for natural visual-linguistic inference.","inclusion":"Assign work whose main objective is compositional generalization or multi-step reasoning over natural images, video, dialog, or language-conditioned visual inputs using persistent structured representations or neuro-symbolic/internal modular mechanisms.","exclusion":"If the task is purely abstract visual puzzle rule induction, assign to abstract visual rule induction; if decomposition is mainly external tool orchestration, assign to tool-augmented orchestration; if no explicit compositional structure is central, assign to broad model integration.","nearest_boundary":"Broad model integration and adaptation","representative_topics":["neuro-symbolic visual dialog with program execution for long-distance co-reference and multi-round reasoning","multimodal visual reasoning via temporal relation construction and visual-linguistic association frameworks","generalization robustness of end-to-end versus neuro-symbolic vision-language reasoning under distribution shift"]},{"direction_index":5,"label":"D5","name":"Broad model integration and adaptation","description":"Improve general multimodal reasoning competence through broadly applicable architectures, representation unification, pretraining, cross-modal fusion, attention, knowledge integration, alignment, routing, foundation-model integration, or domain-shaped adaptation strategies.","inclusion":"Assign work whose primary objective is a general model, representation, pretraining, fusion, alignment, knowledge-enhancement, or adaptation strategy that improves or characterizes reasoning across modalities, including domain-constrained adaptations when the atom’s emphasis is system/model adaptation rather than a separate causal, spatial, or tool program.","exclusion":"If the central mechanism is explicit rationale chains, assign to rationale-based reasoning; if it is symbolic program/module execution over natural scenes, assign to structured compositional reasoning; if it orchestrates external tools or agents, assign to tool-augmented orchestration.","nearest_boundary":"Structured compositional vision-language reasoning","representative_topics":["survey of multimodal machine learning foundations covering representation, alignment, and generative challenges","zero-shot ensemble composition of pre-trained models via iterative consensus optimization","pretraining visual language models for chart and plot numerical reasoning via derendering"]},{"direction_index":6,"label":"D6","name":"Causal and dynamic reasoning","description":"Model causal, counterfactual, temporal, physical, action-effect, event-level, state-change, transformation, and behavior-forecasting relations across multimodal observations.","inclusion":"Assign work whose primary objective is understanding what happened, why it happened, what changed, what physical effect an action caused, or what may happen next in multimodal scenes, videos, image sequences, audio-visual settings, or embodied observations.","exclusion":"If the contribution is mainly a benchmark for such abilities, assign to diagnostic evaluation; if spatial layout or 3D grounding is the dominant objective, assign to spatial and 3D grounding; if event reasoning only validates a broader model integration method, assign to broad model integration.","nearest_boundary":"Spatial and 3D grounding","representative_topics":["active vision theory-inspired guided attention for dynamic visual reasoning and zero-shot generalization","cross-modal causal relational reasoning for event-level video visual question answering","multimodal action outcome prediction for physical commonsense reasoning in planning"]},{"direction_index":7,"label":"D7","name":"Tool-augmented orchestration","description":"Solve complex multimodal reasoning or action tasks by explicit decomposition into subproblems handled by external tools, code execution, specialist models, retrieval systems, agents, or iterative plan-execute-inspect loops.","inclusion":"Assign work whose central objective is orchestration, delegation, routing, program generation, multi-agent collaboration, expert-model selection, or interactive multimodal action planning using multiple components beyond a single internally structured model.","exclusion":"If decomposition is only a natural-language rationale without tool execution, assign to rationale-based reasoning; if modules are internal learned or symbolic structures rather than external tools, assign to structured compositional reasoning; if the system is a broad integrated model without explicit orchestration, assign to broad model integration.","nearest_boundary":"Broad model integration and adaptation","representative_topics":["zero-shot ensemble composition of pre-trained models via iterative consensus optimization","neuro-symbolic compositional visual reasoning via in-context learning and modular program generation","prompt-based integration of language models and vision experts for multimodal reasoning and action"]}],"split":"dev"}
{"episode_id":"F229_T2024-07","field_index":229,"field":"multimodal reasoning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Diagnostic reasoning evaluation","description":"Create, validate, or analyze benchmarks, diagnostic datasets, probes, and evaluation protocols for measuring multimodal reasoning abilities, failures, consistency, robustness, and generalization.","inclusion":"Assign work whose primary objective is measurement of multimodal reasoning capability, including new benchmark suites, evaluation taxonomies, diagnostic tasks, or comparative capability analyses.","exclusion":"If the main contribution is a solver, training method, architecture, rationale method, tool system, or structured representation and evaluation is only validation, assign to that method-oriented direction.","nearest_boundary":"Rationale-based multimodal reasoning","representative_topics":["multimodal word problem solving requiring compositional language, visual, and mathematical reasoning","vision-language benchmarking for compositional generalization in embodied human activity forecasting","benchmarking deep model generalization on elementary visuo-linguistic reasoning tasks"]},{"direction_index":1,"label":"D1","name":"Rationale-based multimodal reasoning","description":"Develop methods that make multimodal inference proceed through explicit intermediate natural-language, quasi-symbolic, or explanation-like reasoning traces such as chains of thought, rationales, self-consistency signals, or visual infillings.","inclusion":"Assign work whose core mechanism is eliciting, supervising, generating, refining, distilling, or using intermediate rationale chains between multimodal inputs and final answers.","exclusion":"If decomposition produces executable programs, tool calls, or external subtask plans, assign to tool-augmented orchestration; if the intermediate structure is a persistent graph, module, or symbolic representation, assign to structured compositional reasoning.","nearest_boundary":"Tool-augmented orchestration","representative_topics":["multimodal chain-of-thought reasoning for interpretable science question answering","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","multimodal chain-of-thought augmentation via recursive visual infillings for temporal reasoning"]},{"direction_index":2,"label":"D2","name":"Abstract visual rule induction","description":"Solve abstract, formal, geometric, analogy, puzzle-like, or visual pattern reasoning tasks by inducing latent rules, transformations, programs, object relations, or algebraic/logical structures from visual stimuli.","inclusion":"Assign work whose primary objective is rule induction for abstract visual reasoning settings such as Raven-style matrices, ARC-like tasks, tangrams, visual analogies, constructible geometry, or abstract object-pattern completion.","exclusion":"If the same symbolic or structural machinery is applied mainly to natural-image vision-language QA, dialog, or grounded language tasks, assign to structured compositional reasoning; if the contribution is only a benchmark, assign to diagnostic evaluation.","nearest_boundary":"Structured compositional vision-language reasoning","representative_topics":["transformer-based spatial token reasoning for raven's progressive matrix rule induction","abstract visual reasoning over part-segmented shape stimuli with multimodal fine-tuning","visual analogy completion over complex scene understanding using situation recognition annotations"]},{"direction_index":3,"label":"D3","name":"Spatial and 3D grounding","description":"Improve multimodal understanding of spatial, regional, depth, orientation, positional, 3D, point-cloud, and spatio-temporal grounding relations.","inclusion":"Assign work whose primary objective is modeling, grounding, evaluating, or enhancing spatial or spatio-temporal relation comprehension, including 3D object grounding, pseudo-3D perception, region-specific prompts, spatial relation graphs, and target-aware audio-visual grounding.","exclusion":"If temporal causality, physical outcome prediction, or state-change dynamics dominate over spatial structure, assign to causal and dynamic reasoning; if spatial relations are merely one part of a general model architecture, assign to broad model integration.","nearest_boundary":"Causal and dynamic reasoning","representative_topics":["language-conditioned spatial relation reasoning for 3d object grounding in point clouds","depth-aware pseudo 3d perception framework for visual commonsense reasoning with multimodal large language models","target-aware spatio-temporal grounding for audio-visual question answering via joint temporal fusion"]},{"direction_index":4,"label":"D4","name":"Structured compositional vision-language reasoning","description":"Build multimodal reasoning systems that explicitly represent, compose, and manipulate objects, attributes, relations, scene graphs, neural modules, programs, queries, or symbolic structures for natural visual-linguistic inference.","inclusion":"Assign work whose main objective is compositional generalization or multi-step reasoning over natural images, video, dialog, or language-conditioned visual inputs using persistent structured representations or neuro-symbolic/internal modular mechanisms.","exclusion":"If the task is purely abstract visual puzzle rule induction, assign to abstract visual rule induction; if decomposition is mainly external tool orchestration, assign to tool-augmented orchestration; if no explicit compositional structure is central, assign to broad model integration.","nearest_boundary":"Broad model integration and adaptation","representative_topics":["neuro-symbolic visual dialog with program execution for long-distance co-reference and multi-round reasoning","multimodal visual reasoning via temporal relation construction and visual-linguistic association frameworks","generalization robustness of end-to-end versus neuro-symbolic vision-language reasoning under distribution shift"]},{"direction_index":5,"label":"D5","name":"Broad model integration and adaptation","description":"Improve general multimodal reasoning competence through broadly applicable architectures, representation unification, pretraining, cross-modal fusion, attention, knowledge integration, alignment, routing, foundation-model integration, or domain-shaped adaptation strategies.","inclusion":"Assign work whose primary objective is a general model, representation, pretraining, fusion, alignment, knowledge-enhancement, or adaptation strategy that improves or characterizes reasoning across modalities, including domain-constrained adaptations when the atom’s emphasis is system/model adaptation rather than a separate causal, spatial, or tool program.","exclusion":"If the central mechanism is explicit rationale chains, assign to rationale-based reasoning; if it is symbolic program/module execution over natural scenes, assign to structured compositional reasoning; if it orchestrates external tools or agents, assign to tool-augmented orchestration.","nearest_boundary":"Structured compositional vision-language reasoning","representative_topics":["survey of multimodal machine learning foundations covering representation, alignment, and generative challenges","zero-shot ensemble composition of pre-trained models via iterative consensus optimization","pretraining visual language models for chart and plot numerical reasoning via derendering"]},{"direction_index":6,"label":"D6","name":"Causal and dynamic reasoning","description":"Model causal, counterfactual, temporal, physical, action-effect, event-level, state-change, transformation, and behavior-forecasting relations across multimodal observations.","inclusion":"Assign work whose primary objective is understanding what happened, why it happened, what changed, what physical effect an action caused, or what may happen next in multimodal scenes, videos, image sequences, audio-visual settings, or embodied observations.","exclusion":"If the contribution is mainly a benchmark for such abilities, assign to diagnostic evaluation; if spatial layout or 3D grounding is the dominant objective, assign to spatial and 3D grounding; if event reasoning only validates a broader model integration method, assign to broad model integration.","nearest_boundary":"Spatial and 3D grounding","representative_topics":["active vision theory-inspired guided attention for dynamic visual reasoning and zero-shot generalization","cross-modal causal relational reasoning for event-level video visual question answering","multimodal action outcome prediction for physical commonsense reasoning in planning"]},{"direction_index":7,"label":"D7","name":"Tool-augmented orchestration","description":"Solve complex multimodal reasoning or action tasks by explicit decomposition into subproblems handled by external tools, code execution, specialist models, retrieval systems, agents, or iterative plan-execute-inspect loops.","inclusion":"Assign work whose central objective is orchestration, delegation, routing, program generation, multi-agent collaboration, expert-model selection, or interactive multimodal action planning using multiple components beyond a single internally structured model.","exclusion":"If decomposition is only a natural-language rationale without tool execution, assign to rationale-based reasoning; if modules are internal learned or symbolic structures rather than external tools, assign to structured compositional reasoning; if the system is a broad integrated model without explicit orchestration, assign to broad model integration.","nearest_boundary":"Broad model integration and adaptation","representative_topics":["zero-shot ensemble composition of pre-trained models via iterative consensus optimization","neuro-symbolic compositional visual reasoning via in-context learning and modular program generation","prompt-based integration of language models and vision experts for multimodal reasoning and action"]}],"split":"dev"}
{"episode_id":"F229_T2025-01","field_index":229,"field":"multimodal reasoning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Diagnostic reasoning evaluation","description":"Create, validate, or analyze benchmarks, diagnostic datasets, probes, and evaluation protocols for measuring multimodal reasoning abilities, failures, consistency, robustness, and generalization.","inclusion":"Assign work whose primary objective is measurement of multimodal reasoning capability, including new benchmark suites, evaluation taxonomies, diagnostic tasks, or comparative capability analyses.","exclusion":"If the main contribution is a solver, training method, architecture, rationale method, tool system, or structured representation and evaluation is only validation, assign to that method-oriented direction.","nearest_boundary":"Rationale-based multimodal reasoning","representative_topics":["multimodal word problem solving requiring compositional language, visual, and mathematical reasoning","vision-language benchmarking for compositional generalization in embodied human activity forecasting","benchmarking deep model generalization on elementary visuo-linguistic reasoning tasks"]},{"direction_index":1,"label":"D1","name":"Rationale-based multimodal reasoning","description":"Develop methods that make multimodal inference proceed through explicit intermediate natural-language, quasi-symbolic, or explanation-like reasoning traces such as chains of thought, rationales, self-consistency signals, or visual infillings.","inclusion":"Assign work whose core mechanism is eliciting, supervising, generating, refining, distilling, or using intermediate rationale chains between multimodal inputs and final answers.","exclusion":"If decomposition produces executable programs, tool calls, or external subtask plans, assign to tool-augmented orchestration; if the intermediate structure is a persistent graph, module, or symbolic representation, assign to structured compositional reasoning.","nearest_boundary":"Tool-augmented orchestration","representative_topics":["multimodal chain-of-thought reasoning for interpretable science question answering","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","multimodal chain-of-thought augmentation via recursive visual infillings for temporal reasoning"]},{"direction_index":2,"label":"D2","name":"Abstract visual rule induction","description":"Solve abstract, formal, geometric, analogy, puzzle-like, or visual pattern reasoning tasks by inducing latent rules, transformations, programs, object relations, or algebraic/logical structures from visual stimuli.","inclusion":"Assign work whose primary objective is rule induction for abstract visual reasoning settings such as Raven-style matrices, ARC-like tasks, tangrams, visual analogies, constructible geometry, or abstract object-pattern completion.","exclusion":"If the same symbolic or structural machinery is applied mainly to natural-image vision-language QA, dialog, or grounded language tasks, assign to structured compositional reasoning; if the contribution is only a benchmark, assign to diagnostic evaluation.","nearest_boundary":"Structured compositional vision-language reasoning","representative_topics":["transformer-based spatial token reasoning for raven's progressive matrix rule induction","abstract visual reasoning over part-segmented shape stimuli with multimodal fine-tuning","visual analogy completion over complex scene understanding using situation recognition annotations"]},{"direction_index":3,"label":"D3","name":"Spatial and 3D grounding","description":"Improve multimodal understanding of spatial, regional, depth, orientation, positional, 3D, point-cloud, and spatio-temporal grounding relations.","inclusion":"Assign work whose primary objective is modeling, grounding, evaluating, or enhancing spatial or spatio-temporal relation comprehension, including 3D object grounding, pseudo-3D perception, region-specific prompts, spatial relation graphs, and target-aware audio-visual grounding.","exclusion":"If temporal causality, physical outcome prediction, or state-change dynamics dominate over spatial structure, assign to causal and dynamic reasoning; if spatial relations are merely one part of a general model architecture, assign to broad model integration.","nearest_boundary":"Causal and dynamic reasoning","representative_topics":["language-conditioned spatial relation reasoning for 3d object grounding in point clouds","depth-aware pseudo 3d perception framework for visual commonsense reasoning with multimodal large language models","target-aware spatio-temporal grounding for audio-visual question answering via joint temporal fusion"]},{"direction_index":4,"label":"D4","name":"Structured compositional vision-language reasoning","description":"Build multimodal reasoning systems that explicitly represent, compose, and manipulate objects, attributes, relations, scene graphs, neural modules, programs, queries, or symbolic structures for natural visual-linguistic inference.","inclusion":"Assign work whose main objective is compositional generalization or multi-step reasoning over natural images, video, dialog, or language-conditioned visual inputs using persistent structured representations or neuro-symbolic/internal modular mechanisms.","exclusion":"If the task is purely abstract visual puzzle rule induction, assign to abstract visual rule induction; if decomposition is mainly external tool orchestration, assign to tool-augmented orchestration; if no explicit compositional structure is central, assign to broad model integration.","nearest_boundary":"Broad model integration and adaptation","representative_topics":["neuro-symbolic visual dialog with program execution for long-distance co-reference and multi-round reasoning","multimodal visual reasoning via temporal relation construction and visual-linguistic association frameworks","generalization robustness of end-to-end versus neuro-symbolic vision-language reasoning under distribution shift"]},{"direction_index":5,"label":"D5","name":"Broad model integration and adaptation","description":"Improve general multimodal reasoning competence through broadly applicable architectures, representation unification, pretraining, cross-modal fusion, attention, knowledge integration, alignment, routing, foundation-model integration, or domain-shaped adaptation strategies.","inclusion":"Assign work whose primary objective is a general model, representation, pretraining, fusion, alignment, knowledge-enhancement, or adaptation strategy that improves or characterizes reasoning across modalities, including domain-constrained adaptations when the atom’s emphasis is system/model adaptation rather than a separate causal, spatial, or tool program.","exclusion":"If the central mechanism is explicit rationale chains, assign to rationale-based reasoning; if it is symbolic program/module execution over natural scenes, assign to structured compositional reasoning; if it orchestrates external tools or agents, assign to tool-augmented orchestration.","nearest_boundary":"Structured compositional vision-language reasoning","representative_topics":["survey of multimodal machine learning foundations covering representation, alignment, and generative challenges","zero-shot ensemble composition of pre-trained models via iterative consensus optimization","pretraining visual language models for chart and plot numerical reasoning via derendering"]},{"direction_index":6,"label":"D6","name":"Causal and dynamic reasoning","description":"Model causal, counterfactual, temporal, physical, action-effect, event-level, state-change, transformation, and behavior-forecasting relations across multimodal observations.","inclusion":"Assign work whose primary objective is understanding what happened, why it happened, what changed, what physical effect an action caused, or what may happen next in multimodal scenes, videos, image sequences, audio-visual settings, or embodied observations.","exclusion":"If the contribution is mainly a benchmark for such abilities, assign to diagnostic evaluation; if spatial layout or 3D grounding is the dominant objective, assign to spatial and 3D grounding; if event reasoning only validates a broader model integration method, assign to broad model integration.","nearest_boundary":"Spatial and 3D grounding","representative_topics":["active vision theory-inspired guided attention for dynamic visual reasoning and zero-shot generalization","cross-modal causal relational reasoning for event-level video visual question answering","multimodal action outcome prediction for physical commonsense reasoning in planning"]},{"direction_index":7,"label":"D7","name":"Tool-augmented orchestration","description":"Solve complex multimodal reasoning or action tasks by explicit decomposition into subproblems handled by external tools, code execution, specialist models, retrieval systems, agents, or iterative plan-execute-inspect loops.","inclusion":"Assign work whose central objective is orchestration, delegation, routing, program generation, multi-agent collaboration, expert-model selection, or interactive multimodal action planning using multiple components beyond a single internally structured model.","exclusion":"If decomposition is only a natural-language rationale without tool execution, assign to rationale-based reasoning; if modules are internal learned or symbolic structures rather than external tools, assign to structured compositional reasoning; if the system is a broad integrated model without explicit orchestration, assign to broad model integration.","nearest_boundary":"Broad model integration and adaptation","representative_topics":["zero-shot ensemble composition of pre-trained models via iterative consensus optimization","neuro-symbolic compositional visual reasoning via in-context learning and modular program generation","prompt-based integration of language models and vision experts for multimodal reasoning and action"]}],"split":"dev"}
{"episode_id":"F229_T2025-07","field_index":229,"field":"multimodal reasoning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Diagnostic reasoning evaluation","description":"Create, validate, or analyze benchmarks, diagnostic datasets, probes, and evaluation protocols for measuring multimodal reasoning abilities, failures, consistency, robustness, and generalization.","inclusion":"Assign work whose primary objective is measurement of multimodal reasoning capability, including new benchmark suites, evaluation taxonomies, diagnostic tasks, or comparative capability analyses.","exclusion":"If the main contribution is a solver, training method, architecture, rationale method, tool system, or structured representation and evaluation is only validation, assign to that method-oriented direction.","nearest_boundary":"Rationale-based multimodal reasoning","representative_topics":["multimodal word problem solving requiring compositional language, visual, and mathematical reasoning","vision-language benchmarking for compositional generalization in embodied human activity forecasting","benchmarking deep model generalization on elementary visuo-linguistic reasoning tasks"]},{"direction_index":1,"label":"D1","name":"Rationale-based multimodal reasoning","description":"Develop methods that make multimodal inference proceed through explicit intermediate natural-language, quasi-symbolic, or explanation-like reasoning traces such as chains of thought, rationales, self-consistency signals, or visual infillings.","inclusion":"Assign work whose core mechanism is eliciting, supervising, generating, refining, distilling, or using intermediate rationale chains between multimodal inputs and final answers.","exclusion":"If decomposition produces executable programs, tool calls, or external subtask plans, assign to tool-augmented orchestration; if the intermediate structure is a persistent graph, module, or symbolic representation, assign to structured compositional reasoning.","nearest_boundary":"Tool-augmented orchestration","representative_topics":["multimodal chain-of-thought reasoning for interpretable science question answering","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","multimodal chain-of-thought augmentation via recursive visual infillings for temporal reasoning"]},{"direction_index":2,"label":"D2","name":"Abstract visual rule induction","description":"Solve abstract, formal, geometric, analogy, puzzle-like, or visual pattern reasoning tasks by inducing latent rules, transformations, programs, object relations, or algebraic/logical structures from visual stimuli.","inclusion":"Assign work whose primary objective is rule induction for abstract visual reasoning settings such as Raven-style matrices, ARC-like tasks, tangrams, visual analogies, constructible geometry, or abstract object-pattern completion.","exclusion":"If the same symbolic or structural machinery is applied mainly to natural-image vision-language QA, dialog, or grounded language tasks, assign to structured compositional reasoning; if the contribution is only a benchmark, assign to diagnostic evaluation.","nearest_boundary":"Structured compositional vision-language reasoning","representative_topics":["transformer-based spatial token reasoning for raven's progressive matrix rule induction","abstract visual reasoning over part-segmented shape stimuli with multimodal fine-tuning","visual analogy completion over complex scene understanding using situation recognition annotations"]},{"direction_index":3,"label":"D3","name":"Spatial and 3D grounding","description":"Improve multimodal understanding of spatial, regional, depth, orientation, positional, 3D, point-cloud, and spatio-temporal grounding relations.","inclusion":"Assign work whose primary objective is modeling, grounding, evaluating, or enhancing spatial or spatio-temporal relation comprehension, including 3D object grounding, pseudo-3D perception, region-specific prompts, spatial relation graphs, and target-aware audio-visual grounding.","exclusion":"If temporal causality, physical outcome prediction, or state-change dynamics dominate over spatial structure, assign to causal and dynamic reasoning; if spatial relations are merely one part of a general model architecture, assign to broad model integration.","nearest_boundary":"Causal and dynamic reasoning","representative_topics":["language-conditioned spatial relation reasoning for 3d object grounding in point clouds","depth-aware pseudo 3d perception framework for visual commonsense reasoning with multimodal large language models","target-aware spatio-temporal grounding for audio-visual question answering via joint temporal fusion"]},{"direction_index":4,"label":"D4","name":"Structured compositional vision-language reasoning","description":"Build multimodal reasoning systems that explicitly represent, compose, and manipulate objects, attributes, relations, scene graphs, neural modules, programs, queries, or symbolic structures for natural visual-linguistic inference.","inclusion":"Assign work whose main objective is compositional generalization or multi-step reasoning over natural images, video, dialog, or language-conditioned visual inputs using persistent structured representations or neuro-symbolic/internal modular mechanisms.","exclusion":"If the task is purely abstract visual puzzle rule induction, assign to abstract visual rule induction; if decomposition is mainly external tool orchestration, assign to tool-augmented orchestration; if no explicit compositional structure is central, assign to broad model integration.","nearest_boundary":"Broad model integration and adaptation","representative_topics":["neuro-symbolic visual dialog with program execution for long-distance co-reference and multi-round reasoning","multimodal visual reasoning via temporal relation construction and visual-linguistic association frameworks","generalization robustness of end-to-end versus neuro-symbolic vision-language reasoning under distribution shift"]},{"direction_index":5,"label":"D5","name":"Broad model integration and adaptation","description":"Improve general multimodal reasoning competence through broadly applicable architectures, representation unification, pretraining, cross-modal fusion, attention, knowledge integration, alignment, routing, foundation-model integration, or domain-shaped adaptation strategies.","inclusion":"Assign work whose primary objective is a general model, representation, pretraining, fusion, alignment, knowledge-enhancement, or adaptation strategy that improves or characterizes reasoning across modalities, including domain-constrained adaptations when the atom’s emphasis is system/model adaptation rather than a separate causal, spatial, or tool program.","exclusion":"If the central mechanism is explicit rationale chains, assign to rationale-based reasoning; if it is symbolic program/module execution over natural scenes, assign to structured compositional reasoning; if it orchestrates external tools or agents, assign to tool-augmented orchestration.","nearest_boundary":"Structured compositional vision-language reasoning","representative_topics":["survey of multimodal machine learning foundations covering representation, alignment, and generative challenges","zero-shot ensemble composition of pre-trained models via iterative consensus optimization","pretraining visual language models for chart and plot numerical reasoning via derendering"]},{"direction_index":6,"label":"D6","name":"Causal and dynamic reasoning","description":"Model causal, counterfactual, temporal, physical, action-effect, event-level, state-change, transformation, and behavior-forecasting relations across multimodal observations.","inclusion":"Assign work whose primary objective is understanding what happened, why it happened, what changed, what physical effect an action caused, or what may happen next in multimodal scenes, videos, image sequences, audio-visual settings, or embodied observations.","exclusion":"If the contribution is mainly a benchmark for such abilities, assign to diagnostic evaluation; if spatial layout or 3D grounding is the dominant objective, assign to spatial and 3D grounding; if event reasoning only validates a broader model integration method, assign to broad model integration.","nearest_boundary":"Spatial and 3D grounding","representative_topics":["active vision theory-inspired guided attention for dynamic visual reasoning and zero-shot generalization","cross-modal causal relational reasoning for event-level video visual question answering","multimodal action outcome prediction for physical commonsense reasoning in planning"]},{"direction_index":7,"label":"D7","name":"Tool-augmented orchestration","description":"Solve complex multimodal reasoning or action tasks by explicit decomposition into subproblems handled by external tools, code execution, specialist models, retrieval systems, agents, or iterative plan-execute-inspect loops.","inclusion":"Assign work whose central objective is orchestration, delegation, routing, program generation, multi-agent collaboration, expert-model selection, or interactive multimodal action planning using multiple components beyond a single internally structured model.","exclusion":"If decomposition is only a natural-language rationale without tool execution, assign to rationale-based reasoning; if modules are internal learned or symbolic structures rather than external tools, assign to structured compositional reasoning; if the system is a broad integrated model without explicit orchestration, assign to broad model integration.","nearest_boundary":"Broad model integration and adaptation","representative_topics":["zero-shot ensemble composition of pre-trained models via iterative consensus optimization","neuro-symbolic compositional visual reasoning via in-context learning and modular program generation","prompt-based integration of language models and vision experts for multimodal reasoning and action"]}],"split":"dev"}
{"episode_id":"F229_T2026-01","field_index":229,"field":"multimodal reasoning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Diagnostic reasoning evaluation","description":"Create, validate, or analyze benchmarks, diagnostic datasets, probes, and evaluation protocols for measuring multimodal reasoning abilities, failures, consistency, robustness, and generalization.","inclusion":"Assign work whose primary objective is measurement of multimodal reasoning capability, including new benchmark suites, evaluation taxonomies, diagnostic tasks, or comparative capability analyses.","exclusion":"If the main contribution is a solver, training method, architecture, rationale method, tool system, or structured representation and evaluation is only validation, assign to that method-oriented direction.","nearest_boundary":"Rationale-based multimodal reasoning","representative_topics":["multimodal word problem solving requiring compositional language, visual, and mathematical reasoning","vision-language benchmarking for compositional generalization in embodied human activity forecasting","benchmarking deep model generalization on elementary visuo-linguistic reasoning tasks"]},{"direction_index":1,"label":"D1","name":"Rationale-based multimodal reasoning","description":"Develop methods that make multimodal inference proceed through explicit intermediate natural-language, quasi-symbolic, or explanation-like reasoning traces such as chains of thought, rationales, self-consistency signals, or visual infillings.","inclusion":"Assign work whose core mechanism is eliciting, supervising, generating, refining, distilling, or using intermediate rationale chains between multimodal inputs and final answers.","exclusion":"If decomposition produces executable programs, tool calls, or external subtask plans, assign to tool-augmented orchestration; if the intermediate structure is a persistent graph, module, or symbolic representation, assign to structured compositional reasoning.","nearest_boundary":"Tool-augmented orchestration","representative_topics":["multimodal chain-of-thought reasoning for interpretable science question answering","multimodal chain-of-thought reasoning with decoupled rationale and inference stages","multimodal chain-of-thought augmentation via recursive visual infillings for temporal reasoning"]},{"direction_index":2,"label":"D2","name":"Abstract visual rule induction","description":"Solve abstract, formal, geometric, analogy, puzzle-like, or visual pattern reasoning tasks by inducing latent rules, transformations, programs, object relations, or algebraic/logical structures from visual stimuli.","inclusion":"Assign work whose primary objective is rule induction for abstract visual reasoning settings such as Raven-style matrices, ARC-like tasks, tangrams, visual analogies, constructible geometry, or abstract object-pattern completion.","exclusion":"If the same symbolic or structural machinery is applied mainly to natural-image vision-language QA, dialog, or grounded language tasks, assign to structured compositional reasoning; if the contribution is only a benchmark, assign to diagnostic evaluation.","nearest_boundary":"Structured compositional vision-language reasoning","representative_topics":["transformer-based spatial token reasoning for raven's progressive matrix rule induction","abstract visual reasoning over part-segmented shape stimuli with multimodal fine-tuning","visual analogy completion over complex scene understanding using situation recognition annotations"]},{"direction_index":3,"label":"D3","name":"Spatial and 3D grounding","description":"Improve multimodal understanding of spatial, regional, depth, orientation, positional, 3D, point-cloud, and spatio-temporal grounding relations.","inclusion":"Assign work whose primary objective is modeling, grounding, evaluating, or enhancing spatial or spatio-temporal relation comprehension, including 3D object grounding, pseudo-3D perception, region-specific prompts, spatial relation graphs, and target-aware audio-visual grounding.","exclusion":"If temporal causality, physical outcome prediction, or state-change dynamics dominate over spatial structure, assign to causal and dynamic reasoning; if spatial relations are merely one part of a general model architecture, assign to broad model integration.","nearest_boundary":"Causal and dynamic reasoning","representative_topics":["language-conditioned spatial relation reasoning for 3d object grounding in point clouds","depth-aware pseudo 3d perception framework for visual commonsense reasoning with multimodal large language models","target-aware spatio-temporal grounding for audio-visual question answering via joint temporal fusion"]},{"direction_index":4,"label":"D4","name":"Structured compositional vision-language reasoning","description":"Build multimodal reasoning systems that explicitly represent, compose, and manipulate objects, attributes, relations, scene graphs, neural modules, programs, queries, or symbolic structures for natural visual-linguistic inference.","inclusion":"Assign work whose main objective is compositional generalization or multi-step reasoning over natural images, video, dialog, or language-conditioned visual inputs using persistent structured representations or neuro-symbolic/internal modular mechanisms.","exclusion":"If the task is purely abstract visual puzzle rule induction, assign to abstract visual rule induction; if decomposition is mainly external tool orchestration, assign to tool-augmented orchestration; if no explicit compositional structure is central, assign to broad model integration.","nearest_boundary":"Broad model integration and adaptation","representative_topics":["neuro-symbolic visual dialog with program execution for long-distance co-reference and multi-round reasoning","multimodal visual reasoning via temporal relation construction and visual-linguistic association frameworks","generalization robustness of end-to-end versus neuro-symbolic vision-language reasoning under distribution shift"]},{"direction_index":5,"label":"D5","name":"Broad model integration and adaptation","description":"Improve general multimodal reasoning competence through broadly applicable architectures, representation unification, pretraining, cross-modal fusion, attention, knowledge integration, alignment, routing, foundation-model integration, or domain-shaped adaptation strategies.","inclusion":"Assign work whose primary objective is a general model, representation, pretraining, fusion, alignment, knowledge-enhancement, or adaptation strategy that improves or characterizes reasoning across modalities, including domain-constrained adaptations when the atom’s emphasis is system/model adaptation rather than a separate causal, spatial, or tool program.","exclusion":"If the central mechanism is explicit rationale chains, assign to rationale-based reasoning; if it is symbolic program/module execution over natural scenes, assign to structured compositional reasoning; if it orchestrates external tools or agents, assign to tool-augmented orchestration.","nearest_boundary":"Structured compositional vision-language reasoning","representative_topics":["survey of multimodal machine learning foundations covering representation, alignment, and generative challenges","zero-shot ensemble composition of pre-trained models via iterative consensus optimization","pretraining visual language models for chart and plot numerical reasoning via derendering"]},{"direction_index":6,"label":"D6","name":"Causal and dynamic reasoning","description":"Model causal, counterfactual, temporal, physical, action-effect, event-level, state-change, transformation, and behavior-forecasting relations across multimodal observations.","inclusion":"Assign work whose primary objective is understanding what happened, why it happened, what changed, what physical effect an action caused, or what may happen next in multimodal scenes, videos, image sequences, audio-visual settings, or embodied observations.","exclusion":"If the contribution is mainly a benchmark for such abilities, assign to diagnostic evaluation; if spatial layout or 3D grounding is the dominant objective, assign to spatial and 3D grounding; if event reasoning only validates a broader model integration method, assign to broad model integration.","nearest_boundary":"Spatial and 3D grounding","representative_topics":["active vision theory-inspired guided attention for dynamic visual reasoning and zero-shot generalization","cross-modal causal relational reasoning for event-level video visual question answering","multimodal action outcome prediction for physical commonsense reasoning in planning"]},{"direction_index":7,"label":"D7","name":"Tool-augmented orchestration","description":"Solve complex multimodal reasoning or action tasks by explicit decomposition into subproblems handled by external tools, code execution, specialist models, retrieval systems, agents, or iterative plan-execute-inspect loops.","inclusion":"Assign work whose central objective is orchestration, delegation, routing, program generation, multi-agent collaboration, expert-model selection, or interactive multimodal action planning using multiple components beyond a single internally structured model.","exclusion":"If decomposition is only a natural-language rationale without tool execution, assign to rationale-based reasoning; if modules are internal learned or symbolic structures rather than external tools, assign to structured compositional reasoning; if the system is a broad integrated model without explicit orchestration, assign to broad model integration.","nearest_boundary":"Broad model integration and adaptation","representative_topics":["zero-shot ensemble composition of pre-trained models via iterative consensus optimization","neuro-symbolic compositional visual reasoning via in-context learning and modular program generation","prompt-based integration of language models and vision experts for multimodal reasoning and action"]}],"split":"dev"}
{"episode_id":"F232_T2024-01","field_index":232,"field":"stochastic optimization","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Core SGD and stochastic-approximation theory","description":"General-purpose theory for SGD-like and stochastic-approximation methods, emphasizing convergence, rates, asymptotics, concentration, stability thresholds, inference for iterates, and sample complexity when no specialized structural objective dominates.","inclusion":"Include source atoms whose main program is broadly applicable analysis of SGD, momentum SGD as a basic variant, Polyak-Ruppert averaging, implicit SGD, coupled stochastic approximation, Markovian or decision-dependent stochastic approximation, primal-dual splitting, or ERM/SCO sample complexity under standard or relaxed assumptions.","exclusion":"If the dominant objective is automated step-size or preconditioning design, distributed communication, heavy-tail robustification, finite-sum variance reduction, SDE/sampling dynamics, generalization, privacy, constraints, or saddle/compositional structure, assign to the corresponding specialized direction instead.","nearest_boundary":"D1","representative_topics":["single-timescale convergence analysis for nonlinear stochastic approximation with multiple coupled sequences","statistical inference for implicit sgd via proximal robbins-monro and polyak-ruppert procedures","stochastic approximation with decision-dependent distributions: asymptotic normality and minimax optimality"]},{"direction_index":1,"label":"D1","name":"Adaptive and momentum first-order methods","description":"Design and theory of stochastic first-order optimizers whose main objective is automated learning-rate control, adaptive scaling, diagonal preconditioning, sign updates, momentum mechanisms, or first-order acceleration and stabilization.","inclusion":"Include Adam, AdaGrad, sign-based methods, Polyak/function-value step sizes, parameter-free schedules, preconditioned first-order methods, adaptive momentum, heavy-ball or Nesterov-style stochastic momentum, and related first-order optimizer designs when adaptivity or momentum design is central.","exclusion":"If the method explicitly models Hessian, Fisher, Newton, quasi-Newton, cubic, trust-region, or other curvature structure, assign to D2; if adaptivity is mainly a distributed-compression mechanism, assign to D3.","nearest_boundary":"D2","representative_topics":["preconditioned stochastic gradient methods for ill-conditioned non-convex finite sum minimization","non-monotone adaptive gradient methods with bidirectional step size accumulation","lipschitz-free convergence analysis of adam with hyperparameters close to one"]},{"direction_index":2,"label":"D2","name":"Stochastic curvature and second-order optimization","description":"Algorithms and convergence theory for stochastic optimization methods that explicitly use or approximate curvature information, including Hessian, Fisher, Newton, quasi-Newton, cubic regularization, trust-region, subspace, Kronecker, or diagonal Hessian structures.","inclusion":"Include stochastic Newton, natural-gradient, cubic-regularized Newton, trust-region, quasi-Newton, Hessian-vector, dimension-reduced second-order, scalable curvature approximation, and second-order Polyak-style methods when curvature modeling is the primary scientific objective.","exclusion":"If curvature-like information is only diagonal gradient-moment scaling, ordinary preconditioning, or learning-rate adaptation without an explicit second-order modeling objective, assign to D1.","nearest_boundary":"D1","representative_topics":["second-order stochastic polyak step size using hessian-vector products for non-convex optimization","noise-adaptive second-order methods with unified stochastic and deterministic oracle convergence","stochastic second-order policy optimization via dimension-reduced trust region methods"]},{"direction_index":3,"label":"D3","name":"Distributed and communication-efficient optimization","description":"Algorithms and theory for stochastic optimization across multiple workers, clients, devices, or network nodes, with emphasis on communication, compression, consensus, asynchrony, delays, heterogeneity, client/data sampling, and decentralized or federated execution.","inclusion":"Include distributed, decentralized, federated, asynchronous, delayed, compressed, heterogeneous, or multi-agent stochastic optimization when the distributed systems constraint is the primary contribution, including distributed variants of SGD, minimax, compositional, or variance-reduced methods.","exclusion":"If privacy guarantees are the main objective rather than distributed execution, assign to D7; if saddle/compositional structure dominates and distribution is only an implementation setting, assign to D7.","nearest_boundary":"D7","representative_topics":["sharp sgd convergence rates under flexible client and data sampling in nonconvex optimization","federated gradient compression with random reshuffling and variance reduction for distributed training","asynchronous sgd convergence analysis under arbitrary delays via virtual iterate recursion"]},{"direction_index":4,"label":"D4","name":"Robust optimization under heavy-tailed noise","description":"Stochastic optimization under heavy-tailed, weak-moment, adversarial, outlier-prone, unbounded-variance, or otherwise unreliable stochastic noise, emphasizing robust convergence and high-probability guarantees.","inclusion":"Include clipping, median-of-means, robust aggregation, robust variance reduction, adaptive robust rates, and high-probability analysis when the main scientific objective is reliable optimization under heavy-tailed or corrupted stochastic gradients or losses.","exclusion":"If clipping or moment control is primarily for differential privacy, assign to D7; if heavy tails are primarily studied as a continuous-time limiting law, stationary distribution, or generalization mechanism, assign to D6.","nearest_boundary":"D6","representative_topics":["clipped stochastic extragradient methods for variational inequalities with heavy-tailed noise and unbounded domains","high-probability convergence of stochastic optimization under unbounded variance gradient noise","variance-reduced acceleration in non-convex stochastic optimization with heavy-tailed noise"]},{"direction_index":5,"label":"D5","name":"Nonstandard oracles, geometry, and nonsmoothness","description":"Stochastic optimization when standard smooth Euclidean first-order assumptions fail because gradients are unavailable, indirect, zeroth-order, bandit-like, nonsmooth, nonconvex-composite, Riemannian, manifold-constrained, or geometrically constrained.","inclusion":"Include zeroth-order, gradient-free, simultaneous-perturbation, bandit or semi-bandit, randomized directional, stochastic subgradient, nonsmooth nonconvex proximal, Riemannian, retraction/vector-transport, manifold, and stochastic interior-point methods when oracle limitations, nonsmoothness, or feasible-space geometry is the dominant objective.","exclusion":"If ordinary gradients are available and the main objective is estimator efficiency or sampling order, assign to D6; if constraints are expectation, risk, safety, privacy, or decision-modeling constraints rather than geometry or oracle access, assign to D7.","nearest_boundary":"D7","representative_topics":["structured stochastic zeroth-order descent with orthogonal direction sampling and polyak-łojasiewicz convergence","stochastic approximation convergence on riemannian manifolds via fermi coordinate dynamical systems","riemannian stochastic gradient methods for nested compositional optimization with bias-corrected inner approximation"]},{"direction_index":6,"label":"D6","name":"Estimator efficiency, dynamics, and generalization","description":"Research explaining or improving stochastic optimization through estimator variance structure, finite-sum sampling order, stochastic-process dynamics, sampling behavior, implicit regularization, stability, and generalization properties.","inclusion":"Include variance reduction, finite-sum incremental methods, random reshuffling, sampling without replacement, recursive estimators, proximal finite-sum algorithms, SDE and diffusion limits, Langevin/SGLD methods, stationary laws, process-level approximations, implicit bias, flat-minimum selection, risk trajectories, and stability-based generalization when the central objective is estimator dynamics or statistical behavior beyond basic convergence.","exclusion":"If the main objective is robustifying against heavy-tailed noise, assign to D4; if the main objective is solving a saddle, constrained, private, or decision-structured formulation, assign to D7.","nearest_boundary":"D4","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","high-dimensional sgd limit theorems with critical step-size scaling and effective dynamics","stability and generalization bounds for stochastic gradient methods on nonconvex nonsmooth objectives"]},{"direction_index":7,"label":"D7","name":"Structured, constrained, private, and decision-aware optimization","description":"Stochastic optimization programs whose primary objective is solving structured formulations beyond unconstrained single-level minimization, including saddle-point, variational-inequality, minimax, bilevel, compositional, constrained, risk-sensitive, private, safe, distributionally robust, multistage, online, or decision-focused problems.","inclusion":"Include stochastic minimax, gradient-descent-ascent, extragradient, VI, bilevel, nested or multi-level compositional objectives, expectation or compositional constraints, augmented-Lagrangian, penalty, barrier, safe black-box, risk-measure, DRO, differential privacy, user-level privacy, performative or decision-dependent optimization, multistage stochastic programming, surrogate decision models, bandit decision-making, and estimation-optimization integration when this formulation-level structure is central.","exclusion":"If the dominant contribution is distributed communication rather than the structured formulation, assign to D3; if the dominant issue is manifold geometry, zeroth-order access, or nonsmooth oracle structure rather than constraints/privacy/decisions/saddles, assign to D5.","nearest_boundary":"D5","representative_topics":["differentially private stochastic convex optimization via averaged clipping for heavy-tailed data","variance reduced estimation for coupled compositional optimization with sparse block probing","near-optimal stochastic first-order oracle complexity for minimax optimization via recursive anchored iteration"]}],"split":"dev"}
{"episode_id":"F232_T2024-07","field_index":232,"field":"stochastic optimization","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Core SGD and stochastic-approximation theory","description":"General-purpose theory for SGD-like and stochastic-approximation methods, emphasizing convergence, rates, asymptotics, concentration, stability thresholds, inference for iterates, and sample complexity when no specialized structural objective dominates.","inclusion":"Include source atoms whose main program is broadly applicable analysis of SGD, momentum SGD as a basic variant, Polyak-Ruppert averaging, implicit SGD, coupled stochastic approximation, Markovian or decision-dependent stochastic approximation, primal-dual splitting, or ERM/SCO sample complexity under standard or relaxed assumptions.","exclusion":"If the dominant objective is automated step-size or preconditioning design, distributed communication, heavy-tail robustification, finite-sum variance reduction, SDE/sampling dynamics, generalization, privacy, constraints, or saddle/compositional structure, assign to the corresponding specialized direction instead.","nearest_boundary":"D1","representative_topics":["single-timescale convergence analysis for nonlinear stochastic approximation with multiple coupled sequences","statistical inference for implicit sgd via proximal robbins-monro and polyak-ruppert procedures","stochastic approximation with decision-dependent distributions: asymptotic normality and minimax optimality"]},{"direction_index":1,"label":"D1","name":"Adaptive and momentum first-order methods","description":"Design and theory of stochastic first-order optimizers whose main objective is automated learning-rate control, adaptive scaling, diagonal preconditioning, sign updates, momentum mechanisms, or first-order acceleration and stabilization.","inclusion":"Include Adam, AdaGrad, sign-based methods, Polyak/function-value step sizes, parameter-free schedules, preconditioned first-order methods, adaptive momentum, heavy-ball or Nesterov-style stochastic momentum, and related first-order optimizer designs when adaptivity or momentum design is central.","exclusion":"If the method explicitly models Hessian, Fisher, Newton, quasi-Newton, cubic, trust-region, or other curvature structure, assign to D2; if adaptivity is mainly a distributed-compression mechanism, assign to D3.","nearest_boundary":"D2","representative_topics":["preconditioned stochastic gradient methods for ill-conditioned non-convex finite sum minimization","non-monotone adaptive gradient methods with bidirectional step size accumulation","lipschitz-free convergence analysis of adam with hyperparameters close to one"]},{"direction_index":2,"label":"D2","name":"Stochastic curvature and second-order optimization","description":"Algorithms and convergence theory for stochastic optimization methods that explicitly use or approximate curvature information, including Hessian, Fisher, Newton, quasi-Newton, cubic regularization, trust-region, subspace, Kronecker, or diagonal Hessian structures.","inclusion":"Include stochastic Newton, natural-gradient, cubic-regularized Newton, trust-region, quasi-Newton, Hessian-vector, dimension-reduced second-order, scalable curvature approximation, and second-order Polyak-style methods when curvature modeling is the primary scientific objective.","exclusion":"If curvature-like information is only diagonal gradient-moment scaling, ordinary preconditioning, or learning-rate adaptation without an explicit second-order modeling objective, assign to D1.","nearest_boundary":"D1","representative_topics":["second-order stochastic polyak step size using hessian-vector products for non-convex optimization","noise-adaptive second-order methods with unified stochastic and deterministic oracle convergence","stochastic second-order policy optimization via dimension-reduced trust region methods"]},{"direction_index":3,"label":"D3","name":"Distributed and communication-efficient optimization","description":"Algorithms and theory for stochastic optimization across multiple workers, clients, devices, or network nodes, with emphasis on communication, compression, consensus, asynchrony, delays, heterogeneity, client/data sampling, and decentralized or federated execution.","inclusion":"Include distributed, decentralized, federated, asynchronous, delayed, compressed, heterogeneous, or multi-agent stochastic optimization when the distributed systems constraint is the primary contribution, including distributed variants of SGD, minimax, compositional, or variance-reduced methods.","exclusion":"If privacy guarantees are the main objective rather than distributed execution, assign to D7; if saddle/compositional structure dominates and distribution is only an implementation setting, assign to D7.","nearest_boundary":"D7","representative_topics":["sharp sgd convergence rates under flexible client and data sampling in nonconvex optimization","federated gradient compression with random reshuffling and variance reduction for distributed training","asynchronous sgd convergence analysis under arbitrary delays via virtual iterate recursion"]},{"direction_index":4,"label":"D4","name":"Robust optimization under heavy-tailed noise","description":"Stochastic optimization under heavy-tailed, weak-moment, adversarial, outlier-prone, unbounded-variance, or otherwise unreliable stochastic noise, emphasizing robust convergence and high-probability guarantees.","inclusion":"Include clipping, median-of-means, robust aggregation, robust variance reduction, adaptive robust rates, and high-probability analysis when the main scientific objective is reliable optimization under heavy-tailed or corrupted stochastic gradients or losses.","exclusion":"If clipping or moment control is primarily for differential privacy, assign to D7; if heavy tails are primarily studied as a continuous-time limiting law, stationary distribution, or generalization mechanism, assign to D6.","nearest_boundary":"D6","representative_topics":["clipped stochastic extragradient methods for variational inequalities with heavy-tailed noise and unbounded domains","high-probability convergence of stochastic optimization under unbounded variance gradient noise","variance-reduced acceleration in non-convex stochastic optimization with heavy-tailed noise"]},{"direction_index":5,"label":"D5","name":"Nonstandard oracles, geometry, and nonsmoothness","description":"Stochastic optimization when standard smooth Euclidean first-order assumptions fail because gradients are unavailable, indirect, zeroth-order, bandit-like, nonsmooth, nonconvex-composite, Riemannian, manifold-constrained, or geometrically constrained.","inclusion":"Include zeroth-order, gradient-free, simultaneous-perturbation, bandit or semi-bandit, randomized directional, stochastic subgradient, nonsmooth nonconvex proximal, Riemannian, retraction/vector-transport, manifold, and stochastic interior-point methods when oracle limitations, nonsmoothness, or feasible-space geometry is the dominant objective.","exclusion":"If ordinary gradients are available and the main objective is estimator efficiency or sampling order, assign to D6; if constraints are expectation, risk, safety, privacy, or decision-modeling constraints rather than geometry or oracle access, assign to D7.","nearest_boundary":"D7","representative_topics":["structured stochastic zeroth-order descent with orthogonal direction sampling and polyak-łojasiewicz convergence","stochastic approximation convergence on riemannian manifolds via fermi coordinate dynamical systems","riemannian stochastic gradient methods for nested compositional optimization with bias-corrected inner approximation"]},{"direction_index":6,"label":"D6","name":"Estimator efficiency, dynamics, and generalization","description":"Research explaining or improving stochastic optimization through estimator variance structure, finite-sum sampling order, stochastic-process dynamics, sampling behavior, implicit regularization, stability, and generalization properties.","inclusion":"Include variance reduction, finite-sum incremental methods, random reshuffling, sampling without replacement, recursive estimators, proximal finite-sum algorithms, SDE and diffusion limits, Langevin/SGLD methods, stationary laws, process-level approximations, implicit bias, flat-minimum selection, risk trajectories, and stability-based generalization when the central objective is estimator dynamics or statistical behavior beyond basic convergence.","exclusion":"If the main objective is robustifying against heavy-tailed noise, assign to D4; if the main objective is solving a saddle, constrained, private, or decision-structured formulation, assign to D7.","nearest_boundary":"D4","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","high-dimensional sgd limit theorems with critical step-size scaling and effective dynamics","stability and generalization bounds for stochastic gradient methods on nonconvex nonsmooth objectives"]},{"direction_index":7,"label":"D7","name":"Structured, constrained, private, and decision-aware optimization","description":"Stochastic optimization programs whose primary objective is solving structured formulations beyond unconstrained single-level minimization, including saddle-point, variational-inequality, minimax, bilevel, compositional, constrained, risk-sensitive, private, safe, distributionally robust, multistage, online, or decision-focused problems.","inclusion":"Include stochastic minimax, gradient-descent-ascent, extragradient, VI, bilevel, nested or multi-level compositional objectives, expectation or compositional constraints, augmented-Lagrangian, penalty, barrier, safe black-box, risk-measure, DRO, differential privacy, user-level privacy, performative or decision-dependent optimization, multistage stochastic programming, surrogate decision models, bandit decision-making, and estimation-optimization integration when this formulation-level structure is central.","exclusion":"If the dominant contribution is distributed communication rather than the structured formulation, assign to D3; if the dominant issue is manifold geometry, zeroth-order access, or nonsmooth oracle structure rather than constraints/privacy/decisions/saddles, assign to D5.","nearest_boundary":"D5","representative_topics":["differentially private stochastic convex optimization via averaged clipping for heavy-tailed data","variance reduced estimation for coupled compositional optimization with sparse block probing","near-optimal stochastic first-order oracle complexity for minimax optimization via recursive anchored iteration"]}],"split":"dev"}
{"episode_id":"F232_T2025-01","field_index":232,"field":"stochastic optimization","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Core SGD and stochastic-approximation theory","description":"General-purpose theory for SGD-like and stochastic-approximation methods, emphasizing convergence, rates, asymptotics, concentration, stability thresholds, inference for iterates, and sample complexity when no specialized structural objective dominates.","inclusion":"Include source atoms whose main program is broadly applicable analysis of SGD, momentum SGD as a basic variant, Polyak-Ruppert averaging, implicit SGD, coupled stochastic approximation, Markovian or decision-dependent stochastic approximation, primal-dual splitting, or ERM/SCO sample complexity under standard or relaxed assumptions.","exclusion":"If the dominant objective is automated step-size or preconditioning design, distributed communication, heavy-tail robustification, finite-sum variance reduction, SDE/sampling dynamics, generalization, privacy, constraints, or saddle/compositional structure, assign to the corresponding specialized direction instead.","nearest_boundary":"D1","representative_topics":["single-timescale convergence analysis for nonlinear stochastic approximation with multiple coupled sequences","statistical inference for implicit sgd via proximal robbins-monro and polyak-ruppert procedures","stochastic approximation with decision-dependent distributions: asymptotic normality and minimax optimality"]},{"direction_index":1,"label":"D1","name":"Adaptive and momentum first-order methods","description":"Design and theory of stochastic first-order optimizers whose main objective is automated learning-rate control, adaptive scaling, diagonal preconditioning, sign updates, momentum mechanisms, or first-order acceleration and stabilization.","inclusion":"Include Adam, AdaGrad, sign-based methods, Polyak/function-value step sizes, parameter-free schedules, preconditioned first-order methods, adaptive momentum, heavy-ball or Nesterov-style stochastic momentum, and related first-order optimizer designs when adaptivity or momentum design is central.","exclusion":"If the method explicitly models Hessian, Fisher, Newton, quasi-Newton, cubic, trust-region, or other curvature structure, assign to D2; if adaptivity is mainly a distributed-compression mechanism, assign to D3.","nearest_boundary":"D2","representative_topics":["preconditioned stochastic gradient methods for ill-conditioned non-convex finite sum minimization","non-monotone adaptive gradient methods with bidirectional step size accumulation","lipschitz-free convergence analysis of adam with hyperparameters close to one"]},{"direction_index":2,"label":"D2","name":"Stochastic curvature and second-order optimization","description":"Algorithms and convergence theory for stochastic optimization methods that explicitly use or approximate curvature information, including Hessian, Fisher, Newton, quasi-Newton, cubic regularization, trust-region, subspace, Kronecker, or diagonal Hessian structures.","inclusion":"Include stochastic Newton, natural-gradient, cubic-regularized Newton, trust-region, quasi-Newton, Hessian-vector, dimension-reduced second-order, scalable curvature approximation, and second-order Polyak-style methods when curvature modeling is the primary scientific objective.","exclusion":"If curvature-like information is only diagonal gradient-moment scaling, ordinary preconditioning, or learning-rate adaptation without an explicit second-order modeling objective, assign to D1.","nearest_boundary":"D1","representative_topics":["second-order stochastic polyak step size using hessian-vector products for non-convex optimization","noise-adaptive second-order methods with unified stochastic and deterministic oracle convergence","stochastic second-order policy optimization via dimension-reduced trust region methods"]},{"direction_index":3,"label":"D3","name":"Distributed and communication-efficient optimization","description":"Algorithms and theory for stochastic optimization across multiple workers, clients, devices, or network nodes, with emphasis on communication, compression, consensus, asynchrony, delays, heterogeneity, client/data sampling, and decentralized or federated execution.","inclusion":"Include distributed, decentralized, federated, asynchronous, delayed, compressed, heterogeneous, or multi-agent stochastic optimization when the distributed systems constraint is the primary contribution, including distributed variants of SGD, minimax, compositional, or variance-reduced methods.","exclusion":"If privacy guarantees are the main objective rather than distributed execution, assign to D7; if saddle/compositional structure dominates and distribution is only an implementation setting, assign to D7.","nearest_boundary":"D7","representative_topics":["sharp sgd convergence rates under flexible client and data sampling in nonconvex optimization","federated gradient compression with random reshuffling and variance reduction for distributed training","asynchronous sgd convergence analysis under arbitrary delays via virtual iterate recursion"]},{"direction_index":4,"label":"D4","name":"Robust optimization under heavy-tailed noise","description":"Stochastic optimization under heavy-tailed, weak-moment, adversarial, outlier-prone, unbounded-variance, or otherwise unreliable stochastic noise, emphasizing robust convergence and high-probability guarantees.","inclusion":"Include clipping, median-of-means, robust aggregation, robust variance reduction, adaptive robust rates, and high-probability analysis when the main scientific objective is reliable optimization under heavy-tailed or corrupted stochastic gradients or losses.","exclusion":"If clipping or moment control is primarily for differential privacy, assign to D7; if heavy tails are primarily studied as a continuous-time limiting law, stationary distribution, or generalization mechanism, assign to D6.","nearest_boundary":"D6","representative_topics":["clipped stochastic extragradient methods for variational inequalities with heavy-tailed noise and unbounded domains","high-probability convergence of stochastic optimization under unbounded variance gradient noise","variance-reduced acceleration in non-convex stochastic optimization with heavy-tailed noise"]},{"direction_index":5,"label":"D5","name":"Nonstandard oracles, geometry, and nonsmoothness","description":"Stochastic optimization when standard smooth Euclidean first-order assumptions fail because gradients are unavailable, indirect, zeroth-order, bandit-like, nonsmooth, nonconvex-composite, Riemannian, manifold-constrained, or geometrically constrained.","inclusion":"Include zeroth-order, gradient-free, simultaneous-perturbation, bandit or semi-bandit, randomized directional, stochastic subgradient, nonsmooth nonconvex proximal, Riemannian, retraction/vector-transport, manifold, and stochastic interior-point methods when oracle limitations, nonsmoothness, or feasible-space geometry is the dominant objective.","exclusion":"If ordinary gradients are available and the main objective is estimator efficiency or sampling order, assign to D6; if constraints are expectation, risk, safety, privacy, or decision-modeling constraints rather than geometry or oracle access, assign to D7.","nearest_boundary":"D7","representative_topics":["structured stochastic zeroth-order descent with orthogonal direction sampling and polyak-łojasiewicz convergence","stochastic approximation convergence on riemannian manifolds via fermi coordinate dynamical systems","riemannian stochastic gradient methods for nested compositional optimization with bias-corrected inner approximation"]},{"direction_index":6,"label":"D6","name":"Estimator efficiency, dynamics, and generalization","description":"Research explaining or improving stochastic optimization through estimator variance structure, finite-sum sampling order, stochastic-process dynamics, sampling behavior, implicit regularization, stability, and generalization properties.","inclusion":"Include variance reduction, finite-sum incremental methods, random reshuffling, sampling without replacement, recursive estimators, proximal finite-sum algorithms, SDE and diffusion limits, Langevin/SGLD methods, stationary laws, process-level approximations, implicit bias, flat-minimum selection, risk trajectories, and stability-based generalization when the central objective is estimator dynamics or statistical behavior beyond basic convergence.","exclusion":"If the main objective is robustifying against heavy-tailed noise, assign to D4; if the main objective is solving a saddle, constrained, private, or decision-structured formulation, assign to D7.","nearest_boundary":"D4","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","high-dimensional sgd limit theorems with critical step-size scaling and effective dynamics","stability and generalization bounds for stochastic gradient methods on nonconvex nonsmooth objectives"]},{"direction_index":7,"label":"D7","name":"Structured, constrained, private, and decision-aware optimization","description":"Stochastic optimization programs whose primary objective is solving structured formulations beyond unconstrained single-level minimization, including saddle-point, variational-inequality, minimax, bilevel, compositional, constrained, risk-sensitive, private, safe, distributionally robust, multistage, online, or decision-focused problems.","inclusion":"Include stochastic minimax, gradient-descent-ascent, extragradient, VI, bilevel, nested or multi-level compositional objectives, expectation or compositional constraints, augmented-Lagrangian, penalty, barrier, safe black-box, risk-measure, DRO, differential privacy, user-level privacy, performative or decision-dependent optimization, multistage stochastic programming, surrogate decision models, bandit decision-making, and estimation-optimization integration when this formulation-level structure is central.","exclusion":"If the dominant contribution is distributed communication rather than the structured formulation, assign to D3; if the dominant issue is manifold geometry, zeroth-order access, or nonsmooth oracle structure rather than constraints/privacy/decisions/saddles, assign to D5.","nearest_boundary":"D5","representative_topics":["differentially private stochastic convex optimization via averaged clipping for heavy-tailed data","variance reduced estimation for coupled compositional optimization with sparse block probing","near-optimal stochastic first-order oracle complexity for minimax optimization via recursive anchored iteration"]}],"split":"dev"}
{"episode_id":"F232_T2025-07","field_index":232,"field":"stochastic optimization","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Core SGD and stochastic-approximation theory","description":"General-purpose theory for SGD-like and stochastic-approximation methods, emphasizing convergence, rates, asymptotics, concentration, stability thresholds, inference for iterates, and sample complexity when no specialized structural objective dominates.","inclusion":"Include source atoms whose main program is broadly applicable analysis of SGD, momentum SGD as a basic variant, Polyak-Ruppert averaging, implicit SGD, coupled stochastic approximation, Markovian or decision-dependent stochastic approximation, primal-dual splitting, or ERM/SCO sample complexity under standard or relaxed assumptions.","exclusion":"If the dominant objective is automated step-size or preconditioning design, distributed communication, heavy-tail robustification, finite-sum variance reduction, SDE/sampling dynamics, generalization, privacy, constraints, or saddle/compositional structure, assign to the corresponding specialized direction instead.","nearest_boundary":"D1","representative_topics":["single-timescale convergence analysis for nonlinear stochastic approximation with multiple coupled sequences","statistical inference for implicit sgd via proximal robbins-monro and polyak-ruppert procedures","stochastic approximation with decision-dependent distributions: asymptotic normality and minimax optimality"]},{"direction_index":1,"label":"D1","name":"Adaptive and momentum first-order methods","description":"Design and theory of stochastic first-order optimizers whose main objective is automated learning-rate control, adaptive scaling, diagonal preconditioning, sign updates, momentum mechanisms, or first-order acceleration and stabilization.","inclusion":"Include Adam, AdaGrad, sign-based methods, Polyak/function-value step sizes, parameter-free schedules, preconditioned first-order methods, adaptive momentum, heavy-ball or Nesterov-style stochastic momentum, and related first-order optimizer designs when adaptivity or momentum design is central.","exclusion":"If the method explicitly models Hessian, Fisher, Newton, quasi-Newton, cubic, trust-region, or other curvature structure, assign to D2; if adaptivity is mainly a distributed-compression mechanism, assign to D3.","nearest_boundary":"D2","representative_topics":["preconditioned stochastic gradient methods for ill-conditioned non-convex finite sum minimization","non-monotone adaptive gradient methods with bidirectional step size accumulation","lipschitz-free convergence analysis of adam with hyperparameters close to one"]},{"direction_index":2,"label":"D2","name":"Stochastic curvature and second-order optimization","description":"Algorithms and convergence theory for stochastic optimization methods that explicitly use or approximate curvature information, including Hessian, Fisher, Newton, quasi-Newton, cubic regularization, trust-region, subspace, Kronecker, or diagonal Hessian structures.","inclusion":"Include stochastic Newton, natural-gradient, cubic-regularized Newton, trust-region, quasi-Newton, Hessian-vector, dimension-reduced second-order, scalable curvature approximation, and second-order Polyak-style methods when curvature modeling is the primary scientific objective.","exclusion":"If curvature-like information is only diagonal gradient-moment scaling, ordinary preconditioning, or learning-rate adaptation without an explicit second-order modeling objective, assign to D1.","nearest_boundary":"D1","representative_topics":["second-order stochastic polyak step size using hessian-vector products for non-convex optimization","noise-adaptive second-order methods with unified stochastic and deterministic oracle convergence","stochastic second-order policy optimization via dimension-reduced trust region methods"]},{"direction_index":3,"label":"D3","name":"Distributed and communication-efficient optimization","description":"Algorithms and theory for stochastic optimization across multiple workers, clients, devices, or network nodes, with emphasis on communication, compression, consensus, asynchrony, delays, heterogeneity, client/data sampling, and decentralized or federated execution.","inclusion":"Include distributed, decentralized, federated, asynchronous, delayed, compressed, heterogeneous, or multi-agent stochastic optimization when the distributed systems constraint is the primary contribution, including distributed variants of SGD, minimax, compositional, or variance-reduced methods.","exclusion":"If privacy guarantees are the main objective rather than distributed execution, assign to D7; if saddle/compositional structure dominates and distribution is only an implementation setting, assign to D7.","nearest_boundary":"D7","representative_topics":["sharp sgd convergence rates under flexible client and data sampling in nonconvex optimization","federated gradient compression with random reshuffling and variance reduction for distributed training","asynchronous sgd convergence analysis under arbitrary delays via virtual iterate recursion"]},{"direction_index":4,"label":"D4","name":"Robust optimization under heavy-tailed noise","description":"Stochastic optimization under heavy-tailed, weak-moment, adversarial, outlier-prone, unbounded-variance, or otherwise unreliable stochastic noise, emphasizing robust convergence and high-probability guarantees.","inclusion":"Include clipping, median-of-means, robust aggregation, robust variance reduction, adaptive robust rates, and high-probability analysis when the main scientific objective is reliable optimization under heavy-tailed or corrupted stochastic gradients or losses.","exclusion":"If clipping or moment control is primarily for differential privacy, assign to D7; if heavy tails are primarily studied as a continuous-time limiting law, stationary distribution, or generalization mechanism, assign to D6.","nearest_boundary":"D6","representative_topics":["clipped stochastic extragradient methods for variational inequalities with heavy-tailed noise and unbounded domains","high-probability convergence of stochastic optimization under unbounded variance gradient noise","variance-reduced acceleration in non-convex stochastic optimization with heavy-tailed noise"]},{"direction_index":5,"label":"D5","name":"Nonstandard oracles, geometry, and nonsmoothness","description":"Stochastic optimization when standard smooth Euclidean first-order assumptions fail because gradients are unavailable, indirect, zeroth-order, bandit-like, nonsmooth, nonconvex-composite, Riemannian, manifold-constrained, or geometrically constrained.","inclusion":"Include zeroth-order, gradient-free, simultaneous-perturbation, bandit or semi-bandit, randomized directional, stochastic subgradient, nonsmooth nonconvex proximal, Riemannian, retraction/vector-transport, manifold, and stochastic interior-point methods when oracle limitations, nonsmoothness, or feasible-space geometry is the dominant objective.","exclusion":"If ordinary gradients are available and the main objective is estimator efficiency or sampling order, assign to D6; if constraints are expectation, risk, safety, privacy, or decision-modeling constraints rather than geometry or oracle access, assign to D7.","nearest_boundary":"D7","representative_topics":["structured stochastic zeroth-order descent with orthogonal direction sampling and polyak-łojasiewicz convergence","stochastic approximation convergence on riemannian manifolds via fermi coordinate dynamical systems","riemannian stochastic gradient methods for nested compositional optimization with bias-corrected inner approximation"]},{"direction_index":6,"label":"D6","name":"Estimator efficiency, dynamics, and generalization","description":"Research explaining or improving stochastic optimization through estimator variance structure, finite-sum sampling order, stochastic-process dynamics, sampling behavior, implicit regularization, stability, and generalization properties.","inclusion":"Include variance reduction, finite-sum incremental methods, random reshuffling, sampling without replacement, recursive estimators, proximal finite-sum algorithms, SDE and diffusion limits, Langevin/SGLD methods, stationary laws, process-level approximations, implicit bias, flat-minimum selection, risk trajectories, and stability-based generalization when the central objective is estimator dynamics or statistical behavior beyond basic convergence.","exclusion":"If the main objective is robustifying against heavy-tailed noise, assign to D4; if the main objective is solving a saddle, constrained, private, or decision-structured formulation, assign to D7.","nearest_boundary":"D4","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","high-dimensional sgd limit theorems with critical step-size scaling and effective dynamics","stability and generalization bounds for stochastic gradient methods on nonconvex nonsmooth objectives"]},{"direction_index":7,"label":"D7","name":"Structured, constrained, private, and decision-aware optimization","description":"Stochastic optimization programs whose primary objective is solving structured formulations beyond unconstrained single-level minimization, including saddle-point, variational-inequality, minimax, bilevel, compositional, constrained, risk-sensitive, private, safe, distributionally robust, multistage, online, or decision-focused problems.","inclusion":"Include stochastic minimax, gradient-descent-ascent, extragradient, VI, bilevel, nested or multi-level compositional objectives, expectation or compositional constraints, augmented-Lagrangian, penalty, barrier, safe black-box, risk-measure, DRO, differential privacy, user-level privacy, performative or decision-dependent optimization, multistage stochastic programming, surrogate decision models, bandit decision-making, and estimation-optimization integration when this formulation-level structure is central.","exclusion":"If the dominant contribution is distributed communication rather than the structured formulation, assign to D3; if the dominant issue is manifold geometry, zeroth-order access, or nonsmooth oracle structure rather than constraints/privacy/decisions/saddles, assign to D5.","nearest_boundary":"D5","representative_topics":["differentially private stochastic convex optimization via averaged clipping for heavy-tailed data","variance reduced estimation for coupled compositional optimization with sparse block probing","near-optimal stochastic first-order oracle complexity for minimax optimization via recursive anchored iteration"]}],"split":"dev"}
{"episode_id":"F232_T2026-01","field_index":232,"field":"stochastic optimization","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Core SGD and stochastic-approximation theory","description":"General-purpose theory for SGD-like and stochastic-approximation methods, emphasizing convergence, rates, asymptotics, concentration, stability thresholds, inference for iterates, and sample complexity when no specialized structural objective dominates.","inclusion":"Include source atoms whose main program is broadly applicable analysis of SGD, momentum SGD as a basic variant, Polyak-Ruppert averaging, implicit SGD, coupled stochastic approximation, Markovian or decision-dependent stochastic approximation, primal-dual splitting, or ERM/SCO sample complexity under standard or relaxed assumptions.","exclusion":"If the dominant objective is automated step-size or preconditioning design, distributed communication, heavy-tail robustification, finite-sum variance reduction, SDE/sampling dynamics, generalization, privacy, constraints, or saddle/compositional structure, assign to the corresponding specialized direction instead.","nearest_boundary":"D1","representative_topics":["single-timescale convergence analysis for nonlinear stochastic approximation with multiple coupled sequences","statistical inference for implicit sgd via proximal robbins-monro and polyak-ruppert procedures","stochastic approximation with decision-dependent distributions: asymptotic normality and minimax optimality"]},{"direction_index":1,"label":"D1","name":"Adaptive and momentum first-order methods","description":"Design and theory of stochastic first-order optimizers whose main objective is automated learning-rate control, adaptive scaling, diagonal preconditioning, sign updates, momentum mechanisms, or first-order acceleration and stabilization.","inclusion":"Include Adam, AdaGrad, sign-based methods, Polyak/function-value step sizes, parameter-free schedules, preconditioned first-order methods, adaptive momentum, heavy-ball or Nesterov-style stochastic momentum, and related first-order optimizer designs when adaptivity or momentum design is central.","exclusion":"If the method explicitly models Hessian, Fisher, Newton, quasi-Newton, cubic, trust-region, or other curvature structure, assign to D2; if adaptivity is mainly a distributed-compression mechanism, assign to D3.","nearest_boundary":"D2","representative_topics":["preconditioned stochastic gradient methods for ill-conditioned non-convex finite sum minimization","non-monotone adaptive gradient methods with bidirectional step size accumulation","lipschitz-free convergence analysis of adam with hyperparameters close to one"]},{"direction_index":2,"label":"D2","name":"Stochastic curvature and second-order optimization","description":"Algorithms and convergence theory for stochastic optimization methods that explicitly use or approximate curvature information, including Hessian, Fisher, Newton, quasi-Newton, cubic regularization, trust-region, subspace, Kronecker, or diagonal Hessian structures.","inclusion":"Include stochastic Newton, natural-gradient, cubic-regularized Newton, trust-region, quasi-Newton, Hessian-vector, dimension-reduced second-order, scalable curvature approximation, and second-order Polyak-style methods when curvature modeling is the primary scientific objective.","exclusion":"If curvature-like information is only diagonal gradient-moment scaling, ordinary preconditioning, or learning-rate adaptation without an explicit second-order modeling objective, assign to D1.","nearest_boundary":"D1","representative_topics":["second-order stochastic polyak step size using hessian-vector products for non-convex optimization","noise-adaptive second-order methods with unified stochastic and deterministic oracle convergence","stochastic second-order policy optimization via dimension-reduced trust region methods"]},{"direction_index":3,"label":"D3","name":"Distributed and communication-efficient optimization","description":"Algorithms and theory for stochastic optimization across multiple workers, clients, devices, or network nodes, with emphasis on communication, compression, consensus, asynchrony, delays, heterogeneity, client/data sampling, and decentralized or federated execution.","inclusion":"Include distributed, decentralized, federated, asynchronous, delayed, compressed, heterogeneous, or multi-agent stochastic optimization when the distributed systems constraint is the primary contribution, including distributed variants of SGD, minimax, compositional, or variance-reduced methods.","exclusion":"If privacy guarantees are the main objective rather than distributed execution, assign to D7; if saddle/compositional structure dominates and distribution is only an implementation setting, assign to D7.","nearest_boundary":"D7","representative_topics":["sharp sgd convergence rates under flexible client and data sampling in nonconvex optimization","federated gradient compression with random reshuffling and variance reduction for distributed training","asynchronous sgd convergence analysis under arbitrary delays via virtual iterate recursion"]},{"direction_index":4,"label":"D4","name":"Robust optimization under heavy-tailed noise","description":"Stochastic optimization under heavy-tailed, weak-moment, adversarial, outlier-prone, unbounded-variance, or otherwise unreliable stochastic noise, emphasizing robust convergence and high-probability guarantees.","inclusion":"Include clipping, median-of-means, robust aggregation, robust variance reduction, adaptive robust rates, and high-probability analysis when the main scientific objective is reliable optimization under heavy-tailed or corrupted stochastic gradients or losses.","exclusion":"If clipping or moment control is primarily for differential privacy, assign to D7; if heavy tails are primarily studied as a continuous-time limiting law, stationary distribution, or generalization mechanism, assign to D6.","nearest_boundary":"D6","representative_topics":["clipped stochastic extragradient methods for variational inequalities with heavy-tailed noise and unbounded domains","high-probability convergence of stochastic optimization under unbounded variance gradient noise","variance-reduced acceleration in non-convex stochastic optimization with heavy-tailed noise"]},{"direction_index":5,"label":"D5","name":"Nonstandard oracles, geometry, and nonsmoothness","description":"Stochastic optimization when standard smooth Euclidean first-order assumptions fail because gradients are unavailable, indirect, zeroth-order, bandit-like, nonsmooth, nonconvex-composite, Riemannian, manifold-constrained, or geometrically constrained.","inclusion":"Include zeroth-order, gradient-free, simultaneous-perturbation, bandit or semi-bandit, randomized directional, stochastic subgradient, nonsmooth nonconvex proximal, Riemannian, retraction/vector-transport, manifold, and stochastic interior-point methods when oracle limitations, nonsmoothness, or feasible-space geometry is the dominant objective.","exclusion":"If ordinary gradients are available and the main objective is estimator efficiency or sampling order, assign to D6; if constraints are expectation, risk, safety, privacy, or decision-modeling constraints rather than geometry or oracle access, assign to D7.","nearest_boundary":"D7","representative_topics":["structured stochastic zeroth-order descent with orthogonal direction sampling and polyak-łojasiewicz convergence","stochastic approximation convergence on riemannian manifolds via fermi coordinate dynamical systems","riemannian stochastic gradient methods for nested compositional optimization with bias-corrected inner approximation"]},{"direction_index":6,"label":"D6","name":"Estimator efficiency, dynamics, and generalization","description":"Research explaining or improving stochastic optimization through estimator variance structure, finite-sum sampling order, stochastic-process dynamics, sampling behavior, implicit regularization, stability, and generalization properties.","inclusion":"Include variance reduction, finite-sum incremental methods, random reshuffling, sampling without replacement, recursive estimators, proximal finite-sum algorithms, SDE and diffusion limits, Langevin/SGLD methods, stationary laws, process-level approximations, implicit bias, flat-minimum selection, risk trajectories, and stability-based generalization when the central objective is estimator dynamics or statistical behavior beyond basic convergence.","exclusion":"If the main objective is robustifying against heavy-tailed noise, assign to D4; if the main objective is solving a saddle, constrained, private, or decision-structured formulation, assign to D7.","nearest_boundary":"D4","representative_topics":["variance analysis of shuffling stochastic gradient methods via power spectral density","high-dimensional sgd limit theorems with critical step-size scaling and effective dynamics","stability and generalization bounds for stochastic gradient methods on nonconvex nonsmooth objectives"]},{"direction_index":7,"label":"D7","name":"Structured, constrained, private, and decision-aware optimization","description":"Stochastic optimization programs whose primary objective is solving structured formulations beyond unconstrained single-level minimization, including saddle-point, variational-inequality, minimax, bilevel, compositional, constrained, risk-sensitive, private, safe, distributionally robust, multistage, online, or decision-focused problems.","inclusion":"Include stochastic minimax, gradient-descent-ascent, extragradient, VI, bilevel, nested or multi-level compositional objectives, expectation or compositional constraints, augmented-Lagrangian, penalty, barrier, safe black-box, risk-measure, DRO, differential privacy, user-level privacy, performative or decision-dependent optimization, multistage stochastic programming, surrogate decision models, bandit decision-making, and estimation-optimization integration when this formulation-level structure is central.","exclusion":"If the dominant contribution is distributed communication rather than the structured formulation, assign to D3; if the dominant issue is manifold geometry, zeroth-order access, or nonsmooth oracle structure rather than constraints/privacy/decisions/saddles, assign to D5.","nearest_boundary":"D5","representative_topics":["differentially private stochastic convex optimization via averaged clipping for heavy-tailed data","variance reduced estimation for coupled compositional optimization with sparse block probing","near-optimal stochastic first-order oracle complexity for minimax optimization via recursive anchored iteration"]}],"split":"dev"}
{"episode_id":"F234_T2024-01","field_index":234,"field":"remote sensing object detection","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Oriented object localization","description":"Improving detection of arbitrarily oriented objects in aerial or satellite imagery through rotation-aware localization, oriented bounding boxes, angle representations, rotation-equivariant features, assignment, or rotated-box losses.","inclusion":"Include work whose main scientific objective is better oriented or rotated bounding-box prediction for remote sensing objects, regardless of backbone, detector family, feature pyramid, transformer, loss, or label-assignment mechanism.","exclusion":"If rotated boxes are incidental and the primary objective is label efficiency, small-object recall, compression, multimodal adaptation, or an application-specific monitoring task, assign to that nearer direction.","nearest_boundary":"Small and dense target detection","representative_topics":["task-wise sampling convolutions for arbitrary-oriented object detection in aerial images","multi-grained angle representation for arbitrary-oriented remote sensing object detection","rotation-equivariant attention feature fusion in pyramid networks for aerial object detection"]},{"direction_index":1,"label":"D1","name":"Small and dense target detection","description":"Improving detection of tiny, small, dim, densely distributed, low-pixel, cluttered, or strongly scale-varying targets in aerial, UAV, SAR, infrared, satellite, or Earth-observation imagery.","inclusion":"Include work whose central problem is overcoming object scale, target density, dim signatures, clutter, limited pixel support, or extreme scale variation using scale-adaptive localization, coarse-to-fine processing, feature aggregation, attention, frequency modeling, or small-target-specific assignment/losses.","exclusion":"If the main contribution is oriented-box representation, edge deployment efficiency, scarce-label learning, or a domain monitoring pipeline where small targets are only the application target, assign to that nearer direction.","nearest_boundary":"Oriented object localization","representative_topics":["axial-attention and multi-scale patch fusion for infrared small target detection","tiny object detection in aerial images via normalized wasserstein distance and ranking-based label assignment","frequency-aware feature pyramid networks for small object detection in remote sensing images"]},{"direction_index":2,"label":"D2","name":"Annotation-efficient learning","description":"Learning remote sensing detectors under limited, weak, noisy, sparse, synthetic, transferred, or unlabeled supervision, including semi-supervised learning, self-supervised pretraining, few-shot detection, point supervision, pseudo-labeling, soft labels, and noisy geo-annotations.","inclusion":"Include work whose primary objective is reducing manual annotation burden, improving detection in scarce-data regimes, exploiting weak labels, using unlabeled data, or making detection robust to incomplete or noisy labels.","exclusion":"If limited labels are secondary to a target-specific operational system, small-target methodology, oriented localization, or cross-sensor/domain adaptation, assign to that nearer direction.","nearest_boundary":"Domain-targeted monitoring detection","representative_topics":["soft-labeling strategies for iterative self-training in satellite image sub-type classification","centerpoint-only annotation for overhead imagery object detection with reduced labeling cost","self-supervised pretraining for label-efficient vehicle detection on satellite imagery"]},{"direction_index":3,"label":"D3","name":"Domain-targeted monitoring detection","description":"Adapting object detection to specific monitored entities, environments, operational domains, or specialized benchmarks in remote sensing, such as maritime surveillance, agriculture, wildlife, trees, environmental compliance, infrastructure, industrial assets, or solar panels.","inclusion":"Include work whose primary scientific objective is solving a particular real-world detection task or constructing a domain-specific dataset/benchmark, with domain labels, sensing conditions, and application-driven evaluation central to the contribution.","exclusion":"If the main contribution is a generic method for scarce labels, small targets, orientation, compression, saliency, temporal change, or geometric vectorization, assign to that methodological direction.","nearest_boundary":"Annotation-efficient learning","representative_topics":["dark vessel detection in sar imagery using large-scale annotated maritime object datasets","patch-wise scum detection on river surfaces using mixture augmentation for sparse object classification","uav-based volunteer weed detection in corn fields for precision spot-spray pest management"]},{"direction_index":4,"label":"D4","name":"Region, contour, and saliency extraction","description":"Detecting or delineating remote sensing objects as salient regions, precise contours, polygons, boundaries, masks, outlines, or shape-aware geometric representations rather than ordinary category-level bounding boxes.","inclusion":"Include work whose declared objective is salient object detection in optical remote sensing imagery or geometry-first extraction such as building polygon recovery, vectorized contours, boundary-semantic saliency, active contours, or explicit outline delineation.","exclusion":"If structures are merely standard bounding-box categories, assign to domain-targeted or generic detection; if the main objective is before-after damage or temporal change, assign to change and contextual state detection.","nearest_boundary":"Change and contextual state detection","representative_topics":["attention-guided salient object detection for optical remote sensing images with progressive refinement","hierarchical supervision for reversible building mask-to-polygon mapping in satellite imagery","learnable polygon contour extraction for vectorized building delineation from remote sensing"]},{"direction_index":5,"label":"D5","name":"Efficient deployable detection","description":"Reducing computational cost, memory use, model size, latency, patch-processing burden, or deployment complexity of remote sensing object detectors for large images, embedded platforms, UAVs, drones, or edge devices.","inclusion":"Include work whose main objective is efficient, lightweight, real-time, compressed, quantized, distilled, sparse, patch-filtered, or deployment-ready object detection in aerial or remote sensing imagery.","exclusion":"If efficiency is secondary to small-target recall, oriented localization, scarce-label learning, multimodal adaptation, or a target-specific monitoring task, assign to that primary objective instead.","nearest_boundary":"Small and dense target detection","representative_topics":["knowledge distillation for lightweight oriented object detection on aerial images","lightweight oriented object detection with diagonal support constraints in remote sensing","efficient ship detection in sar imagery via classification model and statistical optimization"]},{"direction_index":6,"label":"D6","name":"Cross-domain and multimodal robust detection","description":"Improving detector robustness and accuracy across imaging conditions, sensors, modalities, domains, illumination, compression, corruptions, uncertainty, or distribution shifts, including RGB-infrared, SAR, hyperspectral, thermal, synthetic-to-real, and multimodal fusion settings.","inclusion":"Include work whose primary contribution is handling domain shift, cross-modal misalignment, heterogeneous sensor fusion, spectral-spatial integration, transformation-invariant representation, or generic robustness to remote sensing acquisition and distributional variation.","exclusion":"If robustness is primarily about scarce labels, saliency robustness, small-target visibility, oriented boxes, efficiency, or a specific operational monitoring use case, assign to that nearer direction.","nearest_boundary":"Efficient deployable detection","representative_topics":["uncertainty-aware graph neural network refinement for aerial image object detection","cross-modal alignment for aerial rgb-infrared vehicle detection under weak misalignment","robust object recognition under illumination and viewing angle variation using cnn bounding box analysis"]},{"direction_index":7,"label":"D7","name":"Change and contextual state detection","description":"Detecting temporal changes, altered regions, damage states, before-after differences, or broad contextual scene states in remote sensing imagery, including bitemporal change detection, disaster damage inference, and general context/foundation modeling when not reducible to orientation, scale, supervision, efficiency, saliency, or domain adaptation.","inclusion":"Include work whose primary objective is identifying changed or damaged regions over time, modeling post-event state, or developing broad context/foundation detection strategies that serve general remote sensing detection rather than a more specific program.","exclusion":"If the work extracts static building geometry, detects a fixed monitored target, improves oriented boxes, targets tiny objects, reduces annotation needs, or focuses on deployment efficiency, assign to that nearer direction.","nearest_boundary":"Region, contour, and saliency extraction","representative_topics":["plain vision transformer with rotated varied-size window attention for remote sensing foundation models","attention-guided remote sensing change detection using image difference and edge difference feature maps","uncertainty-aware graph neural network refinement for aerial image object detection"]}],"split":"dev"}
{"episode_id":"F234_T2024-07","field_index":234,"field":"remote sensing object detection","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Oriented object localization","description":"Improving detection of arbitrarily oriented objects in aerial or satellite imagery through rotation-aware localization, oriented bounding boxes, angle representations, rotation-equivariant features, assignment, or rotated-box losses.","inclusion":"Include work whose main scientific objective is better oriented or rotated bounding-box prediction for remote sensing objects, regardless of backbone, detector family, feature pyramid, transformer, loss, or label-assignment mechanism.","exclusion":"If rotated boxes are incidental and the primary objective is label efficiency, small-object recall, compression, multimodal adaptation, or an application-specific monitoring task, assign to that nearer direction.","nearest_boundary":"Small and dense target detection","representative_topics":["task-wise sampling convolutions for arbitrary-oriented object detection in aerial images","multi-grained angle representation for arbitrary-oriented remote sensing object detection","rotation-equivariant attention feature fusion in pyramid networks for aerial object detection"]},{"direction_index":1,"label":"D1","name":"Small and dense target detection","description":"Improving detection of tiny, small, dim, densely distributed, low-pixel, cluttered, or strongly scale-varying targets in aerial, UAV, SAR, infrared, satellite, or Earth-observation imagery.","inclusion":"Include work whose central problem is overcoming object scale, target density, dim signatures, clutter, limited pixel support, or extreme scale variation using scale-adaptive localization, coarse-to-fine processing, feature aggregation, attention, frequency modeling, or small-target-specific assignment/losses.","exclusion":"If the main contribution is oriented-box representation, edge deployment efficiency, scarce-label learning, or a domain monitoring pipeline where small targets are only the application target, assign to that nearer direction.","nearest_boundary":"Oriented object localization","representative_topics":["axial-attention and multi-scale patch fusion for infrared small target detection","tiny object detection in aerial images via normalized wasserstein distance and ranking-based label assignment","frequency-aware feature pyramid networks for small object detection in remote sensing images"]},{"direction_index":2,"label":"D2","name":"Annotation-efficient learning","description":"Learning remote sensing detectors under limited, weak, noisy, sparse, synthetic, transferred, or unlabeled supervision, including semi-supervised learning, self-supervised pretraining, few-shot detection, point supervision, pseudo-labeling, soft labels, and noisy geo-annotations.","inclusion":"Include work whose primary objective is reducing manual annotation burden, improving detection in scarce-data regimes, exploiting weak labels, using unlabeled data, or making detection robust to incomplete or noisy labels.","exclusion":"If limited labels are secondary to a target-specific operational system, small-target methodology, oriented localization, or cross-sensor/domain adaptation, assign to that nearer direction.","nearest_boundary":"Domain-targeted monitoring detection","representative_topics":["soft-labeling strategies for iterative self-training in satellite image sub-type classification","centerpoint-only annotation for overhead imagery object detection with reduced labeling cost","self-supervised pretraining for label-efficient vehicle detection on satellite imagery"]},{"direction_index":3,"label":"D3","name":"Domain-targeted monitoring detection","description":"Adapting object detection to specific monitored entities, environments, operational domains, or specialized benchmarks in remote sensing, such as maritime surveillance, agriculture, wildlife, trees, environmental compliance, infrastructure, industrial assets, or solar panels.","inclusion":"Include work whose primary scientific objective is solving a particular real-world detection task or constructing a domain-specific dataset/benchmark, with domain labels, sensing conditions, and application-driven evaluation central to the contribution.","exclusion":"If the main contribution is a generic method for scarce labels, small targets, orientation, compression, saliency, temporal change, or geometric vectorization, assign to that methodological direction.","nearest_boundary":"Annotation-efficient learning","representative_topics":["dark vessel detection in sar imagery using large-scale annotated maritime object datasets","patch-wise scum detection on river surfaces using mixture augmentation for sparse object classification","uav-based volunteer weed detection in corn fields for precision spot-spray pest management"]},{"direction_index":4,"label":"D4","name":"Region, contour, and saliency extraction","description":"Detecting or delineating remote sensing objects as salient regions, precise contours, polygons, boundaries, masks, outlines, or shape-aware geometric representations rather than ordinary category-level bounding boxes.","inclusion":"Include work whose declared objective is salient object detection in optical remote sensing imagery or geometry-first extraction such as building polygon recovery, vectorized contours, boundary-semantic saliency, active contours, or explicit outline delineation.","exclusion":"If structures are merely standard bounding-box categories, assign to domain-targeted or generic detection; if the main objective is before-after damage or temporal change, assign to change and contextual state detection.","nearest_boundary":"Change and contextual state detection","representative_topics":["attention-guided salient object detection for optical remote sensing images with progressive refinement","hierarchical supervision for reversible building mask-to-polygon mapping in satellite imagery","learnable polygon contour extraction for vectorized building delineation from remote sensing"]},{"direction_index":5,"label":"D5","name":"Efficient deployable detection","description":"Reducing computational cost, memory use, model size, latency, patch-processing burden, or deployment complexity of remote sensing object detectors for large images, embedded platforms, UAVs, drones, or edge devices.","inclusion":"Include work whose main objective is efficient, lightweight, real-time, compressed, quantized, distilled, sparse, patch-filtered, or deployment-ready object detection in aerial or remote sensing imagery.","exclusion":"If efficiency is secondary to small-target recall, oriented localization, scarce-label learning, multimodal adaptation, or a target-specific monitoring task, assign to that primary objective instead.","nearest_boundary":"Small and dense target detection","representative_topics":["knowledge distillation for lightweight oriented object detection on aerial images","lightweight oriented object detection with diagonal support constraints in remote sensing","efficient ship detection in sar imagery via classification model and statistical optimization"]},{"direction_index":6,"label":"D6","name":"Cross-domain and multimodal robust detection","description":"Improving detector robustness and accuracy across imaging conditions, sensors, modalities, domains, illumination, compression, corruptions, uncertainty, or distribution shifts, including RGB-infrared, SAR, hyperspectral, thermal, synthetic-to-real, and multimodal fusion settings.","inclusion":"Include work whose primary contribution is handling domain shift, cross-modal misalignment, heterogeneous sensor fusion, spectral-spatial integration, transformation-invariant representation, or generic robustness to remote sensing acquisition and distributional variation.","exclusion":"If robustness is primarily about scarce labels, saliency robustness, small-target visibility, oriented boxes, efficiency, or a specific operational monitoring use case, assign to that nearer direction.","nearest_boundary":"Efficient deployable detection","representative_topics":["uncertainty-aware graph neural network refinement for aerial image object detection","cross-modal alignment for aerial rgb-infrared vehicle detection under weak misalignment","robust object recognition under illumination and viewing angle variation using cnn bounding box analysis"]},{"direction_index":7,"label":"D7","name":"Change and contextual state detection","description":"Detecting temporal changes, altered regions, damage states, before-after differences, or broad contextual scene states in remote sensing imagery, including bitemporal change detection, disaster damage inference, and general context/foundation modeling when not reducible to orientation, scale, supervision, efficiency, saliency, or domain adaptation.","inclusion":"Include work whose primary objective is identifying changed or damaged regions over time, modeling post-event state, or developing broad context/foundation detection strategies that serve general remote sensing detection rather than a more specific program.","exclusion":"If the work extracts static building geometry, detects a fixed monitored target, improves oriented boxes, targets tiny objects, reduces annotation needs, or focuses on deployment efficiency, assign to that nearer direction.","nearest_boundary":"Region, contour, and saliency extraction","representative_topics":["plain vision transformer with rotated varied-size window attention for remote sensing foundation models","attention-guided remote sensing change detection using image difference and edge difference feature maps","uncertainty-aware graph neural network refinement for aerial image object detection"]}],"split":"dev"}
{"episode_id":"F234_T2025-01","field_index":234,"field":"remote sensing object detection","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Oriented object localization","description":"Improving detection of arbitrarily oriented objects in aerial or satellite imagery through rotation-aware localization, oriented bounding boxes, angle representations, rotation-equivariant features, assignment, or rotated-box losses.","inclusion":"Include work whose main scientific objective is better oriented or rotated bounding-box prediction for remote sensing objects, regardless of backbone, detector family, feature pyramid, transformer, loss, or label-assignment mechanism.","exclusion":"If rotated boxes are incidental and the primary objective is label efficiency, small-object recall, compression, multimodal adaptation, or an application-specific monitoring task, assign to that nearer direction.","nearest_boundary":"Small and dense target detection","representative_topics":["task-wise sampling convolutions for arbitrary-oriented object detection in aerial images","multi-grained angle representation for arbitrary-oriented remote sensing object detection","rotation-equivariant attention feature fusion in pyramid networks for aerial object detection"]},{"direction_index":1,"label":"D1","name":"Small and dense target detection","description":"Improving detection of tiny, small, dim, densely distributed, low-pixel, cluttered, or strongly scale-varying targets in aerial, UAV, SAR, infrared, satellite, or Earth-observation imagery.","inclusion":"Include work whose central problem is overcoming object scale, target density, dim signatures, clutter, limited pixel support, or extreme scale variation using scale-adaptive localization, coarse-to-fine processing, feature aggregation, attention, frequency modeling, or small-target-specific assignment/losses.","exclusion":"If the main contribution is oriented-box representation, edge deployment efficiency, scarce-label learning, or a domain monitoring pipeline where small targets are only the application target, assign to that nearer direction.","nearest_boundary":"Oriented object localization","representative_topics":["axial-attention and multi-scale patch fusion for infrared small target detection","tiny object detection in aerial images via normalized wasserstein distance and ranking-based label assignment","frequency-aware feature pyramid networks for small object detection in remote sensing images"]},{"direction_index":2,"label":"D2","name":"Annotation-efficient learning","description":"Learning remote sensing detectors under limited, weak, noisy, sparse, synthetic, transferred, or unlabeled supervision, including semi-supervised learning, self-supervised pretraining, few-shot detection, point supervision, pseudo-labeling, soft labels, and noisy geo-annotations.","inclusion":"Include work whose primary objective is reducing manual annotation burden, improving detection in scarce-data regimes, exploiting weak labels, using unlabeled data, or making detection robust to incomplete or noisy labels.","exclusion":"If limited labels are secondary to a target-specific operational system, small-target methodology, oriented localization, or cross-sensor/domain adaptation, assign to that nearer direction.","nearest_boundary":"Domain-targeted monitoring detection","representative_topics":["soft-labeling strategies for iterative self-training in satellite image sub-type classification","centerpoint-only annotation for overhead imagery object detection with reduced labeling cost","self-supervised pretraining for label-efficient vehicle detection on satellite imagery"]},{"direction_index":3,"label":"D3","name":"Domain-targeted monitoring detection","description":"Adapting object detection to specific monitored entities, environments, operational domains, or specialized benchmarks in remote sensing, such as maritime surveillance, agriculture, wildlife, trees, environmental compliance, infrastructure, industrial assets, or solar panels.","inclusion":"Include work whose primary scientific objective is solving a particular real-world detection task or constructing a domain-specific dataset/benchmark, with domain labels, sensing conditions, and application-driven evaluation central to the contribution.","exclusion":"If the main contribution is a generic method for scarce labels, small targets, orientation, compression, saliency, temporal change, or geometric vectorization, assign to that methodological direction.","nearest_boundary":"Annotation-efficient learning","representative_topics":["dark vessel detection in sar imagery using large-scale annotated maritime object datasets","patch-wise scum detection on river surfaces using mixture augmentation for sparse object classification","uav-based volunteer weed detection in corn fields for precision spot-spray pest management"]},{"direction_index":4,"label":"D4","name":"Region, contour, and saliency extraction","description":"Detecting or delineating remote sensing objects as salient regions, precise contours, polygons, boundaries, masks, outlines, or shape-aware geometric representations rather than ordinary category-level bounding boxes.","inclusion":"Include work whose declared objective is salient object detection in optical remote sensing imagery or geometry-first extraction such as building polygon recovery, vectorized contours, boundary-semantic saliency, active contours, or explicit outline delineation.","exclusion":"If structures are merely standard bounding-box categories, assign to domain-targeted or generic detection; if the main objective is before-after damage or temporal change, assign to change and contextual state detection.","nearest_boundary":"Change and contextual state detection","representative_topics":["attention-guided salient object detection for optical remote sensing images with progressive refinement","hierarchical supervision for reversible building mask-to-polygon mapping in satellite imagery","learnable polygon contour extraction for vectorized building delineation from remote sensing"]},{"direction_index":5,"label":"D5","name":"Efficient deployable detection","description":"Reducing computational cost, memory use, model size, latency, patch-processing burden, or deployment complexity of remote sensing object detectors for large images, embedded platforms, UAVs, drones, or edge devices.","inclusion":"Include work whose main objective is efficient, lightweight, real-time, compressed, quantized, distilled, sparse, patch-filtered, or deployment-ready object detection in aerial or remote sensing imagery.","exclusion":"If efficiency is secondary to small-target recall, oriented localization, scarce-label learning, multimodal adaptation, or a target-specific monitoring task, assign to that primary objective instead.","nearest_boundary":"Small and dense target detection","representative_topics":["knowledge distillation for lightweight oriented object detection on aerial images","lightweight oriented object detection with diagonal support constraints in remote sensing","efficient ship detection in sar imagery via classification model and statistical optimization"]},{"direction_index":6,"label":"D6","name":"Cross-domain and multimodal robust detection","description":"Improving detector robustness and accuracy across imaging conditions, sensors, modalities, domains, illumination, compression, corruptions, uncertainty, or distribution shifts, including RGB-infrared, SAR, hyperspectral, thermal, synthetic-to-real, and multimodal fusion settings.","inclusion":"Include work whose primary contribution is handling domain shift, cross-modal misalignment, heterogeneous sensor fusion, spectral-spatial integration, transformation-invariant representation, or generic robustness to remote sensing acquisition and distributional variation.","exclusion":"If robustness is primarily about scarce labels, saliency robustness, small-target visibility, oriented boxes, efficiency, or a specific operational monitoring use case, assign to that nearer direction.","nearest_boundary":"Efficient deployable detection","representative_topics":["uncertainty-aware graph neural network refinement for aerial image object detection","cross-modal alignment for aerial rgb-infrared vehicle detection under weak misalignment","robust object recognition under illumination and viewing angle variation using cnn bounding box analysis"]},{"direction_index":7,"label":"D7","name":"Change and contextual state detection","description":"Detecting temporal changes, altered regions, damage states, before-after differences, or broad contextual scene states in remote sensing imagery, including bitemporal change detection, disaster damage inference, and general context/foundation modeling when not reducible to orientation, scale, supervision, efficiency, saliency, or domain adaptation.","inclusion":"Include work whose primary objective is identifying changed or damaged regions over time, modeling post-event state, or developing broad context/foundation detection strategies that serve general remote sensing detection rather than a more specific program.","exclusion":"If the work extracts static building geometry, detects a fixed monitored target, improves oriented boxes, targets tiny objects, reduces annotation needs, or focuses on deployment efficiency, assign to that nearer direction.","nearest_boundary":"Region, contour, and saliency extraction","representative_topics":["plain vision transformer with rotated varied-size window attention for remote sensing foundation models","attention-guided remote sensing change detection using image difference and edge difference feature maps","uncertainty-aware graph neural network refinement for aerial image object detection"]}],"split":"dev"}
{"episode_id":"F234_T2025-07","field_index":234,"field":"remote sensing object detection","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Oriented object localization","description":"Improving detection of arbitrarily oriented objects in aerial or satellite imagery through rotation-aware localization, oriented bounding boxes, angle representations, rotation-equivariant features, assignment, or rotated-box losses.","inclusion":"Include work whose main scientific objective is better oriented or rotated bounding-box prediction for remote sensing objects, regardless of backbone, detector family, feature pyramid, transformer, loss, or label-assignment mechanism.","exclusion":"If rotated boxes are incidental and the primary objective is label efficiency, small-object recall, compression, multimodal adaptation, or an application-specific monitoring task, assign to that nearer direction.","nearest_boundary":"Small and dense target detection","representative_topics":["task-wise sampling convolutions for arbitrary-oriented object detection in aerial images","multi-grained angle representation for arbitrary-oriented remote sensing object detection","rotation-equivariant attention feature fusion in pyramid networks for aerial object detection"]},{"direction_index":1,"label":"D1","name":"Small and dense target detection","description":"Improving detection of tiny, small, dim, densely distributed, low-pixel, cluttered, or strongly scale-varying targets in aerial, UAV, SAR, infrared, satellite, or Earth-observation imagery.","inclusion":"Include work whose central problem is overcoming object scale, target density, dim signatures, clutter, limited pixel support, or extreme scale variation using scale-adaptive localization, coarse-to-fine processing, feature aggregation, attention, frequency modeling, or small-target-specific assignment/losses.","exclusion":"If the main contribution is oriented-box representation, edge deployment efficiency, scarce-label learning, or a domain monitoring pipeline where small targets are only the application target, assign to that nearer direction.","nearest_boundary":"Oriented object localization","representative_topics":["axial-attention and multi-scale patch fusion for infrared small target detection","tiny object detection in aerial images via normalized wasserstein distance and ranking-based label assignment","frequency-aware feature pyramid networks for small object detection in remote sensing images"]},{"direction_index":2,"label":"D2","name":"Annotation-efficient learning","description":"Learning remote sensing detectors under limited, weak, noisy, sparse, synthetic, transferred, or unlabeled supervision, including semi-supervised learning, self-supervised pretraining, few-shot detection, point supervision, pseudo-labeling, soft labels, and noisy geo-annotations.","inclusion":"Include work whose primary objective is reducing manual annotation burden, improving detection in scarce-data regimes, exploiting weak labels, using unlabeled data, or making detection robust to incomplete or noisy labels.","exclusion":"If limited labels are secondary to a target-specific operational system, small-target methodology, oriented localization, or cross-sensor/domain adaptation, assign to that nearer direction.","nearest_boundary":"Domain-targeted monitoring detection","representative_topics":["soft-labeling strategies for iterative self-training in satellite image sub-type classification","centerpoint-only annotation for overhead imagery object detection with reduced labeling cost","self-supervised pretraining for label-efficient vehicle detection on satellite imagery"]},{"direction_index":3,"label":"D3","name":"Domain-targeted monitoring detection","description":"Adapting object detection to specific monitored entities, environments, operational domains, or specialized benchmarks in remote sensing, such as maritime surveillance, agriculture, wildlife, trees, environmental compliance, infrastructure, industrial assets, or solar panels.","inclusion":"Include work whose primary scientific objective is solving a particular real-world detection task or constructing a domain-specific dataset/benchmark, with domain labels, sensing conditions, and application-driven evaluation central to the contribution.","exclusion":"If the main contribution is a generic method for scarce labels, small targets, orientation, compression, saliency, temporal change, or geometric vectorization, assign to that methodological direction.","nearest_boundary":"Annotation-efficient learning","representative_topics":["dark vessel detection in sar imagery using large-scale annotated maritime object datasets","patch-wise scum detection on river surfaces using mixture augmentation for sparse object classification","uav-based volunteer weed detection in corn fields for precision spot-spray pest management"]},{"direction_index":4,"label":"D4","name":"Region, contour, and saliency extraction","description":"Detecting or delineating remote sensing objects as salient regions, precise contours, polygons, boundaries, masks, outlines, or shape-aware geometric representations rather than ordinary category-level bounding boxes.","inclusion":"Include work whose declared objective is salient object detection in optical remote sensing imagery or geometry-first extraction such as building polygon recovery, vectorized contours, boundary-semantic saliency, active contours, or explicit outline delineation.","exclusion":"If structures are merely standard bounding-box categories, assign to domain-targeted or generic detection; if the main objective is before-after damage or temporal change, assign to change and contextual state detection.","nearest_boundary":"Change and contextual state detection","representative_topics":["attention-guided salient object detection for optical remote sensing images with progressive refinement","hierarchical supervision for reversible building mask-to-polygon mapping in satellite imagery","learnable polygon contour extraction for vectorized building delineation from remote sensing"]},{"direction_index":5,"label":"D5","name":"Efficient deployable detection","description":"Reducing computational cost, memory use, model size, latency, patch-processing burden, or deployment complexity of remote sensing object detectors for large images, embedded platforms, UAVs, drones, or edge devices.","inclusion":"Include work whose main objective is efficient, lightweight, real-time, compressed, quantized, distilled, sparse, patch-filtered, or deployment-ready object detection in aerial or remote sensing imagery.","exclusion":"If efficiency is secondary to small-target recall, oriented localization, scarce-label learning, multimodal adaptation, or a target-specific monitoring task, assign to that primary objective instead.","nearest_boundary":"Small and dense target detection","representative_topics":["knowledge distillation for lightweight oriented object detection on aerial images","lightweight oriented object detection with diagonal support constraints in remote sensing","efficient ship detection in sar imagery via classification model and statistical optimization"]},{"direction_index":6,"label":"D6","name":"Cross-domain and multimodal robust detection","description":"Improving detector robustness and accuracy across imaging conditions, sensors, modalities, domains, illumination, compression, corruptions, uncertainty, or distribution shifts, including RGB-infrared, SAR, hyperspectral, thermal, synthetic-to-real, and multimodal fusion settings.","inclusion":"Include work whose primary contribution is handling domain shift, cross-modal misalignment, heterogeneous sensor fusion, spectral-spatial integration, transformation-invariant representation, or generic robustness to remote sensing acquisition and distributional variation.","exclusion":"If robustness is primarily about scarce labels, saliency robustness, small-target visibility, oriented boxes, efficiency, or a specific operational monitoring use case, assign to that nearer direction.","nearest_boundary":"Efficient deployable detection","representative_topics":["uncertainty-aware graph neural network refinement for aerial image object detection","cross-modal alignment for aerial rgb-infrared vehicle detection under weak misalignment","robust object recognition under illumination and viewing angle variation using cnn bounding box analysis"]},{"direction_index":7,"label":"D7","name":"Change and contextual state detection","description":"Detecting temporal changes, altered regions, damage states, before-after differences, or broad contextual scene states in remote sensing imagery, including bitemporal change detection, disaster damage inference, and general context/foundation modeling when not reducible to orientation, scale, supervision, efficiency, saliency, or domain adaptation.","inclusion":"Include work whose primary objective is identifying changed or damaged regions over time, modeling post-event state, or developing broad context/foundation detection strategies that serve general remote sensing detection rather than a more specific program.","exclusion":"If the work extracts static building geometry, detects a fixed monitored target, improves oriented boxes, targets tiny objects, reduces annotation needs, or focuses on deployment efficiency, assign to that nearer direction.","nearest_boundary":"Region, contour, and saliency extraction","representative_topics":["plain vision transformer with rotated varied-size window attention for remote sensing foundation models","attention-guided remote sensing change detection using image difference and edge difference feature maps","uncertainty-aware graph neural network refinement for aerial image object detection"]}],"split":"dev"}
{"episode_id":"F234_T2026-01","field_index":234,"field":"remote sensing object detection","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Oriented object localization","description":"Improving detection of arbitrarily oriented objects in aerial or satellite imagery through rotation-aware localization, oriented bounding boxes, angle representations, rotation-equivariant features, assignment, or rotated-box losses.","inclusion":"Include work whose main scientific objective is better oriented or rotated bounding-box prediction for remote sensing objects, regardless of backbone, detector family, feature pyramid, transformer, loss, or label-assignment mechanism.","exclusion":"If rotated boxes are incidental and the primary objective is label efficiency, small-object recall, compression, multimodal adaptation, or an application-specific monitoring task, assign to that nearer direction.","nearest_boundary":"Small and dense target detection","representative_topics":["task-wise sampling convolutions for arbitrary-oriented object detection in aerial images","multi-grained angle representation for arbitrary-oriented remote sensing object detection","rotation-equivariant attention feature fusion in pyramid networks for aerial object detection"]},{"direction_index":1,"label":"D1","name":"Small and dense target detection","description":"Improving detection of tiny, small, dim, densely distributed, low-pixel, cluttered, or strongly scale-varying targets in aerial, UAV, SAR, infrared, satellite, or Earth-observation imagery.","inclusion":"Include work whose central problem is overcoming object scale, target density, dim signatures, clutter, limited pixel support, or extreme scale variation using scale-adaptive localization, coarse-to-fine processing, feature aggregation, attention, frequency modeling, or small-target-specific assignment/losses.","exclusion":"If the main contribution is oriented-box representation, edge deployment efficiency, scarce-label learning, or a domain monitoring pipeline where small targets are only the application target, assign to that nearer direction.","nearest_boundary":"Oriented object localization","representative_topics":["axial-attention and multi-scale patch fusion for infrared small target detection","tiny object detection in aerial images via normalized wasserstein distance and ranking-based label assignment","frequency-aware feature pyramid networks for small object detection in remote sensing images"]},{"direction_index":2,"label":"D2","name":"Annotation-efficient learning","description":"Learning remote sensing detectors under limited, weak, noisy, sparse, synthetic, transferred, or unlabeled supervision, including semi-supervised learning, self-supervised pretraining, few-shot detection, point supervision, pseudo-labeling, soft labels, and noisy geo-annotations.","inclusion":"Include work whose primary objective is reducing manual annotation burden, improving detection in scarce-data regimes, exploiting weak labels, using unlabeled data, or making detection robust to incomplete or noisy labels.","exclusion":"If limited labels are secondary to a target-specific operational system, small-target methodology, oriented localization, or cross-sensor/domain adaptation, assign to that nearer direction.","nearest_boundary":"Domain-targeted monitoring detection","representative_topics":["soft-labeling strategies for iterative self-training in satellite image sub-type classification","centerpoint-only annotation for overhead imagery object detection with reduced labeling cost","self-supervised pretraining for label-efficient vehicle detection on satellite imagery"]},{"direction_index":3,"label":"D3","name":"Domain-targeted monitoring detection","description":"Adapting object detection to specific monitored entities, environments, operational domains, or specialized benchmarks in remote sensing, such as maritime surveillance, agriculture, wildlife, trees, environmental compliance, infrastructure, industrial assets, or solar panels.","inclusion":"Include work whose primary scientific objective is solving a particular real-world detection task or constructing a domain-specific dataset/benchmark, with domain labels, sensing conditions, and application-driven evaluation central to the contribution.","exclusion":"If the main contribution is a generic method for scarce labels, small targets, orientation, compression, saliency, temporal change, or geometric vectorization, assign to that methodological direction.","nearest_boundary":"Annotation-efficient learning","representative_topics":["dark vessel detection in sar imagery using large-scale annotated maritime object datasets","patch-wise scum detection on river surfaces using mixture augmentation for sparse object classification","uav-based volunteer weed detection in corn fields for precision spot-spray pest management"]},{"direction_index":4,"label":"D4","name":"Region, contour, and saliency extraction","description":"Detecting or delineating remote sensing objects as salient regions, precise contours, polygons, boundaries, masks, outlines, or shape-aware geometric representations rather than ordinary category-level bounding boxes.","inclusion":"Include work whose declared objective is salient object detection in optical remote sensing imagery or geometry-first extraction such as building polygon recovery, vectorized contours, boundary-semantic saliency, active contours, or explicit outline delineation.","exclusion":"If structures are merely standard bounding-box categories, assign to domain-targeted or generic detection; if the main objective is before-after damage or temporal change, assign to change and contextual state detection.","nearest_boundary":"Change and contextual state detection","representative_topics":["attention-guided salient object detection for optical remote sensing images with progressive refinement","hierarchical supervision for reversible building mask-to-polygon mapping in satellite imagery","learnable polygon contour extraction for vectorized building delineation from remote sensing"]},{"direction_index":5,"label":"D5","name":"Efficient deployable detection","description":"Reducing computational cost, memory use, model size, latency, patch-processing burden, or deployment complexity of remote sensing object detectors for large images, embedded platforms, UAVs, drones, or edge devices.","inclusion":"Include work whose main objective is efficient, lightweight, real-time, compressed, quantized, distilled, sparse, patch-filtered, or deployment-ready object detection in aerial or remote sensing imagery.","exclusion":"If efficiency is secondary to small-target recall, oriented localization, scarce-label learning, multimodal adaptation, or a target-specific monitoring task, assign to that primary objective instead.","nearest_boundary":"Small and dense target detection","representative_topics":["knowledge distillation for lightweight oriented object detection on aerial images","lightweight oriented object detection with diagonal support constraints in remote sensing","efficient ship detection in sar imagery via classification model and statistical optimization"]},{"direction_index":6,"label":"D6","name":"Cross-domain and multimodal robust detection","description":"Improving detector robustness and accuracy across imaging conditions, sensors, modalities, domains, illumination, compression, corruptions, uncertainty, or distribution shifts, including RGB-infrared, SAR, hyperspectral, thermal, synthetic-to-real, and multimodal fusion settings.","inclusion":"Include work whose primary contribution is handling domain shift, cross-modal misalignment, heterogeneous sensor fusion, spectral-spatial integration, transformation-invariant representation, or generic robustness to remote sensing acquisition and distributional variation.","exclusion":"If robustness is primarily about scarce labels, saliency robustness, small-target visibility, oriented boxes, efficiency, or a specific operational monitoring use case, assign to that nearer direction.","nearest_boundary":"Efficient deployable detection","representative_topics":["uncertainty-aware graph neural network refinement for aerial image object detection","cross-modal alignment for aerial rgb-infrared vehicle detection under weak misalignment","robust object recognition under illumination and viewing angle variation using cnn bounding box analysis"]},{"direction_index":7,"label":"D7","name":"Change and contextual state detection","description":"Detecting temporal changes, altered regions, damage states, before-after differences, or broad contextual scene states in remote sensing imagery, including bitemporal change detection, disaster damage inference, and general context/foundation modeling when not reducible to orientation, scale, supervision, efficiency, saliency, or domain adaptation.","inclusion":"Include work whose primary objective is identifying changed or damaged regions over time, modeling post-event state, or developing broad context/foundation detection strategies that serve general remote sensing detection rather than a more specific program.","exclusion":"If the work extracts static building geometry, detects a fixed monitored target, improves oriented boxes, targets tiny objects, reduces annotation needs, or focuses on deployment efficiency, assign to that nearer direction.","nearest_boundary":"Region, contour, and saliency extraction","representative_topics":["plain vision transformer with rotated varied-size window attention for remote sensing foundation models","attention-guided remote sensing change detection using image difference and edge difference feature maps","uncertainty-aware graph neural network refinement for aerial image object detection"]}],"split":"dev"}
{"episode_id":"F259_T2024-01","field_index":259,"field":"affective computing","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Conversational emotion understanding","description":"Infer affective states in multi-turn dialogue by modeling utterance context, speaker roles, turn dependencies, emotion shifts, and conversational multimodal cues.","inclusion":"Include work whose primary objective is emotion or sentiment recognition in conversations, whether text-only, speech-supported, audiovisual, graph-based, prompt-based, or uncertainty-aware, when dialogue structure is central.","exclusion":"Non-dialogue text, speech, or audiovisual affect recognition belongs to the corresponding modality program; empathetic response generation belongs to affective agents and domain interaction.","nearest_boundary":"Affective agents and domain interaction","representative_topics":["multi-modal fusion network for emotion recognition in conversational video","speaker-guided encoder-decoder for dynamic intra- and inter-speaker dependency modeling in conversation emotion recognition","logic-augmented neural emotion recognition in multi-utterance dialogues via probabilistic soft logic"]},{"direction_index":1,"label":"D1","name":"Textual affect and appraisal reasoning","description":"Infer affective meaning from written language, including emotion and sentiment labels, fine-grained opinion structures, implicit sentiment, emotion causes, appraisals, experiencers, and motivation-emotion-action relations.","inclusion":"Include text-centered affect classification, structured or aspect-based sentiment extraction, low-resource and cross-domain sentiment, implicit affect inference, emotion-cause extraction, appraisal prediction, and causal or relational affect understanding in language.","exclusion":"Dialogue-centered affect recognition belongs to conversational emotion understanding; field-level evaluation of foundation models belongs to trustworthy foundations and evaluation; multimodal visual-text media recognition belongs to visual and multimodal media affect recognition.","nearest_boundary":"Conversational emotion understanding","representative_topics":["sequential transfer learning for emotion classification in text messages with forgetting mitigation","persian text emotion detection using transformer transfer learning on low-resource labeled data","contrastive learning for cross-domain sentiment analysis with in-batch negative samples"]},{"direction_index":2,"label":"D2","name":"Acoustic and affective media generation","description":"Model affect in speech, vocal bursts, music, and generated or transformed affective media, including recognition, synthesis, conversion, transfer, and emotion control.","inclusion":"Include audio-primary speech or vocal affect recognition, vocal-burst analysis, affective speech synthesis, emotional voice conversion, emotion-controlled music generation, music or song emotion recognition, affective narration, and media transformation whose main objective is controlling or detecting affective expression.","exclusion":"Conversational emotion recognition with turn dynamics belongs to conversational emotion understanding; empathetic response generation belongs to affective agents and domain interaction; facial or audiovisual recognition without audio/media generation focus belongs to visual and multimodal media affect recognition.","nearest_boundary":"Visual and multimodal media affect recognition","representative_topics":["few-shot speaker enrolment via attention-based feature selection for emotional vocalisation prediction","cross-domain emotion strength assessment in speech via multi-task deep learning","adversarial multi-task learning for vocal burst emotion, age, and origin prediction"]},{"direction_index":3,"label":"D3","name":"Visual and multimodal media affect recognition","description":"Recognize affect from faces, images, video, memes, visual-textual media, body motion, gait, gestures, and non-dialogue multimodal streams by modeling visual, temporal, contextual, and cross-modal cues.","inclusion":"Include facial expression recognition, action-unit and micro-expression analysis, pain or facial affect estimation, audiovisual fusion, image-text or meme sentiment, visual art emotion, temporal emotion localization in video, body/gait/gesture emotion recognition, and visual multimodal datasets when recognition of affect from media is the main objective.","exclusion":"Dialogue-specific multimodal recognition belongs to conversational emotion understanding; biosignal or wearable sensing belongs to biosignal and wearable affect sensing; general weak-supervision methods belong to representation and label learning when the learning problem is primary.","nearest_boundary":"Biosignal and wearable affect sensing","representative_topics":["uncertainty-aware audiovisual fusion via calibrated and ordinal latent distribution constraints","semi-supervised multi-task facial affect recognition with incomplete labels and adaptive thresholds","multi-task video emotion recognition with temporal encoders and inter-task correlation modeling"]},{"direction_index":4,"label":"D4","name":"Biosignal and wearable affect sensing","description":"Infer affective, stress, pain, engagement, or related states from EEG, ECG, EDA, PPG, GSR, thermal, wearable, mobile, or passive physiological sensing.","inclusion":"Include EEG emotion recognition, non-EEG physiological affect recognition, wearable sensor emotion or stress detection, cross-subject biosignal generalization, biosignal representation learning, low-cost sensing, multimodal physiological fusion, and edge deployment for affective sensing.","exclusion":"Speech, facial, text, or audiovisual recognition belongs to the corresponding modality program unless physiological or wearable signals are the main evidence source; specialized population studies belong to affective agents and domain interaction when the population or intervention defines the objective.","nearest_boundary":"Visual and multimodal media affect recognition","representative_topics":["semi-supervised generative adversarial learning for stress detection from wearable physiological signals","multimodal eda and ppg fusion for emotion recognition using low-cost wearable devices","variational autoencoder-based affective eeg representations for classification and synthetic data generation"]},{"direction_index":5,"label":"D5","name":"Representation and label learning","description":"Improve affective models by reformulating supervision, labels, annotations, uncertainty, and representations under scarce, noisy, subjective, imbalanced, incomplete, or heterogeneous data.","inclusion":"Include self-supervised, contrastive, semi-supervised, weakly supervised, pseudo-label, transfer, subject-invariant, and novelty-aware affective representation learning, as well as label distribution learning, soft labels, annotation uncertainty, temporal annotation alignment, and reliability modeling.","exclusion":"Modality-specific recognition belongs to the relevant modality program when the main contribution is a new recognizer; surveys and field-level protocols belong to trustworthy foundations and evaluation.","nearest_boundary":"Trustworthy foundations and evaluation","representative_topics":["grayscale label construction for emotion recognition in conversation without multi-label supervision","unsupervised multimodal emotion recognition via pairwise modality contrastive feature learning","subjectivity-aware visual emotion distribution learning via individual affective memory modeling"]},{"direction_index":6,"label":"D6","name":"Affective agents and domain interaction","description":"Build, regulate, or study affect-aware agents, emotionally supportive systems, affect-driven decision makers, and domain-situated affective applications where interaction, intervention, population, or applied construct defines the research objective.","inclusion":"Include empathetic and supportive dialogue generation, affective chatbots, therapy or coaching agents, affect-driven reinforcement learning, computational emotion models for agents, machine affect modeling, human-machine interaction affect, infant or elderly affect applications, mental-health or psychotherapy affect modeling, autism or education-specific affective systems, and other specialized applied affective-state modeling.","exclusion":"Pure emotion classifiers belong to modality-specific recognition programs when the application setting is incidental; non-interactive affective media generation belongs to acoustic and affective media generation; dialogue emotion classification belongs to conversational emotion understanding.","nearest_boundary":"Conversational emotion understanding","representative_topics":["affective state recognition and confusion modeling in voice-based human-machine interaction failures","positively transitioned sentiment corpus construction for emotion-affective open-domain dialogue systems","coarse-to-fine cognition and affection alignment for empathetic dialogue generation"]},{"direction_index":7,"label":"D7","name":"Trustworthy foundations and evaluation","description":"Organize, evaluate, explain, audit, benchmark, and govern affective computing methods, tasks, datasets, foundation-model uses, ethical practices, fairness, robustness, and field-level methodology.","inclusion":"Include surveys, taxonomies, conceptual frameworks, emotion lexicon guidance, evaluation protocols, challenges and benchmarks, reproducibility studies, LLM or API benchmarking for affective tasks, affective explanation methods, interpretability, bias and fairness analysis, security or adversarial studies, and deconfounding or trust-focused evaluation.","exclusion":"Ordinary empirical model improvements belong to the relevant recognition, generation, or interaction program; label-specific uncertainty modeling belongs to representation and label learning unless the aim is field-level evaluation or governance.","nearest_boundary":"Representation and label learning","representative_topics":["cross-lingual minimal emotion taxonomy via word embedding clustering for affective computing","survey of computer vision methods for animal pain and affective state recognition","multibias mitigation and sentiment knowledge enrichment for debiased multimodal conversational emotion recognition"]}],"split":"dev"}
{"episode_id":"F259_T2024-07","field_index":259,"field":"affective computing","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Conversational emotion understanding","description":"Infer affective states in multi-turn dialogue by modeling utterance context, speaker roles, turn dependencies, emotion shifts, and conversational multimodal cues.","inclusion":"Include work whose primary objective is emotion or sentiment recognition in conversations, whether text-only, speech-supported, audiovisual, graph-based, prompt-based, or uncertainty-aware, when dialogue structure is central.","exclusion":"Non-dialogue text, speech, or audiovisual affect recognition belongs to the corresponding modality program; empathetic response generation belongs to affective agents and domain interaction.","nearest_boundary":"Affective agents and domain interaction","representative_topics":["multi-modal fusion network for emotion recognition in conversational video","speaker-guided encoder-decoder for dynamic intra- and inter-speaker dependency modeling in conversation emotion recognition","logic-augmented neural emotion recognition in multi-utterance dialogues via probabilistic soft logic"]},{"direction_index":1,"label":"D1","name":"Textual affect and appraisal reasoning","description":"Infer affective meaning from written language, including emotion and sentiment labels, fine-grained opinion structures, implicit sentiment, emotion causes, appraisals, experiencers, and motivation-emotion-action relations.","inclusion":"Include text-centered affect classification, structured or aspect-based sentiment extraction, low-resource and cross-domain sentiment, implicit affect inference, emotion-cause extraction, appraisal prediction, and causal or relational affect understanding in language.","exclusion":"Dialogue-centered affect recognition belongs to conversational emotion understanding; field-level evaluation of foundation models belongs to trustworthy foundations and evaluation; multimodal visual-text media recognition belongs to visual and multimodal media affect recognition.","nearest_boundary":"Conversational emotion understanding","representative_topics":["sequential transfer learning for emotion classification in text messages with forgetting mitigation","persian text emotion detection using transformer transfer learning on low-resource labeled data","contrastive learning for cross-domain sentiment analysis with in-batch negative samples"]},{"direction_index":2,"label":"D2","name":"Acoustic and affective media generation","description":"Model affect in speech, vocal bursts, music, and generated or transformed affective media, including recognition, synthesis, conversion, transfer, and emotion control.","inclusion":"Include audio-primary speech or vocal affect recognition, vocal-burst analysis, affective speech synthesis, emotional voice conversion, emotion-controlled music generation, music or song emotion recognition, affective narration, and media transformation whose main objective is controlling or detecting affective expression.","exclusion":"Conversational emotion recognition with turn dynamics belongs to conversational emotion understanding; empathetic response generation belongs to affective agents and domain interaction; facial or audiovisual recognition without audio/media generation focus belongs to visual and multimodal media affect recognition.","nearest_boundary":"Visual and multimodal media affect recognition","representative_topics":["few-shot speaker enrolment via attention-based feature selection for emotional vocalisation prediction","cross-domain emotion strength assessment in speech via multi-task deep learning","adversarial multi-task learning for vocal burst emotion, age, and origin prediction"]},{"direction_index":3,"label":"D3","name":"Visual and multimodal media affect recognition","description":"Recognize affect from faces, images, video, memes, visual-textual media, body motion, gait, gestures, and non-dialogue multimodal streams by modeling visual, temporal, contextual, and cross-modal cues.","inclusion":"Include facial expression recognition, action-unit and micro-expression analysis, pain or facial affect estimation, audiovisual fusion, image-text or meme sentiment, visual art emotion, temporal emotion localization in video, body/gait/gesture emotion recognition, and visual multimodal datasets when recognition of affect from media is the main objective.","exclusion":"Dialogue-specific multimodal recognition belongs to conversational emotion understanding; biosignal or wearable sensing belongs to biosignal and wearable affect sensing; general weak-supervision methods belong to representation and label learning when the learning problem is primary.","nearest_boundary":"Biosignal and wearable affect sensing","representative_topics":["uncertainty-aware audiovisual fusion via calibrated and ordinal latent distribution constraints","semi-supervised multi-task facial affect recognition with incomplete labels and adaptive thresholds","multi-task video emotion recognition with temporal encoders and inter-task correlation modeling"]},{"direction_index":4,"label":"D4","name":"Biosignal and wearable affect sensing","description":"Infer affective, stress, pain, engagement, or related states from EEG, ECG, EDA, PPG, GSR, thermal, wearable, mobile, or passive physiological sensing.","inclusion":"Include EEG emotion recognition, non-EEG physiological affect recognition, wearable sensor emotion or stress detection, cross-subject biosignal generalization, biosignal representation learning, low-cost sensing, multimodal physiological fusion, and edge deployment for affective sensing.","exclusion":"Speech, facial, text, or audiovisual recognition belongs to the corresponding modality program unless physiological or wearable signals are the main evidence source; specialized population studies belong to affective agents and domain interaction when the population or intervention defines the objective.","nearest_boundary":"Visual and multimodal media affect recognition","representative_topics":["semi-supervised generative adversarial learning for stress detection from wearable physiological signals","multimodal eda and ppg fusion for emotion recognition using low-cost wearable devices","variational autoencoder-based affective eeg representations for classification and synthetic data generation"]},{"direction_index":5,"label":"D5","name":"Representation and label learning","description":"Improve affective models by reformulating supervision, labels, annotations, uncertainty, and representations under scarce, noisy, subjective, imbalanced, incomplete, or heterogeneous data.","inclusion":"Include self-supervised, contrastive, semi-supervised, weakly supervised, pseudo-label, transfer, subject-invariant, and novelty-aware affective representation learning, as well as label distribution learning, soft labels, annotation uncertainty, temporal annotation alignment, and reliability modeling.","exclusion":"Modality-specific recognition belongs to the relevant modality program when the main contribution is a new recognizer; surveys and field-level protocols belong to trustworthy foundations and evaluation.","nearest_boundary":"Trustworthy foundations and evaluation","representative_topics":["grayscale label construction for emotion recognition in conversation without multi-label supervision","unsupervised multimodal emotion recognition via pairwise modality contrastive feature learning","subjectivity-aware visual emotion distribution learning via individual affective memory modeling"]},{"direction_index":6,"label":"D6","name":"Affective agents and domain interaction","description":"Build, regulate, or study affect-aware agents, emotionally supportive systems, affect-driven decision makers, and domain-situated affective applications where interaction, intervention, population, or applied construct defines the research objective.","inclusion":"Include empathetic and supportive dialogue generation, affective chatbots, therapy or coaching agents, affect-driven reinforcement learning, computational emotion models for agents, machine affect modeling, human-machine interaction affect, infant or elderly affect applications, mental-health or psychotherapy affect modeling, autism or education-specific affective systems, and other specialized applied affective-state modeling.","exclusion":"Pure emotion classifiers belong to modality-specific recognition programs when the application setting is incidental; non-interactive affective media generation belongs to acoustic and affective media generation; dialogue emotion classification belongs to conversational emotion understanding.","nearest_boundary":"Conversational emotion understanding","representative_topics":["affective state recognition and confusion modeling in voice-based human-machine interaction failures","positively transitioned sentiment corpus construction for emotion-affective open-domain dialogue systems","coarse-to-fine cognition and affection alignment for empathetic dialogue generation"]},{"direction_index":7,"label":"D7","name":"Trustworthy foundations and evaluation","description":"Organize, evaluate, explain, audit, benchmark, and govern affective computing methods, tasks, datasets, foundation-model uses, ethical practices, fairness, robustness, and field-level methodology.","inclusion":"Include surveys, taxonomies, conceptual frameworks, emotion lexicon guidance, evaluation protocols, challenges and benchmarks, reproducibility studies, LLM or API benchmarking for affective tasks, affective explanation methods, interpretability, bias and fairness analysis, security or adversarial studies, and deconfounding or trust-focused evaluation.","exclusion":"Ordinary empirical model improvements belong to the relevant recognition, generation, or interaction program; label-specific uncertainty modeling belongs to representation and label learning unless the aim is field-level evaluation or governance.","nearest_boundary":"Representation and label learning","representative_topics":["cross-lingual minimal emotion taxonomy via word embedding clustering for affective computing","survey of computer vision methods for animal pain and affective state recognition","multibias mitigation and sentiment knowledge enrichment for debiased multimodal conversational emotion recognition"]}],"split":"dev"}
{"episode_id":"F259_T2025-01","field_index":259,"field":"affective computing","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Conversational emotion understanding","description":"Infer affective states in multi-turn dialogue by modeling utterance context, speaker roles, turn dependencies, emotion shifts, and conversational multimodal cues.","inclusion":"Include work whose primary objective is emotion or sentiment recognition in conversations, whether text-only, speech-supported, audiovisual, graph-based, prompt-based, or uncertainty-aware, when dialogue structure is central.","exclusion":"Non-dialogue text, speech, or audiovisual affect recognition belongs to the corresponding modality program; empathetic response generation belongs to affective agents and domain interaction.","nearest_boundary":"Affective agents and domain interaction","representative_topics":["multi-modal fusion network for emotion recognition in conversational video","speaker-guided encoder-decoder for dynamic intra- and inter-speaker dependency modeling in conversation emotion recognition","logic-augmented neural emotion recognition in multi-utterance dialogues via probabilistic soft logic"]},{"direction_index":1,"label":"D1","name":"Textual affect and appraisal reasoning","description":"Infer affective meaning from written language, including emotion and sentiment labels, fine-grained opinion structures, implicit sentiment, emotion causes, appraisals, experiencers, and motivation-emotion-action relations.","inclusion":"Include text-centered affect classification, structured or aspect-based sentiment extraction, low-resource and cross-domain sentiment, implicit affect inference, emotion-cause extraction, appraisal prediction, and causal or relational affect understanding in language.","exclusion":"Dialogue-centered affect recognition belongs to conversational emotion understanding; field-level evaluation of foundation models belongs to trustworthy foundations and evaluation; multimodal visual-text media recognition belongs to visual and multimodal media affect recognition.","nearest_boundary":"Conversational emotion understanding","representative_topics":["sequential transfer learning for emotion classification in text messages with forgetting mitigation","persian text emotion detection using transformer transfer learning on low-resource labeled data","contrastive learning for cross-domain sentiment analysis with in-batch negative samples"]},{"direction_index":2,"label":"D2","name":"Acoustic and affective media generation","description":"Model affect in speech, vocal bursts, music, and generated or transformed affective media, including recognition, synthesis, conversion, transfer, and emotion control.","inclusion":"Include audio-primary speech or vocal affect recognition, vocal-burst analysis, affective speech synthesis, emotional voice conversion, emotion-controlled music generation, music or song emotion recognition, affective narration, and media transformation whose main objective is controlling or detecting affective expression.","exclusion":"Conversational emotion recognition with turn dynamics belongs to conversational emotion understanding; empathetic response generation belongs to affective agents and domain interaction; facial or audiovisual recognition without audio/media generation focus belongs to visual and multimodal media affect recognition.","nearest_boundary":"Visual and multimodal media affect recognition","representative_topics":["few-shot speaker enrolment via attention-based feature selection for emotional vocalisation prediction","cross-domain emotion strength assessment in speech via multi-task deep learning","adversarial multi-task learning for vocal burst emotion, age, and origin prediction"]},{"direction_index":3,"label":"D3","name":"Visual and multimodal media affect recognition","description":"Recognize affect from faces, images, video, memes, visual-textual media, body motion, gait, gestures, and non-dialogue multimodal streams by modeling visual, temporal, contextual, and cross-modal cues.","inclusion":"Include facial expression recognition, action-unit and micro-expression analysis, pain or facial affect estimation, audiovisual fusion, image-text or meme sentiment, visual art emotion, temporal emotion localization in video, body/gait/gesture emotion recognition, and visual multimodal datasets when recognition of affect from media is the main objective.","exclusion":"Dialogue-specific multimodal recognition belongs to conversational emotion understanding; biosignal or wearable sensing belongs to biosignal and wearable affect sensing; general weak-supervision methods belong to representation and label learning when the learning problem is primary.","nearest_boundary":"Biosignal and wearable affect sensing","representative_topics":["uncertainty-aware audiovisual fusion via calibrated and ordinal latent distribution constraints","semi-supervised multi-task facial affect recognition with incomplete labels and adaptive thresholds","multi-task video emotion recognition with temporal encoders and inter-task correlation modeling"]},{"direction_index":4,"label":"D4","name":"Biosignal and wearable affect sensing","description":"Infer affective, stress, pain, engagement, or related states from EEG, ECG, EDA, PPG, GSR, thermal, wearable, mobile, or passive physiological sensing.","inclusion":"Include EEG emotion recognition, non-EEG physiological affect recognition, wearable sensor emotion or stress detection, cross-subject biosignal generalization, biosignal representation learning, low-cost sensing, multimodal physiological fusion, and edge deployment for affective sensing.","exclusion":"Speech, facial, text, or audiovisual recognition belongs to the corresponding modality program unless physiological or wearable signals are the main evidence source; specialized population studies belong to affective agents and domain interaction when the population or intervention defines the objective.","nearest_boundary":"Visual and multimodal media affect recognition","representative_topics":["semi-supervised generative adversarial learning for stress detection from wearable physiological signals","multimodal eda and ppg fusion for emotion recognition using low-cost wearable devices","variational autoencoder-based affective eeg representations for classification and synthetic data generation"]},{"direction_index":5,"label":"D5","name":"Representation and label learning","description":"Improve affective models by reformulating supervision, labels, annotations, uncertainty, and representations under scarce, noisy, subjective, imbalanced, incomplete, or heterogeneous data.","inclusion":"Include self-supervised, contrastive, semi-supervised, weakly supervised, pseudo-label, transfer, subject-invariant, and novelty-aware affective representation learning, as well as label distribution learning, soft labels, annotation uncertainty, temporal annotation alignment, and reliability modeling.","exclusion":"Modality-specific recognition belongs to the relevant modality program when the main contribution is a new recognizer; surveys and field-level protocols belong to trustworthy foundations and evaluation.","nearest_boundary":"Trustworthy foundations and evaluation","representative_topics":["grayscale label construction for emotion recognition in conversation without multi-label supervision","unsupervised multimodal emotion recognition via pairwise modality contrastive feature learning","subjectivity-aware visual emotion distribution learning via individual affective memory modeling"]},{"direction_index":6,"label":"D6","name":"Affective agents and domain interaction","description":"Build, regulate, or study affect-aware agents, emotionally supportive systems, affect-driven decision makers, and domain-situated affective applications where interaction, intervention, population, or applied construct defines the research objective.","inclusion":"Include empathetic and supportive dialogue generation, affective chatbots, therapy or coaching agents, affect-driven reinforcement learning, computational emotion models for agents, machine affect modeling, human-machine interaction affect, infant or elderly affect applications, mental-health or psychotherapy affect modeling, autism or education-specific affective systems, and other specialized applied affective-state modeling.","exclusion":"Pure emotion classifiers belong to modality-specific recognition programs when the application setting is incidental; non-interactive affective media generation belongs to acoustic and affective media generation; dialogue emotion classification belongs to conversational emotion understanding.","nearest_boundary":"Conversational emotion understanding","representative_topics":["affective state recognition and confusion modeling in voice-based human-machine interaction failures","positively transitioned sentiment corpus construction for emotion-affective open-domain dialogue systems","coarse-to-fine cognition and affection alignment for empathetic dialogue generation"]},{"direction_index":7,"label":"D7","name":"Trustworthy foundations and evaluation","description":"Organize, evaluate, explain, audit, benchmark, and govern affective computing methods, tasks, datasets, foundation-model uses, ethical practices, fairness, robustness, and field-level methodology.","inclusion":"Include surveys, taxonomies, conceptual frameworks, emotion lexicon guidance, evaluation protocols, challenges and benchmarks, reproducibility studies, LLM or API benchmarking for affective tasks, affective explanation methods, interpretability, bias and fairness analysis, security or adversarial studies, and deconfounding or trust-focused evaluation.","exclusion":"Ordinary empirical model improvements belong to the relevant recognition, generation, or interaction program; label-specific uncertainty modeling belongs to representation and label learning unless the aim is field-level evaluation or governance.","nearest_boundary":"Representation and label learning","representative_topics":["cross-lingual minimal emotion taxonomy via word embedding clustering for affective computing","survey of computer vision methods for animal pain and affective state recognition","multibias mitigation and sentiment knowledge enrichment for debiased multimodal conversational emotion recognition"]}],"split":"dev"}
{"episode_id":"F259_T2025-07","field_index":259,"field":"affective computing","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Conversational emotion understanding","description":"Infer affective states in multi-turn dialogue by modeling utterance context, speaker roles, turn dependencies, emotion shifts, and conversational multimodal cues.","inclusion":"Include work whose primary objective is emotion or sentiment recognition in conversations, whether text-only, speech-supported, audiovisual, graph-based, prompt-based, or uncertainty-aware, when dialogue structure is central.","exclusion":"Non-dialogue text, speech, or audiovisual affect recognition belongs to the corresponding modality program; empathetic response generation belongs to affective agents and domain interaction.","nearest_boundary":"Affective agents and domain interaction","representative_topics":["multi-modal fusion network for emotion recognition in conversational video","speaker-guided encoder-decoder for dynamic intra- and inter-speaker dependency modeling in conversation emotion recognition","logic-augmented neural emotion recognition in multi-utterance dialogues via probabilistic soft logic"]},{"direction_index":1,"label":"D1","name":"Textual affect and appraisal reasoning","description":"Infer affective meaning from written language, including emotion and sentiment labels, fine-grained opinion structures, implicit sentiment, emotion causes, appraisals, experiencers, and motivation-emotion-action relations.","inclusion":"Include text-centered affect classification, structured or aspect-based sentiment extraction, low-resource and cross-domain sentiment, implicit affect inference, emotion-cause extraction, appraisal prediction, and causal or relational affect understanding in language.","exclusion":"Dialogue-centered affect recognition belongs to conversational emotion understanding; field-level evaluation of foundation models belongs to trustworthy foundations and evaluation; multimodal visual-text media recognition belongs to visual and multimodal media affect recognition.","nearest_boundary":"Conversational emotion understanding","representative_topics":["sequential transfer learning for emotion classification in text messages with forgetting mitigation","persian text emotion detection using transformer transfer learning on low-resource labeled data","contrastive learning for cross-domain sentiment analysis with in-batch negative samples"]},{"direction_index":2,"label":"D2","name":"Acoustic and affective media generation","description":"Model affect in speech, vocal bursts, music, and generated or transformed affective media, including recognition, synthesis, conversion, transfer, and emotion control.","inclusion":"Include audio-primary speech or vocal affect recognition, vocal-burst analysis, affective speech synthesis, emotional voice conversion, emotion-controlled music generation, music or song emotion recognition, affective narration, and media transformation whose main objective is controlling or detecting affective expression.","exclusion":"Conversational emotion recognition with turn dynamics belongs to conversational emotion understanding; empathetic response generation belongs to affective agents and domain interaction; facial or audiovisual recognition without audio/media generation focus belongs to visual and multimodal media affect recognition.","nearest_boundary":"Visual and multimodal media affect recognition","representative_topics":["few-shot speaker enrolment via attention-based feature selection for emotional vocalisation prediction","cross-domain emotion strength assessment in speech via multi-task deep learning","adversarial multi-task learning for vocal burst emotion, age, and origin prediction"]},{"direction_index":3,"label":"D3","name":"Visual and multimodal media affect recognition","description":"Recognize affect from faces, images, video, memes, visual-textual media, body motion, gait, gestures, and non-dialogue multimodal streams by modeling visual, temporal, contextual, and cross-modal cues.","inclusion":"Include facial expression recognition, action-unit and micro-expression analysis, pain or facial affect estimation, audiovisual fusion, image-text or meme sentiment, visual art emotion, temporal emotion localization in video, body/gait/gesture emotion recognition, and visual multimodal datasets when recognition of affect from media is the main objective.","exclusion":"Dialogue-specific multimodal recognition belongs to conversational emotion understanding; biosignal or wearable sensing belongs to biosignal and wearable affect sensing; general weak-supervision methods belong to representation and label learning when the learning problem is primary.","nearest_boundary":"Biosignal and wearable affect sensing","representative_topics":["uncertainty-aware audiovisual fusion via calibrated and ordinal latent distribution constraints","semi-supervised multi-task facial affect recognition with incomplete labels and adaptive thresholds","multi-task video emotion recognition with temporal encoders and inter-task correlation modeling"]},{"direction_index":4,"label":"D4","name":"Biosignal and wearable affect sensing","description":"Infer affective, stress, pain, engagement, or related states from EEG, ECG, EDA, PPG, GSR, thermal, wearable, mobile, or passive physiological sensing.","inclusion":"Include EEG emotion recognition, non-EEG physiological affect recognition, wearable sensor emotion or stress detection, cross-subject biosignal generalization, biosignal representation learning, low-cost sensing, multimodal physiological fusion, and edge deployment for affective sensing.","exclusion":"Speech, facial, text, or audiovisual recognition belongs to the corresponding modality program unless physiological or wearable signals are the main evidence source; specialized population studies belong to affective agents and domain interaction when the population or intervention defines the objective.","nearest_boundary":"Visual and multimodal media affect recognition","representative_topics":["semi-supervised generative adversarial learning for stress detection from wearable physiological signals","multimodal eda and ppg fusion for emotion recognition using low-cost wearable devices","variational autoencoder-based affective eeg representations for classification and synthetic data generation"]},{"direction_index":5,"label":"D5","name":"Representation and label learning","description":"Improve affective models by reformulating supervision, labels, annotations, uncertainty, and representations under scarce, noisy, subjective, imbalanced, incomplete, or heterogeneous data.","inclusion":"Include self-supervised, contrastive, semi-supervised, weakly supervised, pseudo-label, transfer, subject-invariant, and novelty-aware affective representation learning, as well as label distribution learning, soft labels, annotation uncertainty, temporal annotation alignment, and reliability modeling.","exclusion":"Modality-specific recognition belongs to the relevant modality program when the main contribution is a new recognizer; surveys and field-level protocols belong to trustworthy foundations and evaluation.","nearest_boundary":"Trustworthy foundations and evaluation","representative_topics":["grayscale label construction for emotion recognition in conversation without multi-label supervision","unsupervised multimodal emotion recognition via pairwise modality contrastive feature learning","subjectivity-aware visual emotion distribution learning via individual affective memory modeling"]},{"direction_index":6,"label":"D6","name":"Affective agents and domain interaction","description":"Build, regulate, or study affect-aware agents, emotionally supportive systems, affect-driven decision makers, and domain-situated affective applications where interaction, intervention, population, or applied construct defines the research objective.","inclusion":"Include empathetic and supportive dialogue generation, affective chatbots, therapy or coaching agents, affect-driven reinforcement learning, computational emotion models for agents, machine affect modeling, human-machine interaction affect, infant or elderly affect applications, mental-health or psychotherapy affect modeling, autism or education-specific affective systems, and other specialized applied affective-state modeling.","exclusion":"Pure emotion classifiers belong to modality-specific recognition programs when the application setting is incidental; non-interactive affective media generation belongs to acoustic and affective media generation; dialogue emotion classification belongs to conversational emotion understanding.","nearest_boundary":"Conversational emotion understanding","representative_topics":["affective state recognition and confusion modeling in voice-based human-machine interaction failures","positively transitioned sentiment corpus construction for emotion-affective open-domain dialogue systems","coarse-to-fine cognition and affection alignment for empathetic dialogue generation"]},{"direction_index":7,"label":"D7","name":"Trustworthy foundations and evaluation","description":"Organize, evaluate, explain, audit, benchmark, and govern affective computing methods, tasks, datasets, foundation-model uses, ethical practices, fairness, robustness, and field-level methodology.","inclusion":"Include surveys, taxonomies, conceptual frameworks, emotion lexicon guidance, evaluation protocols, challenges and benchmarks, reproducibility studies, LLM or API benchmarking for affective tasks, affective explanation methods, interpretability, bias and fairness analysis, security or adversarial studies, and deconfounding or trust-focused evaluation.","exclusion":"Ordinary empirical model improvements belong to the relevant recognition, generation, or interaction program; label-specific uncertainty modeling belongs to representation and label learning unless the aim is field-level evaluation or governance.","nearest_boundary":"Representation and label learning","representative_topics":["cross-lingual minimal emotion taxonomy via word embedding clustering for affective computing","survey of computer vision methods for animal pain and affective state recognition","multibias mitigation and sentiment knowledge enrichment for debiased multimodal conversational emotion recognition"]}],"split":"dev"}
{"episode_id":"F259_T2026-01","field_index":259,"field":"affective computing","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Conversational emotion understanding","description":"Infer affective states in multi-turn dialogue by modeling utterance context, speaker roles, turn dependencies, emotion shifts, and conversational multimodal cues.","inclusion":"Include work whose primary objective is emotion or sentiment recognition in conversations, whether text-only, speech-supported, audiovisual, graph-based, prompt-based, or uncertainty-aware, when dialogue structure is central.","exclusion":"Non-dialogue text, speech, or audiovisual affect recognition belongs to the corresponding modality program; empathetic response generation belongs to affective agents and domain interaction.","nearest_boundary":"Affective agents and domain interaction","representative_topics":["multi-modal fusion network for emotion recognition in conversational video","speaker-guided encoder-decoder for dynamic intra- and inter-speaker dependency modeling in conversation emotion recognition","logic-augmented neural emotion recognition in multi-utterance dialogues via probabilistic soft logic"]},{"direction_index":1,"label":"D1","name":"Textual affect and appraisal reasoning","description":"Infer affective meaning from written language, including emotion and sentiment labels, fine-grained opinion structures, implicit sentiment, emotion causes, appraisals, experiencers, and motivation-emotion-action relations.","inclusion":"Include text-centered affect classification, structured or aspect-based sentiment extraction, low-resource and cross-domain sentiment, implicit affect inference, emotion-cause extraction, appraisal prediction, and causal or relational affect understanding in language.","exclusion":"Dialogue-centered affect recognition belongs to conversational emotion understanding; field-level evaluation of foundation models belongs to trustworthy foundations and evaluation; multimodal visual-text media recognition belongs to visual and multimodal media affect recognition.","nearest_boundary":"Conversational emotion understanding","representative_topics":["sequential transfer learning for emotion classification in text messages with forgetting mitigation","persian text emotion detection using transformer transfer learning on low-resource labeled data","contrastive learning for cross-domain sentiment analysis with in-batch negative samples"]},{"direction_index":2,"label":"D2","name":"Acoustic and affective media generation","description":"Model affect in speech, vocal bursts, music, and generated or transformed affective media, including recognition, synthesis, conversion, transfer, and emotion control.","inclusion":"Include audio-primary speech or vocal affect recognition, vocal-burst analysis, affective speech synthesis, emotional voice conversion, emotion-controlled music generation, music or song emotion recognition, affective narration, and media transformation whose main objective is controlling or detecting affective expression.","exclusion":"Conversational emotion recognition with turn dynamics belongs to conversational emotion understanding; empathetic response generation belongs to affective agents and domain interaction; facial or audiovisual recognition without audio/media generation focus belongs to visual and multimodal media affect recognition.","nearest_boundary":"Visual and multimodal media affect recognition","representative_topics":["few-shot speaker enrolment via attention-based feature selection for emotional vocalisation prediction","cross-domain emotion strength assessment in speech via multi-task deep learning","adversarial multi-task learning for vocal burst emotion, age, and origin prediction"]},{"direction_index":3,"label":"D3","name":"Visual and multimodal media affect recognition","description":"Recognize affect from faces, images, video, memes, visual-textual media, body motion, gait, gestures, and non-dialogue multimodal streams by modeling visual, temporal, contextual, and cross-modal cues.","inclusion":"Include facial expression recognition, action-unit and micro-expression analysis, pain or facial affect estimation, audiovisual fusion, image-text or meme sentiment, visual art emotion, temporal emotion localization in video, body/gait/gesture emotion recognition, and visual multimodal datasets when recognition of affect from media is the main objective.","exclusion":"Dialogue-specific multimodal recognition belongs to conversational emotion understanding; biosignal or wearable sensing belongs to biosignal and wearable affect sensing; general weak-supervision methods belong to representation and label learning when the learning problem is primary.","nearest_boundary":"Biosignal and wearable affect sensing","representative_topics":["uncertainty-aware audiovisual fusion via calibrated and ordinal latent distribution constraints","semi-supervised multi-task facial affect recognition with incomplete labels and adaptive thresholds","multi-task video emotion recognition with temporal encoders and inter-task correlation modeling"]},{"direction_index":4,"label":"D4","name":"Biosignal and wearable affect sensing","description":"Infer affective, stress, pain, engagement, or related states from EEG, ECG, EDA, PPG, GSR, thermal, wearable, mobile, or passive physiological sensing.","inclusion":"Include EEG emotion recognition, non-EEG physiological affect recognition, wearable sensor emotion or stress detection, cross-subject biosignal generalization, biosignal representation learning, low-cost sensing, multimodal physiological fusion, and edge deployment for affective sensing.","exclusion":"Speech, facial, text, or audiovisual recognition belongs to the corresponding modality program unless physiological or wearable signals are the main evidence source; specialized population studies belong to affective agents and domain interaction when the population or intervention defines the objective.","nearest_boundary":"Visual and multimodal media affect recognition","representative_topics":["semi-supervised generative adversarial learning for stress detection from wearable physiological signals","multimodal eda and ppg fusion for emotion recognition using low-cost wearable devices","variational autoencoder-based affective eeg representations for classification and synthetic data generation"]},{"direction_index":5,"label":"D5","name":"Representation and label learning","description":"Improve affective models by reformulating supervision, labels, annotations, uncertainty, and representations under scarce, noisy, subjective, imbalanced, incomplete, or heterogeneous data.","inclusion":"Include self-supervised, contrastive, semi-supervised, weakly supervised, pseudo-label, transfer, subject-invariant, and novelty-aware affective representation learning, as well as label distribution learning, soft labels, annotation uncertainty, temporal annotation alignment, and reliability modeling.","exclusion":"Modality-specific recognition belongs to the relevant modality program when the main contribution is a new recognizer; surveys and field-level protocols belong to trustworthy foundations and evaluation.","nearest_boundary":"Trustworthy foundations and evaluation","representative_topics":["grayscale label construction for emotion recognition in conversation without multi-label supervision","unsupervised multimodal emotion recognition via pairwise modality contrastive feature learning","subjectivity-aware visual emotion distribution learning via individual affective memory modeling"]},{"direction_index":6,"label":"D6","name":"Affective agents and domain interaction","description":"Build, regulate, or study affect-aware agents, emotionally supportive systems, affect-driven decision makers, and domain-situated affective applications where interaction, intervention, population, or applied construct defines the research objective.","inclusion":"Include empathetic and supportive dialogue generation, affective chatbots, therapy or coaching agents, affect-driven reinforcement learning, computational emotion models for agents, machine affect modeling, human-machine interaction affect, infant or elderly affect applications, mental-health or psychotherapy affect modeling, autism or education-specific affective systems, and other specialized applied affective-state modeling.","exclusion":"Pure emotion classifiers belong to modality-specific recognition programs when the application setting is incidental; non-interactive affective media generation belongs to acoustic and affective media generation; dialogue emotion classification belongs to conversational emotion understanding.","nearest_boundary":"Conversational emotion understanding","representative_topics":["affective state recognition and confusion modeling in voice-based human-machine interaction failures","positively transitioned sentiment corpus construction for emotion-affective open-domain dialogue systems","coarse-to-fine cognition and affection alignment for empathetic dialogue generation"]},{"direction_index":7,"label":"D7","name":"Trustworthy foundations and evaluation","description":"Organize, evaluate, explain, audit, benchmark, and govern affective computing methods, tasks, datasets, foundation-model uses, ethical practices, fairness, robustness, and field-level methodology.","inclusion":"Include surveys, taxonomies, conceptual frameworks, emotion lexicon guidance, evaluation protocols, challenges and benchmarks, reproducibility studies, LLM or API benchmarking for affective tasks, affective explanation methods, interpretability, bias and fairness analysis, security or adversarial studies, and deconfounding or trust-focused evaluation.","exclusion":"Ordinary empirical model improvements belong to the relevant recognition, generation, or interaction program; label-specific uncertainty modeling belongs to representation and label learning unless the aim is field-level evaluation or governance.","nearest_boundary":"Representation and label learning","representative_topics":["cross-lingual minimal emotion taxonomy via word embedding clustering for affective computing","survey of computer vision methods for animal pain and affective state recognition","multibias mitigation and sentiment knowledge enrichment for debiased multimodal conversational emotion recognition"]}],"split":"dev"}
{"episode_id":"F260_T2024-01","field_index":260,"field":"clinical outcome prediction","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Infectious and respiratory outcomes","description":"Prediction, detection, severity stratification, treatment outcome, adherence, and cost outcomes for COVID-19 and other respiratory or infectious diseases.","inclusion":"Assign atoms whose primary objective is clinical prediction for acute COVID-19, tuberculosis, COPD, sleep apnea, bacterial/viral infection differentiation, liver infection, or related infectious/respiratory outcomes, regardless of modality or model family.","exclusion":"If the central objective is hospital-wide sepsis or ICU deterioration rather than a disease-specific infectious/respiratory program, assign to acute-care trajectory modeling; if the central objective is causal treatment-effect estimation, assign to trustworthy clinical prediction.","nearest_boundary":"Acute-care and longitudinal EHR trajectories","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]},{"direction_index":1,"label":"D1","name":"Cancer prognosis and response","description":"Prediction of oncologic survival, recurrence, metastasis, grade, invasion, progression, risk, and treatment response using imaging, pathology, genomics, clinical text, tabular data, or multimodal cancer data.","inclusion":"Assign atoms whose primary objective is prognostication, response assessment, progression prediction, grading with outcome relevance, or risk stratification for cancer patients or cancer screening populations.","exclusion":"General survival-method atoms without a cancer-centered objective belong to survival methodology; non-cancer imaging or specialty-disease prediction belongs to the relevant disease program.","nearest_boundary":"Survival and time-to-event methodology","representative_topics":["deep learning segmentation of necrosis ratio from whole slide images for osteosarcoma chemotherapy response and survival prediction","multi-task contrastive learning for joint meningioma grading and brain invasion prediction from mri","multimodal fusion of radiology and clinical text for lung cancer survival prediction"]},{"direction_index":2,"label":"D2","name":"Survival methodology","description":"General-purpose clinical time-to-event prediction methodology for censored outcomes, tied events, semi-competing risks, longitudinal risk updates, survival pretraining, and specialized survival losses or architectures.","inclusion":"Assign atoms whose primary scientific objective is methodological development for survival or time-to-event prediction intended to generalize across clinical domains, even when demonstrated on a disease dataset.","exclusion":"Disease-centered survival applications, such as cancer prognosis, CKD deterioration, dementia progression, or COVID mortality, belong to the relevant disease program unless the censored-survival method is the central contribution.","nearest_boundary":"Cancer prognosis and response","representative_topics":["efficient neural network cox proportional hazards estimation with tied-event support","neural expectation-maximization for semi-competing risk prediction in multi-state survival","self-supervised foundation model pretraining for time-to-event prediction in clinical records"]},{"direction_index":3,"label":"D3","name":"EHR trajectory and fusion modeling","description":"Prediction of hospital, ICU, readmission, deterioration, acuity, patient-journey, and general EHR outcomes through longitudinal sequence modeling, graph representation learning, clinical NLP, multimodal fusion, and heterogeneous patient representations.","inclusion":"Assign atoms whose primary objective is modeling patient trajectories or heterogeneous EHR evidence to predict clinical outcomes, including temporal EHR architectures, ICU early warning, graph/patient-similarity learning, clinical-note prediction, and multimodal structured/unstructured fusion.","exclusion":"If portability, fairness, interpretability, or causal safety is the primary objective, assign to trustworthy clinical prediction; if a disease-specific outcome program dominates, assign to the corresponding disease direction.","nearest_boundary":"Trustworthy clinical prediction","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","temporal attention networks for ehr health prediction with short-term and long-term dependency modeling","multimodal ehr trajectory modelling using dense and recurrent architectures for clinical outcome prediction"]},{"direction_index":4,"label":"D4","name":"Cardiovascular and cardiac-signal prediction","description":"Prediction of cardiovascular disease, heart failure, coronary disease, stroke, arrhythmia, ECG abnormalities, electrolyte or hemodynamic states, and related cardiovascular outcomes from clinical, retinal, ECG, wearable, or multimodal data.","inclusion":"Assign atoms whose primary clinical target is cardiovascular or stroke risk, cardiac prognosis, rhythm diagnosis, ECG-derived cardiac/electrolyte/hemodynamic prediction, or cardiovascular decision support.","exclusion":"Retinal, wearable, or physiological-signal studies aimed primarily at non-cardiovascular conditions belong to the relevant neurologic, specialty-disease, or infectious direction; general survival methods belong to survival methodology.","nearest_boundary":"Neurologic, mental-health, and behavioral states","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","retinal image-based deep learning for cardiovascular risk prediction and component attribution","anova-driven automatic attribute selection for cardiovascular disease prognosis prediction"]},{"direction_index":5,"label":"D5","name":"Neurobehavioral state prediction","description":"Prediction, diagnosis, staging, monitoring, relapse detection, severity estimation, and progression modeling for neurologic, neurodegenerative, neurophysiologic, mental-health, and behavioral states.","inclusion":"Assign atoms whose primary objective is Alzheimer/dementia, Parkinson or movement-disorder, EEG seizure or brain-electrophysiology, depression, schizophrenia, psychosis relapse, psychotherapy monitoring, or closely related neurobehavioral clinical outcome prediction; include broad specialized-signal atoms when their objective is detecting clinical functional or neurologic/behavioral states.","exclusion":"Orthopedic injury, pain, metabolic, ophthalmic, or respiratory disease prediction belongs to non-cancer specialty disease or infectious/respiratory directions unless the neurologic or mental-health state is the central target.","nearest_boundary":"Non-cancer specialty disease prognosis","representative_topics":["self-supervised motion forecasting pre-training for few-shot parkinson's gait impairment severity estimation","personalized parkinson's medication status prediction from remote wearable time-series using transformer attention","temporal pattern mining and survival analysis for alzheimer's disease risk factor identification from longitudinal clinical data"]},{"direction_index":6,"label":"D6","name":"Trustworthy clinical prediction","description":"Clinical prediction research whose primary objective is interpretability, explainability, reliability under imperfect data, uncertainty, fairness, causal validity, safe actionability, privacy-preserving collaboration, portability, or cross-site deployment.","inclusion":"Assign atoms centered on transparent risk models, checklist/prototype/concept or knowledge-guided explanations, robustness to missingness/imbalance/shift/leakage, uncertainty estimation, fairness, causal inference, target-trial emulation, federated learning, privacy, transfer, or multi-site generalization.","exclusion":"Routine feature-importance reporting, ordinary imputation, or standard external validation inside a disease-focused prediction study remains in the disease program unless trustworthiness or deployment is the main scientific contribution.","nearest_boundary":"EHR trajectory and fusion modeling","representative_topics":["interpretable clinical risk scores via explicit intervention policy modeling in icu","explainable hierarchical attention-recurrent models for intraoperative complication prediction","automated interpretable checklist learning from continuous icu clinical data via mixed-integer programming"]},{"direction_index":7,"label":"D7","name":"Non-cancer specialty prognosis","description":"Prediction, diagnosis, grading, progression, treatment outcome, and monitoring for non-cancer, non-cardiovascular, non-neurobehavioral specialty conditions including metabolic, renal, hepatic, ophthalmic, musculoskeletal, orthopedic, pain, allergy, and related chronic or specialty diseases.","inclusion":"Assign atoms whose primary objective is disease- or specialty-specific clinical prediction for diabetes, glucose dynamics, CKD, liver/metabolic disease, ophthalmic disease, musculoskeletal injury, pain, orthopedic outcomes, food allergy, autism as a specialty diagnosis, or comparable non-cancer clinical conditions not covered by the other disease programs.","exclusion":"Cancer, cardiovascular, infectious/respiratory, neurobehavioral, and general methodological papers belong to their respective directions even when they use similar standard ML pipelines or specialty modalities.","nearest_boundary":"Neurobehavioral state prediction","representative_topics":["blockchain-secured iot hypoglycemia detection with automated insulin injection using random forest","claims-based dynamic prediction of renal replacement therapy onset in chronic kidney disease","transformer-based personalized glucose forecasting with mixture-of-distributions uncertainty quantification"]}],"split":"dev"}
{"episode_id":"F260_T2024-07","field_index":260,"field":"clinical outcome prediction","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Infectious and respiratory outcomes","description":"Prediction, detection, severity stratification, treatment outcome, adherence, and cost outcomes for COVID-19 and other respiratory or infectious diseases.","inclusion":"Assign atoms whose primary objective is clinical prediction for acute COVID-19, tuberculosis, COPD, sleep apnea, bacterial/viral infection differentiation, liver infection, or related infectious/respiratory outcomes, regardless of modality or model family.","exclusion":"If the central objective is hospital-wide sepsis or ICU deterioration rather than a disease-specific infectious/respiratory program, assign to acute-care trajectory modeling; if the central objective is causal treatment-effect estimation, assign to trustworthy clinical prediction.","nearest_boundary":"Acute-care and longitudinal EHR trajectories","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]},{"direction_index":1,"label":"D1","name":"Cancer prognosis and response","description":"Prediction of oncologic survival, recurrence, metastasis, grade, invasion, progression, risk, and treatment response using imaging, pathology, genomics, clinical text, tabular data, or multimodal cancer data.","inclusion":"Assign atoms whose primary objective is prognostication, response assessment, progression prediction, grading with outcome relevance, or risk stratification for cancer patients or cancer screening populations.","exclusion":"General survival-method atoms without a cancer-centered objective belong to survival methodology; non-cancer imaging or specialty-disease prediction belongs to the relevant disease program.","nearest_boundary":"Survival and time-to-event methodology","representative_topics":["deep learning segmentation of necrosis ratio from whole slide images for osteosarcoma chemotherapy response and survival prediction","multi-task contrastive learning for joint meningioma grading and brain invasion prediction from mri","multimodal fusion of radiology and clinical text for lung cancer survival prediction"]},{"direction_index":2,"label":"D2","name":"Survival methodology","description":"General-purpose clinical time-to-event prediction methodology for censored outcomes, tied events, semi-competing risks, longitudinal risk updates, survival pretraining, and specialized survival losses or architectures.","inclusion":"Assign atoms whose primary scientific objective is methodological development for survival or time-to-event prediction intended to generalize across clinical domains, even when demonstrated on a disease dataset.","exclusion":"Disease-centered survival applications, such as cancer prognosis, CKD deterioration, dementia progression, or COVID mortality, belong to the relevant disease program unless the censored-survival method is the central contribution.","nearest_boundary":"Cancer prognosis and response","representative_topics":["efficient neural network cox proportional hazards estimation with tied-event support","neural expectation-maximization for semi-competing risk prediction in multi-state survival","self-supervised foundation model pretraining for time-to-event prediction in clinical records"]},{"direction_index":3,"label":"D3","name":"EHR trajectory and fusion modeling","description":"Prediction of hospital, ICU, readmission, deterioration, acuity, patient-journey, and general EHR outcomes through longitudinal sequence modeling, graph representation learning, clinical NLP, multimodal fusion, and heterogeneous patient representations.","inclusion":"Assign atoms whose primary objective is modeling patient trajectories or heterogeneous EHR evidence to predict clinical outcomes, including temporal EHR architectures, ICU early warning, graph/patient-similarity learning, clinical-note prediction, and multimodal structured/unstructured fusion.","exclusion":"If portability, fairness, interpretability, or causal safety is the primary objective, assign to trustworthy clinical prediction; if a disease-specific outcome program dominates, assign to the corresponding disease direction.","nearest_boundary":"Trustworthy clinical prediction","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","temporal attention networks for ehr health prediction with short-term and long-term dependency modeling","multimodal ehr trajectory modelling using dense and recurrent architectures for clinical outcome prediction"]},{"direction_index":4,"label":"D4","name":"Cardiovascular and cardiac-signal prediction","description":"Prediction of cardiovascular disease, heart failure, coronary disease, stroke, arrhythmia, ECG abnormalities, electrolyte or hemodynamic states, and related cardiovascular outcomes from clinical, retinal, ECG, wearable, or multimodal data.","inclusion":"Assign atoms whose primary clinical target is cardiovascular or stroke risk, cardiac prognosis, rhythm diagnosis, ECG-derived cardiac/electrolyte/hemodynamic prediction, or cardiovascular decision support.","exclusion":"Retinal, wearable, or physiological-signal studies aimed primarily at non-cardiovascular conditions belong to the relevant neurologic, specialty-disease, or infectious direction; general survival methods belong to survival methodology.","nearest_boundary":"Neurologic, mental-health, and behavioral states","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","retinal image-based deep learning for cardiovascular risk prediction and component attribution","anova-driven automatic attribute selection for cardiovascular disease prognosis prediction"]},{"direction_index":5,"label":"D5","name":"Neurobehavioral state prediction","description":"Prediction, diagnosis, staging, monitoring, relapse detection, severity estimation, and progression modeling for neurologic, neurodegenerative, neurophysiologic, mental-health, and behavioral states.","inclusion":"Assign atoms whose primary objective is Alzheimer/dementia, Parkinson or movement-disorder, EEG seizure or brain-electrophysiology, depression, schizophrenia, psychosis relapse, psychotherapy monitoring, or closely related neurobehavioral clinical outcome prediction; include broad specialized-signal atoms when their objective is detecting clinical functional or neurologic/behavioral states.","exclusion":"Orthopedic injury, pain, metabolic, ophthalmic, or respiratory disease prediction belongs to non-cancer specialty disease or infectious/respiratory directions unless the neurologic or mental-health state is the central target.","nearest_boundary":"Non-cancer specialty disease prognosis","representative_topics":["self-supervised motion forecasting pre-training for few-shot parkinson's gait impairment severity estimation","personalized parkinson's medication status prediction from remote wearable time-series using transformer attention","temporal pattern mining and survival analysis for alzheimer's disease risk factor identification from longitudinal clinical data"]},{"direction_index":6,"label":"D6","name":"Trustworthy clinical prediction","description":"Clinical prediction research whose primary objective is interpretability, explainability, reliability under imperfect data, uncertainty, fairness, causal validity, safe actionability, privacy-preserving collaboration, portability, or cross-site deployment.","inclusion":"Assign atoms centered on transparent risk models, checklist/prototype/concept or knowledge-guided explanations, robustness to missingness/imbalance/shift/leakage, uncertainty estimation, fairness, causal inference, target-trial emulation, federated learning, privacy, transfer, or multi-site generalization.","exclusion":"Routine feature-importance reporting, ordinary imputation, or standard external validation inside a disease-focused prediction study remains in the disease program unless trustworthiness or deployment is the main scientific contribution.","nearest_boundary":"EHR trajectory and fusion modeling","representative_topics":["interpretable clinical risk scores via explicit intervention policy modeling in icu","explainable hierarchical attention-recurrent models for intraoperative complication prediction","automated interpretable checklist learning from continuous icu clinical data via mixed-integer programming"]},{"direction_index":7,"label":"D7","name":"Non-cancer specialty prognosis","description":"Prediction, diagnosis, grading, progression, treatment outcome, and monitoring for non-cancer, non-cardiovascular, non-neurobehavioral specialty conditions including metabolic, renal, hepatic, ophthalmic, musculoskeletal, orthopedic, pain, allergy, and related chronic or specialty diseases.","inclusion":"Assign atoms whose primary objective is disease- or specialty-specific clinical prediction for diabetes, glucose dynamics, CKD, liver/metabolic disease, ophthalmic disease, musculoskeletal injury, pain, orthopedic outcomes, food allergy, autism as a specialty diagnosis, or comparable non-cancer clinical conditions not covered by the other disease programs.","exclusion":"Cancer, cardiovascular, infectious/respiratory, neurobehavioral, and general methodological papers belong to their respective directions even when they use similar standard ML pipelines or specialty modalities.","nearest_boundary":"Neurobehavioral state prediction","representative_topics":["blockchain-secured iot hypoglycemia detection with automated insulin injection using random forest","claims-based dynamic prediction of renal replacement therapy onset in chronic kidney disease","transformer-based personalized glucose forecasting with mixture-of-distributions uncertainty quantification"]}],"split":"dev"}
{"episode_id":"F260_T2025-01","field_index":260,"field":"clinical outcome prediction","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Infectious and respiratory outcomes","description":"Prediction, detection, severity stratification, treatment outcome, adherence, and cost outcomes for COVID-19 and other respiratory or infectious diseases.","inclusion":"Assign atoms whose primary objective is clinical prediction for acute COVID-19, tuberculosis, COPD, sleep apnea, bacterial/viral infection differentiation, liver infection, or related infectious/respiratory outcomes, regardless of modality or model family.","exclusion":"If the central objective is hospital-wide sepsis or ICU deterioration rather than a disease-specific infectious/respiratory program, assign to acute-care trajectory modeling; if the central objective is causal treatment-effect estimation, assign to trustworthy clinical prediction.","nearest_boundary":"Acute-care and longitudinal EHR trajectories","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]},{"direction_index":1,"label":"D1","name":"Cancer prognosis and response","description":"Prediction of oncologic survival, recurrence, metastasis, grade, invasion, progression, risk, and treatment response using imaging, pathology, genomics, clinical text, tabular data, or multimodal cancer data.","inclusion":"Assign atoms whose primary objective is prognostication, response assessment, progression prediction, grading with outcome relevance, or risk stratification for cancer patients or cancer screening populations.","exclusion":"General survival-method atoms without a cancer-centered objective belong to survival methodology; non-cancer imaging or specialty-disease prediction belongs to the relevant disease program.","nearest_boundary":"Survival and time-to-event methodology","representative_topics":["deep learning segmentation of necrosis ratio from whole slide images for osteosarcoma chemotherapy response and survival prediction","multi-task contrastive learning for joint meningioma grading and brain invasion prediction from mri","multimodal fusion of radiology and clinical text for lung cancer survival prediction"]},{"direction_index":2,"label":"D2","name":"Survival methodology","description":"General-purpose clinical time-to-event prediction methodology for censored outcomes, tied events, semi-competing risks, longitudinal risk updates, survival pretraining, and specialized survival losses or architectures.","inclusion":"Assign atoms whose primary scientific objective is methodological development for survival or time-to-event prediction intended to generalize across clinical domains, even when demonstrated on a disease dataset.","exclusion":"Disease-centered survival applications, such as cancer prognosis, CKD deterioration, dementia progression, or COVID mortality, belong to the relevant disease program unless the censored-survival method is the central contribution.","nearest_boundary":"Cancer prognosis and response","representative_topics":["efficient neural network cox proportional hazards estimation with tied-event support","neural expectation-maximization for semi-competing risk prediction in multi-state survival","self-supervised foundation model pretraining for time-to-event prediction in clinical records"]},{"direction_index":3,"label":"D3","name":"EHR trajectory and fusion modeling","description":"Prediction of hospital, ICU, readmission, deterioration, acuity, patient-journey, and general EHR outcomes through longitudinal sequence modeling, graph representation learning, clinical NLP, multimodal fusion, and heterogeneous patient representations.","inclusion":"Assign atoms whose primary objective is modeling patient trajectories or heterogeneous EHR evidence to predict clinical outcomes, including temporal EHR architectures, ICU early warning, graph/patient-similarity learning, clinical-note prediction, and multimodal structured/unstructured fusion.","exclusion":"If portability, fairness, interpretability, or causal safety is the primary objective, assign to trustworthy clinical prediction; if a disease-specific outcome program dominates, assign to the corresponding disease direction.","nearest_boundary":"Trustworthy clinical prediction","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","temporal attention networks for ehr health prediction with short-term and long-term dependency modeling","multimodal ehr trajectory modelling using dense and recurrent architectures for clinical outcome prediction"]},{"direction_index":4,"label":"D4","name":"Cardiovascular and cardiac-signal prediction","description":"Prediction of cardiovascular disease, heart failure, coronary disease, stroke, arrhythmia, ECG abnormalities, electrolyte or hemodynamic states, and related cardiovascular outcomes from clinical, retinal, ECG, wearable, or multimodal data.","inclusion":"Assign atoms whose primary clinical target is cardiovascular or stroke risk, cardiac prognosis, rhythm diagnosis, ECG-derived cardiac/electrolyte/hemodynamic prediction, or cardiovascular decision support.","exclusion":"Retinal, wearable, or physiological-signal studies aimed primarily at non-cardiovascular conditions belong to the relevant neurologic, specialty-disease, or infectious direction; general survival methods belong to survival methodology.","nearest_boundary":"Neurologic, mental-health, and behavioral states","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","retinal image-based deep learning for cardiovascular risk prediction and component attribution","anova-driven automatic attribute selection for cardiovascular disease prognosis prediction"]},{"direction_index":5,"label":"D5","name":"Neurobehavioral state prediction","description":"Prediction, diagnosis, staging, monitoring, relapse detection, severity estimation, and progression modeling for neurologic, neurodegenerative, neurophysiologic, mental-health, and behavioral states.","inclusion":"Assign atoms whose primary objective is Alzheimer/dementia, Parkinson or movement-disorder, EEG seizure or brain-electrophysiology, depression, schizophrenia, psychosis relapse, psychotherapy monitoring, or closely related neurobehavioral clinical outcome prediction; include broad specialized-signal atoms when their objective is detecting clinical functional or neurologic/behavioral states.","exclusion":"Orthopedic injury, pain, metabolic, ophthalmic, or respiratory disease prediction belongs to non-cancer specialty disease or infectious/respiratory directions unless the neurologic or mental-health state is the central target.","nearest_boundary":"Non-cancer specialty disease prognosis","representative_topics":["self-supervised motion forecasting pre-training for few-shot parkinson's gait impairment severity estimation","personalized parkinson's medication status prediction from remote wearable time-series using transformer attention","temporal pattern mining and survival analysis for alzheimer's disease risk factor identification from longitudinal clinical data"]},{"direction_index":6,"label":"D6","name":"Trustworthy clinical prediction","description":"Clinical prediction research whose primary objective is interpretability, explainability, reliability under imperfect data, uncertainty, fairness, causal validity, safe actionability, privacy-preserving collaboration, portability, or cross-site deployment.","inclusion":"Assign atoms centered on transparent risk models, checklist/prototype/concept or knowledge-guided explanations, robustness to missingness/imbalance/shift/leakage, uncertainty estimation, fairness, causal inference, target-trial emulation, federated learning, privacy, transfer, or multi-site generalization.","exclusion":"Routine feature-importance reporting, ordinary imputation, or standard external validation inside a disease-focused prediction study remains in the disease program unless trustworthiness or deployment is the main scientific contribution.","nearest_boundary":"EHR trajectory and fusion modeling","representative_topics":["interpretable clinical risk scores via explicit intervention policy modeling in icu","explainable hierarchical attention-recurrent models for intraoperative complication prediction","automated interpretable checklist learning from continuous icu clinical data via mixed-integer programming"]},{"direction_index":7,"label":"D7","name":"Non-cancer specialty prognosis","description":"Prediction, diagnosis, grading, progression, treatment outcome, and monitoring for non-cancer, non-cardiovascular, non-neurobehavioral specialty conditions including metabolic, renal, hepatic, ophthalmic, musculoskeletal, orthopedic, pain, allergy, and related chronic or specialty diseases.","inclusion":"Assign atoms whose primary objective is disease- or specialty-specific clinical prediction for diabetes, glucose dynamics, CKD, liver/metabolic disease, ophthalmic disease, musculoskeletal injury, pain, orthopedic outcomes, food allergy, autism as a specialty diagnosis, or comparable non-cancer clinical conditions not covered by the other disease programs.","exclusion":"Cancer, cardiovascular, infectious/respiratory, neurobehavioral, and general methodological papers belong to their respective directions even when they use similar standard ML pipelines or specialty modalities.","nearest_boundary":"Neurobehavioral state prediction","representative_topics":["blockchain-secured iot hypoglycemia detection with automated insulin injection using random forest","claims-based dynamic prediction of renal replacement therapy onset in chronic kidney disease","transformer-based personalized glucose forecasting with mixture-of-distributions uncertainty quantification"]}],"split":"dev"}
{"episode_id":"F260_T2025-07","field_index":260,"field":"clinical outcome prediction","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Infectious and respiratory outcomes","description":"Prediction, detection, severity stratification, treatment outcome, adherence, and cost outcomes for COVID-19 and other respiratory or infectious diseases.","inclusion":"Assign atoms whose primary objective is clinical prediction for acute COVID-19, tuberculosis, COPD, sleep apnea, bacterial/viral infection differentiation, liver infection, or related infectious/respiratory outcomes, regardless of modality or model family.","exclusion":"If the central objective is hospital-wide sepsis or ICU deterioration rather than a disease-specific infectious/respiratory program, assign to acute-care trajectory modeling; if the central objective is causal treatment-effect estimation, assign to trustworthy clinical prediction.","nearest_boundary":"Acute-care and longitudinal EHR trajectories","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]},{"direction_index":1,"label":"D1","name":"Cancer prognosis and response","description":"Prediction of oncologic survival, recurrence, metastasis, grade, invasion, progression, risk, and treatment response using imaging, pathology, genomics, clinical text, tabular data, or multimodal cancer data.","inclusion":"Assign atoms whose primary objective is prognostication, response assessment, progression prediction, grading with outcome relevance, or risk stratification for cancer patients or cancer screening populations.","exclusion":"General survival-method atoms without a cancer-centered objective belong to survival methodology; non-cancer imaging or specialty-disease prediction belongs to the relevant disease program.","nearest_boundary":"Survival and time-to-event methodology","representative_topics":["deep learning segmentation of necrosis ratio from whole slide images for osteosarcoma chemotherapy response and survival prediction","multi-task contrastive learning for joint meningioma grading and brain invasion prediction from mri","multimodal fusion of radiology and clinical text for lung cancer survival prediction"]},{"direction_index":2,"label":"D2","name":"Survival methodology","description":"General-purpose clinical time-to-event prediction methodology for censored outcomes, tied events, semi-competing risks, longitudinal risk updates, survival pretraining, and specialized survival losses or architectures.","inclusion":"Assign atoms whose primary scientific objective is methodological development for survival or time-to-event prediction intended to generalize across clinical domains, even when demonstrated on a disease dataset.","exclusion":"Disease-centered survival applications, such as cancer prognosis, CKD deterioration, dementia progression, or COVID mortality, belong to the relevant disease program unless the censored-survival method is the central contribution.","nearest_boundary":"Cancer prognosis and response","representative_topics":["efficient neural network cox proportional hazards estimation with tied-event support","neural expectation-maximization for semi-competing risk prediction in multi-state survival","self-supervised foundation model pretraining for time-to-event prediction in clinical records"]},{"direction_index":3,"label":"D3","name":"EHR trajectory and fusion modeling","description":"Prediction of hospital, ICU, readmission, deterioration, acuity, patient-journey, and general EHR outcomes through longitudinal sequence modeling, graph representation learning, clinical NLP, multimodal fusion, and heterogeneous patient representations.","inclusion":"Assign atoms whose primary objective is modeling patient trajectories or heterogeneous EHR evidence to predict clinical outcomes, including temporal EHR architectures, ICU early warning, graph/patient-similarity learning, clinical-note prediction, and multimodal structured/unstructured fusion.","exclusion":"If portability, fairness, interpretability, or causal safety is the primary objective, assign to trustworthy clinical prediction; if a disease-specific outcome program dominates, assign to the corresponding disease direction.","nearest_boundary":"Trustworthy clinical prediction","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","temporal attention networks for ehr health prediction with short-term and long-term dependency modeling","multimodal ehr trajectory modelling using dense and recurrent architectures for clinical outcome prediction"]},{"direction_index":4,"label":"D4","name":"Cardiovascular and cardiac-signal prediction","description":"Prediction of cardiovascular disease, heart failure, coronary disease, stroke, arrhythmia, ECG abnormalities, electrolyte or hemodynamic states, and related cardiovascular outcomes from clinical, retinal, ECG, wearable, or multimodal data.","inclusion":"Assign atoms whose primary clinical target is cardiovascular or stroke risk, cardiac prognosis, rhythm diagnosis, ECG-derived cardiac/electrolyte/hemodynamic prediction, or cardiovascular decision support.","exclusion":"Retinal, wearable, or physiological-signal studies aimed primarily at non-cardiovascular conditions belong to the relevant neurologic, specialty-disease, or infectious direction; general survival methods belong to survival methodology.","nearest_boundary":"Neurologic, mental-health, and behavioral states","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","retinal image-based deep learning for cardiovascular risk prediction and component attribution","anova-driven automatic attribute selection for cardiovascular disease prognosis prediction"]},{"direction_index":5,"label":"D5","name":"Neurobehavioral state prediction","description":"Prediction, diagnosis, staging, monitoring, relapse detection, severity estimation, and progression modeling for neurologic, neurodegenerative, neurophysiologic, mental-health, and behavioral states.","inclusion":"Assign atoms whose primary objective is Alzheimer/dementia, Parkinson or movement-disorder, EEG seizure or brain-electrophysiology, depression, schizophrenia, psychosis relapse, psychotherapy monitoring, or closely related neurobehavioral clinical outcome prediction; include broad specialized-signal atoms when their objective is detecting clinical functional or neurologic/behavioral states.","exclusion":"Orthopedic injury, pain, metabolic, ophthalmic, or respiratory disease prediction belongs to non-cancer specialty disease or infectious/respiratory directions unless the neurologic or mental-health state is the central target.","nearest_boundary":"Non-cancer specialty disease prognosis","representative_topics":["self-supervised motion forecasting pre-training for few-shot parkinson's gait impairment severity estimation","personalized parkinson's medication status prediction from remote wearable time-series using transformer attention","temporal pattern mining and survival analysis for alzheimer's disease risk factor identification from longitudinal clinical data"]},{"direction_index":6,"label":"D6","name":"Trustworthy clinical prediction","description":"Clinical prediction research whose primary objective is interpretability, explainability, reliability under imperfect data, uncertainty, fairness, causal validity, safe actionability, privacy-preserving collaboration, portability, or cross-site deployment.","inclusion":"Assign atoms centered on transparent risk models, checklist/prototype/concept or knowledge-guided explanations, robustness to missingness/imbalance/shift/leakage, uncertainty estimation, fairness, causal inference, target-trial emulation, federated learning, privacy, transfer, or multi-site generalization.","exclusion":"Routine feature-importance reporting, ordinary imputation, or standard external validation inside a disease-focused prediction study remains in the disease program unless trustworthiness or deployment is the main scientific contribution.","nearest_boundary":"EHR trajectory and fusion modeling","representative_topics":["interpretable clinical risk scores via explicit intervention policy modeling in icu","explainable hierarchical attention-recurrent models for intraoperative complication prediction","automated interpretable checklist learning from continuous icu clinical data via mixed-integer programming"]},{"direction_index":7,"label":"D7","name":"Non-cancer specialty prognosis","description":"Prediction, diagnosis, grading, progression, treatment outcome, and monitoring for non-cancer, non-cardiovascular, non-neurobehavioral specialty conditions including metabolic, renal, hepatic, ophthalmic, musculoskeletal, orthopedic, pain, allergy, and related chronic or specialty diseases.","inclusion":"Assign atoms whose primary objective is disease- or specialty-specific clinical prediction for diabetes, glucose dynamics, CKD, liver/metabolic disease, ophthalmic disease, musculoskeletal injury, pain, orthopedic outcomes, food allergy, autism as a specialty diagnosis, or comparable non-cancer clinical conditions not covered by the other disease programs.","exclusion":"Cancer, cardiovascular, infectious/respiratory, neurobehavioral, and general methodological papers belong to their respective directions even when they use similar standard ML pipelines or specialty modalities.","nearest_boundary":"Neurobehavioral state prediction","representative_topics":["blockchain-secured iot hypoglycemia detection with automated insulin injection using random forest","claims-based dynamic prediction of renal replacement therapy onset in chronic kidney disease","transformer-based personalized glucose forecasting with mixture-of-distributions uncertainty quantification"]}],"split":"dev"}
{"episode_id":"F260_T2026-01","field_index":260,"field":"clinical outcome prediction","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Infectious and respiratory outcomes","description":"Prediction, detection, severity stratification, treatment outcome, adherence, and cost outcomes for COVID-19 and other respiratory or infectious diseases.","inclusion":"Assign atoms whose primary objective is clinical prediction for acute COVID-19, tuberculosis, COPD, sleep apnea, bacterial/viral infection differentiation, liver infection, or related infectious/respiratory outcomes, regardless of modality or model family.","exclusion":"If the central objective is hospital-wide sepsis or ICU deterioration rather than a disease-specific infectious/respiratory program, assign to acute-care trajectory modeling; if the central objective is causal treatment-effect estimation, assign to trustworthy clinical prediction.","nearest_boundary":"Acute-care and longitudinal EHR trajectories","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]},{"direction_index":1,"label":"D1","name":"Cancer prognosis and response","description":"Prediction of oncologic survival, recurrence, metastasis, grade, invasion, progression, risk, and treatment response using imaging, pathology, genomics, clinical text, tabular data, or multimodal cancer data.","inclusion":"Assign atoms whose primary objective is prognostication, response assessment, progression prediction, grading with outcome relevance, or risk stratification for cancer patients or cancer screening populations.","exclusion":"General survival-method atoms without a cancer-centered objective belong to survival methodology; non-cancer imaging or specialty-disease prediction belongs to the relevant disease program.","nearest_boundary":"Survival and time-to-event methodology","representative_topics":["deep learning segmentation of necrosis ratio from whole slide images for osteosarcoma chemotherapy response and survival prediction","multi-task contrastive learning for joint meningioma grading and brain invasion prediction from mri","multimodal fusion of radiology and clinical text for lung cancer survival prediction"]},{"direction_index":2,"label":"D2","name":"Survival methodology","description":"General-purpose clinical time-to-event prediction methodology for censored outcomes, tied events, semi-competing risks, longitudinal risk updates, survival pretraining, and specialized survival losses or architectures.","inclusion":"Assign atoms whose primary scientific objective is methodological development for survival or time-to-event prediction intended to generalize across clinical domains, even when demonstrated on a disease dataset.","exclusion":"Disease-centered survival applications, such as cancer prognosis, CKD deterioration, dementia progression, or COVID mortality, belong to the relevant disease program unless the censored-survival method is the central contribution.","nearest_boundary":"Cancer prognosis and response","representative_topics":["efficient neural network cox proportional hazards estimation with tied-event support","neural expectation-maximization for semi-competing risk prediction in multi-state survival","self-supervised foundation model pretraining for time-to-event prediction in clinical records"]},{"direction_index":3,"label":"D3","name":"EHR trajectory and fusion modeling","description":"Prediction of hospital, ICU, readmission, deterioration, acuity, patient-journey, and general EHR outcomes through longitudinal sequence modeling, graph representation learning, clinical NLP, multimodal fusion, and heterogeneous patient representations.","inclusion":"Assign atoms whose primary objective is modeling patient trajectories or heterogeneous EHR evidence to predict clinical outcomes, including temporal EHR architectures, ICU early warning, graph/patient-similarity learning, clinical-note prediction, and multimodal structured/unstructured fusion.","exclusion":"If portability, fairness, interpretability, or causal safety is the primary objective, assign to trustworthy clinical prediction; if a disease-specific outcome program dominates, assign to the corresponding disease direction.","nearest_boundary":"Trustworthy clinical prediction","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","temporal attention networks for ehr health prediction with short-term and long-term dependency modeling","multimodal ehr trajectory modelling using dense and recurrent architectures for clinical outcome prediction"]},{"direction_index":4,"label":"D4","name":"Cardiovascular and cardiac-signal prediction","description":"Prediction of cardiovascular disease, heart failure, coronary disease, stroke, arrhythmia, ECG abnormalities, electrolyte or hemodynamic states, and related cardiovascular outcomes from clinical, retinal, ECG, wearable, or multimodal data.","inclusion":"Assign atoms whose primary clinical target is cardiovascular or stroke risk, cardiac prognosis, rhythm diagnosis, ECG-derived cardiac/electrolyte/hemodynamic prediction, or cardiovascular decision support.","exclusion":"Retinal, wearable, or physiological-signal studies aimed primarily at non-cardiovascular conditions belong to the relevant neurologic, specialty-disease, or infectious direction; general survival methods belong to survival methodology.","nearest_boundary":"Neurologic, mental-health, and behavioral states","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","retinal image-based deep learning for cardiovascular risk prediction and component attribution","anova-driven automatic attribute selection for cardiovascular disease prognosis prediction"]},{"direction_index":5,"label":"D5","name":"Neurobehavioral state prediction","description":"Prediction, diagnosis, staging, monitoring, relapse detection, severity estimation, and progression modeling for neurologic, neurodegenerative, neurophysiologic, mental-health, and behavioral states.","inclusion":"Assign atoms whose primary objective is Alzheimer/dementia, Parkinson or movement-disorder, EEG seizure or brain-electrophysiology, depression, schizophrenia, psychosis relapse, psychotherapy monitoring, or closely related neurobehavioral clinical outcome prediction; include broad specialized-signal atoms when their objective is detecting clinical functional or neurologic/behavioral states.","exclusion":"Orthopedic injury, pain, metabolic, ophthalmic, or respiratory disease prediction belongs to non-cancer specialty disease or infectious/respiratory directions unless the neurologic or mental-health state is the central target.","nearest_boundary":"Non-cancer specialty disease prognosis","representative_topics":["self-supervised motion forecasting pre-training for few-shot parkinson's gait impairment severity estimation","personalized parkinson's medication status prediction from remote wearable time-series using transformer attention","temporal pattern mining and survival analysis for alzheimer's disease risk factor identification from longitudinal clinical data"]},{"direction_index":6,"label":"D6","name":"Trustworthy clinical prediction","description":"Clinical prediction research whose primary objective is interpretability, explainability, reliability under imperfect data, uncertainty, fairness, causal validity, safe actionability, privacy-preserving collaboration, portability, or cross-site deployment.","inclusion":"Assign atoms centered on transparent risk models, checklist/prototype/concept or knowledge-guided explanations, robustness to missingness/imbalance/shift/leakage, uncertainty estimation, fairness, causal inference, target-trial emulation, federated learning, privacy, transfer, or multi-site generalization.","exclusion":"Routine feature-importance reporting, ordinary imputation, or standard external validation inside a disease-focused prediction study remains in the disease program unless trustworthiness or deployment is the main scientific contribution.","nearest_boundary":"EHR trajectory and fusion modeling","representative_topics":["interpretable clinical risk scores via explicit intervention policy modeling in icu","explainable hierarchical attention-recurrent models for intraoperative complication prediction","automated interpretable checklist learning from continuous icu clinical data via mixed-integer programming"]},{"direction_index":7,"label":"D7","name":"Non-cancer specialty prognosis","description":"Prediction, diagnosis, grading, progression, treatment outcome, and monitoring for non-cancer, non-cardiovascular, non-neurobehavioral specialty conditions including metabolic, renal, hepatic, ophthalmic, musculoskeletal, orthopedic, pain, allergy, and related chronic or specialty diseases.","inclusion":"Assign atoms whose primary objective is disease- or specialty-specific clinical prediction for diabetes, glucose dynamics, CKD, liver/metabolic disease, ophthalmic disease, musculoskeletal injury, pain, orthopedic outcomes, food allergy, autism as a specialty diagnosis, or comparable non-cancer clinical conditions not covered by the other disease programs.","exclusion":"Cancer, cardiovascular, infectious/respiratory, neurobehavioral, and general methodological papers belong to their respective directions even when they use similar standard ML pipelines or specialty modalities.","nearest_boundary":"Neurobehavioral state prediction","representative_topics":["blockchain-secured iot hypoglycemia detection with automated insulin injection using random forest","claims-based dynamic prediction of renal replacement therapy onset in chronic kidney disease","transformer-based personalized glucose forecasting with mixture-of-distributions uncertainty quantification"]}],"split":"dev"}
{"episode_id":"F262_T2024-01","field_index":262,"field":"industrial anomaly detection","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Normality representation and distillation","description":"Learn discriminative representations of normal industrial data and detect anomalies through feature-space deviation, memory/prototype mismatch, contrastive separation, or teacher-student feature discrepancy.","inclusion":"Include work whose main objective is constructing or adapting normal feature manifolds, pretrained/self-supervised embeddings, memory banks, prototypes, contrastive objectives, or knowledge-distillation discrepancies as the primary anomaly signal.","exclusion":"If the decisive signal is an input-versus-restoration residual, use reconstruction and restoration; if explicit likelihood under an invertible or density model is central, use density modeling.","nearest_boundary":"Reconstruction and restoration","representative_topics":["target-adaptive feature adaptation for anomaly localization via coupled hypersphere modeling","unsupervised textural defect detection via memory rearrangement and background reconstruction for edge manufacturing inspection","unsupervised industrial anomaly detection via generative and contrastive vision transformer networks"]},{"direction_index":1,"label":"D1","name":"Reconstruction and restoration","description":"Model normal observations by reconstructing, denoising, inpainting, restoring, or regenerating normal patterns and identify anomalies through residuals or restoration gaps.","inclusion":"Include autoencoder, VAE, GAN, masked reconstruction, diffusion denoising, inpainting, or restoration methods where the primary anomaly evidence is discrepancy between the observed input and a reconstructed or restored normal version.","exclusion":"If anomaly scoring is primarily nearest-neighbor, prototype, memory, contrastive, or teacher-student feature disagreement without residual restoration, use normality representation and distillation; if explicit likelihood is primary, use density modeling.","nearest_boundary":"Normality representation and distillation","representative_topics":["deep autoencoder anomaly detection in textured images using complex wavelet structural similarity loss","self-supervised masked convolutional transformer blocks for reconstruction-based anomaly detection","autoencoder-based unsupervised anomaly detection with explained fault localization in industrial cooling systems"]},{"direction_index":2,"label":"D2","name":"Density and flow modeling","description":"Estimate normal-data probability density or likelihood for industrial anomaly detection, especially with normalizing flows, invertible networks, Gaussian mixtures, or related probabilistic normality models.","inclusion":"Include work whose main scientific objective is explicit density estimation and whose anomaly score is low likelihood, abnormal latent density, or flow-based distributional contrast.","exclusion":"If a generative model is used mainly to produce a reconstruction residual, use reconstruction and restoration; if density estimation is incidental to embedding comparison, use normality representation and distillation.","nearest_boundary":"Reconstruction and restoration","representative_topics":["normalizing flow anomaly detection with spatially-aware convolutional attention for localization","self-supervised texture anomaly detection via normalizing flow and dictionary learning fusion","connecting generative and discriminative paradigms via discriminative-oriented gaussian mixture models for anomaly detection"]},{"direction_index":3,"label":"D3","name":"Data-efficient and robust adaptation","description":"Improve industrial anomaly or defect detection when labels, anomaly examples, clean normal data, domains, categories, or operating conditions are limited, imbalanced, contaminated, shifting, or incrementally changing.","inclusion":"Include few-shot, low-label, active, incremental, synthetic-data, augmentation, class-imbalance, contaminated-training, domain-adaptation, domain-generalization, and cross-category robustness work when this data regime is the primary objective.","exclusion":"If scarce-label performance is only a secondary consequence of a representation, reconstruction, or density model, use that mechanism-specific direction; if the main contribution is non-visual sensor modeling, use sensor and process fault diagnosis.","nearest_boundary":"Normality representation and distillation","representative_topics":["supervised regression-based anomaly detection using continuous labels and augmentation for low-resource settings","boundary-guided semi-push-pull contrastive learning for supervised few-shot anomaly detection","knowledge distillation for catastrophic forgetting in incremental few-shot surface defect detection"]},{"direction_index":4,"label":"D4","name":"Sensor and process fault diagnosis","description":"Detect and diagnose industrial anomalies from non-visual temporal or signal data, including sensors, vibration, acoustics, pressure, IIoT streams, machine operating traces, and process measurements.","inclusion":"Include work whose main objective is non-image industrial fault detection, acoustic or vibration anomaly detection, temporal normality modeling, inter-sensor transfer, graph or clustering over sensor relationships, forecasting-based detection, or process-monitoring diagnosis.","exclusion":"If the work is primarily a generic reconstruction, representation, or density method demonstrated on signals, use the corresponding mechanism direction unless the sensor/process modeling objective dominates; image or x-ray inspection belongs to visual inspection.","nearest_boundary":"Data-efficient and robust adaptation","representative_topics":["hierarchical conditional variational autoencoder for cross-domain acoustic anomaly detection in industrial monitoring","inter-sensor transfer learning for anomaly detection and defect classification in smart manufacturing","few-shot anomaly detection via autoregressive normality learning with positive representation separation"]},{"direction_index":5,"label":"D5","name":"Visual defect inspection","description":"Develop visual inspection systems that detect, classify, localize, or segment visible industrial defects in 2D, x-ray, wafer, electronics, surface, casting, fiber, composite, solar-cell, and manufactured-part imagery, including deployable supervised pipelines.","inclusion":"Include supervised, weakly supervised, detection, segmentation, classification, CNN, transformer, explainable, reference-image, lightweight, edge, or high-throughput visual inspection work when the primary objective is visual defect recognition or deployable visual inspection performance.","exclusion":"If the main objective is general one-class normality modeling, use representation, reconstruction, or density directions; if 3D geometry, RGB-D fusion, components, or compositional logic are central, use structural multimodal reasoning.","nearest_boundary":"Structural multimodal reasoning","representative_topics":["automated defect detection in silicon detector modules using deep learning object detection","lightweight custom architectures for efficient metal casting defect detection via separable convolutions","automated defect recognition in industrial x-ray castings using convolutional neural networks"]},{"direction_index":6,"label":"D6","name":"Structural multimodal reasoning","description":"Detect industrial anomalies by modeling 3D geometry, RGB-D or point-cloud fusion, cross-modal mappings, object parts, component structure, adjustable configurations, or logical and compositional relations beyond ordinary 2D appearance cues.","inclusion":"Include work whose core objective is multimodal 2D-3D fusion, depth or point-cloud anomaly localization, cross-modal feature mapping, part segmentation, component-aware inspection, or logical anomaly reasoning.","exclusion":"If the method relies mainly on single-modality 2D appearance, x-ray imagery, or ordinary segmentation cues, use visual defect inspection; if geometry is merely an input feature inside a general embedding detector, use normality representation and distillation.","nearest_boundary":"Visual defect inspection","representative_topics":["multimodal industrial anomaly detection via hybrid point cloud and rgb feature fusion","component-aware logical anomaly detection for adjustable industrial visual inspection","lightweight multi-modality 3d anomaly segmentation for real-time industrial deployment"]},{"direction_index":7,"label":"D7","name":"Benchmarks and evaluation resources","description":"Create, organize, survey, benchmark, or systematically evaluate industrial anomaly detection datasets, baselines, challenges, metrics, protocols, methods, or application testbeds rather than proposing a new detector mechanism.","inclusion":"Include dataset papers, benchmark suites, challenge or evaluation protocols, surveys, systematic reviews, comparative baseline studies, and resource-oriented application evaluations where mapping or measuring the field is central.","exclusion":"If a new detector, adaptation strategy, representation, reconstruction model, deployment architecture, or multimodal reasoning method is the main contribution, use the corresponding algorithmic direction.","nearest_boundary":"Visual defect inspection","representative_topics":["survey of defect detection methods for die attachment and wire bonding in ic manufacturing","unsupervised anomaly detection dataset for outdoor maintenance inspection with surface and logical anomalies","unified benchmark for industrial image anomaly detection across supervision and learning paradigms"]}],"split":"dev"}
{"episode_id":"F262_T2024-07","field_index":262,"field":"industrial anomaly detection","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Normality representation and distillation","description":"Learn discriminative representations of normal industrial data and detect anomalies through feature-space deviation, memory/prototype mismatch, contrastive separation, or teacher-student feature discrepancy.","inclusion":"Include work whose main objective is constructing or adapting normal feature manifolds, pretrained/self-supervised embeddings, memory banks, prototypes, contrastive objectives, or knowledge-distillation discrepancies as the primary anomaly signal.","exclusion":"If the decisive signal is an input-versus-restoration residual, use reconstruction and restoration; if explicit likelihood under an invertible or density model is central, use density modeling.","nearest_boundary":"Reconstruction and restoration","representative_topics":["target-adaptive feature adaptation for anomaly localization via coupled hypersphere modeling","unsupervised textural defect detection via memory rearrangement and background reconstruction for edge manufacturing inspection","unsupervised industrial anomaly detection via generative and contrastive vision transformer networks"]},{"direction_index":1,"label":"D1","name":"Reconstruction and restoration","description":"Model normal observations by reconstructing, denoising, inpainting, restoring, or regenerating normal patterns and identify anomalies through residuals or restoration gaps.","inclusion":"Include autoencoder, VAE, GAN, masked reconstruction, diffusion denoising, inpainting, or restoration methods where the primary anomaly evidence is discrepancy between the observed input and a reconstructed or restored normal version.","exclusion":"If anomaly scoring is primarily nearest-neighbor, prototype, memory, contrastive, or teacher-student feature disagreement without residual restoration, use normality representation and distillation; if explicit likelihood is primary, use density modeling.","nearest_boundary":"Normality representation and distillation","representative_topics":["deep autoencoder anomaly detection in textured images using complex wavelet structural similarity loss","self-supervised masked convolutional transformer blocks for reconstruction-based anomaly detection","autoencoder-based unsupervised anomaly detection with explained fault localization in industrial cooling systems"]},{"direction_index":2,"label":"D2","name":"Density and flow modeling","description":"Estimate normal-data probability density or likelihood for industrial anomaly detection, especially with normalizing flows, invertible networks, Gaussian mixtures, or related probabilistic normality models.","inclusion":"Include work whose main scientific objective is explicit density estimation and whose anomaly score is low likelihood, abnormal latent density, or flow-based distributional contrast.","exclusion":"If a generative model is used mainly to produce a reconstruction residual, use reconstruction and restoration; if density estimation is incidental to embedding comparison, use normality representation and distillation.","nearest_boundary":"Reconstruction and restoration","representative_topics":["normalizing flow anomaly detection with spatially-aware convolutional attention for localization","self-supervised texture anomaly detection via normalizing flow and dictionary learning fusion","connecting generative and discriminative paradigms via discriminative-oriented gaussian mixture models for anomaly detection"]},{"direction_index":3,"label":"D3","name":"Data-efficient and robust adaptation","description":"Improve industrial anomaly or defect detection when labels, anomaly examples, clean normal data, domains, categories, or operating conditions are limited, imbalanced, contaminated, shifting, or incrementally changing.","inclusion":"Include few-shot, low-label, active, incremental, synthetic-data, augmentation, class-imbalance, contaminated-training, domain-adaptation, domain-generalization, and cross-category robustness work when this data regime is the primary objective.","exclusion":"If scarce-label performance is only a secondary consequence of a representation, reconstruction, or density model, use that mechanism-specific direction; if the main contribution is non-visual sensor modeling, use sensor and process fault diagnosis.","nearest_boundary":"Normality representation and distillation","representative_topics":["supervised regression-based anomaly detection using continuous labels and augmentation for low-resource settings","boundary-guided semi-push-pull contrastive learning for supervised few-shot anomaly detection","knowledge distillation for catastrophic forgetting in incremental few-shot surface defect detection"]},{"direction_index":4,"label":"D4","name":"Sensor and process fault diagnosis","description":"Detect and diagnose industrial anomalies from non-visual temporal or signal data, including sensors, vibration, acoustics, pressure, IIoT streams, machine operating traces, and process measurements.","inclusion":"Include work whose main objective is non-image industrial fault detection, acoustic or vibration anomaly detection, temporal normality modeling, inter-sensor transfer, graph or clustering over sensor relationships, forecasting-based detection, or process-monitoring diagnosis.","exclusion":"If the work is primarily a generic reconstruction, representation, or density method demonstrated on signals, use the corresponding mechanism direction unless the sensor/process modeling objective dominates; image or x-ray inspection belongs to visual inspection.","nearest_boundary":"Data-efficient and robust adaptation","representative_topics":["hierarchical conditional variational autoencoder for cross-domain acoustic anomaly detection in industrial monitoring","inter-sensor transfer learning for anomaly detection and defect classification in smart manufacturing","few-shot anomaly detection via autoregressive normality learning with positive representation separation"]},{"direction_index":5,"label":"D5","name":"Visual defect inspection","description":"Develop visual inspection systems that detect, classify, localize, or segment visible industrial defects in 2D, x-ray, wafer, electronics, surface, casting, fiber, composite, solar-cell, and manufactured-part imagery, including deployable supervised pipelines.","inclusion":"Include supervised, weakly supervised, detection, segmentation, classification, CNN, transformer, explainable, reference-image, lightweight, edge, or high-throughput visual inspection work when the primary objective is visual defect recognition or deployable visual inspection performance.","exclusion":"If the main objective is general one-class normality modeling, use representation, reconstruction, or density directions; if 3D geometry, RGB-D fusion, components, or compositional logic are central, use structural multimodal reasoning.","nearest_boundary":"Structural multimodal reasoning","representative_topics":["automated defect detection in silicon detector modules using deep learning object detection","lightweight custom architectures for efficient metal casting defect detection via separable convolutions","automated defect recognition in industrial x-ray castings using convolutional neural networks"]},{"direction_index":6,"label":"D6","name":"Structural multimodal reasoning","description":"Detect industrial anomalies by modeling 3D geometry, RGB-D or point-cloud fusion, cross-modal mappings, object parts, component structure, adjustable configurations, or logical and compositional relations beyond ordinary 2D appearance cues.","inclusion":"Include work whose core objective is multimodal 2D-3D fusion, depth or point-cloud anomaly localization, cross-modal feature mapping, part segmentation, component-aware inspection, or logical anomaly reasoning.","exclusion":"If the method relies mainly on single-modality 2D appearance, x-ray imagery, or ordinary segmentation cues, use visual defect inspection; if geometry is merely an input feature inside a general embedding detector, use normality representation and distillation.","nearest_boundary":"Visual defect inspection","representative_topics":["multimodal industrial anomaly detection via hybrid point cloud and rgb feature fusion","component-aware logical anomaly detection for adjustable industrial visual inspection","lightweight multi-modality 3d anomaly segmentation for real-time industrial deployment"]},{"direction_index":7,"label":"D7","name":"Benchmarks and evaluation resources","description":"Create, organize, survey, benchmark, or systematically evaluate industrial anomaly detection datasets, baselines, challenges, metrics, protocols, methods, or application testbeds rather than proposing a new detector mechanism.","inclusion":"Include dataset papers, benchmark suites, challenge or evaluation protocols, surveys, systematic reviews, comparative baseline studies, and resource-oriented application evaluations where mapping or measuring the field is central.","exclusion":"If a new detector, adaptation strategy, representation, reconstruction model, deployment architecture, or multimodal reasoning method is the main contribution, use the corresponding algorithmic direction.","nearest_boundary":"Visual defect inspection","representative_topics":["survey of defect detection methods for die attachment and wire bonding in ic manufacturing","unsupervised anomaly detection dataset for outdoor maintenance inspection with surface and logical anomalies","unified benchmark for industrial image anomaly detection across supervision and learning paradigms"]}],"split":"dev"}
{"episode_id":"F262_T2025-01","field_index":262,"field":"industrial anomaly detection","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Normality representation and distillation","description":"Learn discriminative representations of normal industrial data and detect anomalies through feature-space deviation, memory/prototype mismatch, contrastive separation, or teacher-student feature discrepancy.","inclusion":"Include work whose main objective is constructing or adapting normal feature manifolds, pretrained/self-supervised embeddings, memory banks, prototypes, contrastive objectives, or knowledge-distillation discrepancies as the primary anomaly signal.","exclusion":"If the decisive signal is an input-versus-restoration residual, use reconstruction and restoration; if explicit likelihood under an invertible or density model is central, use density modeling.","nearest_boundary":"Reconstruction and restoration","representative_topics":["target-adaptive feature adaptation for anomaly localization via coupled hypersphere modeling","unsupervised textural defect detection via memory rearrangement and background reconstruction for edge manufacturing inspection","unsupervised industrial anomaly detection via generative and contrastive vision transformer networks"]},{"direction_index":1,"label":"D1","name":"Reconstruction and restoration","description":"Model normal observations by reconstructing, denoising, inpainting, restoring, or regenerating normal patterns and identify anomalies through residuals or restoration gaps.","inclusion":"Include autoencoder, VAE, GAN, masked reconstruction, diffusion denoising, inpainting, or restoration methods where the primary anomaly evidence is discrepancy between the observed input and a reconstructed or restored normal version.","exclusion":"If anomaly scoring is primarily nearest-neighbor, prototype, memory, contrastive, or teacher-student feature disagreement without residual restoration, use normality representation and distillation; if explicit likelihood is primary, use density modeling.","nearest_boundary":"Normality representation and distillation","representative_topics":["deep autoencoder anomaly detection in textured images using complex wavelet structural similarity loss","self-supervised masked convolutional transformer blocks for reconstruction-based anomaly detection","autoencoder-based unsupervised anomaly detection with explained fault localization in industrial cooling systems"]},{"direction_index":2,"label":"D2","name":"Density and flow modeling","description":"Estimate normal-data probability density or likelihood for industrial anomaly detection, especially with normalizing flows, invertible networks, Gaussian mixtures, or related probabilistic normality models.","inclusion":"Include work whose main scientific objective is explicit density estimation and whose anomaly score is low likelihood, abnormal latent density, or flow-based distributional contrast.","exclusion":"If a generative model is used mainly to produce a reconstruction residual, use reconstruction and restoration; if density estimation is incidental to embedding comparison, use normality representation and distillation.","nearest_boundary":"Reconstruction and restoration","representative_topics":["normalizing flow anomaly detection with spatially-aware convolutional attention for localization","self-supervised texture anomaly detection via normalizing flow and dictionary learning fusion","connecting generative and discriminative paradigms via discriminative-oriented gaussian mixture models for anomaly detection"]},{"direction_index":3,"label":"D3","name":"Data-efficient and robust adaptation","description":"Improve industrial anomaly or defect detection when labels, anomaly examples, clean normal data, domains, categories, or operating conditions are limited, imbalanced, contaminated, shifting, or incrementally changing.","inclusion":"Include few-shot, low-label, active, incremental, synthetic-data, augmentation, class-imbalance, contaminated-training, domain-adaptation, domain-generalization, and cross-category robustness work when this data regime is the primary objective.","exclusion":"If scarce-label performance is only a secondary consequence of a representation, reconstruction, or density model, use that mechanism-specific direction; if the main contribution is non-visual sensor modeling, use sensor and process fault diagnosis.","nearest_boundary":"Normality representation and distillation","representative_topics":["supervised regression-based anomaly detection using continuous labels and augmentation for low-resource settings","boundary-guided semi-push-pull contrastive learning for supervised few-shot anomaly detection","knowledge distillation for catastrophic forgetting in incremental few-shot surface defect detection"]},{"direction_index":4,"label":"D4","name":"Sensor and process fault diagnosis","description":"Detect and diagnose industrial anomalies from non-visual temporal or signal data, including sensors, vibration, acoustics, pressure, IIoT streams, machine operating traces, and process measurements.","inclusion":"Include work whose main objective is non-image industrial fault detection, acoustic or vibration anomaly detection, temporal normality modeling, inter-sensor transfer, graph or clustering over sensor relationships, forecasting-based detection, or process-monitoring diagnosis.","exclusion":"If the work is primarily a generic reconstruction, representation, or density method demonstrated on signals, use the corresponding mechanism direction unless the sensor/process modeling objective dominates; image or x-ray inspection belongs to visual inspection.","nearest_boundary":"Data-efficient and robust adaptation","representative_topics":["hierarchical conditional variational autoencoder for cross-domain acoustic anomaly detection in industrial monitoring","inter-sensor transfer learning for anomaly detection and defect classification in smart manufacturing","few-shot anomaly detection via autoregressive normality learning with positive representation separation"]},{"direction_index":5,"label":"D5","name":"Visual defect inspection","description":"Develop visual inspection systems that detect, classify, localize, or segment visible industrial defects in 2D, x-ray, wafer, electronics, surface, casting, fiber, composite, solar-cell, and manufactured-part imagery, including deployable supervised pipelines.","inclusion":"Include supervised, weakly supervised, detection, segmentation, classification, CNN, transformer, explainable, reference-image, lightweight, edge, or high-throughput visual inspection work when the primary objective is visual defect recognition or deployable visual inspection performance.","exclusion":"If the main objective is general one-class normality modeling, use representation, reconstruction, or density directions; if 3D geometry, RGB-D fusion, components, or compositional logic are central, use structural multimodal reasoning.","nearest_boundary":"Structural multimodal reasoning","representative_topics":["automated defect detection in silicon detector modules using deep learning object detection","lightweight custom architectures for efficient metal casting defect detection via separable convolutions","automated defect recognition in industrial x-ray castings using convolutional neural networks"]},{"direction_index":6,"label":"D6","name":"Structural multimodal reasoning","description":"Detect industrial anomalies by modeling 3D geometry, RGB-D or point-cloud fusion, cross-modal mappings, object parts, component structure, adjustable configurations, or logical and compositional relations beyond ordinary 2D appearance cues.","inclusion":"Include work whose core objective is multimodal 2D-3D fusion, depth or point-cloud anomaly localization, cross-modal feature mapping, part segmentation, component-aware inspection, or logical anomaly reasoning.","exclusion":"If the method relies mainly on single-modality 2D appearance, x-ray imagery, or ordinary segmentation cues, use visual defect inspection; if geometry is merely an input feature inside a general embedding detector, use normality representation and distillation.","nearest_boundary":"Visual defect inspection","representative_topics":["multimodal industrial anomaly detection via hybrid point cloud and rgb feature fusion","component-aware logical anomaly detection for adjustable industrial visual inspection","lightweight multi-modality 3d anomaly segmentation for real-time industrial deployment"]},{"direction_index":7,"label":"D7","name":"Benchmarks and evaluation resources","description":"Create, organize, survey, benchmark, or systematically evaluate industrial anomaly detection datasets, baselines, challenges, metrics, protocols, methods, or application testbeds rather than proposing a new detector mechanism.","inclusion":"Include dataset papers, benchmark suites, challenge or evaluation protocols, surveys, systematic reviews, comparative baseline studies, and resource-oriented application evaluations where mapping or measuring the field is central.","exclusion":"If a new detector, adaptation strategy, representation, reconstruction model, deployment architecture, or multimodal reasoning method is the main contribution, use the corresponding algorithmic direction.","nearest_boundary":"Visual defect inspection","representative_topics":["survey of defect detection methods for die attachment and wire bonding in ic manufacturing","unsupervised anomaly detection dataset for outdoor maintenance inspection with surface and logical anomalies","unified benchmark for industrial image anomaly detection across supervision and learning paradigms"]}],"split":"dev"}
{"episode_id":"F262_T2025-07","field_index":262,"field":"industrial anomaly detection","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Normality representation and distillation","description":"Learn discriminative representations of normal industrial data and detect anomalies through feature-space deviation, memory/prototype mismatch, contrastive separation, or teacher-student feature discrepancy.","inclusion":"Include work whose main objective is constructing or adapting normal feature manifolds, pretrained/self-supervised embeddings, memory banks, prototypes, contrastive objectives, or knowledge-distillation discrepancies as the primary anomaly signal.","exclusion":"If the decisive signal is an input-versus-restoration residual, use reconstruction and restoration; if explicit likelihood under an invertible or density model is central, use density modeling.","nearest_boundary":"Reconstruction and restoration","representative_topics":["target-adaptive feature adaptation for anomaly localization via coupled hypersphere modeling","unsupervised textural defect detection via memory rearrangement and background reconstruction for edge manufacturing inspection","unsupervised industrial anomaly detection via generative and contrastive vision transformer networks"]},{"direction_index":1,"label":"D1","name":"Reconstruction and restoration","description":"Model normal observations by reconstructing, denoising, inpainting, restoring, or regenerating normal patterns and identify anomalies through residuals or restoration gaps.","inclusion":"Include autoencoder, VAE, GAN, masked reconstruction, diffusion denoising, inpainting, or restoration methods where the primary anomaly evidence is discrepancy between the observed input and a reconstructed or restored normal version.","exclusion":"If anomaly scoring is primarily nearest-neighbor, prototype, memory, contrastive, or teacher-student feature disagreement without residual restoration, use normality representation and distillation; if explicit likelihood is primary, use density modeling.","nearest_boundary":"Normality representation and distillation","representative_topics":["deep autoencoder anomaly detection in textured images using complex wavelet structural similarity loss","self-supervised masked convolutional transformer blocks for reconstruction-based anomaly detection","autoencoder-based unsupervised anomaly detection with explained fault localization in industrial cooling systems"]},{"direction_index":2,"label":"D2","name":"Density and flow modeling","description":"Estimate normal-data probability density or likelihood for industrial anomaly detection, especially with normalizing flows, invertible networks, Gaussian mixtures, or related probabilistic normality models.","inclusion":"Include work whose main scientific objective is explicit density estimation and whose anomaly score is low likelihood, abnormal latent density, or flow-based distributional contrast.","exclusion":"If a generative model is used mainly to produce a reconstruction residual, use reconstruction and restoration; if density estimation is incidental to embedding comparison, use normality representation and distillation.","nearest_boundary":"Reconstruction and restoration","representative_topics":["normalizing flow anomaly detection with spatially-aware convolutional attention for localization","self-supervised texture anomaly detection via normalizing flow and dictionary learning fusion","connecting generative and discriminative paradigms via discriminative-oriented gaussian mixture models for anomaly detection"]},{"direction_index":3,"label":"D3","name":"Data-efficient and robust adaptation","description":"Improve industrial anomaly or defect detection when labels, anomaly examples, clean normal data, domains, categories, or operating conditions are limited, imbalanced, contaminated, shifting, or incrementally changing.","inclusion":"Include few-shot, low-label, active, incremental, synthetic-data, augmentation, class-imbalance, contaminated-training, domain-adaptation, domain-generalization, and cross-category robustness work when this data regime is the primary objective.","exclusion":"If scarce-label performance is only a secondary consequence of a representation, reconstruction, or density model, use that mechanism-specific direction; if the main contribution is non-visual sensor modeling, use sensor and process fault diagnosis.","nearest_boundary":"Normality representation and distillation","representative_topics":["supervised regression-based anomaly detection using continuous labels and augmentation for low-resource settings","boundary-guided semi-push-pull contrastive learning for supervised few-shot anomaly detection","knowledge distillation for catastrophic forgetting in incremental few-shot surface defect detection"]},{"direction_index":4,"label":"D4","name":"Sensor and process fault diagnosis","description":"Detect and diagnose industrial anomalies from non-visual temporal or signal data, including sensors, vibration, acoustics, pressure, IIoT streams, machine operating traces, and process measurements.","inclusion":"Include work whose main objective is non-image industrial fault detection, acoustic or vibration anomaly detection, temporal normality modeling, inter-sensor transfer, graph or clustering over sensor relationships, forecasting-based detection, or process-monitoring diagnosis.","exclusion":"If the work is primarily a generic reconstruction, representation, or density method demonstrated on signals, use the corresponding mechanism direction unless the sensor/process modeling objective dominates; image or x-ray inspection belongs to visual inspection.","nearest_boundary":"Data-efficient and robust adaptation","representative_topics":["hierarchical conditional variational autoencoder for cross-domain acoustic anomaly detection in industrial monitoring","inter-sensor transfer learning for anomaly detection and defect classification in smart manufacturing","few-shot anomaly detection via autoregressive normality learning with positive representation separation"]},{"direction_index":5,"label":"D5","name":"Visual defect inspection","description":"Develop visual inspection systems that detect, classify, localize, or segment visible industrial defects in 2D, x-ray, wafer, electronics, surface, casting, fiber, composite, solar-cell, and manufactured-part imagery, including deployable supervised pipelines.","inclusion":"Include supervised, weakly supervised, detection, segmentation, classification, CNN, transformer, explainable, reference-image, lightweight, edge, or high-throughput visual inspection work when the primary objective is visual defect recognition or deployable visual inspection performance.","exclusion":"If the main objective is general one-class normality modeling, use representation, reconstruction, or density directions; if 3D geometry, RGB-D fusion, components, or compositional logic are central, use structural multimodal reasoning.","nearest_boundary":"Structural multimodal reasoning","representative_topics":["automated defect detection in silicon detector modules using deep learning object detection","lightweight custom architectures for efficient metal casting defect detection via separable convolutions","automated defect recognition in industrial x-ray castings using convolutional neural networks"]},{"direction_index":6,"label":"D6","name":"Structural multimodal reasoning","description":"Detect industrial anomalies by modeling 3D geometry, RGB-D or point-cloud fusion, cross-modal mappings, object parts, component structure, adjustable configurations, or logical and compositional relations beyond ordinary 2D appearance cues.","inclusion":"Include work whose core objective is multimodal 2D-3D fusion, depth or point-cloud anomaly localization, cross-modal feature mapping, part segmentation, component-aware inspection, or logical anomaly reasoning.","exclusion":"If the method relies mainly on single-modality 2D appearance, x-ray imagery, or ordinary segmentation cues, use visual defect inspection; if geometry is merely an input feature inside a general embedding detector, use normality representation and distillation.","nearest_boundary":"Visual defect inspection","representative_topics":["multimodal industrial anomaly detection via hybrid point cloud and rgb feature fusion","component-aware logical anomaly detection for adjustable industrial visual inspection","lightweight multi-modality 3d anomaly segmentation for real-time industrial deployment"]},{"direction_index":7,"label":"D7","name":"Benchmarks and evaluation resources","description":"Create, organize, survey, benchmark, or systematically evaluate industrial anomaly detection datasets, baselines, challenges, metrics, protocols, methods, or application testbeds rather than proposing a new detector mechanism.","inclusion":"Include dataset papers, benchmark suites, challenge or evaluation protocols, surveys, systematic reviews, comparative baseline studies, and resource-oriented application evaluations where mapping or measuring the field is central.","exclusion":"If a new detector, adaptation strategy, representation, reconstruction model, deployment architecture, or multimodal reasoning method is the main contribution, use the corresponding algorithmic direction.","nearest_boundary":"Visual defect inspection","representative_topics":["survey of defect detection methods for die attachment and wire bonding in ic manufacturing","unsupervised anomaly detection dataset for outdoor maintenance inspection with surface and logical anomalies","unified benchmark for industrial image anomaly detection across supervision and learning paradigms"]}],"split":"dev"}
{"episode_id":"F262_T2026-01","field_index":262,"field":"industrial anomaly detection","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Normality representation and distillation","description":"Learn discriminative representations of normal industrial data and detect anomalies through feature-space deviation, memory/prototype mismatch, contrastive separation, or teacher-student feature discrepancy.","inclusion":"Include work whose main objective is constructing or adapting normal feature manifolds, pretrained/self-supervised embeddings, memory banks, prototypes, contrastive objectives, or knowledge-distillation discrepancies as the primary anomaly signal.","exclusion":"If the decisive signal is an input-versus-restoration residual, use reconstruction and restoration; if explicit likelihood under an invertible or density model is central, use density modeling.","nearest_boundary":"Reconstruction and restoration","representative_topics":["target-adaptive feature adaptation for anomaly localization via coupled hypersphere modeling","unsupervised textural defect detection via memory rearrangement and background reconstruction for edge manufacturing inspection","unsupervised industrial anomaly detection via generative and contrastive vision transformer networks"]},{"direction_index":1,"label":"D1","name":"Reconstruction and restoration","description":"Model normal observations by reconstructing, denoising, inpainting, restoring, or regenerating normal patterns and identify anomalies through residuals or restoration gaps.","inclusion":"Include autoencoder, VAE, GAN, masked reconstruction, diffusion denoising, inpainting, or restoration methods where the primary anomaly evidence is discrepancy between the observed input and a reconstructed or restored normal version.","exclusion":"If anomaly scoring is primarily nearest-neighbor, prototype, memory, contrastive, or teacher-student feature disagreement without residual restoration, use normality representation and distillation; if explicit likelihood is primary, use density modeling.","nearest_boundary":"Normality representation and distillation","representative_topics":["deep autoencoder anomaly detection in textured images using complex wavelet structural similarity loss","self-supervised masked convolutional transformer blocks for reconstruction-based anomaly detection","autoencoder-based unsupervised anomaly detection with explained fault localization in industrial cooling systems"]},{"direction_index":2,"label":"D2","name":"Density and flow modeling","description":"Estimate normal-data probability density or likelihood for industrial anomaly detection, especially with normalizing flows, invertible networks, Gaussian mixtures, or related probabilistic normality models.","inclusion":"Include work whose main scientific objective is explicit density estimation and whose anomaly score is low likelihood, abnormal latent density, or flow-based distributional contrast.","exclusion":"If a generative model is used mainly to produce a reconstruction residual, use reconstruction and restoration; if density estimation is incidental to embedding comparison, use normality representation and distillation.","nearest_boundary":"Reconstruction and restoration","representative_topics":["normalizing flow anomaly detection with spatially-aware convolutional attention for localization","self-supervised texture anomaly detection via normalizing flow and dictionary learning fusion","connecting generative and discriminative paradigms via discriminative-oriented gaussian mixture models for anomaly detection"]},{"direction_index":3,"label":"D3","name":"Data-efficient and robust adaptation","description":"Improve industrial anomaly or defect detection when labels, anomaly examples, clean normal data, domains, categories, or operating conditions are limited, imbalanced, contaminated, shifting, or incrementally changing.","inclusion":"Include few-shot, low-label, active, incremental, synthetic-data, augmentation, class-imbalance, contaminated-training, domain-adaptation, domain-generalization, and cross-category robustness work when this data regime is the primary objective.","exclusion":"If scarce-label performance is only a secondary consequence of a representation, reconstruction, or density model, use that mechanism-specific direction; if the main contribution is non-visual sensor modeling, use sensor and process fault diagnosis.","nearest_boundary":"Normality representation and distillation","representative_topics":["supervised regression-based anomaly detection using continuous labels and augmentation for low-resource settings","boundary-guided semi-push-pull contrastive learning for supervised few-shot anomaly detection","knowledge distillation for catastrophic forgetting in incremental few-shot surface defect detection"]},{"direction_index":4,"label":"D4","name":"Sensor and process fault diagnosis","description":"Detect and diagnose industrial anomalies from non-visual temporal or signal data, including sensors, vibration, acoustics, pressure, IIoT streams, machine operating traces, and process measurements.","inclusion":"Include work whose main objective is non-image industrial fault detection, acoustic or vibration anomaly detection, temporal normality modeling, inter-sensor transfer, graph or clustering over sensor relationships, forecasting-based detection, or process-monitoring diagnosis.","exclusion":"If the work is primarily a generic reconstruction, representation, or density method demonstrated on signals, use the corresponding mechanism direction unless the sensor/process modeling objective dominates; image or x-ray inspection belongs to visual inspection.","nearest_boundary":"Data-efficient and robust adaptation","representative_topics":["hierarchical conditional variational autoencoder for cross-domain acoustic anomaly detection in industrial monitoring","inter-sensor transfer learning for anomaly detection and defect classification in smart manufacturing","few-shot anomaly detection via autoregressive normality learning with positive representation separation"]},{"direction_index":5,"label":"D5","name":"Visual defect inspection","description":"Develop visual inspection systems that detect, classify, localize, or segment visible industrial defects in 2D, x-ray, wafer, electronics, surface, casting, fiber, composite, solar-cell, and manufactured-part imagery, including deployable supervised pipelines.","inclusion":"Include supervised, weakly supervised, detection, segmentation, classification, CNN, transformer, explainable, reference-image, lightweight, edge, or high-throughput visual inspection work when the primary objective is visual defect recognition or deployable visual inspection performance.","exclusion":"If the main objective is general one-class normality modeling, use representation, reconstruction, or density directions; if 3D geometry, RGB-D fusion, components, or compositional logic are central, use structural multimodal reasoning.","nearest_boundary":"Structural multimodal reasoning","representative_topics":["automated defect detection in silicon detector modules using deep learning object detection","lightweight custom architectures for efficient metal casting defect detection via separable convolutions","automated defect recognition in industrial x-ray castings using convolutional neural networks"]},{"direction_index":6,"label":"D6","name":"Structural multimodal reasoning","description":"Detect industrial anomalies by modeling 3D geometry, RGB-D or point-cloud fusion, cross-modal mappings, object parts, component structure, adjustable configurations, or logical and compositional relations beyond ordinary 2D appearance cues.","inclusion":"Include work whose core objective is multimodal 2D-3D fusion, depth or point-cloud anomaly localization, cross-modal feature mapping, part segmentation, component-aware inspection, or logical anomaly reasoning.","exclusion":"If the method relies mainly on single-modality 2D appearance, x-ray imagery, or ordinary segmentation cues, use visual defect inspection; if geometry is merely an input feature inside a general embedding detector, use normality representation and distillation.","nearest_boundary":"Visual defect inspection","representative_topics":["multimodal industrial anomaly detection via hybrid point cloud and rgb feature fusion","component-aware logical anomaly detection for adjustable industrial visual inspection","lightweight multi-modality 3d anomaly segmentation for real-time industrial deployment"]},{"direction_index":7,"label":"D7","name":"Benchmarks and evaluation resources","description":"Create, organize, survey, benchmark, or systematically evaluate industrial anomaly detection datasets, baselines, challenges, metrics, protocols, methods, or application testbeds rather than proposing a new detector mechanism.","inclusion":"Include dataset papers, benchmark suites, challenge or evaluation protocols, surveys, systematic reviews, comparative baseline studies, and resource-oriented application evaluations where mapping or measuring the field is central.","exclusion":"If a new detector, adaptation strategy, representation, reconstruction model, deployment architecture, or multimodal reasoning method is the main contribution, use the corresponding algorithmic direction.","nearest_boundary":"Visual defect inspection","representative_topics":["survey of defect detection methods for die attachment and wire bonding in ic manufacturing","unsupervised anomaly detection dataset for outdoor maintenance inspection with surface and logical anomalies","unified benchmark for industrial image anomaly detection across supervision and learning paradigms"]}],"split":"dev"}
{"episode_id":"F263_T2024-01","field_index":263,"field":"clinical predictive modeling","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Fixed-endpoint clinical risk prediction","description":"Supervised prediction of diagnosis, risk, complications, biomarkers, severity, treatment outcome, or other fixed clinical endpoints from structured, imaging, genomic, demographic, or otherwise single-task clinical inputs.","inclusion":"Assign atoms whose main objective is ordinary disease- or organ-specific endpoint prediction, feature selection, tabular risk modeling, biomarker estimation, or specialized task architecture when the endpoint prediction task is primary and not dominated by longitudinal dynamics, censoring, multimodal fusion, language modeling, trustworthiness, COVID, or physiological signal management.","exclusion":"Use time-dependent prognosis when the scientific objective is longitudinal trajectory, disease progression, or censored event-time modeling; use multimodal fusion when cross-modal integration is central; use physiological signal management for ECG, EEG, or glucose-focused monitoring.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","comparative machine learning classification for breast cancer prediction from tabular data","deep learning-based opportunistic hip fracture risk prediction from plain radiographs and ct projections"]},{"direction_index":1,"label":"D1","name":"Time-dependent prognosis and trajectories","description":"Prediction and modeling of clinical outcomes through temporal structure, censored event times, hazards, latent states, disease progression, or patient trajectories.","inclusion":"Assign atoms whose primary objective is modeling longitudinal EHR sequences, ICU trajectories, irregular clinical time series, dynamic disease progression, censored survival outcomes, competing risks, hazards, multistate event processes, or latent progression states.","exclusion":"Use fixed-endpoint clinical risk prediction for non-censored static endpoints; use representation learning under data limitations when pretraining, sparsity, missingness, or schema transfer is the primary scientific objective rather than temporal prognosis itself.","nearest_boundary":"Representation learning under clinical data limitations","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","neural ode-based patient trajectory modeling with clinical code embeddings for early disease prediction","pseudo value-based neural networks for subject-specific multi-state survival prediction under censoring"]},{"direction_index":2,"label":"D2","name":"Multimodal and neurodegenerative fusion","description":"Clinical prediction programs centered on integrating heterogeneous modalities or views, including imaging, structured data, text, genetics, sensors, longitudinal measurements, and neurodegenerative-disease data sources.","inclusion":"Assign atoms where cross-modal fusion, multi-view alignment, missing-modality handling, modality-specific attention, or coordinated use of heterogeneous data is the central objective; include Alzheimer's, MCI, dementia, Parkinson's, and cognitive-decline prediction when the sustained objective is multimodal or specialized neurodegenerative prediction.","exclusion":"Use language and behavioral clinical prediction when unstructured text, dialogue, speech, or behavioral assessment is the dominant data program; use fixed-endpoint prediction when a single modality or static endpoint dominates and fusion is incidental.","nearest_boundary":"Language and behavioral clinical prediction","representative_topics":["multimodal mri glioma survival prediction via cross-scale nonlocal feature fusion","multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","cross-modal attention fusion for multimodal alzheimer's disease and mild cognitive impairment diagnosis"]},{"direction_index":3,"label":"D3","name":"Language and behavioral clinical prediction","description":"Prediction, reasoning, phenotyping, and risk assessment from unstructured clinical text, reports, dialogue, transcripts, language-model representations, speech, behavioral observations, or psychiatric and behavioral-health signals.","inclusion":"Assign atoms whose primary objective is clinical NLP, note-based prognosis, text-derived concepts, transcription-based diagnosis, dialogue-based assessment, language-model clinical reasoning, or mental and behavioral-health prediction from speech, activity, physiological, clinical, or multimodal behavioral data.","exclusion":"Use representation learning under clinical data limitations when the main objective is general pretraining or transferable EHR representation rather than text or behavioral clinical assessment; use multimodal fusion when the dominant contribution is general cross-modal fusion outside a language or behavioral-health objective.","nearest_boundary":"Representation learning under clinical data limitations","representative_topics":["nlp-based automated diagnosis from clinical transcription reports using hybrid deep learning models","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","transformer-based clinical note embedding for mortality prediction via umls concept mapping"]},{"direction_index":4,"label":"D4","name":"Representation learning under data limitations","description":"Learning transferable, sparse, weakly supervised, self-supervised, contrastive, semi-supervised, or foundation-style patient representations to improve prediction under limited labels, missingness, sparsity, imbalance, heterogeneous schemas, or rare features.","inclusion":"Assign atoms where the primary objective is pretraining, contrastive learning, weak supervision, rare-feature learning, class-imbalance handling, transfer, schema harmonization, foundation modeling, or representation design for downstream clinical prediction under data limitations.","exclusion":"Use time-dependent prognosis when ordinary longitudinal sequence modeling or survival estimation is primary; use language and behavioral clinical prediction when the core objective is clinical text reasoning or behavioral-health assessment rather than transferable representation learning.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["weakly supervised pseudo-labeling for vital sign alert classification in clinical monitoring","tree-guided feature selection and logic aggregation for rare binary ehr-based risk prediction","class-imbalanced learning for adverse event severity prediction in immunization data"]},{"direction_index":5,"label":"D5","name":"Trustworthy and equitable prediction","description":"Clinical prediction programs whose primary objective is interpretability, explanation, uncertainty, calibration, stability, robustness, external generalization, bias mitigation, causal validity, or equitable performance.","inclusion":"Assign atoms focused on inherently interpretable models, post-hoc explanations, clinical risk-score transparency, uncertainty quantification, calibration, model instability, leakage and validation critique, distribution shift, fairness mitigation, health-disparity modeling, causal representation, target-trial emulation, or treatment-effect validity.","exclusion":"Use the task-specific predictive program when explanation, robustness, fairness, or validation is only a routine evaluation component; use fixed-endpoint prediction when the main contribution is accuracy on a conventional endpoint rather than trustworthiness or equity.","nearest_boundary":"Fixed-endpoint clinical risk prediction","representative_topics":["bias-aware machine learning pipeline for equitable hospital readmission prediction","interpretable clinical risk scores via explicit intervention policy modeling in icu","target trial emulation for individualized antibiotic treatment effect estimation in resistant infection"]},{"direction_index":6,"label":"D6","name":"Physiological signal and disease-management prediction","description":"Prediction programs centered on physiological signals or continuous disease-management data, including cardiac signals and imaging, EEG and neural signals, glucose monitoring, diabetes management, and related sensor-derived clinical states.","inclusion":"Assign atoms whose primary objective is prediction from ECG, PPG, heart-rate variability, echocardiography, cardiac images or shapes, EEG, cortico-cortical evoked potentials, continuous glucose monitors, glycemic trajectories, hypoglycemia or hyperglycemia risk, meal effects, or diabetes-management signals.","exclusion":"Use fixed-endpoint clinical risk prediction for ordinary tabular cardiovascular or diabetes risk models without a signal-management focus; use multimodal fusion when general cross-modal integration rather than the physiological signal program is central.","nearest_boundary":"Fixed-endpoint clinical risk prediction","representative_topics":["hrv-based ventricular and atrial arrhythmia prediction for implantable cardiac devices","binarized 1d convolutional neural networks for hardware-efficient eeg-based seizure prediction","multitask neural network for macronutrient composition estimation from continuous glucose monitor signals"]},{"direction_index":7,"label":"D7","name":"COVID-specific outcome prediction","description":"Predictive modeling for acute COVID-19 severity, hospitalization, ICU transfer, mortality, disease progression, and post-acute or long-COVID sequelae.","inclusion":"Assign atoms where SARS-CoV-2 infection or its aftermath defines the scientific objective, including pediatric COVID severity, COVID hospitalization, COVID ICU outcome benchmarking, severe progression, and long-COVID risk-factor or sequelae prediction.","exclusion":"Use time-dependent prognosis for pandemic-agnostic ICU or EHR trajectory methods merely evaluated on COVID data; use trustworthy and equitable prediction when reliability, fairness, or causal validity dominates over COVID-specific clinical prediction.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]}],"split":"dev"}
{"episode_id":"F263_T2024-07","field_index":263,"field":"clinical predictive modeling","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Fixed-endpoint clinical risk prediction","description":"Supervised prediction of diagnosis, risk, complications, biomarkers, severity, treatment outcome, or other fixed clinical endpoints from structured, imaging, genomic, demographic, or otherwise single-task clinical inputs.","inclusion":"Assign atoms whose main objective is ordinary disease- or organ-specific endpoint prediction, feature selection, tabular risk modeling, biomarker estimation, or specialized task architecture when the endpoint prediction task is primary and not dominated by longitudinal dynamics, censoring, multimodal fusion, language modeling, trustworthiness, COVID, or physiological signal management.","exclusion":"Use time-dependent prognosis when the scientific objective is longitudinal trajectory, disease progression, or censored event-time modeling; use multimodal fusion when cross-modal integration is central; use physiological signal management for ECG, EEG, or glucose-focused monitoring.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","comparative machine learning classification for breast cancer prediction from tabular data","deep learning-based opportunistic hip fracture risk prediction from plain radiographs and ct projections"]},{"direction_index":1,"label":"D1","name":"Time-dependent prognosis and trajectories","description":"Prediction and modeling of clinical outcomes through temporal structure, censored event times, hazards, latent states, disease progression, or patient trajectories.","inclusion":"Assign atoms whose primary objective is modeling longitudinal EHR sequences, ICU trajectories, irregular clinical time series, dynamic disease progression, censored survival outcomes, competing risks, hazards, multistate event processes, or latent progression states.","exclusion":"Use fixed-endpoint clinical risk prediction for non-censored static endpoints; use representation learning under data limitations when pretraining, sparsity, missingness, or schema transfer is the primary scientific objective rather than temporal prognosis itself.","nearest_boundary":"Representation learning under clinical data limitations","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","neural ode-based patient trajectory modeling with clinical code embeddings for early disease prediction","pseudo value-based neural networks for subject-specific multi-state survival prediction under censoring"]},{"direction_index":2,"label":"D2","name":"Multimodal and neurodegenerative fusion","description":"Clinical prediction programs centered on integrating heterogeneous modalities or views, including imaging, structured data, text, genetics, sensors, longitudinal measurements, and neurodegenerative-disease data sources.","inclusion":"Assign atoms where cross-modal fusion, multi-view alignment, missing-modality handling, modality-specific attention, or coordinated use of heterogeneous data is the central objective; include Alzheimer's, MCI, dementia, Parkinson's, and cognitive-decline prediction when the sustained objective is multimodal or specialized neurodegenerative prediction.","exclusion":"Use language and behavioral clinical prediction when unstructured text, dialogue, speech, or behavioral assessment is the dominant data program; use fixed-endpoint prediction when a single modality or static endpoint dominates and fusion is incidental.","nearest_boundary":"Language and behavioral clinical prediction","representative_topics":["multimodal mri glioma survival prediction via cross-scale nonlocal feature fusion","multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","cross-modal attention fusion for multimodal alzheimer's disease and mild cognitive impairment diagnosis"]},{"direction_index":3,"label":"D3","name":"Language and behavioral clinical prediction","description":"Prediction, reasoning, phenotyping, and risk assessment from unstructured clinical text, reports, dialogue, transcripts, language-model representations, speech, behavioral observations, or psychiatric and behavioral-health signals.","inclusion":"Assign atoms whose primary objective is clinical NLP, note-based prognosis, text-derived concepts, transcription-based diagnosis, dialogue-based assessment, language-model clinical reasoning, or mental and behavioral-health prediction from speech, activity, physiological, clinical, or multimodal behavioral data.","exclusion":"Use representation learning under clinical data limitations when the main objective is general pretraining or transferable EHR representation rather than text or behavioral clinical assessment; use multimodal fusion when the dominant contribution is general cross-modal fusion outside a language or behavioral-health objective.","nearest_boundary":"Representation learning under clinical data limitations","representative_topics":["nlp-based automated diagnosis from clinical transcription reports using hybrid deep learning models","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","transformer-based clinical note embedding for mortality prediction via umls concept mapping"]},{"direction_index":4,"label":"D4","name":"Representation learning under data limitations","description":"Learning transferable, sparse, weakly supervised, self-supervised, contrastive, semi-supervised, or foundation-style patient representations to improve prediction under limited labels, missingness, sparsity, imbalance, heterogeneous schemas, or rare features.","inclusion":"Assign atoms where the primary objective is pretraining, contrastive learning, weak supervision, rare-feature learning, class-imbalance handling, transfer, schema harmonization, foundation modeling, or representation design for downstream clinical prediction under data limitations.","exclusion":"Use time-dependent prognosis when ordinary longitudinal sequence modeling or survival estimation is primary; use language and behavioral clinical prediction when the core objective is clinical text reasoning or behavioral-health assessment rather than transferable representation learning.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["weakly supervised pseudo-labeling for vital sign alert classification in clinical monitoring","tree-guided feature selection and logic aggregation for rare binary ehr-based risk prediction","class-imbalanced learning for adverse event severity prediction in immunization data"]},{"direction_index":5,"label":"D5","name":"Trustworthy and equitable prediction","description":"Clinical prediction programs whose primary objective is interpretability, explanation, uncertainty, calibration, stability, robustness, external generalization, bias mitigation, causal validity, or equitable performance.","inclusion":"Assign atoms focused on inherently interpretable models, post-hoc explanations, clinical risk-score transparency, uncertainty quantification, calibration, model instability, leakage and validation critique, distribution shift, fairness mitigation, health-disparity modeling, causal representation, target-trial emulation, or treatment-effect validity.","exclusion":"Use the task-specific predictive program when explanation, robustness, fairness, or validation is only a routine evaluation component; use fixed-endpoint prediction when the main contribution is accuracy on a conventional endpoint rather than trustworthiness or equity.","nearest_boundary":"Fixed-endpoint clinical risk prediction","representative_topics":["bias-aware machine learning pipeline for equitable hospital readmission prediction","interpretable clinical risk scores via explicit intervention policy modeling in icu","target trial emulation for individualized antibiotic treatment effect estimation in resistant infection"]},{"direction_index":6,"label":"D6","name":"Physiological signal and disease-management prediction","description":"Prediction programs centered on physiological signals or continuous disease-management data, including cardiac signals and imaging, EEG and neural signals, glucose monitoring, diabetes management, and related sensor-derived clinical states.","inclusion":"Assign atoms whose primary objective is prediction from ECG, PPG, heart-rate variability, echocardiography, cardiac images or shapes, EEG, cortico-cortical evoked potentials, continuous glucose monitors, glycemic trajectories, hypoglycemia or hyperglycemia risk, meal effects, or diabetes-management signals.","exclusion":"Use fixed-endpoint clinical risk prediction for ordinary tabular cardiovascular or diabetes risk models without a signal-management focus; use multimodal fusion when general cross-modal integration rather than the physiological signal program is central.","nearest_boundary":"Fixed-endpoint clinical risk prediction","representative_topics":["hrv-based ventricular and atrial arrhythmia prediction for implantable cardiac devices","binarized 1d convolutional neural networks for hardware-efficient eeg-based seizure prediction","multitask neural network for macronutrient composition estimation from continuous glucose monitor signals"]},{"direction_index":7,"label":"D7","name":"COVID-specific outcome prediction","description":"Predictive modeling for acute COVID-19 severity, hospitalization, ICU transfer, mortality, disease progression, and post-acute or long-COVID sequelae.","inclusion":"Assign atoms where SARS-CoV-2 infection or its aftermath defines the scientific objective, including pediatric COVID severity, COVID hospitalization, COVID ICU outcome benchmarking, severe progression, and long-COVID risk-factor or sequelae prediction.","exclusion":"Use time-dependent prognosis for pandemic-agnostic ICU or EHR trajectory methods merely evaluated on COVID data; use trustworthy and equitable prediction when reliability, fairness, or causal validity dominates over COVID-specific clinical prediction.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]}],"split":"dev"}
{"episode_id":"F263_T2025-01","field_index":263,"field":"clinical predictive modeling","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Fixed-endpoint clinical risk prediction","description":"Supervised prediction of diagnosis, risk, complications, biomarkers, severity, treatment outcome, or other fixed clinical endpoints from structured, imaging, genomic, demographic, or otherwise single-task clinical inputs.","inclusion":"Assign atoms whose main objective is ordinary disease- or organ-specific endpoint prediction, feature selection, tabular risk modeling, biomarker estimation, or specialized task architecture when the endpoint prediction task is primary and not dominated by longitudinal dynamics, censoring, multimodal fusion, language modeling, trustworthiness, COVID, or physiological signal management.","exclusion":"Use time-dependent prognosis when the scientific objective is longitudinal trajectory, disease progression, or censored event-time modeling; use multimodal fusion when cross-modal integration is central; use physiological signal management for ECG, EEG, or glucose-focused monitoring.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","comparative machine learning classification for breast cancer prediction from tabular data","deep learning-based opportunistic hip fracture risk prediction from plain radiographs and ct projections"]},{"direction_index":1,"label":"D1","name":"Time-dependent prognosis and trajectories","description":"Prediction and modeling of clinical outcomes through temporal structure, censored event times, hazards, latent states, disease progression, or patient trajectories.","inclusion":"Assign atoms whose primary objective is modeling longitudinal EHR sequences, ICU trajectories, irregular clinical time series, dynamic disease progression, censored survival outcomes, competing risks, hazards, multistate event processes, or latent progression states.","exclusion":"Use fixed-endpoint clinical risk prediction for non-censored static endpoints; use representation learning under data limitations when pretraining, sparsity, missingness, or schema transfer is the primary scientific objective rather than temporal prognosis itself.","nearest_boundary":"Representation learning under clinical data limitations","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","neural ode-based patient trajectory modeling with clinical code embeddings for early disease prediction","pseudo value-based neural networks for subject-specific multi-state survival prediction under censoring"]},{"direction_index":2,"label":"D2","name":"Multimodal and neurodegenerative fusion","description":"Clinical prediction programs centered on integrating heterogeneous modalities or views, including imaging, structured data, text, genetics, sensors, longitudinal measurements, and neurodegenerative-disease data sources.","inclusion":"Assign atoms where cross-modal fusion, multi-view alignment, missing-modality handling, modality-specific attention, or coordinated use of heterogeneous data is the central objective; include Alzheimer's, MCI, dementia, Parkinson's, and cognitive-decline prediction when the sustained objective is multimodal or specialized neurodegenerative prediction.","exclusion":"Use language and behavioral clinical prediction when unstructured text, dialogue, speech, or behavioral assessment is the dominant data program; use fixed-endpoint prediction when a single modality or static endpoint dominates and fusion is incidental.","nearest_boundary":"Language and behavioral clinical prediction","representative_topics":["multimodal mri glioma survival prediction via cross-scale nonlocal feature fusion","multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","cross-modal attention fusion for multimodal alzheimer's disease and mild cognitive impairment diagnosis"]},{"direction_index":3,"label":"D3","name":"Language and behavioral clinical prediction","description":"Prediction, reasoning, phenotyping, and risk assessment from unstructured clinical text, reports, dialogue, transcripts, language-model representations, speech, behavioral observations, or psychiatric and behavioral-health signals.","inclusion":"Assign atoms whose primary objective is clinical NLP, note-based prognosis, text-derived concepts, transcription-based diagnosis, dialogue-based assessment, language-model clinical reasoning, or mental and behavioral-health prediction from speech, activity, physiological, clinical, or multimodal behavioral data.","exclusion":"Use representation learning under clinical data limitations when the main objective is general pretraining or transferable EHR representation rather than text or behavioral clinical assessment; use multimodal fusion when the dominant contribution is general cross-modal fusion outside a language or behavioral-health objective.","nearest_boundary":"Representation learning under clinical data limitations","representative_topics":["nlp-based automated diagnosis from clinical transcription reports using hybrid deep learning models","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","transformer-based clinical note embedding for mortality prediction via umls concept mapping"]},{"direction_index":4,"label":"D4","name":"Representation learning under data limitations","description":"Learning transferable, sparse, weakly supervised, self-supervised, contrastive, semi-supervised, or foundation-style patient representations to improve prediction under limited labels, missingness, sparsity, imbalance, heterogeneous schemas, or rare features.","inclusion":"Assign atoms where the primary objective is pretraining, contrastive learning, weak supervision, rare-feature learning, class-imbalance handling, transfer, schema harmonization, foundation modeling, or representation design for downstream clinical prediction under data limitations.","exclusion":"Use time-dependent prognosis when ordinary longitudinal sequence modeling or survival estimation is primary; use language and behavioral clinical prediction when the core objective is clinical text reasoning or behavioral-health assessment rather than transferable representation learning.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["weakly supervised pseudo-labeling for vital sign alert classification in clinical monitoring","tree-guided feature selection and logic aggregation for rare binary ehr-based risk prediction","class-imbalanced learning for adverse event severity prediction in immunization data"]},{"direction_index":5,"label":"D5","name":"Trustworthy and equitable prediction","description":"Clinical prediction programs whose primary objective is interpretability, explanation, uncertainty, calibration, stability, robustness, external generalization, bias mitigation, causal validity, or equitable performance.","inclusion":"Assign atoms focused on inherently interpretable models, post-hoc explanations, clinical risk-score transparency, uncertainty quantification, calibration, model instability, leakage and validation critique, distribution shift, fairness mitigation, health-disparity modeling, causal representation, target-trial emulation, or treatment-effect validity.","exclusion":"Use the task-specific predictive program when explanation, robustness, fairness, or validation is only a routine evaluation component; use fixed-endpoint prediction when the main contribution is accuracy on a conventional endpoint rather than trustworthiness or equity.","nearest_boundary":"Fixed-endpoint clinical risk prediction","representative_topics":["bias-aware machine learning pipeline for equitable hospital readmission prediction","interpretable clinical risk scores via explicit intervention policy modeling in icu","target trial emulation for individualized antibiotic treatment effect estimation in resistant infection"]},{"direction_index":6,"label":"D6","name":"Physiological signal and disease-management prediction","description":"Prediction programs centered on physiological signals or continuous disease-management data, including cardiac signals and imaging, EEG and neural signals, glucose monitoring, diabetes management, and related sensor-derived clinical states.","inclusion":"Assign atoms whose primary objective is prediction from ECG, PPG, heart-rate variability, echocardiography, cardiac images or shapes, EEG, cortico-cortical evoked potentials, continuous glucose monitors, glycemic trajectories, hypoglycemia or hyperglycemia risk, meal effects, or diabetes-management signals.","exclusion":"Use fixed-endpoint clinical risk prediction for ordinary tabular cardiovascular or diabetes risk models without a signal-management focus; use multimodal fusion when general cross-modal integration rather than the physiological signal program is central.","nearest_boundary":"Fixed-endpoint clinical risk prediction","representative_topics":["hrv-based ventricular and atrial arrhythmia prediction for implantable cardiac devices","binarized 1d convolutional neural networks for hardware-efficient eeg-based seizure prediction","multitask neural network for macronutrient composition estimation from continuous glucose monitor signals"]},{"direction_index":7,"label":"D7","name":"COVID-specific outcome prediction","description":"Predictive modeling for acute COVID-19 severity, hospitalization, ICU transfer, mortality, disease progression, and post-acute or long-COVID sequelae.","inclusion":"Assign atoms where SARS-CoV-2 infection or its aftermath defines the scientific objective, including pediatric COVID severity, COVID hospitalization, COVID ICU outcome benchmarking, severe progression, and long-COVID risk-factor or sequelae prediction.","exclusion":"Use time-dependent prognosis for pandemic-agnostic ICU or EHR trajectory methods merely evaluated on COVID data; use trustworthy and equitable prediction when reliability, fairness, or causal validity dominates over COVID-specific clinical prediction.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]}],"split":"dev"}
{"episode_id":"F263_T2025-07","field_index":263,"field":"clinical predictive modeling","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Fixed-endpoint clinical risk prediction","description":"Supervised prediction of diagnosis, risk, complications, biomarkers, severity, treatment outcome, or other fixed clinical endpoints from structured, imaging, genomic, demographic, or otherwise single-task clinical inputs.","inclusion":"Assign atoms whose main objective is ordinary disease- or organ-specific endpoint prediction, feature selection, tabular risk modeling, biomarker estimation, or specialized task architecture when the endpoint prediction task is primary and not dominated by longitudinal dynamics, censoring, multimodal fusion, language modeling, trustworthiness, COVID, or physiological signal management.","exclusion":"Use time-dependent prognosis when the scientific objective is longitudinal trajectory, disease progression, or censored event-time modeling; use multimodal fusion when cross-modal integration is central; use physiological signal management for ECG, EEG, or glucose-focused monitoring.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","comparative machine learning classification for breast cancer prediction from tabular data","deep learning-based opportunistic hip fracture risk prediction from plain radiographs and ct projections"]},{"direction_index":1,"label":"D1","name":"Time-dependent prognosis and trajectories","description":"Prediction and modeling of clinical outcomes through temporal structure, censored event times, hazards, latent states, disease progression, or patient trajectories.","inclusion":"Assign atoms whose primary objective is modeling longitudinal EHR sequences, ICU trajectories, irregular clinical time series, dynamic disease progression, censored survival outcomes, competing risks, hazards, multistate event processes, or latent progression states.","exclusion":"Use fixed-endpoint clinical risk prediction for non-censored static endpoints; use representation learning under data limitations when pretraining, sparsity, missingness, or schema transfer is the primary scientific objective rather than temporal prognosis itself.","nearest_boundary":"Representation learning under clinical data limitations","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","neural ode-based patient trajectory modeling with clinical code embeddings for early disease prediction","pseudo value-based neural networks for subject-specific multi-state survival prediction under censoring"]},{"direction_index":2,"label":"D2","name":"Multimodal and neurodegenerative fusion","description":"Clinical prediction programs centered on integrating heterogeneous modalities or views, including imaging, structured data, text, genetics, sensors, longitudinal measurements, and neurodegenerative-disease data sources.","inclusion":"Assign atoms where cross-modal fusion, multi-view alignment, missing-modality handling, modality-specific attention, or coordinated use of heterogeneous data is the central objective; include Alzheimer's, MCI, dementia, Parkinson's, and cognitive-decline prediction when the sustained objective is multimodal or specialized neurodegenerative prediction.","exclusion":"Use language and behavioral clinical prediction when unstructured text, dialogue, speech, or behavioral assessment is the dominant data program; use fixed-endpoint prediction when a single modality or static endpoint dominates and fusion is incidental.","nearest_boundary":"Language and behavioral clinical prediction","representative_topics":["multimodal mri glioma survival prediction via cross-scale nonlocal feature fusion","multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","cross-modal attention fusion for multimodal alzheimer's disease and mild cognitive impairment diagnosis"]},{"direction_index":3,"label":"D3","name":"Language and behavioral clinical prediction","description":"Prediction, reasoning, phenotyping, and risk assessment from unstructured clinical text, reports, dialogue, transcripts, language-model representations, speech, behavioral observations, or psychiatric and behavioral-health signals.","inclusion":"Assign atoms whose primary objective is clinical NLP, note-based prognosis, text-derived concepts, transcription-based diagnosis, dialogue-based assessment, language-model clinical reasoning, or mental and behavioral-health prediction from speech, activity, physiological, clinical, or multimodal behavioral data.","exclusion":"Use representation learning under clinical data limitations when the main objective is general pretraining or transferable EHR representation rather than text or behavioral clinical assessment; use multimodal fusion when the dominant contribution is general cross-modal fusion outside a language or behavioral-health objective.","nearest_boundary":"Representation learning under clinical data limitations","representative_topics":["nlp-based automated diagnosis from clinical transcription reports using hybrid deep learning models","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","transformer-based clinical note embedding for mortality prediction via umls concept mapping"]},{"direction_index":4,"label":"D4","name":"Representation learning under data limitations","description":"Learning transferable, sparse, weakly supervised, self-supervised, contrastive, semi-supervised, or foundation-style patient representations to improve prediction under limited labels, missingness, sparsity, imbalance, heterogeneous schemas, or rare features.","inclusion":"Assign atoms where the primary objective is pretraining, contrastive learning, weak supervision, rare-feature learning, class-imbalance handling, transfer, schema harmonization, foundation modeling, or representation design for downstream clinical prediction under data limitations.","exclusion":"Use time-dependent prognosis when ordinary longitudinal sequence modeling or survival estimation is primary; use language and behavioral clinical prediction when the core objective is clinical text reasoning or behavioral-health assessment rather than transferable representation learning.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["weakly supervised pseudo-labeling for vital sign alert classification in clinical monitoring","tree-guided feature selection and logic aggregation for rare binary ehr-based risk prediction","class-imbalanced learning for adverse event severity prediction in immunization data"]},{"direction_index":5,"label":"D5","name":"Trustworthy and equitable prediction","description":"Clinical prediction programs whose primary objective is interpretability, explanation, uncertainty, calibration, stability, robustness, external generalization, bias mitigation, causal validity, or equitable performance.","inclusion":"Assign atoms focused on inherently interpretable models, post-hoc explanations, clinical risk-score transparency, uncertainty quantification, calibration, model instability, leakage and validation critique, distribution shift, fairness mitigation, health-disparity modeling, causal representation, target-trial emulation, or treatment-effect validity.","exclusion":"Use the task-specific predictive program when explanation, robustness, fairness, or validation is only a routine evaluation component; use fixed-endpoint prediction when the main contribution is accuracy on a conventional endpoint rather than trustworthiness or equity.","nearest_boundary":"Fixed-endpoint clinical risk prediction","representative_topics":["bias-aware machine learning pipeline for equitable hospital readmission prediction","interpretable clinical risk scores via explicit intervention policy modeling in icu","target trial emulation for individualized antibiotic treatment effect estimation in resistant infection"]},{"direction_index":6,"label":"D6","name":"Physiological signal and disease-management prediction","description":"Prediction programs centered on physiological signals or continuous disease-management data, including cardiac signals and imaging, EEG and neural signals, glucose monitoring, diabetes management, and related sensor-derived clinical states.","inclusion":"Assign atoms whose primary objective is prediction from ECG, PPG, heart-rate variability, echocardiography, cardiac images or shapes, EEG, cortico-cortical evoked potentials, continuous glucose monitors, glycemic trajectories, hypoglycemia or hyperglycemia risk, meal effects, or diabetes-management signals.","exclusion":"Use fixed-endpoint clinical risk prediction for ordinary tabular cardiovascular or diabetes risk models without a signal-management focus; use multimodal fusion when general cross-modal integration rather than the physiological signal program is central.","nearest_boundary":"Fixed-endpoint clinical risk prediction","representative_topics":["hrv-based ventricular and atrial arrhythmia prediction for implantable cardiac devices","binarized 1d convolutional neural networks for hardware-efficient eeg-based seizure prediction","multitask neural network for macronutrient composition estimation from continuous glucose monitor signals"]},{"direction_index":7,"label":"D7","name":"COVID-specific outcome prediction","description":"Predictive modeling for acute COVID-19 severity, hospitalization, ICU transfer, mortality, disease progression, and post-acute or long-COVID sequelae.","inclusion":"Assign atoms where SARS-CoV-2 infection or its aftermath defines the scientific objective, including pediatric COVID severity, COVID hospitalization, COVID ICU outcome benchmarking, severe progression, and long-COVID risk-factor or sequelae prediction.","exclusion":"Use time-dependent prognosis for pandemic-agnostic ICU or EHR trajectory methods merely evaluated on COVID data; use trustworthy and equitable prediction when reliability, fairness, or causal validity dominates over COVID-specific clinical prediction.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]}],"split":"dev"}
{"episode_id":"F263_T2026-01","field_index":263,"field":"clinical predictive modeling","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Fixed-endpoint clinical risk prediction","description":"Supervised prediction of diagnosis, risk, complications, biomarkers, severity, treatment outcome, or other fixed clinical endpoints from structured, imaging, genomic, demographic, or otherwise single-task clinical inputs.","inclusion":"Assign atoms whose main objective is ordinary disease- or organ-specific endpoint prediction, feature selection, tabular risk modeling, biomarker estimation, or specialized task architecture when the endpoint prediction task is primary and not dominated by longitudinal dynamics, censoring, multimodal fusion, language modeling, trustworthiness, COVID, or physiological signal management.","exclusion":"Use time-dependent prognosis when the scientific objective is longitudinal trajectory, disease progression, or censored event-time modeling; use multimodal fusion when cross-modal integration is central; use physiological signal management for ECG, EEG, or glucose-focused monitoring.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["feature selection for heart disease classification using high-dimensional clinical datasets","comparative machine learning classification for breast cancer prediction from tabular data","deep learning-based opportunistic hip fracture risk prediction from plain radiographs and ct projections"]},{"direction_index":1,"label":"D1","name":"Time-dependent prognosis and trajectories","description":"Prediction and modeling of clinical outcomes through temporal structure, censored event times, hazards, latent states, disease progression, or patient trajectories.","inclusion":"Assign atoms whose primary objective is modeling longitudinal EHR sequences, ICU trajectories, irregular clinical time series, dynamic disease progression, censored survival outcomes, competing risks, hazards, multistate event processes, or latent progression states.","exclusion":"Use fixed-endpoint clinical risk prediction for non-censored static endpoints; use representation learning under data limitations when pretraining, sparsity, missingness, or schema transfer is the primary scientific objective rather than temporal prognosis itself.","nearest_boundary":"Representation learning under clinical data limitations","representative_topics":["bidirectional lstm readmission risk prediction using longitudinal insurance billing sequences","neural ode-based patient trajectory modeling with clinical code embeddings for early disease prediction","pseudo value-based neural networks for subject-specific multi-state survival prediction under censoring"]},{"direction_index":2,"label":"D2","name":"Multimodal and neurodegenerative fusion","description":"Clinical prediction programs centered on integrating heterogeneous modalities or views, including imaging, structured data, text, genetics, sensors, longitudinal measurements, and neurodegenerative-disease data sources.","inclusion":"Assign atoms where cross-modal fusion, multi-view alignment, missing-modality handling, modality-specific attention, or coordinated use of heterogeneous data is the central objective; include Alzheimer's, MCI, dementia, Parkinson's, and cognitive-decline prediction when the sustained objective is multimodal or specialized neurodegenerative prediction.","exclusion":"Use language and behavioral clinical prediction when unstructured text, dialogue, speech, or behavioral assessment is the dominant data program; use fixed-endpoint prediction when a single modality or static endpoint dominates and fusion is incidental.","nearest_boundary":"Language and behavioral clinical prediction","representative_topics":["multimodal mri glioma survival prediction via cross-scale nonlocal feature fusion","multi-view imputation and cross-attention for mci conversion prediction from incomplete longitudinal multimodal data","cross-modal attention fusion for multimodal alzheimer's disease and mild cognitive impairment diagnosis"]},{"direction_index":3,"label":"D3","name":"Language and behavioral clinical prediction","description":"Prediction, reasoning, phenotyping, and risk assessment from unstructured clinical text, reports, dialogue, transcripts, language-model representations, speech, behavioral observations, or psychiatric and behavioral-health signals.","inclusion":"Assign atoms whose primary objective is clinical NLP, note-based prognosis, text-derived concepts, transcription-based diagnosis, dialogue-based assessment, language-model clinical reasoning, or mental and behavioral-health prediction from speech, activity, physiological, clinical, or multimodal behavioral data.","exclusion":"Use representation learning under clinical data limitations when the main objective is general pretraining or transferable EHR representation rather than text or behavioral clinical assessment; use multimodal fusion when the dominant contribution is general cross-modal fusion outside a language or behavioral-health objective.","nearest_boundary":"Representation learning under clinical data limitations","representative_topics":["nlp-based automated diagnosis from clinical transcription reports using hybrid deep learning models","multimodal depression severity estimation using sub-attentional fusion of heterogeneous modalities","transformer-based clinical note embedding for mortality prediction via umls concept mapping"]},{"direction_index":4,"label":"D4","name":"Representation learning under data limitations","description":"Learning transferable, sparse, weakly supervised, self-supervised, contrastive, semi-supervised, or foundation-style patient representations to improve prediction under limited labels, missingness, sparsity, imbalance, heterogeneous schemas, or rare features.","inclusion":"Assign atoms where the primary objective is pretraining, contrastive learning, weak supervision, rare-feature learning, class-imbalance handling, transfer, schema harmonization, foundation modeling, or representation design for downstream clinical prediction under data limitations.","exclusion":"Use time-dependent prognosis when ordinary longitudinal sequence modeling or survival estimation is primary; use language and behavioral clinical prediction when the core objective is clinical text reasoning or behavioral-health assessment rather than transferable representation learning.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["weakly supervised pseudo-labeling for vital sign alert classification in clinical monitoring","tree-guided feature selection and logic aggregation for rare binary ehr-based risk prediction","class-imbalanced learning for adverse event severity prediction in immunization data"]},{"direction_index":5,"label":"D5","name":"Trustworthy and equitable prediction","description":"Clinical prediction programs whose primary objective is interpretability, explanation, uncertainty, calibration, stability, robustness, external generalization, bias mitigation, causal validity, or equitable performance.","inclusion":"Assign atoms focused on inherently interpretable models, post-hoc explanations, clinical risk-score transparency, uncertainty quantification, calibration, model instability, leakage and validation critique, distribution shift, fairness mitigation, health-disparity modeling, causal representation, target-trial emulation, or treatment-effect validity.","exclusion":"Use the task-specific predictive program when explanation, robustness, fairness, or validation is only a routine evaluation component; use fixed-endpoint prediction when the main contribution is accuracy on a conventional endpoint rather than trustworthiness or equity.","nearest_boundary":"Fixed-endpoint clinical risk prediction","representative_topics":["bias-aware machine learning pipeline for equitable hospital readmission prediction","interpretable clinical risk scores via explicit intervention policy modeling in icu","target trial emulation for individualized antibiotic treatment effect estimation in resistant infection"]},{"direction_index":6,"label":"D6","name":"Physiological signal and disease-management prediction","description":"Prediction programs centered on physiological signals or continuous disease-management data, including cardiac signals and imaging, EEG and neural signals, glucose monitoring, diabetes management, and related sensor-derived clinical states.","inclusion":"Assign atoms whose primary objective is prediction from ECG, PPG, heart-rate variability, echocardiography, cardiac images or shapes, EEG, cortico-cortical evoked potentials, continuous glucose monitors, glycemic trajectories, hypoglycemia or hyperglycemia risk, meal effects, or diabetes-management signals.","exclusion":"Use fixed-endpoint clinical risk prediction for ordinary tabular cardiovascular or diabetes risk models without a signal-management focus; use multimodal fusion when general cross-modal integration rather than the physiological signal program is central.","nearest_boundary":"Fixed-endpoint clinical risk prediction","representative_topics":["hrv-based ventricular and atrial arrhythmia prediction for implantable cardiac devices","binarized 1d convolutional neural networks for hardware-efficient eeg-based seizure prediction","multitask neural network for macronutrient composition estimation from continuous glucose monitor signals"]},{"direction_index":7,"label":"D7","name":"COVID-specific outcome prediction","description":"Predictive modeling for acute COVID-19 severity, hospitalization, ICU transfer, mortality, disease progression, and post-acute or long-COVID sequelae.","inclusion":"Assign atoms where SARS-CoV-2 infection or its aftermath defines the scientific objective, including pediatric COVID severity, COVID hospitalization, COVID ICU outcome benchmarking, severe progression, and long-COVID risk-factor or sequelae prediction.","exclusion":"Use time-dependent prognosis for pandemic-agnostic ICU or EHR trajectory methods merely evaluated on COVID data; use trustworthy and equitable prediction when reliability, fairness, or causal validity dominates over COVID-specific clinical prediction.","nearest_boundary":"Time-dependent prognosis and trajectory modeling","representative_topics":["deep learning prediction of pediatric covid-19 severity using large-scale medical feature representations","knowledge-guided graph neural networks for pediatric covid-19 hospitalization and severity prediction","concept learning for severe covid-19 progression risk prediction from ehr data"]}],"split":"dev"}
{"episode_id":"F266_T2024-01","field_index":266,"field":"constrained reinforcement learning","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Forward CMDP policy optimization","description":"Design, analyze, benchmark, or deploy algorithms that optimize rewards subject to known, fixed constraints in CMDPs or directly equivalent constrained policy-optimization problems.","inclusion":"Assign work whose main objective is solving a forward constrained RL problem with pre-specified constraints, including primal-dual, Lagrangian, projection, trust-region, state-augmentation, feasible-action, value-based, model-free, model-based, average-criterion, applied-system, or benchmark contributions centered on standard constrained policy optimization.","exclusion":"If the objective is maintaining safety throughout learning rather than solving the final constrained optimization problem, use Safe online learning and feasible improvement; if the primary novelty is a specialized risk, robustness, or structural constraint formulation, use Specialized constraint criteria or Structured constrained planning.","nearest_boundary":"Safe online learning and feasible improvement","representative_topics":["dual methods for constrained markov decision processes with linear convergence guarantees","natural policy gradient primal-dual methods for constrained mdps with convergence and sample complexity guarantees","primal-dual natural policy gradient for zero constraint violation in continuous cmdps"]},{"direction_index":1,"label":"D1","name":"Specialized constraint criteria","description":"Formulate and solve constrained RL problems whose central objective is a nonstandard safety or constraint criterion, such as distributional risk, chance constraints, quantiles, CVaR, outage probabilities, resource budgets, robustness, or other specialized safety semantics.","inclusion":"Assign work when the main scientific contribution is the constrained objective or safety criterion itself, and algorithms or guarantees are developed around that criterion, even if standard CMDP optimization machinery is used internally.","exclusion":"If constraints are ordinary fixed expected costs and the novelty is the optimizer, use Forward CMDP policy optimization; if the novelty is hierarchy, recursion, partial observability, semi-infinite structure, or long-horizon constrained planning, use Structured constrained planning.","nearest_boundary":"Forward CMDP policy optimization","representative_topics":["episodic iterated cvar reinforcement learning with regret minimization and worst-path guarantees","risk-sensitive safe reinforcement learning via conditional value-at-risk constraints under joint disturbances","risk-sensitive policy optimization using cdf-based objectives and pessimistic exploration"]},{"direction_index":2,"label":"D2","name":"Safe online learning","description":"Develop RL procedures that preserve feasibility, satisfy safety requirements, or bound constraint violations during data collection, exploration, online learning, policy improvement, or offline-to-online transfer.","inclusion":"Assign work whose primary objective is safety during learning rather than only final-policy feasibility, including safe posterior sampling, safe policy improvement, bounded-violation online learning, instantaneous hard-constraint exploration, demonstration-guided safe online improvement, and transfer methods for safe learning.","exclusion":"If safety is enforced mainly by an explicit execution-time shield, reachability certificate, control barrier function, or action filter, use Safety certification and filtering; if safety is only a convergence or final-solution constraint, use Forward CMDP policy optimization.","nearest_boundary":"Safety certification and filtering","representative_topics":["posterior sampling algorithms for efficient exploration in constrained reinforcement learning","safe policy improvement in constrained markov decision processes with formal requirement compliance","safe posterior sampling for online constrained mdps with bounded constraint violation"]},{"direction_index":3,"label":"D3","name":"Safety certification and filtering","description":"Ensure safe RL behavior by certifying, shielding, projecting, or filtering actions and trajectories using models, dynamics knowledge, reachability analysis, control barrier functions, Lyapunov certificates, safety critics, or learned feasible-action generators.","inclusion":"Assign work where the primary mechanism is a safety layer, model-based safe planner, reachability certificate, barrier or Lyapunov filter, trajectory projection, or action feasibility generator that constrains an RL policy during training or execution.","exclusion":"If the main contribution is an optimizer that internalizes constraints without a distinct filter or certificate, use Forward CMDP policy optimization; if the main issue is maintaining learning-time safety through regret, posterior sampling, or safe improvement guarantees, use Safe online learning.","nearest_boundary":"Safe online learning","representative_topics":["model-based safe reinforcement learning with divergent search planners for constrained control","safe model-based reinforcement learning via uncertainty-aware distributional reachability certificates","model-free safe reinforcement learning via disturbance observer-based control barrier function safety filters"]},{"direction_index":4,"label":"D4","name":"Inverse constraint learning","description":"Infer constraints, rewards, safe sets, or implicit safety specifications from expert demonstrations or related behavioral data for subsequent constrained decision making.","inclusion":"Assign work whose primary objective is recovering constrained objectives or safety restrictions from observed behavior, including inverse constrained RL, maximum-causal-entropy formulations, variational constraint inference, benchmarking, identifiability, generalizability, and reward-constraint recovery theory.","exclusion":"If demonstrations are used only to warm-start, imitate, or guide a forward safe learner with known constraints, use Safe online learning; if constraints are hand-specified through logic, scenarios, or domain knowledge, use Constraint specification and adaptation.","nearest_boundary":"Constraint specification and adaptation","representative_topics":["learning soft constraints from expert demonstrations via iterative constrained optimization in inverse reinforcement learning","benchmarking inverse constrained reinforcement learning with variational constraint inference","sample complexity theory for feasible reward set estimation in inverse reinforcement learning"]},{"direction_index":5,"label":"D5","name":"Constraint specification and adaptation","description":"Represent, encode, monitor, relax, condition on, or adapt safety constraints so RL agents can operate under domain knowledge, logical rules, temporal requirements, changing thresholds, or partially known safety requirements.","inclusion":"Assign work whose main objective is how constraints are specified, structured, varied, monitored, or conditioned on, including scenario programs, logic or temporal-logic constraints, domain-knowledge safe sets, adaptive safety budgets, resilient constraint relaxation, joint constraint-and-policy search, and constraint-conditioned policies.","exclusion":"If the main objective is long-horizon search, hierarchy, belief-state planning, or recursive constrained planning induced by the constraint structure, use Structured constrained planning; if constraints are learned from demonstrations, use Inverse constraint learning.","nearest_boundary":"Structured constrained planning","representative_topics":["scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","logic-based domain knowledge constraints for safe exploration in safety-critical reinforcement learning","reward machine-guided self-paced reinforcement learning for long-horizon planning"]},{"direction_index":6,"label":"D6","name":"Structured constrained planning","description":"Solve constrained sequential-decision problems where hierarchy, recursion, partial observability, belief-state search, Monte Carlo tree search, task decomposition, semi-infinite constraint structure, adversarial structure, or other rich long-horizon structure is central to satisfying constraints.","inclusion":"Assign work whose primary objective is constraint-aware planning or solving structurally rich constrained decision processes, including constrained hierarchical RL, constrained POMDPs, constrained MCTS, recursive constraints, hierarchical belief-state planning, semi-infinite constraints, topological ordering, adversarial robust structure, or multi-objective constrained planning.","exclusion":"If the structure mainly serves as a specification language rather than driving the planning or solution architecture, use Constraint specification and adaptation; if the novelty is a scalar risk or distributional safety measure, use Specialized constraint criteria.","nearest_boundary":"Constraint specification and adaptation","representative_topics":["topological constraint-aware policy gradient for multi-objective sequential decision making","hierarchical reinforcement learning for long-horizon tasks with distributional safety constraints","semi-infinitely constrained markov decision processes via semi-infinitely programming and policy optimization"]},{"direction_index":7,"label":"D7","name":"Safe representation and transfer","description":"Learn state, goal, latent, observation, reachability, or transferable representations that enable constrained or safe decision making in high-dimensional, partially observed, goal-conditioned, language-conditioned, or multi-task settings.","inclusion":"Assign work whose primary objective is representation learning, goal-conditioned generalization, reachability abstraction, latent safety evaluation, meta-learning, contextual generalization, or transfer under safety or constraint requirements.","exclusion":"If the learned representation directly functions as an execution-time shield or formal safety certificate, use Safety certification and filtering; if representation learning is only an implementation detail of a standard constrained optimizer, use Forward CMDP policy optimization.","nearest_boundary":"Safety certification and filtering","representative_topics":["contrastive representation learning as goal-conditioned value function for policy optimization","metric residual network architecture for sample-efficient goal-conditioned reinforcement learning","safe reinforcement learning from pixels via stochastic latent actor-critic safety critic"]}],"split":"dev"}
{"episode_id":"F266_T2024-07","field_index":266,"field":"constrained reinforcement learning","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Forward CMDP policy optimization","description":"Design, analyze, benchmark, or deploy algorithms that optimize rewards subject to known, fixed constraints in CMDPs or directly equivalent constrained policy-optimization problems.","inclusion":"Assign work whose main objective is solving a forward constrained RL problem with pre-specified constraints, including primal-dual, Lagrangian, projection, trust-region, state-augmentation, feasible-action, value-based, model-free, model-based, average-criterion, applied-system, or benchmark contributions centered on standard constrained policy optimization.","exclusion":"If the objective is maintaining safety throughout learning rather than solving the final constrained optimization problem, use Safe online learning and feasible improvement; if the primary novelty is a specialized risk, robustness, or structural constraint formulation, use Specialized constraint criteria or Structured constrained planning.","nearest_boundary":"Safe online learning and feasible improvement","representative_topics":["dual methods for constrained markov decision processes with linear convergence guarantees","natural policy gradient primal-dual methods for constrained mdps with convergence and sample complexity guarantees","primal-dual natural policy gradient for zero constraint violation in continuous cmdps"]},{"direction_index":1,"label":"D1","name":"Specialized constraint criteria","description":"Formulate and solve constrained RL problems whose central objective is a nonstandard safety or constraint criterion, such as distributional risk, chance constraints, quantiles, CVaR, outage probabilities, resource budgets, robustness, or other specialized safety semantics.","inclusion":"Assign work when the main scientific contribution is the constrained objective or safety criterion itself, and algorithms or guarantees are developed around that criterion, even if standard CMDP optimization machinery is used internally.","exclusion":"If constraints are ordinary fixed expected costs and the novelty is the optimizer, use Forward CMDP policy optimization; if the novelty is hierarchy, recursion, partial observability, semi-infinite structure, or long-horizon constrained planning, use Structured constrained planning.","nearest_boundary":"Forward CMDP policy optimization","representative_topics":["episodic iterated cvar reinforcement learning with regret minimization and worst-path guarantees","risk-sensitive safe reinforcement learning via conditional value-at-risk constraints under joint disturbances","risk-sensitive policy optimization using cdf-based objectives and pessimistic exploration"]},{"direction_index":2,"label":"D2","name":"Safe online learning","description":"Develop RL procedures that preserve feasibility, satisfy safety requirements, or bound constraint violations during data collection, exploration, online learning, policy improvement, or offline-to-online transfer.","inclusion":"Assign work whose primary objective is safety during learning rather than only final-policy feasibility, including safe posterior sampling, safe policy improvement, bounded-violation online learning, instantaneous hard-constraint exploration, demonstration-guided safe online improvement, and transfer methods for safe learning.","exclusion":"If safety is enforced mainly by an explicit execution-time shield, reachability certificate, control barrier function, or action filter, use Safety certification and filtering; if safety is only a convergence or final-solution constraint, use Forward CMDP policy optimization.","nearest_boundary":"Safety certification and filtering","representative_topics":["posterior sampling algorithms for efficient exploration in constrained reinforcement learning","safe policy improvement in constrained markov decision processes with formal requirement compliance","safe posterior sampling for online constrained mdps with bounded constraint violation"]},{"direction_index":3,"label":"D3","name":"Safety certification and filtering","description":"Ensure safe RL behavior by certifying, shielding, projecting, or filtering actions and trajectories using models, dynamics knowledge, reachability analysis, control barrier functions, Lyapunov certificates, safety critics, or learned feasible-action generators.","inclusion":"Assign work where the primary mechanism is a safety layer, model-based safe planner, reachability certificate, barrier or Lyapunov filter, trajectory projection, or action feasibility generator that constrains an RL policy during training or execution.","exclusion":"If the main contribution is an optimizer that internalizes constraints without a distinct filter or certificate, use Forward CMDP policy optimization; if the main issue is maintaining learning-time safety through regret, posterior sampling, or safe improvement guarantees, use Safe online learning.","nearest_boundary":"Safe online learning","representative_topics":["model-based safe reinforcement learning with divergent search planners for constrained control","safe model-based reinforcement learning via uncertainty-aware distributional reachability certificates","model-free safe reinforcement learning via disturbance observer-based control barrier function safety filters"]},{"direction_index":4,"label":"D4","name":"Inverse constraint learning","description":"Infer constraints, rewards, safe sets, or implicit safety specifications from expert demonstrations or related behavioral data for subsequent constrained decision making.","inclusion":"Assign work whose primary objective is recovering constrained objectives or safety restrictions from observed behavior, including inverse constrained RL, maximum-causal-entropy formulations, variational constraint inference, benchmarking, identifiability, generalizability, and reward-constraint recovery theory.","exclusion":"If demonstrations are used only to warm-start, imitate, or guide a forward safe learner with known constraints, use Safe online learning; if constraints are hand-specified through logic, scenarios, or domain knowledge, use Constraint specification and adaptation.","nearest_boundary":"Constraint specification and adaptation","representative_topics":["learning soft constraints from expert demonstrations via iterative constrained optimization in inverse reinforcement learning","benchmarking inverse constrained reinforcement learning with variational constraint inference","sample complexity theory for feasible reward set estimation in inverse reinforcement learning"]},{"direction_index":5,"label":"D5","name":"Constraint specification and adaptation","description":"Represent, encode, monitor, relax, condition on, or adapt safety constraints so RL agents can operate under domain knowledge, logical rules, temporal requirements, changing thresholds, or partially known safety requirements.","inclusion":"Assign work whose main objective is how constraints are specified, structured, varied, monitored, or conditioned on, including scenario programs, logic or temporal-logic constraints, domain-knowledge safe sets, adaptive safety budgets, resilient constraint relaxation, joint constraint-and-policy search, and constraint-conditioned policies.","exclusion":"If the main objective is long-horizon search, hierarchy, belief-state planning, or recursive constrained planning induced by the constraint structure, use Structured constrained planning; if constraints are learned from demonstrations, use Inverse constraint learning.","nearest_boundary":"Structured constrained planning","representative_topics":["scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","logic-based domain knowledge constraints for safe exploration in safety-critical reinforcement learning","reward machine-guided self-paced reinforcement learning for long-horizon planning"]},{"direction_index":6,"label":"D6","name":"Structured constrained planning","description":"Solve constrained sequential-decision problems where hierarchy, recursion, partial observability, belief-state search, Monte Carlo tree search, task decomposition, semi-infinite constraint structure, adversarial structure, or other rich long-horizon structure is central to satisfying constraints.","inclusion":"Assign work whose primary objective is constraint-aware planning or solving structurally rich constrained decision processes, including constrained hierarchical RL, constrained POMDPs, constrained MCTS, recursive constraints, hierarchical belief-state planning, semi-infinite constraints, topological ordering, adversarial robust structure, or multi-objective constrained planning.","exclusion":"If the structure mainly serves as a specification language rather than driving the planning or solution architecture, use Constraint specification and adaptation; if the novelty is a scalar risk or distributional safety measure, use Specialized constraint criteria.","nearest_boundary":"Constraint specification and adaptation","representative_topics":["topological constraint-aware policy gradient for multi-objective sequential decision making","hierarchical reinforcement learning for long-horizon tasks with distributional safety constraints","semi-infinitely constrained markov decision processes via semi-infinitely programming and policy optimization"]},{"direction_index":7,"label":"D7","name":"Safe representation and transfer","description":"Learn state, goal, latent, observation, reachability, or transferable representations that enable constrained or safe decision making in high-dimensional, partially observed, goal-conditioned, language-conditioned, or multi-task settings.","inclusion":"Assign work whose primary objective is representation learning, goal-conditioned generalization, reachability abstraction, latent safety evaluation, meta-learning, contextual generalization, or transfer under safety or constraint requirements.","exclusion":"If the learned representation directly functions as an execution-time shield or formal safety certificate, use Safety certification and filtering; if representation learning is only an implementation detail of a standard constrained optimizer, use Forward CMDP policy optimization.","nearest_boundary":"Safety certification and filtering","representative_topics":["contrastive representation learning as goal-conditioned value function for policy optimization","metric residual network architecture for sample-efficient goal-conditioned reinforcement learning","safe reinforcement learning from pixels via stochastic latent actor-critic safety critic"]}],"split":"dev"}
{"episode_id":"F266_T2025-01","field_index":266,"field":"constrained reinforcement learning","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Forward CMDP policy optimization","description":"Design, analyze, benchmark, or deploy algorithms that optimize rewards subject to known, fixed constraints in CMDPs or directly equivalent constrained policy-optimization problems.","inclusion":"Assign work whose main objective is solving a forward constrained RL problem with pre-specified constraints, including primal-dual, Lagrangian, projection, trust-region, state-augmentation, feasible-action, value-based, model-free, model-based, average-criterion, applied-system, or benchmark contributions centered on standard constrained policy optimization.","exclusion":"If the objective is maintaining safety throughout learning rather than solving the final constrained optimization problem, use Safe online learning and feasible improvement; if the primary novelty is a specialized risk, robustness, or structural constraint formulation, use Specialized constraint criteria or Structured constrained planning.","nearest_boundary":"Safe online learning and feasible improvement","representative_topics":["dual methods for constrained markov decision processes with linear convergence guarantees","natural policy gradient primal-dual methods for constrained mdps with convergence and sample complexity guarantees","primal-dual natural policy gradient for zero constraint violation in continuous cmdps"]},{"direction_index":1,"label":"D1","name":"Specialized constraint criteria","description":"Formulate and solve constrained RL problems whose central objective is a nonstandard safety or constraint criterion, such as distributional risk, chance constraints, quantiles, CVaR, outage probabilities, resource budgets, robustness, or other specialized safety semantics.","inclusion":"Assign work when the main scientific contribution is the constrained objective or safety criterion itself, and algorithms or guarantees are developed around that criterion, even if standard CMDP optimization machinery is used internally.","exclusion":"If constraints are ordinary fixed expected costs and the novelty is the optimizer, use Forward CMDP policy optimization; if the novelty is hierarchy, recursion, partial observability, semi-infinite structure, or long-horizon constrained planning, use Structured constrained planning.","nearest_boundary":"Forward CMDP policy optimization","representative_topics":["episodic iterated cvar reinforcement learning with regret minimization and worst-path guarantees","risk-sensitive safe reinforcement learning via conditional value-at-risk constraints under joint disturbances","risk-sensitive policy optimization using cdf-based objectives and pessimistic exploration"]},{"direction_index":2,"label":"D2","name":"Safe online learning","description":"Develop RL procedures that preserve feasibility, satisfy safety requirements, or bound constraint violations during data collection, exploration, online learning, policy improvement, or offline-to-online transfer.","inclusion":"Assign work whose primary objective is safety during learning rather than only final-policy feasibility, including safe posterior sampling, safe policy improvement, bounded-violation online learning, instantaneous hard-constraint exploration, demonstration-guided safe online improvement, and transfer methods for safe learning.","exclusion":"If safety is enforced mainly by an explicit execution-time shield, reachability certificate, control barrier function, or action filter, use Safety certification and filtering; if safety is only a convergence or final-solution constraint, use Forward CMDP policy optimization.","nearest_boundary":"Safety certification and filtering","representative_topics":["posterior sampling algorithms for efficient exploration in constrained reinforcement learning","safe policy improvement in constrained markov decision processes with formal requirement compliance","safe posterior sampling for online constrained mdps with bounded constraint violation"]},{"direction_index":3,"label":"D3","name":"Safety certification and filtering","description":"Ensure safe RL behavior by certifying, shielding, projecting, or filtering actions and trajectories using models, dynamics knowledge, reachability analysis, control barrier functions, Lyapunov certificates, safety critics, or learned feasible-action generators.","inclusion":"Assign work where the primary mechanism is a safety layer, model-based safe planner, reachability certificate, barrier or Lyapunov filter, trajectory projection, or action feasibility generator that constrains an RL policy during training or execution.","exclusion":"If the main contribution is an optimizer that internalizes constraints without a distinct filter or certificate, use Forward CMDP policy optimization; if the main issue is maintaining learning-time safety through regret, posterior sampling, or safe improvement guarantees, use Safe online learning.","nearest_boundary":"Safe online learning","representative_topics":["model-based safe reinforcement learning with divergent search planners for constrained control","safe model-based reinforcement learning via uncertainty-aware distributional reachability certificates","model-free safe reinforcement learning via disturbance observer-based control barrier function safety filters"]},{"direction_index":4,"label":"D4","name":"Inverse constraint learning","description":"Infer constraints, rewards, safe sets, or implicit safety specifications from expert demonstrations or related behavioral data for subsequent constrained decision making.","inclusion":"Assign work whose primary objective is recovering constrained objectives or safety restrictions from observed behavior, including inverse constrained RL, maximum-causal-entropy formulations, variational constraint inference, benchmarking, identifiability, generalizability, and reward-constraint recovery theory.","exclusion":"If demonstrations are used only to warm-start, imitate, or guide a forward safe learner with known constraints, use Safe online learning; if constraints are hand-specified through logic, scenarios, or domain knowledge, use Constraint specification and adaptation.","nearest_boundary":"Constraint specification and adaptation","representative_topics":["learning soft constraints from expert demonstrations via iterative constrained optimization in inverse reinforcement learning","benchmarking inverse constrained reinforcement learning with variational constraint inference","sample complexity theory for feasible reward set estimation in inverse reinforcement learning"]},{"direction_index":5,"label":"D5","name":"Constraint specification and adaptation","description":"Represent, encode, monitor, relax, condition on, or adapt safety constraints so RL agents can operate under domain knowledge, logical rules, temporal requirements, changing thresholds, or partially known safety requirements.","inclusion":"Assign work whose main objective is how constraints are specified, structured, varied, monitored, or conditioned on, including scenario programs, logic or temporal-logic constraints, domain-knowledge safe sets, adaptive safety budgets, resilient constraint relaxation, joint constraint-and-policy search, and constraint-conditioned policies.","exclusion":"If the main objective is long-horizon search, hierarchy, belief-state planning, or recursive constrained planning induced by the constraint structure, use Structured constrained planning; if constraints are learned from demonstrations, use Inverse constraint learning.","nearest_boundary":"Structured constrained planning","representative_topics":["scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","logic-based domain knowledge constraints for safe exploration in safety-critical reinforcement learning","reward machine-guided self-paced reinforcement learning for long-horizon planning"]},{"direction_index":6,"label":"D6","name":"Structured constrained planning","description":"Solve constrained sequential-decision problems where hierarchy, recursion, partial observability, belief-state search, Monte Carlo tree search, task decomposition, semi-infinite constraint structure, adversarial structure, or other rich long-horizon structure is central to satisfying constraints.","inclusion":"Assign work whose primary objective is constraint-aware planning or solving structurally rich constrained decision processes, including constrained hierarchical RL, constrained POMDPs, constrained MCTS, recursive constraints, hierarchical belief-state planning, semi-infinite constraints, topological ordering, adversarial robust structure, or multi-objective constrained planning.","exclusion":"If the structure mainly serves as a specification language rather than driving the planning or solution architecture, use Constraint specification and adaptation; if the novelty is a scalar risk or distributional safety measure, use Specialized constraint criteria.","nearest_boundary":"Constraint specification and adaptation","representative_topics":["topological constraint-aware policy gradient for multi-objective sequential decision making","hierarchical reinforcement learning for long-horizon tasks with distributional safety constraints","semi-infinitely constrained markov decision processes via semi-infinitely programming and policy optimization"]},{"direction_index":7,"label":"D7","name":"Safe representation and transfer","description":"Learn state, goal, latent, observation, reachability, or transferable representations that enable constrained or safe decision making in high-dimensional, partially observed, goal-conditioned, language-conditioned, or multi-task settings.","inclusion":"Assign work whose primary objective is representation learning, goal-conditioned generalization, reachability abstraction, latent safety evaluation, meta-learning, contextual generalization, or transfer under safety or constraint requirements.","exclusion":"If the learned representation directly functions as an execution-time shield or formal safety certificate, use Safety certification and filtering; if representation learning is only an implementation detail of a standard constrained optimizer, use Forward CMDP policy optimization.","nearest_boundary":"Safety certification and filtering","representative_topics":["contrastive representation learning as goal-conditioned value function for policy optimization","metric residual network architecture for sample-efficient goal-conditioned reinforcement learning","safe reinforcement learning from pixels via stochastic latent actor-critic safety critic"]}],"split":"dev"}
{"episode_id":"F266_T2025-07","field_index":266,"field":"constrained reinforcement learning","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Forward CMDP policy optimization","description":"Design, analyze, benchmark, or deploy algorithms that optimize rewards subject to known, fixed constraints in CMDPs or directly equivalent constrained policy-optimization problems.","inclusion":"Assign work whose main objective is solving a forward constrained RL problem with pre-specified constraints, including primal-dual, Lagrangian, projection, trust-region, state-augmentation, feasible-action, value-based, model-free, model-based, average-criterion, applied-system, or benchmark contributions centered on standard constrained policy optimization.","exclusion":"If the objective is maintaining safety throughout learning rather than solving the final constrained optimization problem, use Safe online learning and feasible improvement; if the primary novelty is a specialized risk, robustness, or structural constraint formulation, use Specialized constraint criteria or Structured constrained planning.","nearest_boundary":"Safe online learning and feasible improvement","representative_topics":["dual methods for constrained markov decision processes with linear convergence guarantees","natural policy gradient primal-dual methods for constrained mdps with convergence and sample complexity guarantees","primal-dual natural policy gradient for zero constraint violation in continuous cmdps"]},{"direction_index":1,"label":"D1","name":"Specialized constraint criteria","description":"Formulate and solve constrained RL problems whose central objective is a nonstandard safety or constraint criterion, such as distributional risk, chance constraints, quantiles, CVaR, outage probabilities, resource budgets, robustness, or other specialized safety semantics.","inclusion":"Assign work when the main scientific contribution is the constrained objective or safety criterion itself, and algorithms or guarantees are developed around that criterion, even if standard CMDP optimization machinery is used internally.","exclusion":"If constraints are ordinary fixed expected costs and the novelty is the optimizer, use Forward CMDP policy optimization; if the novelty is hierarchy, recursion, partial observability, semi-infinite structure, or long-horizon constrained planning, use Structured constrained planning.","nearest_boundary":"Forward CMDP policy optimization","representative_topics":["episodic iterated cvar reinforcement learning with regret minimization and worst-path guarantees","risk-sensitive safe reinforcement learning via conditional value-at-risk constraints under joint disturbances","risk-sensitive policy optimization using cdf-based objectives and pessimistic exploration"]},{"direction_index":2,"label":"D2","name":"Safe online learning","description":"Develop RL procedures that preserve feasibility, satisfy safety requirements, or bound constraint violations during data collection, exploration, online learning, policy improvement, or offline-to-online transfer.","inclusion":"Assign work whose primary objective is safety during learning rather than only final-policy feasibility, including safe posterior sampling, safe policy improvement, bounded-violation online learning, instantaneous hard-constraint exploration, demonstration-guided safe online improvement, and transfer methods for safe learning.","exclusion":"If safety is enforced mainly by an explicit execution-time shield, reachability certificate, control barrier function, or action filter, use Safety certification and filtering; if safety is only a convergence or final-solution constraint, use Forward CMDP policy optimization.","nearest_boundary":"Safety certification and filtering","representative_topics":["posterior sampling algorithms for efficient exploration in constrained reinforcement learning","safe policy improvement in constrained markov decision processes with formal requirement compliance","safe posterior sampling for online constrained mdps with bounded constraint violation"]},{"direction_index":3,"label":"D3","name":"Safety certification and filtering","description":"Ensure safe RL behavior by certifying, shielding, projecting, or filtering actions and trajectories using models, dynamics knowledge, reachability analysis, control barrier functions, Lyapunov certificates, safety critics, or learned feasible-action generators.","inclusion":"Assign work where the primary mechanism is a safety layer, model-based safe planner, reachability certificate, barrier or Lyapunov filter, trajectory projection, or action feasibility generator that constrains an RL policy during training or execution.","exclusion":"If the main contribution is an optimizer that internalizes constraints without a distinct filter or certificate, use Forward CMDP policy optimization; if the main issue is maintaining learning-time safety through regret, posterior sampling, or safe improvement guarantees, use Safe online learning.","nearest_boundary":"Safe online learning","representative_topics":["model-based safe reinforcement learning with divergent search planners for constrained control","safe model-based reinforcement learning via uncertainty-aware distributional reachability certificates","model-free safe reinforcement learning via disturbance observer-based control barrier function safety filters"]},{"direction_index":4,"label":"D4","name":"Inverse constraint learning","description":"Infer constraints, rewards, safe sets, or implicit safety specifications from expert demonstrations or related behavioral data for subsequent constrained decision making.","inclusion":"Assign work whose primary objective is recovering constrained objectives or safety restrictions from observed behavior, including inverse constrained RL, maximum-causal-entropy formulations, variational constraint inference, benchmarking, identifiability, generalizability, and reward-constraint recovery theory.","exclusion":"If demonstrations are used only to warm-start, imitate, or guide a forward safe learner with known constraints, use Safe online learning; if constraints are hand-specified through logic, scenarios, or domain knowledge, use Constraint specification and adaptation.","nearest_boundary":"Constraint specification and adaptation","representative_topics":["learning soft constraints from expert demonstrations via iterative constrained optimization in inverse reinforcement learning","benchmarking inverse constrained reinforcement learning with variational constraint inference","sample complexity theory for feasible reward set estimation in inverse reinforcement learning"]},{"direction_index":5,"label":"D5","name":"Constraint specification and adaptation","description":"Represent, encode, monitor, relax, condition on, or adapt safety constraints so RL agents can operate under domain knowledge, logical rules, temporal requirements, changing thresholds, or partially known safety requirements.","inclusion":"Assign work whose main objective is how constraints are specified, structured, varied, monitored, or conditioned on, including scenario programs, logic or temporal-logic constraints, domain-knowledge safe sets, adaptive safety budgets, resilient constraint relaxation, joint constraint-and-policy search, and constraint-conditioned policies.","exclusion":"If the main objective is long-horizon search, hierarchy, belief-state planning, or recursive constrained planning induced by the constraint structure, use Structured constrained planning; if constraints are learned from demonstrations, use Inverse constraint learning.","nearest_boundary":"Structured constrained planning","representative_topics":["scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","logic-based domain knowledge constraints for safe exploration in safety-critical reinforcement learning","reward machine-guided self-paced reinforcement learning for long-horizon planning"]},{"direction_index":6,"label":"D6","name":"Structured constrained planning","description":"Solve constrained sequential-decision problems where hierarchy, recursion, partial observability, belief-state search, Monte Carlo tree search, task decomposition, semi-infinite constraint structure, adversarial structure, or other rich long-horizon structure is central to satisfying constraints.","inclusion":"Assign work whose primary objective is constraint-aware planning or solving structurally rich constrained decision processes, including constrained hierarchical RL, constrained POMDPs, constrained MCTS, recursive constraints, hierarchical belief-state planning, semi-infinite constraints, topological ordering, adversarial robust structure, or multi-objective constrained planning.","exclusion":"If the structure mainly serves as a specification language rather than driving the planning or solution architecture, use Constraint specification and adaptation; if the novelty is a scalar risk or distributional safety measure, use Specialized constraint criteria.","nearest_boundary":"Constraint specification and adaptation","representative_topics":["topological constraint-aware policy gradient for multi-objective sequential decision making","hierarchical reinforcement learning for long-horizon tasks with distributional safety constraints","semi-infinitely constrained markov decision processes via semi-infinitely programming and policy optimization"]},{"direction_index":7,"label":"D7","name":"Safe representation and transfer","description":"Learn state, goal, latent, observation, reachability, or transferable representations that enable constrained or safe decision making in high-dimensional, partially observed, goal-conditioned, language-conditioned, or multi-task settings.","inclusion":"Assign work whose primary objective is representation learning, goal-conditioned generalization, reachability abstraction, latent safety evaluation, meta-learning, contextual generalization, or transfer under safety or constraint requirements.","exclusion":"If the learned representation directly functions as an execution-time shield or formal safety certificate, use Safety certification and filtering; if representation learning is only an implementation detail of a standard constrained optimizer, use Forward CMDP policy optimization.","nearest_boundary":"Safety certification and filtering","representative_topics":["contrastive representation learning as goal-conditioned value function for policy optimization","metric residual network architecture for sample-efficient goal-conditioned reinforcement learning","safe reinforcement learning from pixels via stochastic latent actor-critic safety critic"]}],"split":"dev"}
{"episode_id":"F266_T2026-01","field_index":266,"field":"constrained reinforcement learning","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Forward CMDP policy optimization","description":"Design, analyze, benchmark, or deploy algorithms that optimize rewards subject to known, fixed constraints in CMDPs or directly equivalent constrained policy-optimization problems.","inclusion":"Assign work whose main objective is solving a forward constrained RL problem with pre-specified constraints, including primal-dual, Lagrangian, projection, trust-region, state-augmentation, feasible-action, value-based, model-free, model-based, average-criterion, applied-system, or benchmark contributions centered on standard constrained policy optimization.","exclusion":"If the objective is maintaining safety throughout learning rather than solving the final constrained optimization problem, use Safe online learning and feasible improvement; if the primary novelty is a specialized risk, robustness, or structural constraint formulation, use Specialized constraint criteria or Structured constrained planning.","nearest_boundary":"Safe online learning and feasible improvement","representative_topics":["dual methods for constrained markov decision processes with linear convergence guarantees","natural policy gradient primal-dual methods for constrained mdps with convergence and sample complexity guarantees","primal-dual natural policy gradient for zero constraint violation in continuous cmdps"]},{"direction_index":1,"label":"D1","name":"Specialized constraint criteria","description":"Formulate and solve constrained RL problems whose central objective is a nonstandard safety or constraint criterion, such as distributional risk, chance constraints, quantiles, CVaR, outage probabilities, resource budgets, robustness, or other specialized safety semantics.","inclusion":"Assign work when the main scientific contribution is the constrained objective or safety criterion itself, and algorithms or guarantees are developed around that criterion, even if standard CMDP optimization machinery is used internally.","exclusion":"If constraints are ordinary fixed expected costs and the novelty is the optimizer, use Forward CMDP policy optimization; if the novelty is hierarchy, recursion, partial observability, semi-infinite structure, or long-horizon constrained planning, use Structured constrained planning.","nearest_boundary":"Forward CMDP policy optimization","representative_topics":["episodic iterated cvar reinforcement learning with regret minimization and worst-path guarantees","risk-sensitive safe reinforcement learning via conditional value-at-risk constraints under joint disturbances","risk-sensitive policy optimization using cdf-based objectives and pessimistic exploration"]},{"direction_index":2,"label":"D2","name":"Safe online learning","description":"Develop RL procedures that preserve feasibility, satisfy safety requirements, or bound constraint violations during data collection, exploration, online learning, policy improvement, or offline-to-online transfer.","inclusion":"Assign work whose primary objective is safety during learning rather than only final-policy feasibility, including safe posterior sampling, safe policy improvement, bounded-violation online learning, instantaneous hard-constraint exploration, demonstration-guided safe online improvement, and transfer methods for safe learning.","exclusion":"If safety is enforced mainly by an explicit execution-time shield, reachability certificate, control barrier function, or action filter, use Safety certification and filtering; if safety is only a convergence or final-solution constraint, use Forward CMDP policy optimization.","nearest_boundary":"Safety certification and filtering","representative_topics":["posterior sampling algorithms for efficient exploration in constrained reinforcement learning","safe policy improvement in constrained markov decision processes with formal requirement compliance","safe posterior sampling for online constrained mdps with bounded constraint violation"]},{"direction_index":3,"label":"D3","name":"Safety certification and filtering","description":"Ensure safe RL behavior by certifying, shielding, projecting, or filtering actions and trajectories using models, dynamics knowledge, reachability analysis, control barrier functions, Lyapunov certificates, safety critics, or learned feasible-action generators.","inclusion":"Assign work where the primary mechanism is a safety layer, model-based safe planner, reachability certificate, barrier or Lyapunov filter, trajectory projection, or action feasibility generator that constrains an RL policy during training or execution.","exclusion":"If the main contribution is an optimizer that internalizes constraints without a distinct filter or certificate, use Forward CMDP policy optimization; if the main issue is maintaining learning-time safety through regret, posterior sampling, or safe improvement guarantees, use Safe online learning.","nearest_boundary":"Safe online learning","representative_topics":["model-based safe reinforcement learning with divergent search planners for constrained control","safe model-based reinforcement learning via uncertainty-aware distributional reachability certificates","model-free safe reinforcement learning via disturbance observer-based control barrier function safety filters"]},{"direction_index":4,"label":"D4","name":"Inverse constraint learning","description":"Infer constraints, rewards, safe sets, or implicit safety specifications from expert demonstrations or related behavioral data for subsequent constrained decision making.","inclusion":"Assign work whose primary objective is recovering constrained objectives or safety restrictions from observed behavior, including inverse constrained RL, maximum-causal-entropy formulations, variational constraint inference, benchmarking, identifiability, generalizability, and reward-constraint recovery theory.","exclusion":"If demonstrations are used only to warm-start, imitate, or guide a forward safe learner with known constraints, use Safe online learning; if constraints are hand-specified through logic, scenarios, or domain knowledge, use Constraint specification and adaptation.","nearest_boundary":"Constraint specification and adaptation","representative_topics":["learning soft constraints from expert demonstrations via iterative constrained optimization in inverse reinforcement learning","benchmarking inverse constrained reinforcement learning with variational constraint inference","sample complexity theory for feasible reward set estimation in inverse reinforcement learning"]},{"direction_index":5,"label":"D5","name":"Constraint specification and adaptation","description":"Represent, encode, monitor, relax, condition on, or adapt safety constraints so RL agents can operate under domain knowledge, logical rules, temporal requirements, changing thresholds, or partially known safety requirements.","inclusion":"Assign work whose main objective is how constraints are specified, structured, varied, monitored, or conditioned on, including scenario programs, logic or temporal-logic constraints, domain-knowledge safe sets, adaptive safety budgets, resilient constraint relaxation, joint constraint-and-policy search, and constraint-conditioned policies.","exclusion":"If the main objective is long-horizon search, hierarchy, belief-state planning, or recursive constrained planning induced by the constraint structure, use Structured constrained planning; if constraints are learned from demonstrations, use Inverse constraint learning.","nearest_boundary":"Structured constrained planning","representative_topics":["scenario-based expert knowledge integration for constrained deep reinforcement learning in safe robotics","logic-based domain knowledge constraints for safe exploration in safety-critical reinforcement learning","reward machine-guided self-paced reinforcement learning for long-horizon planning"]},{"direction_index":6,"label":"D6","name":"Structured constrained planning","description":"Solve constrained sequential-decision problems where hierarchy, recursion, partial observability, belief-state search, Monte Carlo tree search, task decomposition, semi-infinite constraint structure, adversarial structure, or other rich long-horizon structure is central to satisfying constraints.","inclusion":"Assign work whose primary objective is constraint-aware planning or solving structurally rich constrained decision processes, including constrained hierarchical RL, constrained POMDPs, constrained MCTS, recursive constraints, hierarchical belief-state planning, semi-infinite constraints, topological ordering, adversarial robust structure, or multi-objective constrained planning.","exclusion":"If the structure mainly serves as a specification language rather than driving the planning or solution architecture, use Constraint specification and adaptation; if the novelty is a scalar risk or distributional safety measure, use Specialized constraint criteria.","nearest_boundary":"Constraint specification and adaptation","representative_topics":["topological constraint-aware policy gradient for multi-objective sequential decision making","hierarchical reinforcement learning for long-horizon tasks with distributional safety constraints","semi-infinitely constrained markov decision processes via semi-infinitely programming and policy optimization"]},{"direction_index":7,"label":"D7","name":"Safe representation and transfer","description":"Learn state, goal, latent, observation, reachability, or transferable representations that enable constrained or safe decision making in high-dimensional, partially observed, goal-conditioned, language-conditioned, or multi-task settings.","inclusion":"Assign work whose primary objective is representation learning, goal-conditioned generalization, reachability abstraction, latent safety evaluation, meta-learning, contextual generalization, or transfer under safety or constraint requirements.","exclusion":"If the learned representation directly functions as an execution-time shield or formal safety certificate, use Safety certification and filtering; if representation learning is only an implementation detail of a standard constrained optimizer, use Forward CMDP policy optimization.","nearest_boundary":"Safety certification and filtering","representative_topics":["contrastive representation learning as goal-conditioned value function for policy optimization","metric residual network architecture for sample-efficient goal-conditioned reinforcement learning","safe reinforcement learning from pixels via stochastic latent actor-critic safety critic"]}],"split":"dev"}
{"episode_id":"F271_T2024-01","field_index":271,"field":"robot learning and control","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Object manipulation and interaction","description":"Learning representations, policies, dynamics, and controllers for robots to grasp, contact, deform, articulate, or otherwise manipulate objects and materials.","inclusion":"Include work whose primary objective is object-level manipulation capability or manipulation-specific perception, including rigid and dexterous grasping, in-hand manipulation, affordances, point-cloud/object-centric manipulation representations, deformable-object manipulation, articulated objects, soft bodies, fluids, and soft-robot modeling/control when these mechanics define the task.","exclusion":"If the main objective is sequencing reusable skills across long horizons rather than manipulating a particular object/material class, use D1; if the main objective is language grounding, imitation, or data pretraining rather than manipulation mechanics, use D2.","nearest_boundary":"D1","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","implicit representation learning for cross-gripper grasp transfer via signed distance functions","deep reinforcement learning for robotic grasping using 3d octree representations in lunar simulation"]},{"direction_index":1,"label":"D1","name":"Skill abstraction and motion composition","description":"Learning reusable motor-skill, trajectory, primitive, full-body motion, and temporal-abstraction structures for composing or generalizing robot behavior over time.","inclusion":"Include work whose primary objective is representing, discovering, chaining, composing, or adapting skills or motions, including hierarchical RL, long-horizon task decomposition, recovery skills, movement primitives, dynamical-system skill representations, residual skill spaces, and high-DOF humanoid or character motion synthesis when motion organization is central.","exclusion":"If the main objective is object/material manipulation itself, use D0; if the main objective is language- or foundation-model task reasoning, use D2; if formal safety constraints dominate the hierarchy or primitive, use D3.","nearest_boundary":"D0","representative_topics":["skill-space model-based reinforcement learning for long-horizon sparse reward planning","neural dynamic movement primitives for robot trajectory learning and control","hierarchical reinforcement learning for simultaneous stroke rendering and motor control in robotic sketching"]},{"direction_index":2,"label":"D2","name":"Supervised, data-centric, and language-grounded policy learning","description":"Learning robot policies, objectives, or semantic control interfaces from demonstrations, fixed datasets, human feedback, reward/preference signals, language, or foundation-model representations.","inclusion":"Include work whose primary objective is external supervision or prior knowledge for policy acquisition: offline RL and self-supervised pretraining from accumulated data, imitation from demonstrations or videos, human-feedback or reward learning, language-conditioned control, VLM/LLM-based planning, and foundation-model transfer when semantic grounding or data-driven policy learning is primary.","exclusion":"If the primary objective is the physical manipulation capability rather than the supervision source, use D0; if the objective is reusable skill sequencing or trajectory primitives, use D1; if the objective is cross-domain transfer itself rather than data pretraining, use D5.","nearest_boundary":"D0","representative_topics":["latent policy learning for stable adversarial imitation in high-dimensional robot tasks","language-conditioned meta-learning for accelerated tool affordance-based manipulation policy adaptation","language-guided robotic navigation via pre-trained model composition without annotated trajectory supervision"]},{"direction_index":3,"label":"D3","name":"Predictive, safe, and guaranteed control","description":"Developing learned model-based, predictive, constrained, stable, or formally guaranteed controllers for robot systems.","inclusion":"Include work whose primary objective is learning dynamics, estimators, residual models, MPC components, differentiable optimal-control modules, or controllers with explicit safety, stability, temporal-logic, barrier, Lyapunov, reachability, or verification guarantees.","exclusion":"If robustness is empirical domain adaptation without explicit predictive-control or guarantee structure, use D5; if the main objective is generic RL optimization rather than learned dynamics, MPC, or formal constraints, use D7.","nearest_boundary":"D5","representative_topics":["adaptive model-based advice for stabilizing black-box policies in nonlinear control","physics-constrained temporal convolutional learning of quadrotor dynamics for model predictive control","ltl-constrained motion policy learning from demonstrations with mode-based plan satisfaction guarantees"]},{"direction_index":4,"label":"D4","name":"Locomotion, navigation, and vehicle autonomy","description":"Learning control and decision policies for robots whose primary objective is moving through environments, including legged locomotion, mobile navigation, and vehicle-like robotic platforms.","inclusion":"Include work whose primary objective is gait learning, terrain-adaptive locomotion, map-free or social navigation, path planning, traversability, cleaning/coverage navigation, autonomous driving, aerial/underwater/space vehicle control, or platform-dynamics-centered movement through space.","exclusion":"If inter-agent coordination is the central scientific objective, use D6; if learned dynamics/MPC methodology is primary independent of platform class, use D3; if sim-to-real transfer is the main objective rather than locomotion/navigation capability, use D5.","nearest_boundary":"D6","representative_topics":["map-less pointgoal navigation using visual odometry under realistic noisy conditions","inverse reinforcement learning for autonomous driving trajectory planning in dense urban traffic","deep reinforcement learning for joint acceleration and steering control in autonomous driving"]},{"direction_index":5,"label":"D5","name":"Transfer and domain adaptation","description":"Bridging differences between simulation and reality, source and target tasks, environments, embodiments, users, hardware, or operating conditions for deployed robot learning.","inclusion":"Include work whose primary objective is zero-shot or few-shot transfer, sim-to-real or real-to-sim adaptation, domain randomization, residual error estimation, teacher-student transfer, policy fusion, cross-task reuse, disturbance adaptation, or representation learning specifically to handle domain shift.","exclusion":"If sim-to-real is only an implementation detail of a locomotion, manipulation, or navigation method, use that capability direction; if the transfer mechanism is broad offline pretraining from robot data, use D2; if explicit verification or hard safety guarantees dominate, use D3.","nearest_boundary":"D2","representative_topics":["learning-based variable impedance control for aerial sliding on uneven heterogeneous surfaces","optimal transport-based policy fusion for cross-task transfer in deep reinforcement learning robotics","zero-shot sim-to-real visual mpc via randomized-to-canonical model learning for robotic manipulation"]},{"direction_index":6,"label":"D6","name":"Multi-robot and swarm coordination","description":"Learning policies, games, communication, or latent coordination structures for teams, fleets, swarms, or multi-robot systems.","inclusion":"Include work whose primary objective is inter-agent coordination, cooperation, formation, coverage, task allocation, decentralized/centralized multi-agent learning, swarm control, communication-constrained collaboration, or multi-robot manipulation.","exclusion":"If multiple robots are only data sources or deployment infrastructure without a coordination problem, use the relevant single-robot objective; if the main problem is vehicle-following dynamics or single-platform navigation, use D4.","nearest_boundary":"D4","representative_topics":["interactive fleet learning with scalable human supervision allocation and continual policy improvement","distributed differentiable dynamic game framework for multi-robot coordination and learning","decentralized multi-agent reinforcement learning for coordinated coverage planning in worker-station multi-robot systems"]},{"direction_index":7,"label":"D7","name":"General RL methods, exploration, and evaluation","description":"Advancing broadly applicable robot-control learning algorithms, autonomous exploration mechanisms, and evaluation infrastructure not tied to a single embodiment, capability, transfer setting, or formal-control guarantee.","inclusion":"Include work whose primary objective is generic robot RL or policy-optimization methodology, sample efficiency, value estimation, hyperparameter robustness, interpretability, feature selection, distributed computation, intrinsic motivation, active perception, unsupervised goal or skill discovery, quality-diversity search, task-specific RL algorithm design, or benchmarks/surveys/evaluation environments for robot learning methods.","exclusion":"Do not include application papers where manipulation, navigation, locomotion, transfer, or safety is the dominant objective; use D3 for learned predictive control or formal guarantees, D2 for reward/preference/language/data supervision, and D1 for hierarchical skill composition.","nearest_boundary":"D3","representative_topics":["heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations","active inference for curiosity-driven exploration in sparse-reward robotic manipulation","dual-world embedded feature selection for sensor reduction in deep reinforcement learning control"]}],"split":"dev"}
{"episode_id":"F271_T2024-07","field_index":271,"field":"robot learning and control","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Object manipulation and interaction","description":"Learning representations, policies, dynamics, and controllers for robots to grasp, contact, deform, articulate, or otherwise manipulate objects and materials.","inclusion":"Include work whose primary objective is object-level manipulation capability or manipulation-specific perception, including rigid and dexterous grasping, in-hand manipulation, affordances, point-cloud/object-centric manipulation representations, deformable-object manipulation, articulated objects, soft bodies, fluids, and soft-robot modeling/control when these mechanics define the task.","exclusion":"If the main objective is sequencing reusable skills across long horizons rather than manipulating a particular object/material class, use D1; if the main objective is language grounding, imitation, or data pretraining rather than manipulation mechanics, use D2.","nearest_boundary":"D1","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","implicit representation learning for cross-gripper grasp transfer via signed distance functions","deep reinforcement learning for robotic grasping using 3d octree representations in lunar simulation"]},{"direction_index":1,"label":"D1","name":"Skill abstraction and motion composition","description":"Learning reusable motor-skill, trajectory, primitive, full-body motion, and temporal-abstraction structures for composing or generalizing robot behavior over time.","inclusion":"Include work whose primary objective is representing, discovering, chaining, composing, or adapting skills or motions, including hierarchical RL, long-horizon task decomposition, recovery skills, movement primitives, dynamical-system skill representations, residual skill spaces, and high-DOF humanoid or character motion synthesis when motion organization is central.","exclusion":"If the main objective is object/material manipulation itself, use D0; if the main objective is language- or foundation-model task reasoning, use D2; if formal safety constraints dominate the hierarchy or primitive, use D3.","nearest_boundary":"D0","representative_topics":["skill-space model-based reinforcement learning for long-horizon sparse reward planning","neural dynamic movement primitives for robot trajectory learning and control","hierarchical reinforcement learning for simultaneous stroke rendering and motor control in robotic sketching"]},{"direction_index":2,"label":"D2","name":"Supervised, data-centric, and language-grounded policy learning","description":"Learning robot policies, objectives, or semantic control interfaces from demonstrations, fixed datasets, human feedback, reward/preference signals, language, or foundation-model representations.","inclusion":"Include work whose primary objective is external supervision or prior knowledge for policy acquisition: offline RL and self-supervised pretraining from accumulated data, imitation from demonstrations or videos, human-feedback or reward learning, language-conditioned control, VLM/LLM-based planning, and foundation-model transfer when semantic grounding or data-driven policy learning is primary.","exclusion":"If the primary objective is the physical manipulation capability rather than the supervision source, use D0; if the objective is reusable skill sequencing or trajectory primitives, use D1; if the objective is cross-domain transfer itself rather than data pretraining, use D5.","nearest_boundary":"D0","representative_topics":["latent policy learning for stable adversarial imitation in high-dimensional robot tasks","language-conditioned meta-learning for accelerated tool affordance-based manipulation policy adaptation","language-guided robotic navigation via pre-trained model composition without annotated trajectory supervision"]},{"direction_index":3,"label":"D3","name":"Predictive, safe, and guaranteed control","description":"Developing learned model-based, predictive, constrained, stable, or formally guaranteed controllers for robot systems.","inclusion":"Include work whose primary objective is learning dynamics, estimators, residual models, MPC components, differentiable optimal-control modules, or controllers with explicit safety, stability, temporal-logic, barrier, Lyapunov, reachability, or verification guarantees.","exclusion":"If robustness is empirical domain adaptation without explicit predictive-control or guarantee structure, use D5; if the main objective is generic RL optimization rather than learned dynamics, MPC, or formal constraints, use D7.","nearest_boundary":"D5","representative_topics":["adaptive model-based advice for stabilizing black-box policies in nonlinear control","physics-constrained temporal convolutional learning of quadrotor dynamics for model predictive control","ltl-constrained motion policy learning from demonstrations with mode-based plan satisfaction guarantees"]},{"direction_index":4,"label":"D4","name":"Locomotion, navigation, and vehicle autonomy","description":"Learning control and decision policies for robots whose primary objective is moving through environments, including legged locomotion, mobile navigation, and vehicle-like robotic platforms.","inclusion":"Include work whose primary objective is gait learning, terrain-adaptive locomotion, map-free or social navigation, path planning, traversability, cleaning/coverage navigation, autonomous driving, aerial/underwater/space vehicle control, or platform-dynamics-centered movement through space.","exclusion":"If inter-agent coordination is the central scientific objective, use D6; if learned dynamics/MPC methodology is primary independent of platform class, use D3; if sim-to-real transfer is the main objective rather than locomotion/navigation capability, use D5.","nearest_boundary":"D6","representative_topics":["map-less pointgoal navigation using visual odometry under realistic noisy conditions","inverse reinforcement learning for autonomous driving trajectory planning in dense urban traffic","deep reinforcement learning for joint acceleration and steering control in autonomous driving"]},{"direction_index":5,"label":"D5","name":"Transfer and domain adaptation","description":"Bridging differences between simulation and reality, source and target tasks, environments, embodiments, users, hardware, or operating conditions for deployed robot learning.","inclusion":"Include work whose primary objective is zero-shot or few-shot transfer, sim-to-real or real-to-sim adaptation, domain randomization, residual error estimation, teacher-student transfer, policy fusion, cross-task reuse, disturbance adaptation, or representation learning specifically to handle domain shift.","exclusion":"If sim-to-real is only an implementation detail of a locomotion, manipulation, or navigation method, use that capability direction; if the transfer mechanism is broad offline pretraining from robot data, use D2; if explicit verification or hard safety guarantees dominate, use D3.","nearest_boundary":"D2","representative_topics":["learning-based variable impedance control for aerial sliding on uneven heterogeneous surfaces","optimal transport-based policy fusion for cross-task transfer in deep reinforcement learning robotics","zero-shot sim-to-real visual mpc via randomized-to-canonical model learning for robotic manipulation"]},{"direction_index":6,"label":"D6","name":"Multi-robot and swarm coordination","description":"Learning policies, games, communication, or latent coordination structures for teams, fleets, swarms, or multi-robot systems.","inclusion":"Include work whose primary objective is inter-agent coordination, cooperation, formation, coverage, task allocation, decentralized/centralized multi-agent learning, swarm control, communication-constrained collaboration, or multi-robot manipulation.","exclusion":"If multiple robots are only data sources or deployment infrastructure without a coordination problem, use the relevant single-robot objective; if the main problem is vehicle-following dynamics or single-platform navigation, use D4.","nearest_boundary":"D4","representative_topics":["interactive fleet learning with scalable human supervision allocation and continual policy improvement","distributed differentiable dynamic game framework for multi-robot coordination and learning","decentralized multi-agent reinforcement learning for coordinated coverage planning in worker-station multi-robot systems"]},{"direction_index":7,"label":"D7","name":"General RL methods, exploration, and evaluation","description":"Advancing broadly applicable robot-control learning algorithms, autonomous exploration mechanisms, and evaluation infrastructure not tied to a single embodiment, capability, transfer setting, or formal-control guarantee.","inclusion":"Include work whose primary objective is generic robot RL or policy-optimization methodology, sample efficiency, value estimation, hyperparameter robustness, interpretability, feature selection, distributed computation, intrinsic motivation, active perception, unsupervised goal or skill discovery, quality-diversity search, task-specific RL algorithm design, or benchmarks/surveys/evaluation environments for robot learning methods.","exclusion":"Do not include application papers where manipulation, navigation, locomotion, transfer, or safety is the dominant objective; use D3 for learned predictive control or formal guarantees, D2 for reward/preference/language/data supervision, and D1 for hierarchical skill composition.","nearest_boundary":"D3","representative_topics":["heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations","active inference for curiosity-driven exploration in sparse-reward robotic manipulation","dual-world embedded feature selection for sensor reduction in deep reinforcement learning control"]}],"split":"dev"}
{"episode_id":"F271_T2025-01","field_index":271,"field":"robot learning and control","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Object manipulation and interaction","description":"Learning representations, policies, dynamics, and controllers for robots to grasp, contact, deform, articulate, or otherwise manipulate objects and materials.","inclusion":"Include work whose primary objective is object-level manipulation capability or manipulation-specific perception, including rigid and dexterous grasping, in-hand manipulation, affordances, point-cloud/object-centric manipulation representations, deformable-object manipulation, articulated objects, soft bodies, fluids, and soft-robot modeling/control when these mechanics define the task.","exclusion":"If the main objective is sequencing reusable skills across long horizons rather than manipulating a particular object/material class, use D1; if the main objective is language grounding, imitation, or data pretraining rather than manipulation mechanics, use D2.","nearest_boundary":"D1","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","implicit representation learning for cross-gripper grasp transfer via signed distance functions","deep reinforcement learning for robotic grasping using 3d octree representations in lunar simulation"]},{"direction_index":1,"label":"D1","name":"Skill abstraction and motion composition","description":"Learning reusable motor-skill, trajectory, primitive, full-body motion, and temporal-abstraction structures for composing or generalizing robot behavior over time.","inclusion":"Include work whose primary objective is representing, discovering, chaining, composing, or adapting skills or motions, including hierarchical RL, long-horizon task decomposition, recovery skills, movement primitives, dynamical-system skill representations, residual skill spaces, and high-DOF humanoid or character motion synthesis when motion organization is central.","exclusion":"If the main objective is object/material manipulation itself, use D0; if the main objective is language- or foundation-model task reasoning, use D2; if formal safety constraints dominate the hierarchy or primitive, use D3.","nearest_boundary":"D0","representative_topics":["skill-space model-based reinforcement learning for long-horizon sparse reward planning","neural dynamic movement primitives for robot trajectory learning and control","hierarchical reinforcement learning for simultaneous stroke rendering and motor control in robotic sketching"]},{"direction_index":2,"label":"D2","name":"Supervised, data-centric, and language-grounded policy learning","description":"Learning robot policies, objectives, or semantic control interfaces from demonstrations, fixed datasets, human feedback, reward/preference signals, language, or foundation-model representations.","inclusion":"Include work whose primary objective is external supervision or prior knowledge for policy acquisition: offline RL and self-supervised pretraining from accumulated data, imitation from demonstrations or videos, human-feedback or reward learning, language-conditioned control, VLM/LLM-based planning, and foundation-model transfer when semantic grounding or data-driven policy learning is primary.","exclusion":"If the primary objective is the physical manipulation capability rather than the supervision source, use D0; if the objective is reusable skill sequencing or trajectory primitives, use D1; if the objective is cross-domain transfer itself rather than data pretraining, use D5.","nearest_boundary":"D0","representative_topics":["latent policy learning for stable adversarial imitation in high-dimensional robot tasks","language-conditioned meta-learning for accelerated tool affordance-based manipulation policy adaptation","language-guided robotic navigation via pre-trained model composition without annotated trajectory supervision"]},{"direction_index":3,"label":"D3","name":"Predictive, safe, and guaranteed control","description":"Developing learned model-based, predictive, constrained, stable, or formally guaranteed controllers for robot systems.","inclusion":"Include work whose primary objective is learning dynamics, estimators, residual models, MPC components, differentiable optimal-control modules, or controllers with explicit safety, stability, temporal-logic, barrier, Lyapunov, reachability, or verification guarantees.","exclusion":"If robustness is empirical domain adaptation without explicit predictive-control or guarantee structure, use D5; if the main objective is generic RL optimization rather than learned dynamics, MPC, or formal constraints, use D7.","nearest_boundary":"D5","representative_topics":["adaptive model-based advice for stabilizing black-box policies in nonlinear control","physics-constrained temporal convolutional learning of quadrotor dynamics for model predictive control","ltl-constrained motion policy learning from demonstrations with mode-based plan satisfaction guarantees"]},{"direction_index":4,"label":"D4","name":"Locomotion, navigation, and vehicle autonomy","description":"Learning control and decision policies for robots whose primary objective is moving through environments, including legged locomotion, mobile navigation, and vehicle-like robotic platforms.","inclusion":"Include work whose primary objective is gait learning, terrain-adaptive locomotion, map-free or social navigation, path planning, traversability, cleaning/coverage navigation, autonomous driving, aerial/underwater/space vehicle control, or platform-dynamics-centered movement through space.","exclusion":"If inter-agent coordination is the central scientific objective, use D6; if learned dynamics/MPC methodology is primary independent of platform class, use D3; if sim-to-real transfer is the main objective rather than locomotion/navigation capability, use D5.","nearest_boundary":"D6","representative_topics":["map-less pointgoal navigation using visual odometry under realistic noisy conditions","inverse reinforcement learning for autonomous driving trajectory planning in dense urban traffic","deep reinforcement learning for joint acceleration and steering control in autonomous driving"]},{"direction_index":5,"label":"D5","name":"Transfer and domain adaptation","description":"Bridging differences between simulation and reality, source and target tasks, environments, embodiments, users, hardware, or operating conditions for deployed robot learning.","inclusion":"Include work whose primary objective is zero-shot or few-shot transfer, sim-to-real or real-to-sim adaptation, domain randomization, residual error estimation, teacher-student transfer, policy fusion, cross-task reuse, disturbance adaptation, or representation learning specifically to handle domain shift.","exclusion":"If sim-to-real is only an implementation detail of a locomotion, manipulation, or navigation method, use that capability direction; if the transfer mechanism is broad offline pretraining from robot data, use D2; if explicit verification or hard safety guarantees dominate, use D3.","nearest_boundary":"D2","representative_topics":["learning-based variable impedance control for aerial sliding on uneven heterogeneous surfaces","optimal transport-based policy fusion for cross-task transfer in deep reinforcement learning robotics","zero-shot sim-to-real visual mpc via randomized-to-canonical model learning for robotic manipulation"]},{"direction_index":6,"label":"D6","name":"Multi-robot and swarm coordination","description":"Learning policies, games, communication, or latent coordination structures for teams, fleets, swarms, or multi-robot systems.","inclusion":"Include work whose primary objective is inter-agent coordination, cooperation, formation, coverage, task allocation, decentralized/centralized multi-agent learning, swarm control, communication-constrained collaboration, or multi-robot manipulation.","exclusion":"If multiple robots are only data sources or deployment infrastructure without a coordination problem, use the relevant single-robot objective; if the main problem is vehicle-following dynamics or single-platform navigation, use D4.","nearest_boundary":"D4","representative_topics":["interactive fleet learning with scalable human supervision allocation and continual policy improvement","distributed differentiable dynamic game framework for multi-robot coordination and learning","decentralized multi-agent reinforcement learning for coordinated coverage planning in worker-station multi-robot systems"]},{"direction_index":7,"label":"D7","name":"General RL methods, exploration, and evaluation","description":"Advancing broadly applicable robot-control learning algorithms, autonomous exploration mechanisms, and evaluation infrastructure not tied to a single embodiment, capability, transfer setting, or formal-control guarantee.","inclusion":"Include work whose primary objective is generic robot RL or policy-optimization methodology, sample efficiency, value estimation, hyperparameter robustness, interpretability, feature selection, distributed computation, intrinsic motivation, active perception, unsupervised goal or skill discovery, quality-diversity search, task-specific RL algorithm design, or benchmarks/surveys/evaluation environments for robot learning methods.","exclusion":"Do not include application papers where manipulation, navigation, locomotion, transfer, or safety is the dominant objective; use D3 for learned predictive control or formal guarantees, D2 for reward/preference/language/data supervision, and D1 for hierarchical skill composition.","nearest_boundary":"D3","representative_topics":["heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations","active inference for curiosity-driven exploration in sparse-reward robotic manipulation","dual-world embedded feature selection for sensor reduction in deep reinforcement learning control"]}],"split":"dev"}
{"episode_id":"F271_T2025-07","field_index":271,"field":"robot learning and control","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Object manipulation and interaction","description":"Learning representations, policies, dynamics, and controllers for robots to grasp, contact, deform, articulate, or otherwise manipulate objects and materials.","inclusion":"Include work whose primary objective is object-level manipulation capability or manipulation-specific perception, including rigid and dexterous grasping, in-hand manipulation, affordances, point-cloud/object-centric manipulation representations, deformable-object manipulation, articulated objects, soft bodies, fluids, and soft-robot modeling/control when these mechanics define the task.","exclusion":"If the main objective is sequencing reusable skills across long horizons rather than manipulating a particular object/material class, use D1; if the main objective is language grounding, imitation, or data pretraining rather than manipulation mechanics, use D2.","nearest_boundary":"D1","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","implicit representation learning for cross-gripper grasp transfer via signed distance functions","deep reinforcement learning for robotic grasping using 3d octree representations in lunar simulation"]},{"direction_index":1,"label":"D1","name":"Skill abstraction and motion composition","description":"Learning reusable motor-skill, trajectory, primitive, full-body motion, and temporal-abstraction structures for composing or generalizing robot behavior over time.","inclusion":"Include work whose primary objective is representing, discovering, chaining, composing, or adapting skills or motions, including hierarchical RL, long-horizon task decomposition, recovery skills, movement primitives, dynamical-system skill representations, residual skill spaces, and high-DOF humanoid or character motion synthesis when motion organization is central.","exclusion":"If the main objective is object/material manipulation itself, use D0; if the main objective is language- or foundation-model task reasoning, use D2; if formal safety constraints dominate the hierarchy or primitive, use D3.","nearest_boundary":"D0","representative_topics":["skill-space model-based reinforcement learning for long-horizon sparse reward planning","neural dynamic movement primitives for robot trajectory learning and control","hierarchical reinforcement learning for simultaneous stroke rendering and motor control in robotic sketching"]},{"direction_index":2,"label":"D2","name":"Supervised, data-centric, and language-grounded policy learning","description":"Learning robot policies, objectives, or semantic control interfaces from demonstrations, fixed datasets, human feedback, reward/preference signals, language, or foundation-model representations.","inclusion":"Include work whose primary objective is external supervision or prior knowledge for policy acquisition: offline RL and self-supervised pretraining from accumulated data, imitation from demonstrations or videos, human-feedback or reward learning, language-conditioned control, VLM/LLM-based planning, and foundation-model transfer when semantic grounding or data-driven policy learning is primary.","exclusion":"If the primary objective is the physical manipulation capability rather than the supervision source, use D0; if the objective is reusable skill sequencing or trajectory primitives, use D1; if the objective is cross-domain transfer itself rather than data pretraining, use D5.","nearest_boundary":"D0","representative_topics":["latent policy learning for stable adversarial imitation in high-dimensional robot tasks","language-conditioned meta-learning for accelerated tool affordance-based manipulation policy adaptation","language-guided robotic navigation via pre-trained model composition without annotated trajectory supervision"]},{"direction_index":3,"label":"D3","name":"Predictive, safe, and guaranteed control","description":"Developing learned model-based, predictive, constrained, stable, or formally guaranteed controllers for robot systems.","inclusion":"Include work whose primary objective is learning dynamics, estimators, residual models, MPC components, differentiable optimal-control modules, or controllers with explicit safety, stability, temporal-logic, barrier, Lyapunov, reachability, or verification guarantees.","exclusion":"If robustness is empirical domain adaptation without explicit predictive-control or guarantee structure, use D5; if the main objective is generic RL optimization rather than learned dynamics, MPC, or formal constraints, use D7.","nearest_boundary":"D5","representative_topics":["adaptive model-based advice for stabilizing black-box policies in nonlinear control","physics-constrained temporal convolutional learning of quadrotor dynamics for model predictive control","ltl-constrained motion policy learning from demonstrations with mode-based plan satisfaction guarantees"]},{"direction_index":4,"label":"D4","name":"Locomotion, navigation, and vehicle autonomy","description":"Learning control and decision policies for robots whose primary objective is moving through environments, including legged locomotion, mobile navigation, and vehicle-like robotic platforms.","inclusion":"Include work whose primary objective is gait learning, terrain-adaptive locomotion, map-free or social navigation, path planning, traversability, cleaning/coverage navigation, autonomous driving, aerial/underwater/space vehicle control, or platform-dynamics-centered movement through space.","exclusion":"If inter-agent coordination is the central scientific objective, use D6; if learned dynamics/MPC methodology is primary independent of platform class, use D3; if sim-to-real transfer is the main objective rather than locomotion/navigation capability, use D5.","nearest_boundary":"D6","representative_topics":["map-less pointgoal navigation using visual odometry under realistic noisy conditions","inverse reinforcement learning for autonomous driving trajectory planning in dense urban traffic","deep reinforcement learning for joint acceleration and steering control in autonomous driving"]},{"direction_index":5,"label":"D5","name":"Transfer and domain adaptation","description":"Bridging differences between simulation and reality, source and target tasks, environments, embodiments, users, hardware, or operating conditions for deployed robot learning.","inclusion":"Include work whose primary objective is zero-shot or few-shot transfer, sim-to-real or real-to-sim adaptation, domain randomization, residual error estimation, teacher-student transfer, policy fusion, cross-task reuse, disturbance adaptation, or representation learning specifically to handle domain shift.","exclusion":"If sim-to-real is only an implementation detail of a locomotion, manipulation, or navigation method, use that capability direction; if the transfer mechanism is broad offline pretraining from robot data, use D2; if explicit verification or hard safety guarantees dominate, use D3.","nearest_boundary":"D2","representative_topics":["learning-based variable impedance control for aerial sliding on uneven heterogeneous surfaces","optimal transport-based policy fusion for cross-task transfer in deep reinforcement learning robotics","zero-shot sim-to-real visual mpc via randomized-to-canonical model learning for robotic manipulation"]},{"direction_index":6,"label":"D6","name":"Multi-robot and swarm coordination","description":"Learning policies, games, communication, or latent coordination structures for teams, fleets, swarms, or multi-robot systems.","inclusion":"Include work whose primary objective is inter-agent coordination, cooperation, formation, coverage, task allocation, decentralized/centralized multi-agent learning, swarm control, communication-constrained collaboration, or multi-robot manipulation.","exclusion":"If multiple robots are only data sources or deployment infrastructure without a coordination problem, use the relevant single-robot objective; if the main problem is vehicle-following dynamics or single-platform navigation, use D4.","nearest_boundary":"D4","representative_topics":["interactive fleet learning with scalable human supervision allocation and continual policy improvement","distributed differentiable dynamic game framework for multi-robot coordination and learning","decentralized multi-agent reinforcement learning for coordinated coverage planning in worker-station multi-robot systems"]},{"direction_index":7,"label":"D7","name":"General RL methods, exploration, and evaluation","description":"Advancing broadly applicable robot-control learning algorithms, autonomous exploration mechanisms, and evaluation infrastructure not tied to a single embodiment, capability, transfer setting, or formal-control guarantee.","inclusion":"Include work whose primary objective is generic robot RL or policy-optimization methodology, sample efficiency, value estimation, hyperparameter robustness, interpretability, feature selection, distributed computation, intrinsic motivation, active perception, unsupervised goal or skill discovery, quality-diversity search, task-specific RL algorithm design, or benchmarks/surveys/evaluation environments for robot learning methods.","exclusion":"Do not include application papers where manipulation, navigation, locomotion, transfer, or safety is the dominant objective; use D3 for learned predictive control or formal guarantees, D2 for reward/preference/language/data supervision, and D1 for hierarchical skill composition.","nearest_boundary":"D3","representative_topics":["heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations","active inference for curiosity-driven exploration in sparse-reward robotic manipulation","dual-world embedded feature selection for sensor reduction in deep reinforcement learning control"]}],"split":"dev"}
{"episode_id":"F271_T2026-01","field_index":271,"field":"robot learning and control","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Object manipulation and interaction","description":"Learning representations, policies, dynamics, and controllers for robots to grasp, contact, deform, articulate, or otherwise manipulate objects and materials.","inclusion":"Include work whose primary objective is object-level manipulation capability or manipulation-specific perception, including rigid and dexterous grasping, in-hand manipulation, affordances, point-cloud/object-centric manipulation representations, deformable-object manipulation, articulated objects, soft bodies, fluids, and soft-robot modeling/control when these mechanics define the task.","exclusion":"If the main objective is sequencing reusable skills across long horizons rather than manipulating a particular object/material class, use D1; if the main objective is language grounding, imitation, or data pretraining rather than manipulation mechanics, use D2.","nearest_boundary":"D1","representative_topics":["occlusion-aware mesh-based dynamics for robot cloth unfolding and canonicalization","implicit representation learning for cross-gripper grasp transfer via signed distance functions","deep reinforcement learning for robotic grasping using 3d octree representations in lunar simulation"]},{"direction_index":1,"label":"D1","name":"Skill abstraction and motion composition","description":"Learning reusable motor-skill, trajectory, primitive, full-body motion, and temporal-abstraction structures for composing or generalizing robot behavior over time.","inclusion":"Include work whose primary objective is representing, discovering, chaining, composing, or adapting skills or motions, including hierarchical RL, long-horizon task decomposition, recovery skills, movement primitives, dynamical-system skill representations, residual skill spaces, and high-DOF humanoid or character motion synthesis when motion organization is central.","exclusion":"If the main objective is object/material manipulation itself, use D0; if the main objective is language- or foundation-model task reasoning, use D2; if formal safety constraints dominate the hierarchy or primitive, use D3.","nearest_boundary":"D0","representative_topics":["skill-space model-based reinforcement learning for long-horizon sparse reward planning","neural dynamic movement primitives for robot trajectory learning and control","hierarchical reinforcement learning for simultaneous stroke rendering and motor control in robotic sketching"]},{"direction_index":2,"label":"D2","name":"Supervised, data-centric, and language-grounded policy learning","description":"Learning robot policies, objectives, or semantic control interfaces from demonstrations, fixed datasets, human feedback, reward/preference signals, language, or foundation-model representations.","inclusion":"Include work whose primary objective is external supervision or prior knowledge for policy acquisition: offline RL and self-supervised pretraining from accumulated data, imitation from demonstrations or videos, human-feedback or reward learning, language-conditioned control, VLM/LLM-based planning, and foundation-model transfer when semantic grounding or data-driven policy learning is primary.","exclusion":"If the primary objective is the physical manipulation capability rather than the supervision source, use D0; if the objective is reusable skill sequencing or trajectory primitives, use D1; if the objective is cross-domain transfer itself rather than data pretraining, use D5.","nearest_boundary":"D0","representative_topics":["latent policy learning for stable adversarial imitation in high-dimensional robot tasks","language-conditioned meta-learning for accelerated tool affordance-based manipulation policy adaptation","language-guided robotic navigation via pre-trained model composition without annotated trajectory supervision"]},{"direction_index":3,"label":"D3","name":"Predictive, safe, and guaranteed control","description":"Developing learned model-based, predictive, constrained, stable, or formally guaranteed controllers for robot systems.","inclusion":"Include work whose primary objective is learning dynamics, estimators, residual models, MPC components, differentiable optimal-control modules, or controllers with explicit safety, stability, temporal-logic, barrier, Lyapunov, reachability, or verification guarantees.","exclusion":"If robustness is empirical domain adaptation without explicit predictive-control or guarantee structure, use D5; if the main objective is generic RL optimization rather than learned dynamics, MPC, or formal constraints, use D7.","nearest_boundary":"D5","representative_topics":["adaptive model-based advice for stabilizing black-box policies in nonlinear control","physics-constrained temporal convolutional learning of quadrotor dynamics for model predictive control","ltl-constrained motion policy learning from demonstrations with mode-based plan satisfaction guarantees"]},{"direction_index":4,"label":"D4","name":"Locomotion, navigation, and vehicle autonomy","description":"Learning control and decision policies for robots whose primary objective is moving through environments, including legged locomotion, mobile navigation, and vehicle-like robotic platforms.","inclusion":"Include work whose primary objective is gait learning, terrain-adaptive locomotion, map-free or social navigation, path planning, traversability, cleaning/coverage navigation, autonomous driving, aerial/underwater/space vehicle control, or platform-dynamics-centered movement through space.","exclusion":"If inter-agent coordination is the central scientific objective, use D6; if learned dynamics/MPC methodology is primary independent of platform class, use D3; if sim-to-real transfer is the main objective rather than locomotion/navigation capability, use D5.","nearest_boundary":"D6","representative_topics":["map-less pointgoal navigation using visual odometry under realistic noisy conditions","inverse reinforcement learning for autonomous driving trajectory planning in dense urban traffic","deep reinforcement learning for joint acceleration and steering control in autonomous driving"]},{"direction_index":5,"label":"D5","name":"Transfer and domain adaptation","description":"Bridging differences between simulation and reality, source and target tasks, environments, embodiments, users, hardware, or operating conditions for deployed robot learning.","inclusion":"Include work whose primary objective is zero-shot or few-shot transfer, sim-to-real or real-to-sim adaptation, domain randomization, residual error estimation, teacher-student transfer, policy fusion, cross-task reuse, disturbance adaptation, or representation learning specifically to handle domain shift.","exclusion":"If sim-to-real is only an implementation detail of a locomotion, manipulation, or navigation method, use that capability direction; if the transfer mechanism is broad offline pretraining from robot data, use D2; if explicit verification or hard safety guarantees dominate, use D3.","nearest_boundary":"D2","representative_topics":["learning-based variable impedance control for aerial sliding on uneven heterogeneous surfaces","optimal transport-based policy fusion for cross-task transfer in deep reinforcement learning robotics","zero-shot sim-to-real visual mpc via randomized-to-canonical model learning for robotic manipulation"]},{"direction_index":6,"label":"D6","name":"Multi-robot and swarm coordination","description":"Learning policies, games, communication, or latent coordination structures for teams, fleets, swarms, or multi-robot systems.","inclusion":"Include work whose primary objective is inter-agent coordination, cooperation, formation, coverage, task allocation, decentralized/centralized multi-agent learning, swarm control, communication-constrained collaboration, or multi-robot manipulation.","exclusion":"If multiple robots are only data sources or deployment infrastructure without a coordination problem, use the relevant single-robot objective; if the main problem is vehicle-following dynamics or single-platform navigation, use D4.","nearest_boundary":"D4","representative_topics":["interactive fleet learning with scalable human supervision allocation and continual policy improvement","distributed differentiable dynamic game framework for multi-robot coordination and learning","decentralized multi-agent reinforcement learning for coordinated coverage planning in worker-station multi-robot systems"]},{"direction_index":7,"label":"D7","name":"General RL methods, exploration, and evaluation","description":"Advancing broadly applicable robot-control learning algorithms, autonomous exploration mechanisms, and evaluation infrastructure not tied to a single embodiment, capability, transfer setting, or formal-control guarantee.","inclusion":"Include work whose primary objective is generic robot RL or policy-optimization methodology, sample efficiency, value estimation, hyperparameter robustness, interpretability, feature selection, distributed computation, intrinsic motivation, active perception, unsupervised goal or skill discovery, quality-diversity search, task-specific RL algorithm design, or benchmarks/surveys/evaluation environments for robot learning methods.","exclusion":"Do not include application papers where manipulation, navigation, locomotion, transfer, or safety is the dominant objective; use D3 for learned predictive control or formal guarantees, D2 for reward/preference/language/data supervision, and D1 for hierarchical skill composition.","nearest_boundary":"D3","representative_topics":["heavy-tailed stochastic policy gradient for sparse reward continuous control without demonstrations","active inference for curiosity-driven exploration in sparse-reward robotic manipulation","dual-world embedded feature selection for sensor reduction in deep reinforcement learning control"]}],"split":"dev"}
{"episode_id":"F275_T2024-01","field_index":275,"field":"biomedical natural language processing","forecast_date":"2024-01-01","history_windows":[{"start":"2022-07-01","end":"2023-01-01"},{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"}],"future_window":{"start":"2024-01-01","end":"2024-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structured biomedical text extraction and coding","description":"Methods whose primary objective is converting biomedical or clinical text into structured entities, relations, events, normalized concepts, medication-safety facts, or standardized clinical codes.","inclusion":"Include NER, entity linking, relation/event extraction, temporal or attribute extraction, prompt-based extraction, low-resource entity normalization, adverse-event or medication-change extraction, and ICD/SNOMED/DRG-style code assignment when the scientific target is structured fact or code recovery from text.","exclusion":"If extracted variables are mainly used to characterize patients or predict outcomes, assign to patient phenotyping; if structured knowledge is used mainly as a representation or reasoning substrate, assign to knowledge-enhanced representation.","nearest_boundary":"Patient phenotyping and outcome modeling","representative_topics":["open set medical entity linking and span extraction via label-attention transformer without span annotations","joint learning for causal relation extraction and entity function detection in biomedical text","bert-based extraction of medication change attributes from clinical narrative text"]},{"direction_index":1,"label":"D1","name":"Domain and resource-efficient model adaptation","description":"Research whose primary objective is adapting language-model representations to biomedical or clinical domains, languages, vocabularies, compute constraints, or scarce annotation regimes.","inclusion":"Include domain-specific or continual pretraining, compact biomedical transformers, vocabulary transfer, multilingual or cross-lingual transfer, prompt or parameter-efficient adaptation under scarcity, and model adaptation where the model or transfer strategy is the central contribution rather than a downstream task system.","exclusion":"If the work primarily builds an extraction, QA, summarization, coding, or diagnosis system, assign to that task objective; if it primarily compares model capabilities, assign to evaluation and assurance.","nearest_boundary":"Evaluation, assurance, and privacy-preserving NLP","representative_topics":["domain-specific conversational language pre-training for low-resource clinical dialogue comprehension","low-resource german medical named entity recognition via pretrained language model transfer learning","french clinical language model adaptation via domain-specific pretraining on electronic health records"]},{"direction_index":2,"label":"D2","name":"Medical QA, dialogue, and clinical reasoning","description":"Systems whose primary objective is answering biomedical or clinical questions, conducting medical dialogue, supporting diagnosis or differential reasoning, or producing patient-specific clinical rationales.","inclusion":"Include retrieval-augmented medical QA, clinical semantic parsing, medical dialogue fine-tuning, LLM-based clinical explanation, diagnostic reasoning, guideline- or case-based reasoning, and decision-support workflows where reasoning or question answering dominates.","exclusion":"If the objective is information-access infrastructure without reasoning, assign to data resources and operational systems; if the primary output is a summary rather than an answer, dialogue, or decision rationale, assign to biomedical generation and communication.","nearest_boundary":"Data resources and operational workflow systems","representative_topics":["query-focused extractive summarisation for complex biomedical question answering with retrieval-augmented sentence scoring","multi-dimensional clinical explanation generation using text generation models as implicit knowledge bases","neural semantic parsing for clinical question answering over electronic health records"]},{"direction_index":3,"label":"D3","name":"Evaluation, assurance, and privacy-preserving NLP","description":"Research whose primary objective is measuring, benchmarking, auditing, or safeguarding biomedical NLP systems, including fairness, bias, robustness, privacy, de-identification, anonymization, and federated evaluation.","inclusion":"Include benchmark construction, shared-task evaluation, comparative model studies, reliability and safety assessment, bias and fairness analyses, de-identification and pseudonymization, anonymization-impact studies, and model-to-data or federated evaluation infrastructure.","exclusion":"If privacy or evaluation is only incidental to introducing a task system or adapted model, assign to that system's primary objective; if the work mainly creates training corpora rather than evaluation or assurance mechanisms, assign to data resources and operational systems.","nearest_boundary":"Data resources and operational workflow systems","representative_topics":["federated model-to-data sandbox for unbiased multi-site clinical nlp evaluation","gender bias analysis in clinical word embeddings for stereotypical medical categories","anonymization technique selection for clinical text nlp under re-identification constraints"]},{"direction_index":4,"label":"D4","name":"Biomedical generation, summarization, and simplification","description":"Methods whose primary objective is generating, summarizing, simplifying, or adapting biomedical and clinical text for clinicians, researchers, patients, or lay audiences.","inclusion":"Include clinical note or report summarization, counseling or dialogue summarization, biomedical literature summarization, keyphrase generation, readability-controlled summarization, plain-language adaptation, patient-facing simplification, and biomedical text-generation toolkits when generation quality is central.","exclusion":"If generated text is primarily an answer or decision rationale, assign to medical QA and clinical reasoning; if generated data are primarily for dataset augmentation, assign to data resources and operational systems.","nearest_boundary":"Medical QA, dialogue, and clinical reasoning","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","domain-adaptive seq2seq summarization of patient problems from clinical progress notes","readability-controllable abstractive summarization of biomedical documents for expert and lay audiences"]},{"direction_index":5,"label":"D5","name":"Patient phenotyping and social-behavioral assessment","description":"Research whose primary objective is using health-related language to characterize patients, infer diagnoses or disease states, monitor outcomes, identify social determinants, or assess neurological, mental-health, pain, or communication-related conditions.","inclusion":"Include clinical phenotyping, disease or symptom classification, mortality/readmission/utilization prediction from notes, cohort expansion, SDOH and equity-relevant variable extraction, chronic pain and dementia language assessment, mental-health or counseling state assessment, and language-based assistive assessment for communication disorders.","exclusion":"If the output is a standardized administrative code, assign to structured extraction and coding; if the work centers on LLM diagnostic reasoning or decision rationales rather than phenotype prediction or assessment, assign to medical QA and clinical reasoning.","nearest_boundary":"Structured biomedical text extraction and coding","representative_topics":["nlp-based automated diagnosis from clinical transcription reports using hybrid deep learning models","transformer-based clinical note embedding for mortality prediction via umls concept mapping","language model classification and facet-based summarization for chronic pain subtype detection"]},{"direction_index":6,"label":"D6","name":"Knowledge-enhanced biomedical representation","description":"Methods whose primary objective is incorporating biomedical ontologies, knowledge graphs, controlled vocabularies, or graph-structured medical knowledge into representations, reasoning substrates, semantic similarity, or knowledge discovery.","inclusion":"Include KG- or ontology-enhanced language understanding, concept and code embeddings, ontology description generation, graph-augmented semantic similarity, knowledge-enhanced text classification, literature-based discovery, and representation learning where structured knowledge is central.","exclusion":"If the work merely links mentions to an ontology as an extraction output, assign to structured extraction and coding; if structured knowledge is only a small pretraining signal for a general domain model, assign to domain and resource-efficient model adaptation.","nearest_boundary":"Structured biomedical text extraction and coding","representative_topics":["graph-in-graph network for gene ontology term description generation using nested gene and term graphs","graph-augmented cyclic co-training for clinical text semantic similarity estimation","knowledge graph enhanced bert for multi-type clinical text classification"]},{"direction_index":7,"label":"D7","name":"Data resources and operational workflow systems","description":"Research whose primary objective is building reusable biomedical NLP datasets, corpora, annotation resources, toolkits, retrieval pipelines, trial-matching systems, literature-search systems, or deployed domain-specific NLP applications.","inclusion":"Include corpus construction, dataset harmonization, synthetic or augmented data resources, annotation pipelines, clinical-trial eligibility corpora, systematic-review query generation, biomedical literature retrieval, patient-trial matching, specialized clinical NLP workbenches, fact-checking resources, and operational application pipelines where infrastructure or workflow enablement dominates.","exclusion":"If the workflow principally performs clinical reasoning or medical QA, assign to medical QA and clinical reasoning; if the resource is primarily an evaluation benchmark, assign to evaluation and assurance; if the core contribution is a trained domain model, assign to domain and resource-efficient model adaptation.","nearest_boundary":"Medical QA, dialogue, and clinical reasoning","representative_topics":["data-centric meta-dataset curation for biomedical language model pretraining and evaluation","low-resource clinical information extraction workbench for german biomedical text","corpus construction and information extraction for clinical trial eligibility criteria query generation"]}],"split":"dev"}
{"episode_id":"F275_T2024-07","field_index":275,"field":"biomedical natural language processing","forecast_date":"2024-07-01","history_windows":[{"start":"2023-01-01","end":"2023-07-01"},{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"}],"future_window":{"start":"2024-07-01","end":"2025-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structured biomedical text extraction and coding","description":"Methods whose primary objective is converting biomedical or clinical text into structured entities, relations, events, normalized concepts, medication-safety facts, or standardized clinical codes.","inclusion":"Include NER, entity linking, relation/event extraction, temporal or attribute extraction, prompt-based extraction, low-resource entity normalization, adverse-event or medication-change extraction, and ICD/SNOMED/DRG-style code assignment when the scientific target is structured fact or code recovery from text.","exclusion":"If extracted variables are mainly used to characterize patients or predict outcomes, assign to patient phenotyping; if structured knowledge is used mainly as a representation or reasoning substrate, assign to knowledge-enhanced representation.","nearest_boundary":"Patient phenotyping and outcome modeling","representative_topics":["open set medical entity linking and span extraction via label-attention transformer without span annotations","joint learning for causal relation extraction and entity function detection in biomedical text","bert-based extraction of medication change attributes from clinical narrative text"]},{"direction_index":1,"label":"D1","name":"Domain and resource-efficient model adaptation","description":"Research whose primary objective is adapting language-model representations to biomedical or clinical domains, languages, vocabularies, compute constraints, or scarce annotation regimes.","inclusion":"Include domain-specific or continual pretraining, compact biomedical transformers, vocabulary transfer, multilingual or cross-lingual transfer, prompt or parameter-efficient adaptation under scarcity, and model adaptation where the model or transfer strategy is the central contribution rather than a downstream task system.","exclusion":"If the work primarily builds an extraction, QA, summarization, coding, or diagnosis system, assign to that task objective; if it primarily compares model capabilities, assign to evaluation and assurance.","nearest_boundary":"Evaluation, assurance, and privacy-preserving NLP","representative_topics":["domain-specific conversational language pre-training for low-resource clinical dialogue comprehension","low-resource german medical named entity recognition via pretrained language model transfer learning","french clinical language model adaptation via domain-specific pretraining on electronic health records"]},{"direction_index":2,"label":"D2","name":"Medical QA, dialogue, and clinical reasoning","description":"Systems whose primary objective is answering biomedical or clinical questions, conducting medical dialogue, supporting diagnosis or differential reasoning, or producing patient-specific clinical rationales.","inclusion":"Include retrieval-augmented medical QA, clinical semantic parsing, medical dialogue fine-tuning, LLM-based clinical explanation, diagnostic reasoning, guideline- or case-based reasoning, and decision-support workflows where reasoning or question answering dominates.","exclusion":"If the objective is information-access infrastructure without reasoning, assign to data resources and operational systems; if the primary output is a summary rather than an answer, dialogue, or decision rationale, assign to biomedical generation and communication.","nearest_boundary":"Data resources and operational workflow systems","representative_topics":["query-focused extractive summarisation for complex biomedical question answering with retrieval-augmented sentence scoring","multi-dimensional clinical explanation generation using text generation models as implicit knowledge bases","neural semantic parsing for clinical question answering over electronic health records"]},{"direction_index":3,"label":"D3","name":"Evaluation, assurance, and privacy-preserving NLP","description":"Research whose primary objective is measuring, benchmarking, auditing, or safeguarding biomedical NLP systems, including fairness, bias, robustness, privacy, de-identification, anonymization, and federated evaluation.","inclusion":"Include benchmark construction, shared-task evaluation, comparative model studies, reliability and safety assessment, bias and fairness analyses, de-identification and pseudonymization, anonymization-impact studies, and model-to-data or federated evaluation infrastructure.","exclusion":"If privacy or evaluation is only incidental to introducing a task system or adapted model, assign to that system's primary objective; if the work mainly creates training corpora rather than evaluation or assurance mechanisms, assign to data resources and operational systems.","nearest_boundary":"Data resources and operational workflow systems","representative_topics":["federated model-to-data sandbox for unbiased multi-site clinical nlp evaluation","gender bias analysis in clinical word embeddings for stereotypical medical categories","anonymization technique selection for clinical text nlp under re-identification constraints"]},{"direction_index":4,"label":"D4","name":"Biomedical generation, summarization, and simplification","description":"Methods whose primary objective is generating, summarizing, simplifying, or adapting biomedical and clinical text for clinicians, researchers, patients, or lay audiences.","inclusion":"Include clinical note or report summarization, counseling or dialogue summarization, biomedical literature summarization, keyphrase generation, readability-controlled summarization, plain-language adaptation, patient-facing simplification, and biomedical text-generation toolkits when generation quality is central.","exclusion":"If generated text is primarily an answer or decision rationale, assign to medical QA and clinical reasoning; if generated data are primarily for dataset augmentation, assign to data resources and operational systems.","nearest_boundary":"Medical QA, dialogue, and clinical reasoning","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","domain-adaptive seq2seq summarization of patient problems from clinical progress notes","readability-controllable abstractive summarization of biomedical documents for expert and lay audiences"]},{"direction_index":5,"label":"D5","name":"Patient phenotyping and social-behavioral assessment","description":"Research whose primary objective is using health-related language to characterize patients, infer diagnoses or disease states, monitor outcomes, identify social determinants, or assess neurological, mental-health, pain, or communication-related conditions.","inclusion":"Include clinical phenotyping, disease or symptom classification, mortality/readmission/utilization prediction from notes, cohort expansion, SDOH and equity-relevant variable extraction, chronic pain and dementia language assessment, mental-health or counseling state assessment, and language-based assistive assessment for communication disorders.","exclusion":"If the output is a standardized administrative code, assign to structured extraction and coding; if the work centers on LLM diagnostic reasoning or decision rationales rather than phenotype prediction or assessment, assign to medical QA and clinical reasoning.","nearest_boundary":"Structured biomedical text extraction and coding","representative_topics":["nlp-based automated diagnosis from clinical transcription reports using hybrid deep learning models","transformer-based clinical note embedding for mortality prediction via umls concept mapping","language model classification and facet-based summarization for chronic pain subtype detection"]},{"direction_index":6,"label":"D6","name":"Knowledge-enhanced biomedical representation","description":"Methods whose primary objective is incorporating biomedical ontologies, knowledge graphs, controlled vocabularies, or graph-structured medical knowledge into representations, reasoning substrates, semantic similarity, or knowledge discovery.","inclusion":"Include KG- or ontology-enhanced language understanding, concept and code embeddings, ontology description generation, graph-augmented semantic similarity, knowledge-enhanced text classification, literature-based discovery, and representation learning where structured knowledge is central.","exclusion":"If the work merely links mentions to an ontology as an extraction output, assign to structured extraction and coding; if structured knowledge is only a small pretraining signal for a general domain model, assign to domain and resource-efficient model adaptation.","nearest_boundary":"Structured biomedical text extraction and coding","representative_topics":["graph-in-graph network for gene ontology term description generation using nested gene and term graphs","graph-augmented cyclic co-training for clinical text semantic similarity estimation","knowledge graph enhanced bert for multi-type clinical text classification"]},{"direction_index":7,"label":"D7","name":"Data resources and operational workflow systems","description":"Research whose primary objective is building reusable biomedical NLP datasets, corpora, annotation resources, toolkits, retrieval pipelines, trial-matching systems, literature-search systems, or deployed domain-specific NLP applications.","inclusion":"Include corpus construction, dataset harmonization, synthetic or augmented data resources, annotation pipelines, clinical-trial eligibility corpora, systematic-review query generation, biomedical literature retrieval, patient-trial matching, specialized clinical NLP workbenches, fact-checking resources, and operational application pipelines where infrastructure or workflow enablement dominates.","exclusion":"If the workflow principally performs clinical reasoning or medical QA, assign to medical QA and clinical reasoning; if the resource is primarily an evaluation benchmark, assign to evaluation and assurance; if the core contribution is a trained domain model, assign to domain and resource-efficient model adaptation.","nearest_boundary":"Medical QA, dialogue, and clinical reasoning","representative_topics":["data-centric meta-dataset curation for biomedical language model pretraining and evaluation","low-resource clinical information extraction workbench for german biomedical text","corpus construction and information extraction for clinical trial eligibility criteria query generation"]}],"split":"dev"}
{"episode_id":"F275_T2025-01","field_index":275,"field":"biomedical natural language processing","forecast_date":"2025-01-01","history_windows":[{"start":"2023-07-01","end":"2024-01-01"},{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"}],"future_window":{"start":"2025-01-01","end":"2025-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structured biomedical text extraction and coding","description":"Methods whose primary objective is converting biomedical or clinical text into structured entities, relations, events, normalized concepts, medication-safety facts, or standardized clinical codes.","inclusion":"Include NER, entity linking, relation/event extraction, temporal or attribute extraction, prompt-based extraction, low-resource entity normalization, adverse-event or medication-change extraction, and ICD/SNOMED/DRG-style code assignment when the scientific target is structured fact or code recovery from text.","exclusion":"If extracted variables are mainly used to characterize patients or predict outcomes, assign to patient phenotyping; if structured knowledge is used mainly as a representation or reasoning substrate, assign to knowledge-enhanced representation.","nearest_boundary":"Patient phenotyping and outcome modeling","representative_topics":["open set medical entity linking and span extraction via label-attention transformer without span annotations","joint learning for causal relation extraction and entity function detection in biomedical text","bert-based extraction of medication change attributes from clinical narrative text"]},{"direction_index":1,"label":"D1","name":"Domain and resource-efficient model adaptation","description":"Research whose primary objective is adapting language-model representations to biomedical or clinical domains, languages, vocabularies, compute constraints, or scarce annotation regimes.","inclusion":"Include domain-specific or continual pretraining, compact biomedical transformers, vocabulary transfer, multilingual or cross-lingual transfer, prompt or parameter-efficient adaptation under scarcity, and model adaptation where the model or transfer strategy is the central contribution rather than a downstream task system.","exclusion":"If the work primarily builds an extraction, QA, summarization, coding, or diagnosis system, assign to that task objective; if it primarily compares model capabilities, assign to evaluation and assurance.","nearest_boundary":"Evaluation, assurance, and privacy-preserving NLP","representative_topics":["domain-specific conversational language pre-training for low-resource clinical dialogue comprehension","low-resource german medical named entity recognition via pretrained language model transfer learning","french clinical language model adaptation via domain-specific pretraining on electronic health records"]},{"direction_index":2,"label":"D2","name":"Medical QA, dialogue, and clinical reasoning","description":"Systems whose primary objective is answering biomedical or clinical questions, conducting medical dialogue, supporting diagnosis or differential reasoning, or producing patient-specific clinical rationales.","inclusion":"Include retrieval-augmented medical QA, clinical semantic parsing, medical dialogue fine-tuning, LLM-based clinical explanation, diagnostic reasoning, guideline- or case-based reasoning, and decision-support workflows where reasoning or question answering dominates.","exclusion":"If the objective is information-access infrastructure without reasoning, assign to data resources and operational systems; if the primary output is a summary rather than an answer, dialogue, or decision rationale, assign to biomedical generation and communication.","nearest_boundary":"Data resources and operational workflow systems","representative_topics":["query-focused extractive summarisation for complex biomedical question answering with retrieval-augmented sentence scoring","multi-dimensional clinical explanation generation using text generation models as implicit knowledge bases","neural semantic parsing for clinical question answering over electronic health records"]},{"direction_index":3,"label":"D3","name":"Evaluation, assurance, and privacy-preserving NLP","description":"Research whose primary objective is measuring, benchmarking, auditing, or safeguarding biomedical NLP systems, including fairness, bias, robustness, privacy, de-identification, anonymization, and federated evaluation.","inclusion":"Include benchmark construction, shared-task evaluation, comparative model studies, reliability and safety assessment, bias and fairness analyses, de-identification and pseudonymization, anonymization-impact studies, and model-to-data or federated evaluation infrastructure.","exclusion":"If privacy or evaluation is only incidental to introducing a task system or adapted model, assign to that system's primary objective; if the work mainly creates training corpora rather than evaluation or assurance mechanisms, assign to data resources and operational systems.","nearest_boundary":"Data resources and operational workflow systems","representative_topics":["federated model-to-data sandbox for unbiased multi-site clinical nlp evaluation","gender bias analysis in clinical word embeddings for stereotypical medical categories","anonymization technique selection for clinical text nlp under re-identification constraints"]},{"direction_index":4,"label":"D4","name":"Biomedical generation, summarization, and simplification","description":"Methods whose primary objective is generating, summarizing, simplifying, or adapting biomedical and clinical text for clinicians, researchers, patients, or lay audiences.","inclusion":"Include clinical note or report summarization, counseling or dialogue summarization, biomedical literature summarization, keyphrase generation, readability-controlled summarization, plain-language adaptation, patient-facing simplification, and biomedical text-generation toolkits when generation quality is central.","exclusion":"If generated text is primarily an answer or decision rationale, assign to medical QA and clinical reasoning; if generated data are primarily for dataset augmentation, assign to data resources and operational systems.","nearest_boundary":"Medical QA, dialogue, and clinical reasoning","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","domain-adaptive seq2seq summarization of patient problems from clinical progress notes","readability-controllable abstractive summarization of biomedical documents for expert and lay audiences"]},{"direction_index":5,"label":"D5","name":"Patient phenotyping and social-behavioral assessment","description":"Research whose primary objective is using health-related language to characterize patients, infer diagnoses or disease states, monitor outcomes, identify social determinants, or assess neurological, mental-health, pain, or communication-related conditions.","inclusion":"Include clinical phenotyping, disease or symptom classification, mortality/readmission/utilization prediction from notes, cohort expansion, SDOH and equity-relevant variable extraction, chronic pain and dementia language assessment, mental-health or counseling state assessment, and language-based assistive assessment for communication disorders.","exclusion":"If the output is a standardized administrative code, assign to structured extraction and coding; if the work centers on LLM diagnostic reasoning or decision rationales rather than phenotype prediction or assessment, assign to medical QA and clinical reasoning.","nearest_boundary":"Structured biomedical text extraction and coding","representative_topics":["nlp-based automated diagnosis from clinical transcription reports using hybrid deep learning models","transformer-based clinical note embedding for mortality prediction via umls concept mapping","language model classification and facet-based summarization for chronic pain subtype detection"]},{"direction_index":6,"label":"D6","name":"Knowledge-enhanced biomedical representation","description":"Methods whose primary objective is incorporating biomedical ontologies, knowledge graphs, controlled vocabularies, or graph-structured medical knowledge into representations, reasoning substrates, semantic similarity, or knowledge discovery.","inclusion":"Include KG- or ontology-enhanced language understanding, concept and code embeddings, ontology description generation, graph-augmented semantic similarity, knowledge-enhanced text classification, literature-based discovery, and representation learning where structured knowledge is central.","exclusion":"If the work merely links mentions to an ontology as an extraction output, assign to structured extraction and coding; if structured knowledge is only a small pretraining signal for a general domain model, assign to domain and resource-efficient model adaptation.","nearest_boundary":"Structured biomedical text extraction and coding","representative_topics":["graph-in-graph network for gene ontology term description generation using nested gene and term graphs","graph-augmented cyclic co-training for clinical text semantic similarity estimation","knowledge graph enhanced bert for multi-type clinical text classification"]},{"direction_index":7,"label":"D7","name":"Data resources and operational workflow systems","description":"Research whose primary objective is building reusable biomedical NLP datasets, corpora, annotation resources, toolkits, retrieval pipelines, trial-matching systems, literature-search systems, or deployed domain-specific NLP applications.","inclusion":"Include corpus construction, dataset harmonization, synthetic or augmented data resources, annotation pipelines, clinical-trial eligibility corpora, systematic-review query generation, biomedical literature retrieval, patient-trial matching, specialized clinical NLP workbenches, fact-checking resources, and operational application pipelines where infrastructure or workflow enablement dominates.","exclusion":"If the workflow principally performs clinical reasoning or medical QA, assign to medical QA and clinical reasoning; if the resource is primarily an evaluation benchmark, assign to evaluation and assurance; if the core contribution is a trained domain model, assign to domain and resource-efficient model adaptation.","nearest_boundary":"Medical QA, dialogue, and clinical reasoning","representative_topics":["data-centric meta-dataset curation for biomedical language model pretraining and evaluation","low-resource clinical information extraction workbench for german biomedical text","corpus construction and information extraction for clinical trial eligibility criteria query generation"]}],"split":"dev"}
{"episode_id":"F275_T2025-07","field_index":275,"field":"biomedical natural language processing","forecast_date":"2025-07-01","history_windows":[{"start":"2024-01-01","end":"2024-07-01"},{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"}],"future_window":{"start":"2025-07-01","end":"2026-01-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structured biomedical text extraction and coding","description":"Methods whose primary objective is converting biomedical or clinical text into structured entities, relations, events, normalized concepts, medication-safety facts, or standardized clinical codes.","inclusion":"Include NER, entity linking, relation/event extraction, temporal or attribute extraction, prompt-based extraction, low-resource entity normalization, adverse-event or medication-change extraction, and ICD/SNOMED/DRG-style code assignment when the scientific target is structured fact or code recovery from text.","exclusion":"If extracted variables are mainly used to characterize patients or predict outcomes, assign to patient phenotyping; if structured knowledge is used mainly as a representation or reasoning substrate, assign to knowledge-enhanced representation.","nearest_boundary":"Patient phenotyping and outcome modeling","representative_topics":["open set medical entity linking and span extraction via label-attention transformer without span annotations","joint learning for causal relation extraction and entity function detection in biomedical text","bert-based extraction of medication change attributes from clinical narrative text"]},{"direction_index":1,"label":"D1","name":"Domain and resource-efficient model adaptation","description":"Research whose primary objective is adapting language-model representations to biomedical or clinical domains, languages, vocabularies, compute constraints, or scarce annotation regimes.","inclusion":"Include domain-specific or continual pretraining, compact biomedical transformers, vocabulary transfer, multilingual or cross-lingual transfer, prompt or parameter-efficient adaptation under scarcity, and model adaptation where the model or transfer strategy is the central contribution rather than a downstream task system.","exclusion":"If the work primarily builds an extraction, QA, summarization, coding, or diagnosis system, assign to that task objective; if it primarily compares model capabilities, assign to evaluation and assurance.","nearest_boundary":"Evaluation, assurance, and privacy-preserving NLP","representative_topics":["domain-specific conversational language pre-training for low-resource clinical dialogue comprehension","low-resource german medical named entity recognition via pretrained language model transfer learning","french clinical language model adaptation via domain-specific pretraining on electronic health records"]},{"direction_index":2,"label":"D2","name":"Medical QA, dialogue, and clinical reasoning","description":"Systems whose primary objective is answering biomedical or clinical questions, conducting medical dialogue, supporting diagnosis or differential reasoning, or producing patient-specific clinical rationales.","inclusion":"Include retrieval-augmented medical QA, clinical semantic parsing, medical dialogue fine-tuning, LLM-based clinical explanation, diagnostic reasoning, guideline- or case-based reasoning, and decision-support workflows where reasoning or question answering dominates.","exclusion":"If the objective is information-access infrastructure without reasoning, assign to data resources and operational systems; if the primary output is a summary rather than an answer, dialogue, or decision rationale, assign to biomedical generation and communication.","nearest_boundary":"Data resources and operational workflow systems","representative_topics":["query-focused extractive summarisation for complex biomedical question answering with retrieval-augmented sentence scoring","multi-dimensional clinical explanation generation using text generation models as implicit knowledge bases","neural semantic parsing for clinical question answering over electronic health records"]},{"direction_index":3,"label":"D3","name":"Evaluation, assurance, and privacy-preserving NLP","description":"Research whose primary objective is measuring, benchmarking, auditing, or safeguarding biomedical NLP systems, including fairness, bias, robustness, privacy, de-identification, anonymization, and federated evaluation.","inclusion":"Include benchmark construction, shared-task evaluation, comparative model studies, reliability and safety assessment, bias and fairness analyses, de-identification and pseudonymization, anonymization-impact studies, and model-to-data or federated evaluation infrastructure.","exclusion":"If privacy or evaluation is only incidental to introducing a task system or adapted model, assign to that system's primary objective; if the work mainly creates training corpora rather than evaluation or assurance mechanisms, assign to data resources and operational systems.","nearest_boundary":"Data resources and operational workflow systems","representative_topics":["federated model-to-data sandbox for unbiased multi-site clinical nlp evaluation","gender bias analysis in clinical word embeddings for stereotypical medical categories","anonymization technique selection for clinical text nlp under re-identification constraints"]},{"direction_index":4,"label":"D4","name":"Biomedical generation, summarization, and simplification","description":"Methods whose primary objective is generating, summarizing, simplifying, or adapting biomedical and clinical text for clinicians, researchers, patients, or lay audiences.","inclusion":"Include clinical note or report summarization, counseling or dialogue summarization, biomedical literature summarization, keyphrase generation, readability-controlled summarization, plain-language adaptation, patient-facing simplification, and biomedical text-generation toolkits when generation quality is central.","exclusion":"If generated text is primarily an answer or decision rationale, assign to medical QA and clinical reasoning; if generated data are primarily for dataset augmentation, assign to data resources and operational systems.","nearest_boundary":"Medical QA, dialogue, and clinical reasoning","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","domain-adaptive seq2seq summarization of patient problems from clinical progress notes","readability-controllable abstractive summarization of biomedical documents for expert and lay audiences"]},{"direction_index":5,"label":"D5","name":"Patient phenotyping and social-behavioral assessment","description":"Research whose primary objective is using health-related language to characterize patients, infer diagnoses or disease states, monitor outcomes, identify social determinants, or assess neurological, mental-health, pain, or communication-related conditions.","inclusion":"Include clinical phenotyping, disease or symptom classification, mortality/readmission/utilization prediction from notes, cohort expansion, SDOH and equity-relevant variable extraction, chronic pain and dementia language assessment, mental-health or counseling state assessment, and language-based assistive assessment for communication disorders.","exclusion":"If the output is a standardized administrative code, assign to structured extraction and coding; if the work centers on LLM diagnostic reasoning or decision rationales rather than phenotype prediction or assessment, assign to medical QA and clinical reasoning.","nearest_boundary":"Structured biomedical text extraction and coding","representative_topics":["nlp-based automated diagnosis from clinical transcription reports using hybrid deep learning models","transformer-based clinical note embedding for mortality prediction via umls concept mapping","language model classification and facet-based summarization for chronic pain subtype detection"]},{"direction_index":6,"label":"D6","name":"Knowledge-enhanced biomedical representation","description":"Methods whose primary objective is incorporating biomedical ontologies, knowledge graphs, controlled vocabularies, or graph-structured medical knowledge into representations, reasoning substrates, semantic similarity, or knowledge discovery.","inclusion":"Include KG- or ontology-enhanced language understanding, concept and code embeddings, ontology description generation, graph-augmented semantic similarity, knowledge-enhanced text classification, literature-based discovery, and representation learning where structured knowledge is central.","exclusion":"If the work merely links mentions to an ontology as an extraction output, assign to structured extraction and coding; if structured knowledge is only a small pretraining signal for a general domain model, assign to domain and resource-efficient model adaptation.","nearest_boundary":"Structured biomedical text extraction and coding","representative_topics":["graph-in-graph network for gene ontology term description generation using nested gene and term graphs","graph-augmented cyclic co-training for clinical text semantic similarity estimation","knowledge graph enhanced bert for multi-type clinical text classification"]},{"direction_index":7,"label":"D7","name":"Data resources and operational workflow systems","description":"Research whose primary objective is building reusable biomedical NLP datasets, corpora, annotation resources, toolkits, retrieval pipelines, trial-matching systems, literature-search systems, or deployed domain-specific NLP applications.","inclusion":"Include corpus construction, dataset harmonization, synthetic or augmented data resources, annotation pipelines, clinical-trial eligibility corpora, systematic-review query generation, biomedical literature retrieval, patient-trial matching, specialized clinical NLP workbenches, fact-checking resources, and operational application pipelines where infrastructure or workflow enablement dominates.","exclusion":"If the workflow principally performs clinical reasoning or medical QA, assign to medical QA and clinical reasoning; if the resource is primarily an evaluation benchmark, assign to evaluation and assurance; if the core contribution is a trained domain model, assign to domain and resource-efficient model adaptation.","nearest_boundary":"Medical QA, dialogue, and clinical reasoning","representative_topics":["data-centric meta-dataset curation for biomedical language model pretraining and evaluation","low-resource clinical information extraction workbench for german biomedical text","corpus construction and information extraction for clinical trial eligibility criteria query generation"]}],"split":"dev"}
{"episode_id":"F275_T2026-01","field_index":275,"field":"biomedical natural language processing","forecast_date":"2026-01-01","history_windows":[{"start":"2024-07-01","end":"2025-01-01"},{"start":"2025-01-01","end":"2025-07-01"},{"start":"2025-07-01","end":"2026-01-01"}],"future_window":{"start":"2026-01-01","end":"2026-07-01"},"directions":[{"direction_index":0,"label":"D0","name":"Structured biomedical text extraction and coding","description":"Methods whose primary objective is converting biomedical or clinical text into structured entities, relations, events, normalized concepts, medication-safety facts, or standardized clinical codes.","inclusion":"Include NER, entity linking, relation/event extraction, temporal or attribute extraction, prompt-based extraction, low-resource entity normalization, adverse-event or medication-change extraction, and ICD/SNOMED/DRG-style code assignment when the scientific target is structured fact or code recovery from text.","exclusion":"If extracted variables are mainly used to characterize patients or predict outcomes, assign to patient phenotyping; if structured knowledge is used mainly as a representation or reasoning substrate, assign to knowledge-enhanced representation.","nearest_boundary":"Patient phenotyping and outcome modeling","representative_topics":["open set medical entity linking and span extraction via label-attention transformer without span annotations","joint learning for causal relation extraction and entity function detection in biomedical text","bert-based extraction of medication change attributes from clinical narrative text"]},{"direction_index":1,"label":"D1","name":"Domain and resource-efficient model adaptation","description":"Research whose primary objective is adapting language-model representations to biomedical or clinical domains, languages, vocabularies, compute constraints, or scarce annotation regimes.","inclusion":"Include domain-specific or continual pretraining, compact biomedical transformers, vocabulary transfer, multilingual or cross-lingual transfer, prompt or parameter-efficient adaptation under scarcity, and model adaptation where the model or transfer strategy is the central contribution rather than a downstream task system.","exclusion":"If the work primarily builds an extraction, QA, summarization, coding, or diagnosis system, assign to that task objective; if it primarily compares model capabilities, assign to evaluation and assurance.","nearest_boundary":"Evaluation, assurance, and privacy-preserving NLP","representative_topics":["domain-specific conversational language pre-training for low-resource clinical dialogue comprehension","low-resource german medical named entity recognition via pretrained language model transfer learning","french clinical language model adaptation via domain-specific pretraining on electronic health records"]},{"direction_index":2,"label":"D2","name":"Medical QA, dialogue, and clinical reasoning","description":"Systems whose primary objective is answering biomedical or clinical questions, conducting medical dialogue, supporting diagnosis or differential reasoning, or producing patient-specific clinical rationales.","inclusion":"Include retrieval-augmented medical QA, clinical semantic parsing, medical dialogue fine-tuning, LLM-based clinical explanation, diagnostic reasoning, guideline- or case-based reasoning, and decision-support workflows where reasoning or question answering dominates.","exclusion":"If the objective is information-access infrastructure without reasoning, assign to data resources and operational systems; if the primary output is a summary rather than an answer, dialogue, or decision rationale, assign to biomedical generation and communication.","nearest_boundary":"Data resources and operational workflow systems","representative_topics":["query-focused extractive summarisation for complex biomedical question answering with retrieval-augmented sentence scoring","multi-dimensional clinical explanation generation using text generation models as implicit knowledge bases","neural semantic parsing for clinical question answering over electronic health records"]},{"direction_index":3,"label":"D3","name":"Evaluation, assurance, and privacy-preserving NLP","description":"Research whose primary objective is measuring, benchmarking, auditing, or safeguarding biomedical NLP systems, including fairness, bias, robustness, privacy, de-identification, anonymization, and federated evaluation.","inclusion":"Include benchmark construction, shared-task evaluation, comparative model studies, reliability and safety assessment, bias and fairness analyses, de-identification and pseudonymization, anonymization-impact studies, and model-to-data or federated evaluation infrastructure.","exclusion":"If privacy or evaluation is only incidental to introducing a task system or adapted model, assign to that system's primary objective; if the work mainly creates training corpora rather than evaluation or assurance mechanisms, assign to data resources and operational systems.","nearest_boundary":"Data resources and operational workflow systems","representative_topics":["federated model-to-data sandbox for unbiased multi-site clinical nlp evaluation","gender bias analysis in clinical word embeddings for stereotypical medical categories","anonymization technique selection for clinical text nlp under re-identification constraints"]},{"direction_index":4,"label":"D4","name":"Biomedical generation, summarization, and simplification","description":"Methods whose primary objective is generating, summarizing, simplifying, or adapting biomedical and clinical text for clinicians, researchers, patients, or lay audiences.","inclusion":"Include clinical note or report summarization, counseling or dialogue summarization, biomedical literature summarization, keyphrase generation, readability-controlled summarization, plain-language adaptation, patient-facing simplification, and biomedical text-generation toolkits when generation quality is central.","exclusion":"If generated text is primarily an answer or decision rationale, assign to medical QA and clinical reasoning; if generated data are primarily for dataset augmentation, assign to data resources and operational systems.","nearest_boundary":"Medical QA, dialogue, and clinical reasoning","representative_topics":["mental health knowledge-guided utterance filtering for counseling dialogue summarization","domain-adaptive seq2seq summarization of patient problems from clinical progress notes","readability-controllable abstractive summarization of biomedical documents for expert and lay audiences"]},{"direction_index":5,"label":"D5","name":"Patient phenotyping and social-behavioral assessment","description":"Research whose primary objective is using health-related language to characterize patients, infer diagnoses or disease states, monitor outcomes, identify social determinants, or assess neurological, mental-health, pain, or communication-related conditions.","inclusion":"Include clinical phenotyping, disease or symptom classification, mortality/readmission/utilization prediction from notes, cohort expansion, SDOH and equity-relevant variable extraction, chronic pain and dementia language assessment, mental-health or counseling state assessment, and language-based assistive assessment for communication disorders.","exclusion":"If the output is a standardized administrative code, assign to structured extraction and coding; if the work centers on LLM diagnostic reasoning or decision rationales rather than phenotype prediction or assessment, assign to medical QA and clinical reasoning.","nearest_boundary":"Structured biomedical text extraction and coding","representative_topics":["nlp-based automated diagnosis from clinical transcription reports using hybrid deep learning models","transformer-based clinical note embedding for mortality prediction via umls concept mapping","language model classification and facet-based summarization for chronic pain subtype detection"]},{"direction_index":6,"label":"D6","name":"Knowledge-enhanced biomedical representation","description":"Methods whose primary objective is incorporating biomedical ontologies, knowledge graphs, controlled vocabularies, or graph-structured medical knowledge into representations, reasoning substrates, semantic similarity, or knowledge discovery.","inclusion":"Include KG- or ontology-enhanced language understanding, concept and code embeddings, ontology description generation, graph-augmented semantic similarity, knowledge-enhanced text classification, literature-based discovery, and representation learning where structured knowledge is central.","exclusion":"If the work merely links mentions to an ontology as an extraction output, assign to structured extraction and coding; if structured knowledge is only a small pretraining signal for a general domain model, assign to domain and resource-efficient model adaptation.","nearest_boundary":"Structured biomedical text extraction and coding","representative_topics":["graph-in-graph network for gene ontology term description generation using nested gene and term graphs","graph-augmented cyclic co-training for clinical text semantic similarity estimation","knowledge graph enhanced bert for multi-type clinical text classification"]},{"direction_index":7,"label":"D7","name":"Data resources and operational workflow systems","description":"Research whose primary objective is building reusable biomedical NLP datasets, corpora, annotation resources, toolkits, retrieval pipelines, trial-matching systems, literature-search systems, or deployed domain-specific NLP applications.","inclusion":"Include corpus construction, dataset harmonization, synthetic or augmented data resources, annotation pipelines, clinical-trial eligibility corpora, systematic-review query generation, biomedical literature retrieval, patient-trial matching, specialized clinical NLP workbenches, fact-checking resources, and operational application pipelines where infrastructure or workflow enablement dominates.","exclusion":"If the workflow principally performs clinical reasoning or medical QA, assign to medical QA and clinical reasoning; if the resource is primarily an evaluation benchmark, assign to evaluation and assurance; if the core contribution is a trained domain model, assign to domain and resource-efficient model adaptation.","nearest_boundary":"Medical QA, dialogue, and clinical reasoning","representative_topics":["data-centric meta-dataset curation for biomedical language model pretraining and evaluation","low-resource clinical information extraction workbench for german biomedical text","corpus construction and information extraction for clinical trial eligibility criteria query generation"]}],"split":"dev"}