| {"dataset":"CreativeEval","example_id":"1JPfHljXL4","group_id":"1JPfHljXL4","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"When minimizing a function, stochastic gradient descent (SGD) iteratively updates the current point by subtracting the product of a learning rate and a (possibly stochastic) sub-gradient of the function. The choice of learning rates plays a central role in optimization, and the learning rate sequence is typically divided into two parts: a baseline rate that represents the maximum value and a schedule that multiplies this baseline to yield the final sequence of rates. In our work, we focus exclusively on the scheduling aspect.\nChoosing an optimal learning rate schedule is challenging, and practitioners often resort to exploring a range of standardized schedules through hyper-parameter tuning. Theoretical guidance on how to set the learning rates is complicated by various assumptions that lead to different schedules—for example, schedules that decay as the reciprocal of the square root of time, as the reciprocal of time, or remain constant. In practice, however, these theoretical schedules tend to perform suboptimally for deep learning, making them less popular in real applications.\nWe identify two major reasons for this discrepancy between theory and practice. First, traditional analyses typically consider the average iterate or a randomly selected one, whereas in practical settings the final iterate is used. Second, existing theories for the final iterate rely on coarse, constant bounds on gradient norms or curvature, whereas our approach incorporates the entire sequence of gradient norms. This leads to learning rate schedules that more effectively adapt to the structure of the problem.\nOur method is a refinement approach that uses information from a previous training run—specifically, the sequence of gradient norms—to produce an improved learning rate schedule for future runs. In our approach, the new schedule is derived by first smoothing the recorded gradient norms and then defining weights as the inverse squared values of these smoothed norms. At each step, the learning rate is determined by accumulating the contributions from these weights over the remaining iterations. This process leads to a schedule that is highly adaptive and represents a significant departure from traditional methods.\nMoreover, our analytical framework is not restricted to SGD. Rather than relying on the specific form of the SGD iterates, our technique generalizes to any base optimization algorithm, including widely used methods like Adam. Technically, we design a step-size schedule that transforms any algorithm with vanishing regret into one that guarantees a strong performance for the final iterate.\nIn summary, our contributions are both theoretical and practical. On the theory side, we provide a general analysis of learning rate schedules that not only recovers the optimal convergence rates for SGD but also yields task-specific, refined schedules. On the practical side, we demonstrate that a schedule incorporating an initial warm-up phase followed by nearly linear decay consistently matches or outperforms commonly used alternatives such as cosine decay. Notably, our theory directly motivates the inclusion of a warm-up phase, which has traditionally been used only as an empirical heuristic.\nIn the following sections, we introduce our notation, describe our analytical approach and results in detail, and discuss how our method applies to various optimization algorithms."},"source":{"paper_id":"1JPfHljXL4","cohort_label":"uncreative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0,3.0,4.0],"soundness":[2.0,3.0,3.0,4.0,4.0],"presentation":[3.0,3.0,3.0,3.0,3.0],"overall_score":[5.0,6.0,5.0,5.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.8,"soundness_mean":3.2,"presentation_mean":3.0,"overall_score_mean":5.8},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"1XDG1Z5Nhk","group_id":"1XDG1Z5Nhk","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Large-scale pre-training across various applications has demonstrated the importance of scalable, cost-effective models. Recent advances in sparsely activated networks, known as Mixture-of-Experts (MoE), have attracted significant attention. Unlike traditional networks that activate all modules for every input, MoE selectively activates only portions of the network through expert routing, resulting in notable efficiency improvements. However, this efficiency comes with a challenge: the discrete nature of expert routing produces non-differentiable outputs, while backpropagation relies on differentiable functions to compute gradients.\nSeveral methods have been proposed to bridge the gap between discrete functions and backpropagation, often using a straight-through approach. Unfortunately, these estimators typically require activating all experts for gradient computation, which eliminates the efficiency gains of MoE. As a result, standard MoE training usually omits the gradient computation for routing, sacrificing some training signals for the sake of sparse computation. This trade-off, while improving scalability, may lead to slower convergence and suboptimally trained models.\nTo address these issues, we introduce SparseMixer, a novel approach designed to bridge the gap between sparse MoE routing and backpropagation. Drawing inspiration from numerical methods for ordinary differential equations, SparseMixer offers a reliable gradient approximation for expert routing even when only a subset of experts is activated. It employs the mid-point method—a second-order numerical ODE solver—that accurately matches the Taylor expansion of the gradient up to the second order, all without requiring the Hessian or other second-order derivatives.\nWe demonstrate the effectiveness of SparseMixer by applying it to the Switch Transformer in both pretraining and neural machine translation tasks. SparseMixer not only accelerates training convergence by up to twice as fast, but it also leads to better-trained expert routing. Remarkably, even though the Switch Transformer typically underperforms compared to dense models in various pretraining setups, adding SparseMixer as the gradient estimator consistently enables the MoE models to surpass the dense baseline."},"source":{"paper_id":"1XDG1Z5Nhk","cohort_label":"uncreative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,3.0,2.0],"soundness":[3.0,3.0,3.0,3.0],"presentation":[3.0,2.0,3.0,3.0],"overall_score":[3.0,6.0,5.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":3.0,"presentation_mean":2.75,"overall_score_mean":4.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"2XkTz7gdpc","group_id":"2XkTz7gdpc","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Graphs are mathematical structures used to represent relational data through sets of nodes and edges. This abstraction is widely employed in various domains such as social networks, program synthesis, and even origami design. A key task in this area is generating new graphs that maintain the characteristics observed in real-world examples—ranging from modeling the structure of proteins to representing molecules in drug discovery.\nTraditional methods for graph generation rely on estimating parameters of well-known models like the Stochastic Block Model or Erdos-Renyi graphs. However, these methods often fail to capture the complexity of real-world data. Deep learning offers a compelling alternative, with techniques that either build graphs incrementally by predicting edges with each new node or generate the entire graph at once using methods such as variational autoencoders, generative adversarial networks, normalizing flows, and diffusion models. Despite their success with graphs containing several hundred nodes, these approaches face challenges when scaling to larger graphs. In particular, predicting edges between all pairs of nodes incurs a computational cost that grows quadratically, an issue for sparse graphs. Additionally, autoregressive methods must contend with the vast number of potential node orderings, and one-shot methods often struggle to simultaneously capture both global and local structures. Moreover, neither type of method has demonstrated a reliable ability to generalize to larger, unseen graph sizes, and many neural architectures either lack sufficient expressiveness or are computationally expensive.\nOur work introduces a novel graph generation approach based on iterative local expansion. In each step, selected nodes are expanded into small subgraphs, and a diffusion model is employed to reconstruct the appropriate local structure. The model is trained to reverse a graph coarsening process applied to the input data. This strategy is inherently suited for graph generation because it first establishes an approximate global structure and then refines it by adding local details, effectively mimicking a type of network growth that is more robust to variations in graph size. By modeling edge distributions locally rather than representing the entire joint distribution over all node pairs, our method maintains scalability for larger graphs. Theoretical analysis shows that, under mild conditions, our approach achieves subquadratic sampling complexity with respect to the number of nodes for sparse graphs. \nFurthermore, we propose Local PPGN, a more efficient variant of the Provably Powerful Graph Network, which preserves high expressive power for local subgraphs while improving computational efficiency. Our experiments on benchmark datasets demonstrate that our model not only achieves state-of-the-art performance in standard graph distribution tasks—delivering the highest Validity-Uniqueness-Novelty scores on planar and tree datasets—but also generates graphs that closely match the structural statistics of protein and point cloud datasets. In tests involving the generation of graphs with an unseen number of nodes, our method uniquely preserves the defining characteristics of the training data. Empirical results further confirm that our model exhibits subquadratic sampling complexity for sparse graphs. Our implementation is available for further exploration and validation."},"source":{"paper_id":"2XkTz7gdpc","cohort_label":"creative","pairing_group":"4","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,3.0,2.0],"soundness":[3.0,3.0,3.0,2.0],"presentation":[2.0,2.0,2.0,2.0],"overall_score":[6.0,5.0,8.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":2.75,"presentation_mean":2.0,"overall_score_mean":6.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"4FUa5dxiiA","group_id":"4FUa5dxiiA","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Casting reinforcement learning (RL) as probabilistic inference offers a formal framework for developing novel RL algorithms in complex and continuous domains. This approach, known as RL-as-inference, addresses the exploration–exploitation trade-off by adapting the policy to account for posterior uncertainty while providing flexible models that incorporate task knowledge and offer effective tools for inference over quantities of interest. These properties are especially valuable when an agent must handle partial observability and when data collection is costly, making sample efficiency crucial.\nA key difference between classical RL and RL-as-inference lies in their optimization goals. Classical RL seeks a policy that produces trajectories with the maximum expected return, whereas RL-as-inference searches for trajectories that maximize the probability of being optimal. However, the latter objective can sometimes lead to risk-seeking behavior, as the agent tends to favor high potential returns by overlooking low-probability transitions.\nTo mitigate this risk-seeking behavior, two main variational approaches have emerged within the RL-as-inference framework. The first approach removes control from the variational dynamics, leading to model-free algorithms that directly map states to actions using only samples. This maximum entropy RL method constrains the variational approximation of the posterior dynamics, though it provides looser bounds on the learning objective. The advantage of this model-free strategy is that it avoids the need to explicitly evaluate model dynamics, even though it often requires many interactions with the environment. Numerous algorithms have been developed based on this approach, including methods based on expectation maximization and variational policy search.\nThe second approach, in contrast, does not restrict the variational dynamics, allowing for more flexible approximations and tighter bounds on the learning objective. These methods are model-based since they learn a predictive model for the posterior dynamics. For example, variational model-based policy optimization (VMBPO) optimizes a variational bound using an expectation maximization style procedure, and another approach—termed “Mismatched No More”—optimizes a similar objective but modifies the reward function. Although these methods tend to be sample efficient, the learned dynamics can be overly optimistic and risk-seeking, often resulting in policies that fail to generalize well to the real environment.\nThe advancement of model-based approaches such as VMBPO has been limited, partly due to insufficient study of the risk-seeking behavior in the learned policies. In response, we present a comprehensive investigation of risk sensitivity within the model-based RL-as-inference framework. Specifically, we introduce β-VMBPO, a generalization of VMBPO that adjusts policy risk by modulating the allowable divergence between posterior and prior dynamics during learning. This method augments the conventional log-likelihood objective with a hyperparameter β, which interpolates between traditional RL and the RL-as-inference objective. We analyze the impact of β on learning variational dynamics and on policies that generalize to the actual environment, and propose an automatic tuning method based on Lagrangian optimization. We also compare our method to other RL-as-inference algorithms, including VMBPO, a model-based policy optimization algorithm, and soft actor critic.\nTo further support development in this area, we provide the first open-source implementation of both VMBPO and our generalized β-VMBPO. Additionally, we offer a thorough theoretical analysis of the risk-sensitive RL-as-inference framework, demonstrating that the proposed Bellman operator is a contraction and establishing a generalized policy improvement theorem. Based on these findings, we prove that an optimal deterministic policy exists and show that such a deterministic policy is optimal under the β-VMBPO objective."},"source":{"paper_id":"4FUa5dxiiA","cohort_label":"uncreative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,2.0],"soundness":[3.0,3.0,2.0],"presentation":[4.0,3.0,3.0],"overall_score":[6.0,6.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6666666666666665,"soundness_mean":2.6666666666666665,"presentation_mean":3.3333333333333335,"overall_score_mean":5.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"4L0xnS4GQM","group_id":"4L0xnS4GQM","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Tables are a widely used data format in daily life, and understanding their structure with language models can benefit a variety of tasks such as fact verification and question answering based on tabular data. Unlike plain text, tables deliver rich information through the interaction of rows and columns. This increased data capacity enhances the amount of information available but also makes it more challenging for language models to interpret them correctly. As a result, reasoning over tabular data has become an important area in natural language processing, attracting significant attention from both academia and industry.\nRecent work has explored different approaches to improve table understanding by training language models. One common method involves incorporating specialized embedding layers or attention mechanisms into the models and pre-training them to recover table cells or segments, making the models aware of the underlying table structure. Another approach relies on synthesizing SQL query-response pairs to pre-train an encoder-decoder model that functions as a neural SQL executor.\nLarge language models have recently achieved remarkable performance across a range of tasks using only prompting, thanks to their extensive pre-training. Advances in prompting techniques, such as those that design reasoning chains (for example, Chain-of-Thought, Least-to-Most, Program-of-Thought, and Tree-of-Thought), have further enhanced the reliability of these models. Although some studies have explored using large language models to solve table-based problems, many of these approaches rely on free-form text or code to represent reasoning steps. This can be problematic when dealing with complex tables, as illustrated by cases where the reasoning fails either because the multi-step process cannot handle the table’s complexity or because generated programs cannot correctly parse merged cell contents.\nIn contrast, reasoning on tables typically requires a series of intermediate steps that correspond to specific tabular operations. To address this, we propose an approach called Chain-of-Table. This method conducts step-by-step reasoning by applying a series of tabular operations that transform the table gradually. Each resulting table in the chain reflects intermediate reasoning results, similar to the idea of Chain-of-Thought. We define a set of common table operations—such as adding columns, selecting rows, and grouping data—that are frequently used in SQL and DataFrame development. At each step, the language model dynamically generates an operation along with its necessary arguments, and the operation is then executed on the table programmatically. This can either enrich the table with detailed intermediate results or condense it by removing irrelevant information. Visualizing these intermediate tables is crucial for arriving at the correct prediction, as the updated table is fed back into the process for further processing. Ultimately, the Chain-of-Table method produces a final table from which the language model can derive the correct answer.\nWe validate the Chain-of-Table approach using three tabular benchmarks for tasks such as fact verification and question answering. Experiments demonstrate that our method generalizes well across different large language model options and achieves state-of-the-art performance on these datasets.\nOur contributions are as follows:\n• We extend the Chain-of-Thought concept to tabular data by transforming the input table to store intermediate results, thereby enabling a multi-step reasoning process that leads to more accurate table understanding.\n• Extensive experiments on table-based fact verification and question answering benchmarks show that the Chain-of-Table method achieves state-of-the-art performance."},"source":{"paper_id":"4L0xnS4GQM","cohort_label":"creative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0,4.0],"soundness":[3.0,3.0,3.0,3.0],"presentation":[3.0,3.0,3.0,3.0],"overall_score":[5.0,6.0,5.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.75,"soundness_mean":3.0,"presentation_mean":3.0,"overall_score_mean":5.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"4r2ybzJnmN","group_id":"4r2ybzJnmN","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Spiking neurons act as coincidence detectors—they respond more strongly to synchronous input spikes than to asynchronous ones. It is important to note that what matters for detection is the spike arrival time, not the time at which the spike is emitted. Because transmission is not instantaneous, there is a delay between when a spike is emitted and when it is received. This delay varies across connections and results from the sum of axonal, synaptic, and dendritic delays. Depending on the specific factors involved, this delay can be as short as 1 millisecond or extend to several tens of milliseconds. For example, axonal delays can be reduced through myelination, an adaptive process critical for learning. Thus, learning in the brain involves more than just synaptic plasticity; delay learning is also essential.\nTheoretical studies have shown that a spiking neural network with adjustable delays can compute a much richer class of functions than one limited to adjustable weights. Moreover, many neuromorphic chips today feature synapses with programmable delays.\nThese observations have motivated us and others to propose learning rules that incorporate delays. In our work, we demonstrate that delays can be learned along with synaptic weights using backpropagation in arbitrarily deep spiking neural networks. We first reveal a mathematical equivalence between one-dimensional temporal convolutions and connection delays. Building on this equivalence, we then show that delays can be learned using a method called Dilated Convolution with Learnable Spacings, originally developed to increase receptive field sizes in conventional convolutional neural networks. Our approach is fully integrated with PyTorch and leverages its automatic differentiation engine."},"source":{"paper_id":"4r2ybzJnmN","cohort_label":"creative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,4.0,2.0,2.0],"soundness":[3.0,3.0,3.0,2.0],"presentation":[3.0,3.0,3.0,3.0],"overall_score":[6.0,8.0,8.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":2.75,"presentation_mean":3.0,"overall_score_mean":7.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"5e0yWSNGIc","group_id":"5e0yWSNGIc","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Recent advancements in deep neural networks have expanded their use across various areas such as image classification, natural language processing, speech recognition, and self-learning systems. In particular, the emergence of deep reinforcement learning—where deep neural networks serve as function approximators—has broadened their application from pharmaceuticals to self-driving cars.\nAs research progressed, attention turned to the vulnerabilities of these networks to specifically crafted perturbations derived from the training cost function. Some studies focused on generating optimal, norm-bounded perturbations to inflict maximum damage, while a significant body of work developed methods to enhance network robustness. Deep neural policies, like their classification counterparts, are also susceptible to these adversarial perturbations, prompting the machine learning community to seek ways to improve their resilience.\nIn this paper, we examine adversarially trained deep neural policies and the state-action value functions they learn when under attack. We address several key questions: How accurate is the state-action value function in estimating values for state-action pairs in environments with high-dimensional representations? How does adversarial training affect these estimates? What are the effects of training under worst-case distributional shifts on the representation of optimal actions? And are there inherent trade-offs associated with explicit worst-case regularization during policy training?\nTo answer these questions, we compare the state-action values learned by adversarially trained policies with those learned by conventionally trained policies. Our analysis provides a solid theoretical explanation of the subtle impacts that certified adversarial training has on the state-action value function. Through experiments on Atari games from the Arcade Learning Environment, we demonstrate that conventionally trained policies offer a more accurate representation of sub-optimal actions. We also identify inconsistencies in the action ranking produced by adversarially trained policies, highlighting a fundamental trade-off—these policies tend to overestimate state-action values and lose critical information. Ultimately, our findings challenge previous hypotheses on the relationship between the action gap and overestimation."},"source":{"paper_id":"5e0yWSNGIc","cohort_label":"uncreative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,2.0],"soundness":[3.0,3.0,3.0],"presentation":[3.0,3.0,3.0],"overall_score":[6.0,5.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.3333333333333335,"soundness_mean":3.0,"presentation_mean":3.0,"overall_score_mean":5.333333333333333},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"6iwg437CZs","group_id":"6iwg437CZs","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"In this work we focus on improving multivariate time series prediction by incorporating additional task-specific information, an important goal given the wide range of real-world applications. Multivariate time series prediction presents a unique challenge because of its multi-dimensional sequential structure and noise sensitivity. A robust model must not only capture the correlations present among different series at each time step, but also understand the complex dynamics within each individual series over time. Moreover, many existing time series models struggle to respond quickly to sudden or rare events—such as financial crises or extreme climate changes—which can significantly degrade their performance.\nTo address these challenges, we introduce STanHop-Net (Sparse Tandem Hopfield Network), a novel deep learning model based on Hopfield networks that incorporates optional memory-enhanced capabilities. Our design is inspired by the connection between associative memory models in the human brain and the attention mechanism. By leveraging the memory retrieval properties of modern Hopfield models, our approach integrates external information, such as real-time or auxiliary data, directly into the prediction process via a differentiable memory module.\nAt the core of our method, we extend the deep-learning-compatible Hopfield layers. Unlike traditional transformer-based architectures, these layers not only replace conventional attention mechanisms but also function as memory modules that can retrieve relevant external cues to enhance predictions. Specifically, we first introduce a set of generalized sparse Hopfield layers—an extension of the sparse modern Hopfield model. Building on these layers, we propose the STanHop block, which consists of two sequential sub-blocks of these layers. This tandem design enables the model to learn and store both temporal and cross-series representations in a sparse, step-by-step manner.\nWe further develop STanHop-Net by stacking multiple STanHop blocks to enable multi-resolution representation learning. In this architecture, each block is capable of incorporating additional external information via the Hopfield memory retrieval mechanism. This means that at every level of resolution, the model can inject useful external memory, leading to improved predictive performance with very little additional effort from the user.\nOur contributions are as follows:\n• Theoretically, we introduce a unified, sparsity-aware modern Hopfield model—the generalized sparse Hopfield model. This model not only provides a tighter error bound during memory retrieval compared to its dense counterpart, but also retains essential theoretical properties such as fast fixed-point convergence and exponential memory capacity. Furthermore, it subsumes both the sparse and dense models as special cases.\n• Computationally, we demonstrate that a one-step approximation of the retrieval dynamics of this model is connected to sparse attention mechanisms. This insight allows us to develop learnable sparsity within our Hopfield layers, resulting in faster memory retrieval and enhanced robustness to noise.\n• Methodologically, we propose the STanHop block—a hierarchical design that captures the intrinsic multi-resolution structure of both temporal and cross-series dimensions in time series data, with resolution-specific sparsity at each level. In addition, we introduce the concept of pseudo-label retrieval, and we present two schemes for incorporating external memory: the Plug-and-Play and Tune-and-Play modules.\n• Experimentally, we validate STanHop-Net on multivariate time series prediction tasks, both with and without external memory integration. In scenarios without external memory, our model consistently meets or exceeds the performance of well-known baselines on diverse real-world datasets. When external memory is incorporated, STanHop-Net achieves even greater performance improvements.\nThis work provides visual overviews of both STanHop-Net and the STanHop block to aid in understanding their architectures and functionalities."},"source":{"paper_id":"6iwg437CZs","cohort_label":"creative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,2.0,3.0],"soundness":[3.0,3.0,2.0,3.0],"presentation":[2.0,3.0,2.0,3.0],"overall_score":[5.0,8.0,5.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.75,"presentation_mean":2.5,"overall_score_mean":6.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"6ssOs9BBxa","group_id":"6ssOs9BBxa","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Deep reinforcement learning (DRL) has shown remarkable success in solving complex, multi-step problems, including playing video games, managing continuous control tasks, and handling real-time strategy (RTS) games like StarCraft II. Training challenges in RTS games arise from several factors: the observation and action spaces are large and varied due to differing terrains and unit types; each unit type has distinct actions and abilities; individual actions may control multiple units simultaneously; rewards are both sparse and delayed, sometimes taking thousands of timesteps to appear; success depends on a blend of tactical (micro) and strategic (macro) decisions; decisions must be made in real time; the agent may have only partial visibility of the game state (as with fog of war); and game events can be non-deterministic.\nmicroRTS (stylized as µRTS) is a minimalist, open-source, two-player, zero-sum RTS game testbed developed for research. It incorporates many core aspects of traditional RTS games—including different unit types with specific actions, terrain effects, resource gathering, unit production, and combat scenarios where units have distinct strengths and weaknesses—while simplifying them. Although microRTS supports features like fog of war and non-determinism, these were disabled for the IEEE-CoG2023 microRTS competition.\nThe IEEE microRTS competitions have been held at the Conference on Games in recent years and previously at the Conference on Computational Intelligence and Games. In these tournaments, competitors submit agents that face off against others in a round-robin format on 12 maps—8 available beforehand and 4 hidden until after the results are released. Agents are required to select their actions within 100 milliseconds each step, which poses a significant challenge for deep neural network-based approaches without GPU acceleration.\nThis paper details how the AnonymizedAI agent was developed and trained to become the first DRL agent to win the microRTS competition at CoG in 2023. The agent comprises seven trained neural networks and required 70 GPU-days to train. Although the long training times and inherent difficulties in debugging and fine-tuning DRL implementations may help explain why DRL methods have lagged in competitiveness, our findings highlight that transfer learning to specific maps was critical to our success. We believe this training framework and transfer learning strategy can serve as a foundation for future research and competitive agents.\nEven though microRTS does not support human players, past competitions have provided several agents for imitation learning. Subsequent work indicates that combining behavior cloning with fine-tuning using DRL can yield competitive agents more efficiently. By training critic components using solely win-loss rewards from the same playthroughs, we demonstrate that DRL can be effectively trained with sparse rewards, thereby eliminating the need for a handcrafted reward function."},"source":{"paper_id":"6ssOs9BBxa","cohort_label":"uncreative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,3.0,2.0],"soundness":[2.0,4.0,3.0,3.0],"presentation":[2.0,3.0,2.0,3.0],"overall_score":[3.0,5.0,5.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":3.0,"presentation_mean":2.5,"overall_score_mean":4.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"92KV9xAMhF","group_id":"92KV9xAMhF","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Generative models create data from noise by learning a mapping from a noisy latent space to a structured data space. Different models employ different mappings and learning procedures. For example, Normalizing Flows use a one-to-one mapping optimized via maximum likelihood, while Generative Adversarial Networks rely on a non-bijective mapping trained using a divergence between the model and data distributions. Recently, diffusion models have achieved remarkable success across various fields, even surpassing GANs in generating visually high-quality images. Unlike traditional models that learn a single latent-to-data mapping, diffusion models gradually introduce structure through an infinite series of denoising steps.\nThis process can be described using stochastic differential equations; starting with structured data, noise is incrementally added until the data is unstructured, and then the process is reversed by using the gradient of the true score function of the underlying stochastic process. In practice, a diffusion model is implemented as a neural network that approximates this score function. Even though the true score is derived from the gradient of the log density, the neural network is typically not restricted to be the gradient of any scalar function—that is, it is generally not a conservative vector field. This raises the question of whether the diffusion model should be constrained to be conservative by design.\nWe suggest that focusing on conservativity misses the point for density estimation and sampling tasks. Instead, one should ask whether it is sufficient to learn a broader family of vector fields—essentially, whether adding an additional function to the score function can still yield exact sampling and precise density estimation. We refer to this flexibility as the gauge freedom in diffusion models.\nThis gauge freedom leads to two key insights for practitioners. First, for the purposes of density estimation or sampling, there is no need to force the diffusion model to be conservative, provided the gauge freedom conditions are met exactly. Second, when it comes to analyzing local properties of the data manifold, such as estimating its intrinsic dimensionality, a conservative diffusion model is guaranteed to produce correct conclusions.\nThere has been considerable debate about whether diffusion models should be conservative. Some studies have found that constraining the model to be conservative yields performance similar to that of an unconstrained model, though doing so typically incurs extra computational cost. Other research shows that an unconstrained model might offer slight advantages in density estimation, yet enforcing conservativity can improve sampling performance in certain cases. Practical considerations include the ease of combining likelihood models and the ability to use more accurate numerical samplers, both of which benefit from a conservative approach. Interestingly, one study even found that for an unconstrained model to behave conservatively, the neural network’s internal structure must satisfy very specific conditions, suggesting a strong case for explicitly constraining the model when necessary.\nOverall, while there is no strict need to enforce conservative properties in diffusion models for density estimation and sampling, doing so can provide important benefits in specific applications, especially when investigating the fine-grained features of the data manifold."},"source":{"paper_id":"92KV9xAMhF","cohort_label":"creative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,3.0,2.0],"soundness":[2.0,3.0,4.0,2.0],"presentation":[2.0,2.0,4.0,3.0],"overall_score":[6.0,5.0,8.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.75,"presentation_mean":2.75,"overall_score_mean":6.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"9bmTbVaA2A","group_id":"9bmTbVaA2A","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Modern deep network architectures have become increasingly complex, raising concerns about the opacity of their decision-making processes. Early efforts to explain these networks involved post-hoc methods that assigned scores to input features based on their contribution to a prediction. However, such explanations often fail to align with human expectations, which typically involve high-level semantic concepts, and their reliability in accurately reflecting feature importance has been questioned. This has fueled a demand for deep learning methods that are interpretable by design.\nOne promising direction is Information Pursuit (IP). In this framework, users define a set of task-specific image queries that represent clear, semantically meaningful questions about the image. When an image is observed, IP selects queries one by one until the gathered responses are sufficient to make a prediction with high confidence. Importantly, the decision is made solely on these query answers rather than directly from the image, ensuring that the explanation faithfully reflects the model's reasoning. To keep the explanations concise, IP chooses each query to maximize the information gained about the prediction, given the answers obtained so far.\nA significant challenge with IP is its dependence on mutual information, which is difficult to estimate in high-dimensional settings. To address this, a variational approach known as Variational Information Pursuit (V-IP) was introduced. V-IP employs a deep network to directly learn the most informative next query from the data, thus bypassing the need for explicit mutual information estimation. However, V-IP still requires users to specify a relevant set of queries for the task and to rely on datasets with manually annotated query answers—a limitation that restricts its broader application.\nIn this paper, we focus on image classification tasks and overcome this limitation by leveraging large language models (LLMs) to generate interpretable queries and vision-language models (VLMs) to answer them. This strategy enables us to extend V-IP to any image classification task without the need for manually defining query sets or collecting detailed annotations. Inspired by recent advances in concept-based image classification, where LLMs like GPT have proven effective at generating discriminative semantic concepts, we integrate these concepts as interpretable queries into the V-IP framework. Nevertheless, a reliable mechanism is still needed to answer these queries.\nPrior work has used models like CLIP to associate semantic concepts with images by measuring the similarity between image and text embeddings. However, the continuous output scores provided by CLIP do not directly suit the V-IP framework because they can undermine the clarity of the resulting query-answer chains. For example, a score of 0.24 for the concept “mammal” leaves ambiguity about its presence in the image, complicating the interpretation of subsequent queries. Simply thresholding these scores into binary values often leads to inaccurate and noisy results.\nOur main contribution is the development of a Concept Question-Answering (Concept-QA) system. This system is trained to deliver a clear binary response—indicating whether a given concept is present in an image—by using soft pseudo-labels derived from both GPT and CLIP. We demonstrate that GPT complements the concept-image association scores from CLIP by drawing on its extensive text training to recognize which concepts are salient for identifying particular classes. Experiments on several image classification datasets show that our Concept-QA system performs competitively with current pretrained VLM models used for visual question answering, while being much more computationally efficient. Moreover, we find that when V-IP is trained on the binary answers provided by Concept-QA, it produces shorter and more interpretable query chains compared to when it relies on continuous CLIP scores.\nIn summary, our work makes the following contributions. First, we introduce a method to train a Concept-QA system that outputs binary answers regarding the presence of a concept in an image using pseudo-labels from GPT and CLIP without needing manually annotated training data. Second, we empirically validate Concept-QA’s effectiveness in accurately representing true concepts. Third, we extend V-IP’s applicability to image classification tasks—even those without pre-specified query sets or annotations—such as ImageNet. Finally, we demonstrate that V-IP, when provided with Concept-QA responses, yields more concise and interpretable query-answer chains."},"source":{"paper_id":"9bmTbVaA2A","cohort_label":"creative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,2.0,2.0],"soundness":[2.0,3.0,3.0,3.0],"presentation":[3.0,3.0,2.0,3.0],"overall_score":[6.0,5.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.75,"presentation_mean":2.75,"overall_score_mean":5.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"9ceadCJY4B","group_id":"9ceadCJY4B","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Recent advances in conversational large language models (LLMs) like ChatGPT have revolutionized artificial intelligence and natural language processing. These models are increasingly used as virtual assistants across various fields due to their proficiency in generating meaningful and relevant responses. However, despite their impressive language generation capabilities, LLMs can sometimes deliver inconsistent or inaccurate answers, which undermines the security and trustworthiness of applications built upon them.\nObservation reveals that while LLMs often provide accurate and reasonable responses early in a conversation, their performance may deteriorate when users express skepticism or disagree with their initial judgments. This raises an important question: How consistent are LLMs in their judgments when challenged with questioning, disagreement, or misleading input? In our work, we define judgment consistency as the coherence of a model’s answers to objective questions with clear-cut answers. Consistent responses are crucial for establishing user trust, ensuring predictable real-world performance, and confirming that the model truly understands the subject matter. They help prevent the dissemination of misinformation and avoid reinforcing biases in sensitive domains.\nInspired by educational “questioning strategies,” we introduce a follow-up questioning mechanism to evaluate the judgment consistency of conversational LLMs. Much like teachers who continue to question students to assess their real understanding, our approach begins with a correct response from the model and continues with multi-turn dialogues that include challenges, negations, or misleading prompts. A notable drop in performance following these follow-up queries indicates poor judgment consistency.\nOur framework consists of three types of follow-up questions—closed-ended, open-ended, and leading—organized into two forms: Direct and Progressive. The Direct Form employs a single type of question for further inquiry, similar to a teacher asking an additional or challenging question after an initial correct answer. The Progressive Form, on the other hand, uses all three types sequentially to determine whether the model’s correct answer is due to genuine understanding or mere chance.\nWe conducted extensive experiments on eight benchmarks spanning arithmetic, commonsense, symbolic, and knowledge reasoning tasks. In further tests, we applied the same settings to other models to assess the generality of the issue. Our empirical results show that these LLMs are highly sensitive to questioning; even after providing an accurate answer, a simple follow-up such as “Are you sure?” can cause significant performance drops. Through our observations, we found that LLMs often tend to flatter users, which in turn reduces their judgment consistency when faced with negations or misleading prompts.\nWe also investigated the effect of different temperature and prompt settings on judgment consistency, evaluating how the tone of a prompt might impact outcomes. Additionally, a detailed error analysis provided further insights into the underlying behaviors of these models. To address the observed issue, we explored various prompting strategies that can improve judgment consistency, though the level of enhancement varies across methods."},"source":{"paper_id":"9ceadCJY4B","cohort_label":"uncreative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0],"soundness":[3.0,3.0,3.0],"presentation":[3.0,4.0,3.0],"overall_score":[6.0,6.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.3333333333333335,"soundness_mean":3.0,"presentation_mean":3.3333333333333335,"overall_score_mean":5.666666666666667},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"9nT8ouPui8","group_id":"9nT8ouPui8","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"In recent years, diffusion models have achieved remarkable success across various areas of generative modeling—including image creation, text-to-image synthesis, audio and speech generation, graph construction, and 3D content production. These models have demonstrated the ability to generate diverse, high-quality, and often novel samples by effectively capturing and abstracting the characteristics of the training data.\nThe basic idea behind diffusion models is to use a forward process that progressively adds Gaussian noise to a sample until the data are nearly indistinguishable from pure noise. The process is designed so that the initial noisy distribution aligns closely with the actual data distribution, while the final distribution becomes approximately Gaussian. To generate samples, one reverses this noising process, which requires estimating the gradient of the log-density (or score) of the noisy data. This estimation is typically accomplished through a score model learned via denoising score matching.\nIt turns out that if one were to perfectly optimize this score model, the reverse process would merely reproduce the training data itself, essentially causing the model to memorize the data rather than generalize from it. Although this theoretical optimum is never exactly achieved in practice—since state-of-the-art models show strong generalization—this discrepancy raises important questions. Under what conditions do diffusion models inadvertently resemble the theoretical optimum and start memorizing the training data? Moreover, what factors influence such memorization behavior?\nThese concerns are not only academic; they have practical implications for issues like privacy and copyright infringement. For instance, there have been demonstrations of extracting training images from widely used models, which highlights the risks of unintentional data replication.\nTo address these questions, the paper presents an extensive empirical investigation into memorization in diffusion models. An analysis of one prominent model on a common dataset reveals that memorization tends to occur when training data sets are small, while the effect is largely undetectable with larger datasets. This observation led to the introduction of a metric called effective model memorization (EMM), which quantifies the number of training examples for which a diffusion model exhibits the same memorization behavior as the theoretical optimum after training.\nFurthermore, the study examines how different factors—such as the underlying data distribution, the model architecture, and the training procedure—influence memorization. A particularly surprising finding is that conditioning the training data on completely random and uninformative labels can trigger memorization. Under this conditioning, more than 65% of the samples generated by a diffusion model trained on a 50,000-image dataset turned out to be exact replicas of the training data, compared to virtually no memorization with the original labels.\nThe study offers both practical guidance for users of diffusion models and valuable insights for theoretical research into deep generative models."},"source":{"paper_id":"9nT8ouPui8","cohort_label":"uncreative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[1.0,3.0,2.0,2.0,2.0],"soundness":[3.0,3.0,3.0,2.0,3.0],"presentation":[3.0,4.0,3.0,2.0,3.0],"overall_score":[3.0,5.0,6.0,5.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.0,"soundness_mean":2.8,"presentation_mean":3.0,"overall_score_mean":4.8},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"AHgc5SMdtd","group_id":"AHgc5SMdtd","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Industrial anomaly classification (AC) and anomaly segmentation (AS) are essential computer vision tasks with a wide range of applications. Real-world anomalies can occur on various objects and textures, making the task challenging due to their high diversity. Many unsupervised AC/AS approaches have been introduced that rely on an entire dataset of normal images for training—often referred to as full-shot methods. In contrast, some methods use only a few normal images in what is known as the few-shot setting and still achieve promising accuracy. These techniques, which are based on one-class AC/AS strategies, generally follow a similar pipeline.\nMore recently, some approaches have explored zero-shot AC/AS by utilizing text prompts for anomaly measurement, opening a new direction in this field. In all the aforementioned methods, test images are compared against additional sources of information, such as normal training images or text prompts. However, the vast amount of normal content inherently present in the unlabeled test images is seldom exploited; for example, normal pixels constitute over 97% of the test images in one dataset and nearly 99.5% in another. Moreover, since anomalies are random and unpredictable, abnormal regions may vary significantly even within the same anomaly type. This observation suggests that both normal and abnormal cues available in unlabeled test images can be utilized for anomaly detection.\nInspired by these findings, we propose a novel zero-shot AC/AS approach named MuSc, which relies solely on the unlabeled test images, without requiring any normal training images or text prompts. Our method consists of a patch-level representation extractor, a corresponding anomaly scoring approach, and an image-level optimization step. First, we apply local neighborhood aggregation with multiple degrees on the patch tokens to represent each image patch—a process that effectively models anomalies of various sizes, making it suitable for both AC and AS tasks. Next, we implement a mutual scoring mechanism that leverages the unlabeled test images to assign anomaly scores. This straightforward mechanism produces high-quality anomaly scores for each image region based on the observation that normal image patches tend to find many similar patches across images, while abnormal patches do not. Furthermore, we explore the relationship between image-level features and propose a constrained image-level neighborhood re-scoring process to refine anomaly classification.\nExperiments on benchmark datasets demonstrate the effectiveness of our approach. Compared to state-of-the-art zero-shot methods, our approach significantly improves performance in anomaly classification and segmentation. Notably, it even surpasses many few-shot methods and rivals several full-shot methods.\nIn summary, our contributions are threefold:\n1. We propose the first method for industrial anomaly classification and segmentation that uses only unlabeled test images.\n2. We reveal and exploit the potential of normal and abnormal patches inherent in unlabeled images, introducing a novel mutual scoring mechanism.\n3. We achieve significantly improved performance over existing zero-shot methods, while also outperforming most few-shot approaches and matching some full-shot methods."},"source":{"paper_id":"AHgc5SMdtd","cohort_label":"creative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,3.0,1.0,3.0],"soundness":[4.0,3.0,3.0,1.0,2.0],"presentation":[4.0,3.0,3.0,1.0,2.0],"overall_score":[6.0,5.0,6.0,3.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.4,"soundness_mean":2.6,"presentation_mean":2.6,"overall_score_mean":5.2},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"BLGQ3oqldb","group_id":"BLGQ3oqldb","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Deep learning has advanced tremendously over the past decade thanks to the development of neural networks. These networks usually employ a feed-forward architecture, where interactions between variables mostly occur implicitly within internal layers using various neural mechanisms. Unfortunately, such interactions do not impose explicit logical constraints on prediction variables, and as a result, the predictions sometimes fail to meet certain structural requirements.\nFor instance, in the document understanding task, the goal is to segment tokens in a document image into blocks. This task can be formalized as predicting a binary coexistence value for every pair of tokens, where a value of one indicates that both tokens belong to the same block. A key logical constraint in this scenario is the transitivity of coexistence: if token A coexists with token B and token B coexists with token C, then token A must coexist with token C. However, typical neural networks predict each pair independently, often violating this constraint. Even more advanced regularization methods have difficulty enforcing it.\nThis paper investigates how to incorporate first-order logic constraints (FOLCs) into neural networks. Such logical constraints are essential not only for document understanding but also for various real-world tasks, including collective classification in graphs and structured text prediction. The challenge of modeling FOLCs lies in the need to represent complex dependencies among a vast number of propositional groundings. For example, with 512 tokens, there can be more than 262,000 coexistence variables that are interdependent through over a hundred million logical groundings. This problem fundamentally relates to weighted first-order model counting (WFOMC), which is known to be computationally intractable for moderately complex logic rules.\nOne common approach to modeling these constraints is using Markov Logic Networks (MLNs). MLNs combine logical rules with probabilistic modeling by using joint potentials to quantify constraint satisfaction. However, exact inference in MLNs is challenging because of the underlying computational complexity, and even approximate methods have traditionally suffered from inefficiencies. Previous attempts to integrate MLNs with neural networks still end up processing groundings individually and do not fully exploit the inherent structure and symmetries to speed up computation.\nTo address these issues, we propose a novel method called Logical Message Passing (LogicMP) for encoding first-order logic constraints. LogicMP operates as an efficient MLN inference algorithm that can be seamlessly integrated with any off-the-shelf neural network, effectively functioning as a modular neuro-symbolic layer. The key innovation is an accelerated mean-field iteration algorithm that approximates MLN inference while leveraging the structure and symmetries inherent in the logic framework. This iterative algorithm is designed to take advantage of parallel tensor computations, drastically reducing computational complexity compared to traditional approaches.\nWe evaluate LogicMP across three different domains: visual document understanding, collective graph classification, and textual sequence labeling. In the document understanding task, LogicMP is demonstrated on a challenging benchmark dataset featuring a massive number of interdependent variables, outperforming previous state-of-the-art methods. In collective classification tasks on established datasets, LogicMP shows significant speed improvements, running roughly ten times faster than other competitive MLN inference methods. Finally, in a text sequence labeling task, incorporating specific logical rules with LogicMP leads to improved performance over existing techniques.\nOur contributions are threefold. First, we introduce LogicMP—a modular, differentiable neural layer that effectively encodes first-order logic constraints into neural network predictions. Second, we design an efficient mean-field algorithm that exploits the structural properties of MLNs, enabling large-scale inference through parallel tensor computation. Third, we demonstrate the versatility and effectiveness of LogicMP on challenging real-world tasks spanning images, graphs, and text, consistently achieving performance improvements over current neuro-symbolic approaches."},"source":{"paper_id":"BLGQ3oqldb","cohort_label":"creative","pairing_group":"4","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,1.0,3.0,2.0],"soundness":[2.0,3.0,3.0,2.0],"presentation":[3.0,2.0,3.0,1.0],"overall_score":[6.0,5.0,8.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.0,"soundness_mean":2.5,"presentation_mean":2.25,"overall_score_mean":6.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"BoMvv7ypDF","group_id":"BoMvv7ypDF","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Sampling problems involve generating samples from a target distribution proportional to exp(–f*). A popular approach to addressing these problems is to use gradient-based Markov chain Monte Carlo methods, such as the Unadjusted Langevin Algorithm (ULA), Underdamped Langevin Dynamics (ULD), the Metropolis-Adjusted Langevin Algorithm (MALA), and Hamiltonian Monte Carlo (HMC). These methods can be interpreted as discretizations of the continuous Langevin dynamics or its variants, which, under appropriate regularity conditions on the energy function f*(x), converge to a unique stationary distribution matching our target.\nHowever, the convergence rate of these Langevin-based algorithms crucially depends on the target distribution. Ensuring convergence in polynomial time typically requires the target to possess favorable properties, such as strong log-concavity or compliance with log-Sobolev or Poincaré inequalities with a large constant. In contrast, for more general non-log-concave distributions, the convergence rate might depend exponentially on the problem dimension, or even fail to converge to the desired target, possibly leading to convergence only to a locally stationary distribution. This inefficiency of the Langevin-based methods raises the question: Can we develop a new sampling algorithm that achieves a non-exponential convergence rate for sampling from general non-log-concave distributions?\nInspiration for an alternative approach comes from recent studies that design samplers based on diffusion models. One such approach uses the reverse process of an Ornstein-Uhlenbeck (OU) process. In this framework, instead of letting a process that converges slowly to a stationary distribution starting from a simple initialization, we reverse an OU process that begins at the desired target distribution and converges to a standard Gaussian distribution in a much shorter time. Essentially, if one can accurately recover the reverse process of the OU process, the slow mixing issue of the standard Langevin dynamics can be avoided, paving the way for more efficient sampling algorithms.\nA critical step in recovering the reverse process is obtaining a reliable estimate for the score function (the gradient of the log-density) of the distribution at every intermediate time. Although a method called reverse diffusion sampling (RDS) was proposed for this purpose, it still suffers from an exponential dependency on the target sampling error, presenting significant challenges in practical applications."},"source":{"paper_id":"BoMvv7ypDF","cohort_label":"uncreative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[4.0,2.0,4.0,4.0,3.0],"soundness":[3.0,2.0,4.0,4.0,3.0],"presentation":[3.0,2.0,2.0,2.0,4.0],"overall_score":[8.0,5.0,8.0,5.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":3.4,"soundness_mean":3.2,"presentation_mean":2.6,"overall_score_mean":5.8},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"C36v8541Ns","group_id":"C36v8541Ns","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Deep neural networks are vulnerable to adversarial attacks—small, carefully crafted perturbations that can cause the model to produce incorrect predictions. This vulnerability is particularly concerning in high-stakes applications such as autonomous driving and medical diagnostics. To mitigate these risks, researchers have developed various defense strategies, including methods that enforce smoothness in model behavior and techniques known as randomized smoothing. In both approaches, an important measure of robustness is the certified robust radius, which quantifies how much one can perturb an input while ensuring that the model’s prediction remains unchanged. Similarly, the prediction margin—reflecting the model’s confidence in its prediction—serves as an indicator of how resilient the classifier is to input changes.\nOne way to enhance robustness is by controlling the sensitivity of the classifier through Lipschitz continuity. Essentially, a classifier with a small Lipschitz constant will exhibit slower changes in its output when the input is perturbed. By combining the concepts of prediction margin and Lipschitz continuity, earlier work established a bound on the certified robust radius that effectively mediates the trade-off between the classifier's confidence and its sensitivity to input variations.\nRandomized smoothing takes a different approach by convolving the classifier with a Gaussian noise distribution. This process yields a probabilistic classifier whose predictions account for random variations in the input. While Monte Carlo methods are used to approximate these predictions, they introduce randomness that must be carefully controlled. Various techniques, such as using confidence intervals or concentration inequalities, have been employed to ensure that the classifier reliably predicts the correct label despite this randomness. Traditional classification typically relies on an arg max decision rule; however, randomized smoothing involves applying the arg max after averaging the outputs of the base classifier, a subtle but important difference.\nIn our work, we revisit the entire decision-making process in randomized smoothing to better understand and exploit the interplay between prediction margin and variance. We propose a refined approach that uses alternative methods for mapping logits to probability vectors, thereby achieving a more favorable balance between margin and smoothness. Importantly, we examine how enforcing Lipschitz regularity on the base classifier impacts the performance of randomized smoothing and, consequently, the certified robust radius.\nOur contributions are as follows:\n• First, we employ a Gaussian functional inequality to elucidate how the Lipschitz constant of the base classifier affects the variance introduced by Monte Carlo sampling. This insight motivates our use of an empirical variance-based concentration inequality to manage uncertainty.\n• Second, we introduce an alternative mapping strategy in the form of an r-simplex, which leads to improved margins and a better-controlled Lipschitz constant in the smoothed classifier. We derive a new limit on the Lipschitz constant that depends on the noise variance, the simplex parameter, and the base classifier’s original Lipschitz constant, thereby linking robustness certificates obtained through randomized smoothing with those from deterministic methods.\n• Finally, we present the Lipschitz-Variance-Margin Randomized Smoothing (LVM-RS) method, which optimally balances Monte Carlo variance and prediction margin using our proposed mapping techniques. Our experimental results on datasets such as CIFAR-10 and ImageNet demonstrate that this approach achieves state-of-the-art robustness results.\nThis work provides a comprehensive reexamination of the decision process in randomized smoothing, offering both theoretical insights and practical improvements for enhancing the robustness of deep neural networks."},"source":{"paper_id":"C36v8541Ns","cohort_label":"creative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,3.0,3.0],"soundness":[3.0,3.0,2.0,3.0],"presentation":[3.0,2.0,2.0,3.0],"overall_score":[6.0,8.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":3.0,"soundness_mean":2.75,"presentation_mean":2.5,"overall_score_mean":6.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"DLJznSp6X3","group_id":"DLJznSp6X3","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Over the past decade, the field of machine learning has focused on training overparameterized networks and simply adding more layers. The definition of a “large” network has shifted from models with around 100 million parameters to those with hundreds of billions, making the computational cost of training these models prohibitive for most research groups. Despite the common practice of training models with many more parameters than available training examples, the theoretical justification for this overparameterization remains unclear.\nAlternative approaches to scaling have been explored, such as more compute-efficient optimization methods, retrieval-augmented models, and even training smaller models for longer periods. However, while these methods offer new trade-offs, they do not explain the need for overparameterized models nor do they make training these models more accessible. Some methods require complex training setups or significant computational resources, whereas others still demand hundreds of GPUs even when scaled down.\nTechniques like zero-redundancy optimizers, 16-bit training, 8-bit inference, and parameter-efficient fine-tuning have played a crucial role in reducing these barriers. In particular, parameter-efficient fine-tuning now allows billion-scale language or diffusion models to be fine-tuned on consumer hardware, raising the question of whether similar approaches can be applied to pre-training.\nIn this study, we introduce ReLoRA, a method that uses individually low-rank updates which accumulate over the course of training to effectively produce a high-rank network. Our experiments show that ReLoRA, despite using low-rank updates, achieves performance comparable to that of conventional training methods. The ReLoRA method includes an initial phase of full-rank network training, followed by low-rank updates, restarts, a variable learning rate schedule, and partial optimizer resets. We evaluate ReLoRA on transformer language models with up to 1.3 billion parameters and observe that its efficiency improves with increasing model size, making it a promising approach for training networks with multiple billions of parameters."},"source":{"paper_id":"DLJznSp6X3","cohort_label":"creative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,2.0,3.0],"soundness":[3.0,3.0,2.0,2.0],"presentation":[2.0,3.0,3.0,2.0],"overall_score":[5.0,6.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":2.5,"presentation_mean":2.5,"overall_score_mean":5.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"DP4NkPZOpD","group_id":"DP4NkPZOpD","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Machine learning innovations in structural and computational biology—especially the ability to predict protein structure from its amino acid sequence (protein folding) and to infer sequence from structure (inverse folding)—have led to groundbreaking advances in our understanding of protein form and function. While studying individual proteins provides essential insights, analyzing protein complexes, where multiple proteins interact, is crucial for deciphering the complexities of cellular processes and disease mechanisms. In drug development, for example, a key challenge is to design a protein that can bind to a specific target to form a functional complex.\nProtein generation encompasses several demanding tasks, including folding, inverse folding, and docking. Traditionally, these tasks have been treated as independent problems, with each tackled by specialized techniques that focus either on structural or sequence information. Recent approaches have applied methods such as manifold diffusion and regression-based models for protein-protein docking, while autoregressive models have predominantly addressed inverse folding. Additionally, existing diffusion models for protein generation often decouple the design of structure and sequence by using an inverse folding model after generating the structure. As a result, these separate approaches do not fully exploit the interconnected relationship between protein sequence and three-dimensional structure.\nIn this work, we introduce OMNIPROT, a generative model that inherently captures the relationship between sequence and structure and addresses the diverse challenges in protein generation in a unified framework. OMNIPROT consists of two main components: an autoencoder with a geometrically structured latent space and a diffusion model that operates within this latent space. The autoencoder is specifically designed for protein complexes, using roto-translational invariant features to produce latent representations that preserve both sequence and structural information. By performing diffusion in this latent space, OMNIPROT can be applied universally to conditional generative tasks in protein design—ranging from full-atom flexible protein-protein docking to inverse folding—simply by altering the conditioning information provided to the model.\nWe evaluated OMNIPROT on two common tasks in conditional generative protein design: protein-protein docking and inverse folding. In protein-protein docking, OMNIPROT achieved a 70% success rate, an improvement over the 63.3% success rate of the next-best machine learning-based approach. For inverse folding, it achieved a sequence recovery rate of 46.8% on a standard benchmark, which is close to the 48.8% obtained by the state-of-the-art method designed specifically for this task. We believe that OMNIPROT represents a significant step toward a unified approach for protein generation and design."},"source":{"paper_id":"DP4NkPZOpD","cohort_label":"uncreative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,1.0,3.0,4.0],"soundness":[3.0,2.0,3.0,4.0],"presentation":[3.0,2.0,3.0,3.0],"overall_score":[3.0,3.0,6.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":3.0,"presentation_mean":2.75,"overall_score_mean":5.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"DZ6B5u4vfe","group_id":"DZ6B5u4vfe","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Instruction-tuning is a widely adopted approach for fine-tuning large language models (LLMs) on datasets that include task-specific instructions. This method not only enables LLMs to generalize effectively to unseen tasks by learning to follow provided instructions but also significantly boosts reasoning and problem-solving abilities, sometimes approaching human performance—even when few or no training examples are used. Moreover, instruction-tuning teaches LLMs to mimic ground-truth outputs, often human-written, resulting in more controllable, human-aligned, and stylistically similar outputs.\nThis refined capability raises an intriguing question from a neuroscience perspective: Does instruction-tuning make LLMs more similar to the human language system? Previous research has shown that models with higher task performance tend to better align with human neural activity, sometimes even reaching the estimated noise ceiling. However, no studies have specifically examined how instruction-tuning, a key factor behind powerful LLMs like ChatGPT, influences this alignment.\nIn our work, we explore this by examining two aspects of alignment. The first, brain alignment, assesses how closely the internal representations of LLMs match neural activity patterns in the human language system. The second, behavioral alignment, evaluates the similarity between LLM behavior and human behavior. To make this comparison, both LLMs and human participants are presented with the same naturalistic language stimuli in the form of stories and sentences. For LLMs, we analyze their internal representations and compute per-word perplexity; for humans, we draw on functional magnetic resonance imaging (fMRI) data as well as measurements of per-word reading times.\nWe measure brain alignment by using a linear predictivity metric that evaluates how well the representations of LLMs can predict human brain responses to the same language stimuli, using multiple neural datasets. Behavioral alignment is measured by correlating LLM per-word perplexity with human per-word reading times, since both metrics offer insights into comprehension difficulty. Because different models vary in their alignment across architectures and training goals, we analyze 17 instruction-tuned LLMs alongside 8 standard (or \"vanilla\") models and find that instruction-tuned models exhibit, on average, a 6.2% increase in brain alignment compared to the vanilla models.\nTo understand why instruction-tuning enhances alignment with human brain activity, we further examine which LLM properties contribute to this improvement. We compute correlations between brain alignment and various properties, including next-word prediction ability, model size, problem-solving skills, and world knowledge spanning different domains. Our results reveal three key findings:\n1. Instruction-tuning generally enhances the alignment of LLM representations to human brain activity.\n2. World knowledge and model size show strong correlations with brain alignment.\n3. Contrary to expectations, instruction-tuning does not significantly improve the alignment of LLM behavior with human reading times. In fact, behavioral alignment on this task shows weak correlations with other measures such as task performance and model size.\nOverall, our findings indicate that while instruction-tuning makes LLM representations more brain-like, it does not necessarily bring their behavioral responses closer to those of humans."},"source":{"paper_id":"DZ6B5u4vfe","cohort_label":"uncreative","pairing_group":"1","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,3.0,1.0],"soundness":[2.0,2.0,3.0,2.0],"presentation":[3.0,3.0,3.0,1.0],"overall_score":[5.0,3.0,6.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.0,"soundness_mean":2.25,"presentation_mean":2.5,"overall_score_mean":4.25},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"Dgc5RWZwTR","group_id":"Dgc5RWZwTR","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Although a generic neural solver for multiple combinatorial optimization problems (COPs) is appealing, this challenge has received limited attention, and training such a solver can be prohibitively expensive, particularly given the current scale of models. To alleviate the training burden and optimize resource allocation, we propose a novel training paradigm that leverages multi-armed bandits (MAB) from a multi-task learning (MTL) perspective. This approach efficiently trains a multi-task combinatorial neural solver even when working under constrained training budgets.\nIn our framework, each COP at a specific problem scale is treated as a separate task, and we develop a universal solver that can handle a set of tasks simultaneously. Instead of standard joint training in MTL, where balancing losses from multiple tasks can be complex, we adopt MAB algorithms to select a single task in each training round. To guide the MAB algorithms effectively, we design a reward strategy based on a theoretical loss decomposition for the widely used encoder-decoder architecture in MTL. This decomposition also leads to an influence matrix that reveals the mutual impacts between tasks, offering valuable insights into common practices within COPs.\nNotably, our method is the first to focus on training a generic neural solver for various types of COPs. This sets our approach apart from prior work that has concentrated on either solution construction or heuristic improvement. While recent studies have attempted to generalize neural solvers to different scales or distributions, they have not addressed the challenge of handling multiple COP types at the same time.\nWe test our approach on 12 tasks covering four types of COPs—the Travelling Salesman Problem, the Capacitated Vehicle Routing Problem, the Orienteering Problem, and the Knapsack Problem—each represented at three different problem scales. Our experiments compare our method with single-task training as well as various multi-task learning baselines, all under identical training budgets and epochs. Compared to single-task training, our approach requires no prior knowledge about the tasks and automatically focuses on the more challenging ones to maximize the utilization of the training budget. Furthermore, even when evaluated over the same number of training epochs, our method not only reduces training costs—remaining lower than that of the most expensive single task—but also demonstrates robust generalization by providing a model that accommodates different COP types. Relative to other MTL methods, our approach selects the most impactful task to train at each iteration, significantly enhancing training efficiency without the need to balance multiple loss functions explicitly.\nIn summary, our contributions are threefold. First, we introduce a novel framework for efficiently training a neural solver for multiple COPs using MAB, achieving notable performance improvements under limited training resources and offering insights for training other large models. Second, we develop a theoretical loss decomposition for the encoder-decoder architecture that produces an influence matrix capturing inherent task relationships and informs a tailored reward system for MAB updates. Third, we validate several empirical observations from previous studies through the influence matrix, thereby confirming the effectiveness and soundness of our approach."},"source":{"paper_id":"Dgc5RWZwTR","cohort_label":"uncreative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,2.0,2.0],"soundness":[3.0,2.0,3.0,3.0],"presentation":[3.0,3.0,2.0,2.0],"overall_score":[5.0,6.0,5.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.75,"presentation_mean":2.5,"overall_score_mean":4.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"Dnc3paMqDE","group_id":"Dnc3paMqDE","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Many 3D applications—such as virtual and augmented reality—traditionally depend on manually designed indoor environments. However, reconstructing such environments from 3D point clouds has been gaining attention as a promising alternative to manual design. Advances in 3D vision sensors have significantly improved the quality of point clouds, yet the absence of explicit surface information means that these improvements do not automatically translate to better 3D content applications.\nIn this work, we focus on reconstructing indoor environments from point clouds using Scan-to-CAD (S2C) methods. By retrieving existing CAD models from a public database like ShapeNet, we can align these models with scanned data to achieve high-quality reconstruction of objects and scenes. The tasks involved include registration, retrieval, and completion.\nRecent developments in architectures for learning features from 3D data have paved the way for advanced methods. Nevertheless, three key issues remain unresolved. First, most current networks do not adequately account for the different poses of point clouds in the latent space—a critical factor in tasks such as registration and retrieval. Although equivariant networks have been introduced to extract rotation-equivariant features from point clouds, they often treat local and global equivariance similarly, which results in insufficient separation between global and local structural features. Second, the definition of local structures is typically based on fixed, pre-defined patches that cannot adapt during learning. Third, there is no point-to-point correspondence between point clouds acquired from different sensing devices. Although some representation methods have overlooked local information, they have shown that spherical representations can enhance registration tasks.\nTo address these challenges, we propose a novel point network tailored for instance-based applications including registration, retrieval, and completion. Our approach introduces Spherical Patch Fields (SPF), a representation that maps a 3D point cloud to numerous spherical patches, each represented by a Spherical Gaussian that provides a rotation-equivariant feature. To further enhance the representation, we introduce the Patch Gaussian Layer (PG-Layer), a convolutional layer that retains the original mathematical properties of Spherical Gaussians. This retention allows for deeper networks and improved learning capacity. Additionally, the PG-Layer incorporates low-frequency analysis for better shape recognition while learning the optimal size for each spherical patch to flexibly capture both local and global information depending on the task.\nIn summary, our contributions are as follows:\n• We present SPF, a new representation that converts a 3D point cloud into multiple spherical patches described by Spherical Gaussians, while learning both inter- and intra-patch relations in a graph-based manner.\n• We introduce the PG-Layer, a convolutional layer based on Spherical Gaussians that supports deep networks, enhances shape recognition through low-frequency analysis, and learns the appropriate patch size for capturing local or global information.\n• We demonstrate the effectiveness of DeepSPF as an easily integrable backbone, showing significant improvements in S2C tasks, including point cloud registration, retrieval, and completion, through extensive evaluation."},"source":{"paper_id":"Dnc3paMqDE","cohort_label":"creative","pairing_group":"4","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,2.0],"soundness":[3.0,3.0,2.0],"presentation":[3.0,3.0,1.0],"overall_score":[8.0,6.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.3333333333333335,"soundness_mean":2.6666666666666665,"presentation_mean":2.3333333333333335,"overall_score_mean":6.333333333333333},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"E296x0YpML","group_id":"E296x0YpML","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Over the past decade, deep neural networks have achieved remarkable performance in natural language processing, with models such as recurrent neural networks, transformers, and pretrained language models setting new benchmarks. At the same time, research has shown that these models can be tricked by adversarial examples—inputs that are subtly perturbed to widen the gap between the model’s predictions and the true labels while remaining unnoticeable to human readers. This situation has raised concerns about the security and robustness of NLP systems.\nAdversarial examples are crafted by altering various aspects of an input sentence. Some methods introduce character-level errors like typos or visually similar characters, while others replace words with alternatives that retain the original meaning. There are also techniques that generate paraphrased sentences with similar semantics. Among these strategies, adversarial word substitutions have proven to be particularly effective, as they repeatedly query the model to identify and perturb key words.\nIn response to these challenges, many defenses against textual adversarial attacks have been proposed. Popular approaches include adversarial training, which augments the training data with adversarial examples through extra optimization steps, and randomized smoothing, which creates a stochastic ensemble by introducing random perturbations. Due to the discrete nature of language, these defenses often replace words with adversarial alternatives selected from predefined sets designed around potential attacks, all while trying to keep the original meaning intact. However, assuming prior knowledge of all potential attacks is neither realistic nor practical.\nIn our work, we introduce AdvFooler, a lightweight and attack-agnostic defensive method that enhances the robustness of NLP models against textual adversarial attacks. AdvFooler randomizes the latent representation of the input at test time, thus confusing adversaries as they iteratively sample discrete perturbations to create adversarial examples. This test-time defense adds negligible computational overhead and requires no additional training. Moreover, because it operates within the model’s latent space, it does not depend on any predetermined perturbation sets associated with particular attacks. Unlike some existing randomization defenses that require training modifications to reduce output variance, AdvFooler offers a simple, plug-and-play solution.\nOur contributions are threefold. First, we present a defense method that disrupts the adversary’s ability to optimize for effective perturbations, thereby substantially lowering the attack success rate. Second, we provide both theoretical insights and empirical evidence demonstrating how randomizing latent representations can confound adversarial attacks. Finally, we conduct extensive experiments on benchmark datasets and various attacks, showing that AdvFooler is a competitive defense that imposes minimal computational overhead and fewer modeling constraints compared to current alternatives."},"source":{"paper_id":"E296x0YpML","cohort_label":"uncreative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[4.0,2.0,2.0,2.0],"soundness":[4.0,2.0,3.0,3.0],"presentation":[3.0,4.0,3.0,3.0],"overall_score":[8.0,3.0,6.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":3.0,"presentation_mean":3.25,"overall_score_mean":5.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"ElykcDu5YK","group_id":"ElykcDu5YK","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Real-world applications—ranging from improving manufacturing lines and enhancing crop yields to optimizing machine learning models—often involve blackbox objective functions that lack closed-form expressions, convexity, or derivative information. Moreover, evaluating these functions is usually expensive, time-consuming, and subject to noise. These challenges have led researchers to adopt frameworks like Bayesian optimization, especially when addressing risk arising from uncontrollable environmental factors such as sensor errors, weather conditions, or perturbations in training data. In these scenarios, common risk measures include worst-case outcomes, value-at-risk (VaR), conditional value-at-risk (CVaR), probability thresholds, and variance.\nWhile Bayesian optimization with risk measures has received considerable attention, the idea of leveraging experiences from previous tasks to accelerate current optimization has not been explored extensively. Prior work on meta-Bayesian optimization has shown that transferring knowledge—from probabilistic models to optimization outcomes—can significantly reduce the number of costly function evaluations. However, existing methods that transfer Gaussian process models are often computationally intensive and not specifically designed for locating the function’s maximizer, while those transferring optimization outcomes face challenges in defining a clear framework and handling issues such as scaling, vertical shifting, and the presence of harmful prior tasks.\nIn our work, we introduce meta-VBO, a novel approach that transfers optimization outcomes while explicitly accounting for the risk introduced by environmental uncertainties. Focusing on two well-known risk measures, VaR and CVaR, our method also extends to adversarially robust Bayesian optimization. We assume that the distribution of environmental variables is known—through historical data or domain expertise—and that noisy evaluations can be obtained under controlled conditions (for example, by simulating manufacturing errors or regulating weather in laboratory settings). Once the optimization process is complete, the recommended decision variable is implemented in the real-world environment where environmental factors cannot be controlled.\nSeveral challenges arise in transferring knowledge for the Bayesian optimization of risk measures. First, it is essential to transfer information that remains invariant under scaling or vertical shifting of the function to ensure that the optimization focuses on identifying the optimum. This requirement eliminates certain similarity measures that depend on the magnitude of function evaluations. Second, as more prior tasks are incorporated, there is a risk of introducing detrimental information from tasks that are not aligned with the current optimization goal. Thus, our method is specifically designed to be resilient against such harmful influences.\nTo address these challenges, we propose a new perspective: rather than selecting a single query point at each iteration (as is common in traditional methods), we construct a versatile query set (V-set). Remarkably, any input chosen from this set ensures a no-regret outcome. This flexibility not only allows for robust performance even in the presence of detrimental prior tasks but also provides a mechanism to transfer information from previous tasks by prioritizing inputs that have been promising in the past. Two hyperparameters control the size of the V-set, balancing the use of prior information with the regret bound for the current task. As iterations progress and more data are collected, the V-set is updated adaptively, naturally reducing the influence of any harmful prior tasks. Additionally, we propose a strategy for recommending inputs that increasingly approximate the optimal solution, with theoretical guarantees that the instantaneous regret diminishes over time.\nFinally, we demonstrate the effectiveness of meta-VBO on both synthetic and real-world problems involving the optimization of VaR and CVaR, showing that our approach successfully accelerates the optimization process while effectively managing risk."},"source":{"paper_id":"ElykcDu5YK","cohort_label":"creative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,3.0],"soundness":[4.0,2.0,3.0],"presentation":[4.0,1.0,3.0],"overall_score":[8.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6666666666666665,"soundness_mean":3.0,"presentation_mean":2.6666666666666665,"overall_score_mean":6.666666666666667},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"FDQF6A1s6M","group_id":"FDQF6A1s6M","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Reinforcement learning (RL) and multi-agent reinforcement learning (MARL) face a major challenge: the non-stationary nature of the environment. In such settings, each agent’s outcome depends not only on its own actions but also on those of other agents. This interdependency often prevents traditional algorithms from converging to desirable solutions. In general-sum games, when each RL agent optimizes its own utility, they tend to settle on sub-optimal outcomes from a Pareto perspective. This mirrors many real-world scenarios where individuals’ pursuit of self-interest leads to worse outcomes compared to cooperation. Consequently, a key objective in MARL research is to design decentralized agents that can cooperate while protecting themselves from exploitation in partially competitive environments, a strategy we refer to as reciprocity-based cooperation.\nEarlier approaches have tackled this by training cooperative agents either by differentiating through an opponent’s learning steps or by framing opponent shaping as a meta-game over agent policies. However, both methods suffer from significant computational inefficiencies. Differentiating through even a few of an opponent’s learning steps must be done sequentially and requires constructing large computational graphs, which becomes especially costly with complex opponent policies. Alternatively, the meta-learning approach formulates the problem as a meta-state over the joint policy space of both agents, and learns a meta-policy mapping this state to updated policies. This scales badly due to the high number of parameters typically present in neural network policy representations.\nIn this paper, we introduce Learning with Opponent Q-Learning Awareness (LOQA), a method that avoids computing gradients with respect to optimization steps or the dynamics of a meta-game. This significantly enhances computational efficiency. LOQA shapes opponent behavior by assuming that the opponent is guided by an internal action-value function. This assumption enables LOQA agents to model and influence the opponent’s policy by affecting the returns associated with different actions. A fundamental concept in RL, controlling returns by differentiating through stochastic objectives, is employed using techniques like the REINFORCE estimator."},"source":{"paper_id":"FDQF6A1s6M","cohort_label":"creative","pairing_group":"1","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0,2.0,2.0],"soundness":[3.0,2.0,3.0,2.0,2.0],"presentation":[3.0,3.0,3.0,2.0,2.0],"overall_score":[3.0,6.0,3.0,5.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.2,"soundness_mean":2.4,"presentation_mean":2.6,"overall_score_mean":4.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"FsVxd9CIlb","group_id":"FsVxd9CIlb","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Deep neural networks (DNNs) have recently achieved state-of-the-art performance in various areas such as medical diagnostics, autonomous driving, and sentiment analysis. In safety‐critical applications, the trustworthiness of these models is paramount, especially given that real-world noise and human-induced perturbations can affect their reliability. Because DNNs are inherently nonlinear with complex parameters, ensuring both high performance and an interpretable decision-making process is essential. Understanding how data flows from input to output is a crucial step toward explainable artificial intelligence.\nThere are two main strategies for interpreting DNNs. One approach approximates the model’s behavior in a local neighborhood to explain decisions at a specific point in the input space, while the other uses gradients related to the inputs to determine the importance of individual features. In our work, we focus on gradient-based attribution methods that provide pixel-level explanations by identifying which input features are most critical to the model’s decisions. The underlying assumption is that small changes to important features may cause the output to change, indicating their role in moving the input across the decision boundary.\nExisting methods such as Integrated Gradients, Boundary-based Integrated Gradient, and Adversarial Gradient Integration have laid the groundwork for feature attribution. However, these methods often depend on a baseline or a predetermined integration path, and may even cross multiple decision boundaries in a targeted adversarial attack, leading to inaccuracies—especially when there is overlap between the boundaries of different categories. This observation leads us to our first research question: How can we construct a more general decision boundary exploration approach to ensure that features effectively probe the current decision boundary?\nTo illustrate the concept, consider modifying a subset of features so that they cross the decision boundary; minimal changes can reveal the most important features. We observed that transferable attacks, typically used for generating adversarial examples for black-box attacks, inherently explore the model parameters to produce examples that cross several decision boundaries. The decision boundaries obtained through transferable attacks are generally less prone to overfitting and are more accurate. This insight motivates us to combine transferable attack techniques with the attribution process. We introduce a novel method, based on model parameter exploration (MPE), to generate the necessary feature changes.\nEvaluating the attribution algorithm involves verifying that the model continues to make correct decisions when only the essential features are retained. This evaluation is challenging because many model parameters that should be activated remain silent when unimportant features are removed. These inactive parameters indicate that the decision boundary has shifted, but not drastically. Consequently, an effective attribution algorithm must be both stable and adaptable across different decision boundaries. This consideration leads us to our second research question: How can we develop a robust approach for making minimal feature alterations via MPE to ensure high attribution performance?\nRecent research has shown that DNNs respond differently to perturbations across various frequency domains. By applying spectral transformations to inputs, we gain new insights into how model parameters influence the decision boundary. Our findings indicate that incorporating frequency information can significantly enhance feature attribution, leading to more precise identification of important features. As a result, we employ frequency-based transferable attacks to create minimally altered adversarial examples, with experimental results confirming the effectiveness of our approach. Notably, our work is the first to introduce MPE as a generalizable method for exploring the decision boundaries of different models. Moreover, because transferable attacks can vary in how they explore these boundaries, our method can be used alongside other state-of-the-art techniques to uncover differences in attribution performance.\nThis paper makes several key contributions. First, we demonstrate for the first time that the strategies for exploring decision boundaries in attribution and transferable attacks are aligned. Second, we propose an innovative attribution algorithm—AttEXplore—which integrates model parameter exploration with transferable attacks. Third, we provide extensive experimental validation of AttEXplore’s effectiveness. Finally, we have made the code for AttEXplore publicly available."},"source":{"paper_id":"FsVxd9CIlb","cohort_label":"creative","pairing_group":"4","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,3.0,2.0],"soundness":[3.0,2.0,3.0,2.0],"presentation":[2.0,2.0,3.0,2.0],"overall_score":[6.0,6.0,6.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":2.5,"presentation_mean":2.25,"overall_score_mean":5.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"GACjMj1MS1","group_id":"GACjMj1MS1","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Machine learning methods are increasingly used to assist human decision-making, but concerns have emerged about the fairness of algorithmic decision systems. To address these concerns, recent studies have proposed mechanisms to ensure fairness. One such measure is the covariance between sensitive attributes and the distance from subjects’ feature vectors to the decision boundary of a classifier. Earlier work introduced the idea of using this covariance to design classifiers free of both disparate impact and disparate mistreatment, and subsequent extensions applied similar ideas to decision trees. Other methods for addressing fairness have included measures based on correlation and covariance operators, along with fairness constraints based on equalized correlations. However, these methods typically provide only point estimates of the dependence between sensitive variables and predictions, neglecting the uncertainty inherent in estimation. In addition, some fairness measures are not dimensionless, which complicates the process of setting an appropriate threshold.\nOur work addresses these issues by utilizing empirical likelihood (EL), a nonparametric method that does not assume a particular underlying distribution and has been successfully applied in various problems. Rather than developing a new fairness criterion, we adopt the covariance as a tractable proxy for measuring fairness, focusing on group fairness – reducing the differences in favorable outcome proportions among different sensitive groups. This covariance measure can handle both discrete and continuous sensitive attributes. By using EL, we construct a confidence region for the fairness criterion, which allows us to account for the uncertainty in estimating the dependence between sensitive variables and predictions—something that is not possible with point estimates alone. We use the significance level as a threshold for the fairness constraint. Because this threshold is dimensionless, it provides a clear and practical way to assess fairness. With the confidence region, fairness is assessed by checking whether zero is included within it, while controlling for the Type I error rate. By enforcing a fairness constraint such that zero lies within the confidence region of the covariance vector, we eliminate the need to set an arbitrary threshold based on the vector's dimension. In our algorithm, the trade-off between estimation accuracy and the fairness constraint is balanced through the significance level.\nSimulation studies demonstrate that our method achieves good coverage probability and a reasonable average length of the confidence interval for the covariance. In some cases, it outperforms previous methods in balancing estimation accuracy and fairness. Our analysis of real data shows that our approach is robust to distributional shifts and capable of ensuring fairness when both discrete and continuous features are considered as protected attributes.\nIn summary, our contributions are as follows. First, we propose an EL-based estimator of the covariance and show that its limiting distribution is the standard chi-square distribution. This result is achieved without making assumptions about the underlying distribution. Based on this result, we construct a confidence region for the covariance between sensitive attributes and the decision boundary, which can then be used in statistical hypothesis tests for group fairness. Second, we articulate the fairness constraint through a confidence region where the significance level serves as the fairness threshold. This approach is more practical compared to previous methods that rely on dimensional thresholds, as the significance level can be easily determined in practice. Moreover, our method can handle multiple sensitive features simultaneously, accommodating both categorical and continuous variables, while integrating uncertainty into the fairness measurement to enhance robustness. Finally, our work introduces a new direction in fairness constraints by incorporating uncertainty into both the assessment and enforcement of fairness—a factor that has often been overlooked. Our framework is also extendable to other fairness measures.\nThe remainder of the paper is organized as follows. After a review of related work, we explain the use of covariance as a fairness constraint, develop the confidence region using empirical likelihood, and present the overall framework. Simulation studies are then conducted to evaluate our method compared to others. Next, we assess our method using real data, and finally, we provide conclusions and discuss potential extensions. Additional proofs and supplementary results are included in the appendix."},"source":{"paper_id":"GACjMj1MS1","cohort_label":"creative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,3.0],"soundness":[3.0,3.0,3.0],"presentation":[2.0,3.0,3.0],"overall_score":[5.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6666666666666665,"soundness_mean":3.0,"presentation_mean":2.6666666666666665,"overall_score_mean":5.666666666666667},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"GGPyzKsHZ1","group_id":"GGPyzKsHZ1","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Multimodal learning is vital for many real-world applications because data often come in multiple modalities—for example, text-image, text-video, and audio-video pairs. In language-vision tasks, aligning these modalities usually relies on human-annotated descriptions. However, most videos naturally include audio, which allows audio-video models to be trained on vast amounts of raw web data with minimal preprocessing, leading to effective multimodal representations.\nDespite this advantage, current audio-video models typically assume a static data distribution and therefore struggle in scenarios where the distribution shifts over time. In real life, environmental conditions can change continuously, rendering past data less relevant. Imagine a situation where initial data depict people conversing outdoors with clear human voices and interaction visuals. Suddenly, due to a volcanic eruption nearby, the audio shifts to erupting sounds and the visuals to images of mountains.\nThis continuous change poses two main challenges. The first is that only a few objects or regions in a video are strongly correlated with the audio, resulting in sparse spatiotemporal connections between the modalities. The second challenge is representational forgetting, where the model not only loses the representations for individual modalities but also the learned correlations between them. For example, a model might correctly learn the association of a car’s engine sound with its corresponding visual cues, only to forget this relationship when exposed to a new sequence of tasks.\nTo overcome these challenges, we introduce the Forget-robust Localized Audio-Video Alignments (FLAVA) method. FLAVA uses a small, trainable multimodal encoder to generate attention maps that capture the interrelated audio-video patches. By comparing these maps from the current task with those from previous time steps, the method identifies stable patches that preserve previously learned correlations even as new data are introduced. This approach helps maintain the accuracy of audio-video relationships over time.\nWe validate our approach on VG-GSound and AudioSet by comparing it with strong baseline models on audio-to-video and video-to-audio retrieval and audiovisual classification tasks. Our method outperforms these baselines, demonstrating its effectiveness in handling continuously changing data distributions.\nIn summary, our contributions are threefold:\n1. We introduce the problem of audio-video representation learning under continuously evolving data distributions, highlighting challenges such as sparse spatiotemporal correlations and representational forgetting.\n2. We propose a novel method, FLAVA, that adaptively captures sparse audio-video attention and preserves historical correlations without the need for explicit task identification.\n3. We provide extensive experimental validation and in-depth analyses, including visualizations, that demonstrate the superiority of our approach over recent methods."},"source":{"paper_id":"GGPyzKsHZ1","cohort_label":"uncreative","pairing_group":"4","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,3.0,3.0,2.0],"soundness":[2.0,3.0,3.0,3.0,2.0],"presentation":[2.0,3.0,2.0,2.0,1.0],"overall_score":[5.0,5.0,8.0,5.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6,"soundness_mean":2.6,"presentation_mean":2.0,"overall_score_mean":5.6},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"Gdm87rRjep","group_id":"Gdm87rRjep","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Large Language Models (LLMs) have transformed natural language processing by demonstrating remarkable in-context learning, which allows them to generalize to unseen tasks without additional fine-tuning. LLMs operate through user-provided natural language prompts that specify how to complete a task, enabling a single large model to efficiently address a wide range of tasks. Despite these advantages, deploying LLMs can be prohibitively expensive because their inference often requires multiple powerful GPUs—a challenge for many users. Therefore, it is essential to make LLM inference accessible on more affordable hardware, such as a single GPU with limited computational and memory resources.\nOne common solution to this problem is the use of model compression techniques like quantization and pruning. These methods reduce model size and inference latency by trading off some accuracy and overall performance. For example, while the full model (LLaMA-7B) provides accurate answers to simple questions, a pruned version of the same model may generate unrelated or off-topic responses.\nConsider a series of examples where the full model and its compressed version are asked questions such as \"Where is Long Beach?\" or \"Where is Tulsa, Oklahoma?\" The full model reliably produces precise answers, whereas the compressed model sometimes exhibits inaccuracies. To address this issue, a manually designed prompt can be attached to the original question. This prompt advises the model to carefully examine its weight matrix for potential errors and adjust as needed, which significantly improves the responses of the compressed model. Although this prompt only suggests the possibility of weight inaccuracies without considering specific datasets, compression methods, or tasks, it demonstrates notable transferability across various scenarios.\nInspired by earlier work on learnable prompts, we propose that involving the compressed weights in the prompt learning process might yield a prompt that performs even better while maintaining its broad applicability. We introduce a new prompt learning paradigm that trains additive prompt tokens directly on a compressed LLM to enhance its accuracy. Unlike previous methods that learn task-specific prompts with limited generalization, our approach produces a prompt that is as transferable as the manually designed version.\nExperimental results show that our method greatly improves the performance of an eight times compressed Llama-7B model, which has undergone both joint four-bit quantization and 50% weight pruning, allowing it to match the performance of its uncompressed counterpart on several standard benchmarks. Furthermore, these learned prompts exhibit a degree of transferability across different datasets, tasks, and compression levels. Compared to other parameter-efficient fine-tuning approaches, our method requires less overhead while effectively recovering the performance of compressed LLMs."},"source":{"paper_id":"Gdm87rRjep","cohort_label":"uncreative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,2.0,3.0],"soundness":[3.0,3.0,3.0,3.0],"presentation":[3.0,2.0,3.0,3.0],"overall_score":[5.0,5.0,5.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":3.0,"presentation_mean":2.75,"overall_score_mean":5.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"HANfmG6tQK","group_id":"HANfmG6tQK","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Large Batch Learning (LBL) is essential in modern deep learning for its efficiency gains through parallel processing, enhanced generalization by exposure to diverse samples, memory efficiency, and effective hardware utilization. These features make LBL especially suitable for training large deep neural network models and self-supervised learning tasks, where increased model capacity and robust representation learning are critical. However, combining LBL with conventional gradient-based methods often requires heuristic strategies and can lead to compromised generalization accuracy. Various methods have been explored to address the challenges associated with large-batch training. One widely popular approach is Layer-wise Adaptive Rate Scaling (LARS), which employs adaptive rate scaling on a per-layer basis to enhance training stability. Despite its benefits, LARS can be unstable during the early stages of large-batch training, resulting in slow convergence when using large batches.\nA common remedy for this instability is to implement a warm-up strategy that gradually reduces the LARS adaptive rate, thereby stabilizing the learning process. Yet, this approach relies on vague tuning strategies and lacks a strong theoretical foundation, leaving room for improved adaptive rate scaling algorithms.\nEmpirical experiments revealed that in the initial phase of LARS training, the layer-wise learning rate tends to be high due to the layer-wise normalization rate, which is significantly affected by the near-zero value of the layer-wise gradient norm. This near-zero gradient norm is a consequence of the model getting trapped in sharp minimizers, leading to an explosion of the scaled gradient. With the current warm-up technique, many unnecessary steps are required to scale the gradient to a level that allows the model to escape from these sharp minimizers. Moreover, because the learning rate decays constantly during this process, the warm-up does not effectively facilitate gradient exploration or adapt well to diverse datasets.\nTo address these issues, we introduce a new algorithm called Time-Varying Layer-wise Adaptive Rate Scaling (TVLARS). This approach promotes gradient exploration in the initial phase while preserving the stability of traditional LARS methods in later stages. Rather than relying on the problematic warm-up strategy, TVLARS leverages a high initial target learning rate along with an inverted sigmoid function to enhance training stability. This design aligns with theoretical insights regarding sharp minimizers in large-batch learning.\nOur contributions are as follows: We first conducted experiments on two canonical large-batch learning techniques—LARS and LAMB—to better understand their performance enhancements. We then examined the necessity of using warm-up for LARS and uncovered potential issues stemming from a lack of understanding about the behavior of sharp minimizers in large-batch learning. Building on these insights, we propose TVLARS as a simple yet effective alternative that avoids the drawbacks of the warm-up approach. Finally, extensive experimental evaluations demonstrate that, under the same delay step and target learning rate, TVLARS significantly outperforms state-of-the-art benchmarks, particularly when using very large batch sizes."},"source":{"paper_id":"HANfmG6tQK","cohort_label":"uncreative","pairing_group":"14","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,4.0,2.0],"soundness":[4.0,4.0,3.0],"presentation":[4.0,3.0,2.0],"overall_score":[6.0,6.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6666666666666665,"soundness_mean":3.6666666666666665,"presentation_mean":3.0,"overall_score_mean":5.666666666666667},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"I2mIxuXA72","group_id":"I2mIxuXA72","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"A common assumption in machine learning is that the training and test data are independent and identically distributed (i.i.d.) samples. In practice, however, unseen test data are often drawn from distributions that differ from those used during training. For example, while a camel is typically associated with desert environments, it might appear in a grassland during testing. Overparameterized deep neural networks tend to rely on spurious correlations present only in the training distribution, and as a result, they often fail to generalize under distribution shifts. This is particularly concerning in safety-critical applications such as autonomous driving and medical diagnosis.\nMany studies have aimed to improve the generalization of deep learning models in the face of distribution shifts, approaching the problem from various angles. One significant line of research involves partitioning the training data according to their respective domains or environments, such as collecting bird photos with cameras placed in different settings (e.g., near water versus on land). This additional environmental information has spurred the development of domain generalization algorithms. These approaches share a common goal: to discover invariant representations that minimize loss across all environments, thereby discouraging the learning of spurious correlations that are effective only for certain data subsets. Nevertheless, some recent research has shown that domain generalization algorithms do not consistently outperform standard empirical risk minimization (ERM), suggesting that ERM might already be learning sufficiently good representations. This observation raises important questions about whether, when, and why domain generalization methods might actually surpass ERM.\nIn our work, we explore the conditions under which domain generalization methods outperform ERM. We demonstrate that the success of some domain generalization algorithms can be attributed to their robustness against label noise in the presence of subpopulation shifts. Our theoretical analysis shows that, when using overparameterized models trained on finite samples with label noise, ERM is more likely to converge to suboptimal solutions that primarily rely on spurious correlations. This analysis is supported by experiments on synthetic datasets. In contrast, certain domain generalization algorithms exhibit implicit noise robustness that offers an additional performance advantage. By analyzing the optimization process of several domain generalization methods, we illustrate why algorithms such as IRM, V-REx, and GroupDRO benefit from this robustness, while methods like ERM and Mixup, which do not explicitly modify the objective function, lack this advantage. Empirically, our findings indicate that noise robustness can be beneficial in synthetic scenarios with significant label noise and strong spurious correlations. However, for noisy real-world data and pretrained models, there is currently no clear evidence that these domain generalization techniques yield better performance.\nLabel noise is an important consideration because, in some cases, achieving a completely accurate decision boundary from the data may be impossible due to missing critical information. Any inaccuracies can be viewed as a form of label noise. Additionally, common analyses of ERM's failure modes assume that classifiers based solely on invariant features are only partially predictive, which fits within our framework where a fully predictive invariant classifier exists for noise-free data, and label noise is added subsequently. More broadly, we examine ERM's failure mode without assuming that spurious features are inherently less noisy than invariant features. By focusing on domain generalization through the perspective of noise robustness, we gain a unique understanding of when and why ERM and domain generalization algorithms succeed or fail.\nOur main contributions are as follows:\n1. We theoretically demonstrate that for overparameterized models trained on finite samples, the tendency to learn spurious correlations instead of invariant features is influenced by both the strength of the spurious correlation and the amount of label noise.\n2. We show that several domain generalization algorithms possess noise-robust properties that enable models to learn invariant features despite the presence of label noise.\n3. We perform extensive experiments comparing domain generalization algorithms on both synthetic and real-life datasets with injected label noise. Our findings suggest that while noise robustness can be beneficial in specific controlled scenarios, there is still no clear evidence that it leads to better overall performance in general applications. We also discuss the challenges of meeting the theoretical conditions in practice and propose potential mitigators."},"source":{"paper_id":"I2mIxuXA72","cohort_label":"creative","pairing_group":"1","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0,2.0],"soundness":[2.0,2.0,2.0,2.0],"presentation":[2.0,3.0,3.0,3.0],"overall_score":[5.0,6.0,3.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.0,"presentation_mean":2.75,"overall_score_mean":4.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"Js5PJPHDyY","group_id":"Js5PJPHDyY","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Contrastive Language-Image Pretraining (CLIP) has garnered significant attention as a powerful method for aligning vision and language representations. By leveraging a dataset of 400 million image-text pairs, CLIP encodes images and texts into a shared semantic space, allowing for effective comparisons of similarities and differences between the two modalities. One remarkable aspect of CLIP is its ability to perform image classification without any additional training. This is achieved by using the language encoder to generate classifier weights based on a simple prompt (for example, \"a photo of a [class]\"), enabling the classification of images into diverse categories directly.\nWhile the zero-shot capability of CLIP is impressive, recent works have focused on designing efficient fine-tuning methods for downstream classification tasks. These approaches have shown considerable improvements over the original zero-shot method, even when trained on few samples and with a minimal number of parameters. For example, one method achieves a 15% improvement with only 16 samples per class by fine-tuning a small number of parameters. However, these improvements come at the cost of additional computational resources, which can be challenging on devices with limited capabilities.\nTo address this challenge, we propose a method that eliminates the need for additional training while achieving comparable or even superior performance to training-required techniques. Our approach revisits Gaussian Discriminant Analysis (GDA), a classical algorithm that assumes features from different classes follow Gaussian distributions with identical covariance. Through Bayes’ rule, the classification probability can be expressed in a way that allows us to compute the necessary classifier parameters directly from the training data. By integrating this classifier with the CLIP zero-shot classifier, we effectively combine visual and textual modalities without further training.\nWe further extend our method to handle scenarios where it cannot be directly applied, such as base-to-new generalization and unsupervised learning. In the base-to-new scenario, a model trained on a base dataset is adapted to a new dataset with different classes. Observing that similar samples possess similar statistical properties, we use a K-Nearest-Neighbor algorithm to synthesize data for new classes and derive the corresponding classifier. For unsupervised learning, where labels are unavailable, we assume that the data follow a Gaussian mixture distribution and employ the Expectation-Maximization algorithm to estimate the required statistical parameters. Both of these variants are designed to be simple, avoiding additional complexity while maintaining competitive performance.\nWe evaluate our approach on 17 widely used datasets. Despite its simplicity, our method establishes a robust baseline. In few-shot classification, it outperforms previous training-free methods by an average margin of 2.82% over 11 datasets and achieves performance comparable to methods that require training. In imbalanced learning, our approach also exceeds the performance of fully fine-tuned state-of-the-art methods. Additionally, our variants for base-to-new generalization and unsupervised learning match the performance of more complex previous approaches. These results highlight the effectiveness and practicality of our method."},"source":{"paper_id":"Js5PJPHDyY","cohort_label":"creative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,3.0],"soundness":[3.0,3.0,2.0],"presentation":[3.0,3.0,3.0],"overall_score":[6.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6666666666666665,"soundness_mean":2.6666666666666665,"presentation_mean":3.0,"overall_score_mean":6.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"JzAuFCKiov","group_id":"JzAuFCKiov","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Large Language Models have made significant progress and deeply influenced the AI community. However, because they rely on vast amounts of internet data—including a large portion of low-quality content—they can sometimes exhibit unintended behavior such as making up facts, generating biased or toxic text, or even producing harmful content. This makes it essential to align these models with human values like being helpful, honest, and harmless.\nOne widely used approach for aligning these models is called Reinforcement Learning with Human Feedback (RLHF), which typically involves three steps: supervised fine-tuning, reward learning, and reinforcement learning fine-tuning. In the final step, an optimizer known as Proximal Policy Optimization (PPO) is commonly used. Although PPO is effective, recent research has raised several issues that merit further attention.\nOne concern is the instability of PPO. Various factors—such as reward normalization, scaling, clipping, KL control, advantage normalization, and critic initialization—can contribute to this instability. Additionally, there is another source of instability related to how the reward is optimized. The reward model is typically trained using a comparative loss (the Bradley-Terry Loss) that is invariant to constant shifts. However, PPO is not invariant in the same way, meaning that even if two reward models convey the same information about human preferences, optimizing them with PPO may result in different outcomes.\nAnother challenge in fine-tuning these models is deciding whether to perform optimization on a trajectory-wise or token-wise basis. The reward model is trained using datasets where human labelers rank multiple completions for a single prompt, naturally leading to a trajectory-wise comparison. This approach is often favored because it is difficult for labelers to assign direct scores or make detailed comparisons at the token level—except in some less ambiguous tasks like math problems, where partial credit can be given. The challenge is compounded by the fact that the reward model provides only one scalar reward after reading an entire response. As a result, two main viewpoints emerge: should language generation be treated as a Contextual Bandit or as a Markov Decision Process? The common practice of using token-wise updates with PPO adds complexity, as it requires learning a value function and using techniques to redistribute the sparse reward over all tokens.\nIn this work, we address these issues with several key contributions. First, we define an equivalence relation for reward functions trained from human preferences and show that while the Bradley-Terry Loss is invariant under this relation, PPO is not. Second, we find that Direct Preference Optimization (DPO) can be combined with a trained reward model in what we call online-DPO, which also remains invariant under the reward equivalence relation. Our experiments suggest that online-DPO may achieve higher overall rewards, although it might be less efficient in maintaining a balance between the KL divergence and reward compared to PPO. Third, we propose a novel policy gradient algorithm called Pairwise Proximal Policy Optimization (P3O), which operates within a framework of Reinforcement Learning using Relative Feedback. This new approach is perfectly aligned with how the reward model is constructed, thereby avoiding additional complications like value function estimation, complex redistribution of sparse rewards, and various normalization steps. Empirical evaluations indicate that P3O consistently outperforms both PPO and DPO in balancing the KL divergence with the reward signal, as confirmed via GPT-4 Evaluation."},"source":{"paper_id":"JzAuFCKiov","cohort_label":"uncreative","pairing_group":"8","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,4.0,2.0],"soundness":[3.0,3.0,4.0,2.0],"presentation":[1.0,3.0,3.0,2.0],"overall_score":[5.0,6.0,6.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.75,"soundness_mean":3.0,"presentation_mean":2.25,"overall_score_mean":5.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"K98byXpOpU","group_id":"K98byXpOpU","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Bilevel optimization plays a central role in many machine learning applications, ranging from hyperparameter tuning and meta-learning to reinforcement learning. In essence, this class of problems involves a hierarchical structure where the upper-level objective depends on the solution of a lower-level problem; one decision affects the optimal response of another.\nRecent advances in gradient-based algorithms have yielded effective solutions for large-scale bilevel problems when the lower-level objective is unconstrained. In these cases, researchers have developed double-loop strategies that first approximate the lower-level solution using gradient techniques and then compute an approximate hypergradient—either through implicit or explicit differentiation—to update the upper-level variable. However, in certain real-world scenarios, such as sequential games where both levels must be updated simultaneously, these methods fall short. An alternative two-timescale approach has been proposed where the lower- and upper-level variables are updated alternately with different step sizes; although this method guarantees convergence, it does so with a suboptimal iteration complexity.\nTo further improve efficiency, single-loop single-timescale methods have been introduced. In these approaches, the ratio of the upper-level to lower-level step sizes is kept constant, leading to better iteration complexities when approaching a stationary point. Despite these advances, current methods are typically designed for unconstrained lower-level problems and assume that the upper-level objective is differentiable. When constraints exist in the lower-level problem, the upper-level objective becomes non-smooth, and these algorithms cannot be directly applied.\nAddressing these challenges, our work proposes a novel single-loop single-timescale method that comes with a proven convergence guarantee for lower-level constrained bilevel optimization problems. By employing Gaussian smoothing to design a new hypergradient estimator, our approach updates both levels simultaneously using a combination of a double-momentum technique and an adaptive step size strategy. We rigorously show that our method converges to a stationary point within a theoretically optimal number of iterations. In experiments on problems such as data hypercleaning and training data poisoning attacks, our algorithm outperforms several state-of-the-art alternatives in both training time and accuracy.\nIn summary, our contributions are threefold. First, we introduce a new hypergradient approximation based on Gaussian smoothing that enables a single-loop algorithm for constrained bilevel problems—unlike existing double-loop methods. Second, we provide a detailed convergence rate analysis demonstrating that our approach reaches a stationary point within an optimal iteration complexity. Finally, through comprehensive experiments in challenging application settings, we demonstrate the efficiency and superiority of our method compared to current state-of-the-art techniques."},"source":{"paper_id":"K98byXpOpU","cohort_label":"uncreative","pairing_group":"1","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,2.0,2.0],"soundness":[2.0,2.0,2.0,2.0],"presentation":[2.0,3.0,3.0,3.0],"overall_score":[5.0,5.0,5.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":2.0,"presentation_mean":2.75,"overall_score_mean":5.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"KszBlT26pl","group_id":"KszBlT26pl","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Understanding incentive mechanisms in complex decision-making systems is challenging, especially when little or no information is available about the system’s environment or agents. This challenge appears in fields such as cancer biology, where the mechanisms behind cancer cell metastasis are still unclear, and in human behavior, where agents can act unpredictably and enter unfamiliar regimes. The intrinsic randomness of these systems and the diverse characteristics of their individual elements further complicate the problem. Despite these difficulties, uncovering the underlying incentives could be key to better understanding these systems and designing targeted interventions to control them.\nInverse Reinforcement Learning (IRL) has proven to be a powerful tool in recovering the incentive mechanisms behind observed behaviors in various domains, including human behavior, robotics, and biology. However, IRL typically depends on sampling the next state of the environment using a known or empirically estimated transition model. This dependency is problematic when the dynamics of the environment are unknown or when it is challenging to sample transitions, which is often the case in interactions involving cancer cells or human agents. Empirical transition models may struggle to generalize to unobserved state and action spaces, particularly in high-dimensional settings with limited and noisy data. Additionally, recent IRL methods that address unknown transitions often rely on deep learning techniques, which, while effective, tend to lack interpretability and thereby make it difficult to gain scientific insights into the system’s behavior.\nMany complex systems, such as swarms or crowd dynamics, are based on mechanistic foundations that, if properly exploited, could lead to a more efficient and insightful understanding of their incentive structures. Motivated by this potential, we propose a new method called physics-constrained IRL. This approach simultaneously estimates the transition and reward functions using only trajectory data, while also inferring and incorporating the physical principles that govern the system. Our key contribution is a conjecture regarding the structural similarity between the physics governed by the optimal transport model—the Fokker-Planck equation—and the framework of a Markov Decision Process. By leveraging these physical principles, we develop computationally efficient models for decision processes and integrate them into our proposed physics-based FP-IRL algorithm. We demonstrate the effectiveness of FP-IRL through numerical experiments on both synthetic and real-world examples."},"source":{"paper_id":"KszBlT26pl","cohort_label":"uncreative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,4.0,3.0],"soundness":[3.0,1.0,4.0,3.0],"presentation":[3.0,2.0,4.0,3.0],"overall_score":[3.0,3.0,8.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.75,"soundness_mean":2.75,"presentation_mean":3.0,"overall_score_mean":4.25},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"LCpayOuqBx","group_id":"LCpayOuqBx","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Machine learning models perform well in closed-set scenarios where the training and testing data come from the same distribution. However, in open-world settings—especially in high-stakes applications like autonomous driving, where errors can be catastrophic—models often encounter out-of-distribution (OOD) samples that lie outside the training label space, resulting in unpredictable and often erroneous behavior. This challenge has sparked growing interest in OOD detection, which aims to distinguish OOD samples from test data while preserving classification accuracy.\nMost current OOD detection methods rely on a well-trained in-distribution (ID) classifier to identify OOD samples and are often restricted to settings where the label spaces between ID and OOD datasets differ. Additionally, these methods typically depend solely on visual patterns, overlooking the relationship between images and their textual labels. Recently, a new setting known as zero-shot OOD detection has emerged. This approach leverages large-scale vision-language models, such as CLIP, to detect OOD samples across various ID datasets without needing training samples. By constructing a textual classifier using only ID class labels, these methods have shown impressive performance compared to traditional techniques. However, relying exclusively on ID class labels tends to struggle with hard OOD samples.\nIntriguingly, our findings indicate that models like CLIP can achieve far superior OOD detection performance when actual OOD class labels are incorporated into the process. This suggests that using only closed-set ID labels is insufficient for reliably identifying hard OOD samples. In real-world open-world scenarios, however, the actual OOD label space is generally unknown, raising an important question: Is it possible to generate potential outlier class labels for OOD detection without accessing test-time data?\nTo address this challenge, we explore the use of large language models (LLMs), leveraging their expert knowledge and reasoning abilities to \"dream\" potential outlier semantics, a method we call DOS. Our approach does not rely on any actual or auxiliary OOD data. Instead, we design LLM prompts to generate potential outlier class labels tailored for OOD detection based on visual similarity. For example, given an ID class like \"horse,\" a prompt might ask for three categories that visually resemble a horse. We also introduce a novel scoring function that measures the proportionality between potential outlier and ID class labels to effectively distinguish hard OOD samples.\nUnlike previous methods that generate “not-ID” categories by requiring additional training on image description generators or extra datasets for adapted architectures, our DOS approach is entirely zero-shot. It offers several advantages: it is OOD-agnostic (requiring no prior knowledge of unknown OOD data), zero-shot (usable across diverse ID datasets with a single pre-trained model), and scalable and generalizable, even to large-scale datasets such as ImageNet-1K across far, near, and fine-grained OOD detection tasks.\nIn summary, our contributions are as follows:\n1. We introduce DOS, a new paradigm that leverages expert knowledge from LLMs to generate potential outlier class labels for zero-shot OOD detection.\n2. We propose three LLM prompts that can dream up potential outlier class labels applicable to a variety of datasets and OOD detection tasks.\n3. We design a new scoring function based on the proportionality between potential outlier and ID class labels, facilitating more effective differentiation between ID and OOD samples.\n4. Our DOS method significantly outperforms strong baselines in various OOD detection tasks without increasing inference time, showing notable improvements on far, near, fine-grained OOD, and large-scale datasets."},"source":{"paper_id":"LCpayOuqBx","cohort_label":"uncreative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,2.0,3.0],"soundness":[2.0,4.0,2.0,4.0],"presentation":[3.0,4.0,3.0,3.0],"overall_score":[3.0,8.0,3.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":3.0,"presentation_mean":3.25,"overall_score_mean":5.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"LqRGsGWOTX","group_id":"LqRGsGWOTX","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Bilevel optimization involves solving an optimization problem that contains another nested optimization problem. This approach has gained significant attention in machine learning tasks such as meta-learning, hyperparameter optimization, continual learning, reinforcement learning, and neural network architecture search. In the bilevel formulation, the objective is to minimize an upper-level function with respect to a variable x while ensuring that another variable y is chosen as the unique minimizer of a strongly convex lower-level function given x. For example, in hyper-representation learning with deep neural networks, x represents a shared representation across layers and y corresponds to a classifier in the final layer. In practice, one only accesses noisy estimates of the true functions based on underlying data distributions.\nExisting bilevel optimization methods typically assume that the upper-level function has bounded smoothness, meaning that its gradients are Lipschitz continuous. However, this assumption excludes many popular neural network models such as recurrent neural networks, LSTMs, and Transformers, which are known to exhibit unbounded smooth behavior. Although relaxed smoothness conditions have been proposed for single-level problems by linking the Hessian to the gradient norm, extending these ideas to bilevel problems is nontrivial.\nAddressing bilevel optimization with unbounded smooth upper-level functions presents two main challenges. First, the hypergradient, which represents the gradient of the overall bilevel objective, is extremely sensitive to the accuracy of the lower-level solution; even small errors in y can significantly amplify hypergradient estimation errors. Second, the bias in this hypergradient is intertwined with both the error in approximating the lower-level solution and the hypergradient itself, making it difficult to analyze and control.\nTo overcome these issues, we introduce a new algorithm called BO-REP. This approach is the first to tackle bilevel problems with potentially unbounded smooth upper-level functions by incorporating two innovative techniques for updating the lower-level variable: initialization refinement and periodic updates. Specifically, when a new x is introduced, the algorithm refines the corresponding y using a first-order method, and it then updates y periodically rather than at every iteration. Moreover, the upper-level variable is updated using normalized momentum, a strategy that helps manage stochastic gradient noise and large gradients.\nOur contributions are threefold. First, we design BO-REP, which leverages novel techniques for updating the lower-level variable in bilevel optimization problems with unbounded smooth upper-level functions. Second, we prove that when the upper-level function is nonconvex and unbounded smooth while the lower-level function is strongly convex, BO-REP can find an ϵ-stationary point within a number of iterations that matches the state-of-the-art complexity results (up to logarithmic factors) achieved under bounded smoothness assumptions. Lastly, through experiments on hyper-representation learning, hyperparameter optimization, and data hyper-cleaning for text classification, we demonstrate that BO-REP consistently outperforms existing bilevel optimization algorithms."},"source":{"paper_id":"LqRGsGWOTX","cohort_label":"creative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0,3.0,3.0],"soundness":[3.0,2.0,2.0,3.0,4.0],"presentation":[2.0,3.0,2.0,2.0,4.0],"overall_score":[6.0,6.0,6.0,8.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6,"soundness_mean":2.8,"presentation_mean":2.6,"overall_score_mean":6.8},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"Mgq6kxl115","group_id":"Mgq6kxl115","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Deep Ensemble (DE) is a straightforward yet powerful technique for improving the performance of deep neural networks. This method involves averaging the outputs of multiple models that are trained independently with different random initializations and data orders. Essentially, DE is a form of bagging where models are trained on fixed datasets, and it can be seen as an approximation of Bayesian Model Averaging. DE has been shown to significantly enhance prediction accuracy, improve uncertainty quantification, and boost out-of-distribution robustness across various tasks.\nHowever, a major drawback of DE is that during inference, the computational overhead and memory usage increase linearly with the number of ensemble members. This can be problematic in resource-limited environments or when dealing with very large models. To address this issue, several approaches have been proposed to reduce the inference cost—ideally down to the cost of a single forward pass—without severely degrading predictive performance. One popular method is ensemble distillation, where the average output of multiple ensemble members is used as the teacher output, and a single student model is trained to mimic this output. Other techniques include using multi-head models, shared weights, learning ensemble distributions, diversity-promoting augmentation, or training generators to simulate ensemble predictions. Despite these efforts, many of these methods either underperform compared to DE or require inference costs similar to that of DE to achieve comparable results.\nA recent, alternative approach reduces inference costs by linking ensemble members through low-loss subspaces. This approach constructs Bezier curves between ensemble members, creating low-loss subspaces from which parameters are sampled (typically from the center). To streamline the inference process, a lightweight neural network called Bridge Network (BN) is introduced. BN takes an intermediate feature from an ensemble member and directly predicts the output from the low-loss subspace. Ultimately, the final output is approximated by averaging the ensemble member’s output with the BN prediction. Although BN significantly reduces inference costs, it has two main limitations. First, BN predicts only for models on low-loss subspaces, which necessitates additional training to learn these subspaces between every pair of ensemble members. Second, a single BN can only connect a pair of ensemble members, so as the number of ensemble members increases, the number of required BNs grows quadratically. This structure leaves room for further improvements despite the sub-linear scaling in inference cost.\nIn our work, we propose an improved method that addresses BN’s limitations while further reducing inference costs. We start by designating one ensemble member as the starting point and then train a mapping to transport that model's output to the average output of the entire ensemble. Unlike BN, our approach does not require learning low-loss subspaces or constructing mappings for every pair of ensemble members, though it does involve training a more complex mapping. To achieve this, we use the Diffusion Schrödinger Bridge (DSB), a powerful model that builds a stochastic path between two probability distributions. By leveraging DSB, we create a sequence of predictions that progressively transition from those of the starting point model to the ensemble's output, effectively capturing the correlations among the predictions. To further lower inference costs, we design the DSB’s score network to be lightweight and incorporate diffusion step distillation. We call our approach the Diffusion Bridge Network (DBN), and our experimental results show that it significantly improves ensemble inference efficiency, surpassing BN and other ensemble distillation techniques."},"source":{"paper_id":"Mgq6kxl115","cohort_label":"creative","pairing_group":"1","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0,3.0,2.0],"soundness":[2.0,3.0,2.0,3.0,2.0],"presentation":[2.0,3.0,2.0,3.0,4.0],"overall_score":[6.0,8.0,6.0,8.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.4,"soundness_mean":2.4,"presentation_mean":2.8,"overall_score_mean":6.6},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"N0isTh3rml","group_id":"N0isTh3rml","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Graph neural networks (GNNs) have achieved remarkable success in various applications, such as molecular generation and link prediction, by learning informative representations from graph-structured data. A common learning paradigm in GNNs is the message-passing mechanism, where node and edge features are used to propagate information through the graph's topology. In a typical GNN, the model is constructed by stacking several layers that update node features based on messages aggregated from neighboring nodes. Essentially, while different GNN architectures may vary, they primarily differ in the design of the topological layouts used during message propagation.\nEarly GNN models directly used the original connectivity of the input graph, which sometimes led to suboptimal performance. More recent approaches focus on designing sophisticated sets of layouts that decouple the computational structure from the raw graph, thereby providing benefits for downstream tasks. Researchers have explored various aspects in this design, including spectral information, expressivity, and attention mechanisms. Particularly, attention-based GNNs have demonstrated state-of-the-art performance on many public datasets. In parallel, methods that incorporate subgraph sampling—using techniques such as random walks or edge rewiring—have also been used to generate useful topological layouts by introducing a degree of controlled randomness.\nHowever, most existing GNNs obtain their topological supports either through deterministic mappings or by introducing localized randomness. In realistic scenarios, the potential layouts associated with a graph may span a wide distribution when considered under global criteria. For instance, the different three-dimensional conformations of a molecule, defined by its atom-bond graph, tend to follow a Boltzmann distribution governed by a global free energy. Relying on a single conformation can therefore be insufficient to fully capture the underlying energy landscape and the global information of the graph.\nMotivated by this observation, we propose an approach that considers a distribution over topological layouts. Inspired by energy-based methods in graph visualization, where diverse and viable configurations are generated through optimization over physics-driven energy functions, we adopt a similar idea. These energy functions depend solely on the graph's topology and can be coupled with various GNN models. By employing a Boltzmann distribution defined on the energy surface, we can sample diverse layouts, for example using Langevin Dynamics. In practice, starting from a sufficiently random configuration, even noise-free optimization can yield diverse layouts that enhance downstream performance. Importantly, these layouts can be computed in advance and then integrated into standard GNNs.\nWe introduce Distributional Edge Layouts (DELs), a strategy that samples layouts from the designated distribution based on a global energy function, and then incorporates this information into GNNs as an additional supplement. Experimental results show that DELs can significantly improve the performance of selected GNNs across various datasets, demonstrating both their flexibility and broad applicability.\nIn summary, our contributions are as follows:\n• We propose a generic GNN framework that samples topological layouts from a Boltzmann distribution with an explicit energy surface, enabling the capture of a wide spectrum of global graph information.\n• The computation of DELs is independent of the GNN architecture, which allows for high flexibility and broad applicability.\n• As a plug-in component, DELs substantially enhance the performance of existing GNNs, achieving state-of-the-art results on multiple datasets."},"source":{"paper_id":"N0isTh3rml","cohort_label":"uncreative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,4.0,1.0],"soundness":[3.0,4.0,2.0],"presentation":[3.0,4.0,2.0],"overall_score":[5.0,6.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.3333333333333335,"soundness_mean":3.0,"presentation_mean":3.0,"overall_score_mean":4.666666666666667},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"OHpvivXrQr","group_id":"OHpvivXrQr","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Recent advances in deep learning have led to the development of AlphaFold 2, a groundbreaking method for predicting protein three-dimensional structures. With minor modifications, AlphaFold 2 can be extended to predict the structure of multimers—proteins made up of multiple chains—which is essential for understanding molecular functions and cellular signaling in various biological processes. Although this extended method accurately predicts the structures of small-scale multimers, its performance rapidly declines as the number of chains increases.\nAnother line of research in multimer structure prediction follows the idea of step-wise assembly, where each step represents a protein–protein interaction. In this approach, the multimer is built by sequentially adding the chain with the highest docking probability. This method leverages recent breakthroughs in dimer structure prediction, making it more effective for handling larger multimers.\nHowever, the current state-of-the-art assembly-based methods score interactions using independent protein–protein interactions (I-PPI) without considering the influence of additional proteins. For instance, if chain x has already docked to chain y, the interface on x available for an interaction with chain z is partially occupied, potentially lowering the docking probability for the (x, z) pair compared to (y, z). We refer to this additional influence as conditional protein–protein interaction (C-PPI), and its neglect can lead to poor generalization.\nIn our work, we address assembly-based multimer structure prediction by incorporating C-PPI knowledge along with I-PPI. There are two main challenges in effectively learning C-PPI: first, there are significant gaps in C-PPI information among multimers of different sizes, suggesting that their biological assembly processes may vary; and second, experimental data for large-scale multimers is extremely limited, further complicating model generalization.\nInspired by recent advancements in prompt learning—which have enhanced model generalization to new tasks and datasets—we explore whether a model can be prompted to predict C-PPIs for multimers of arbitrary scale. Our core idea is to design learnable prompts that transform multimers of any size into a fixed-scale format. Specifically, we define a target training task as conditional link prediction, and we introduce an auxiliary pre-training task where the model learns to assess the correctness of an arbitrarily assembled multimer. By transforming two query chains into a virtual assembled multimer and feeding it into the pre-trained model, we obtain a correctness score that serves as the linking probability between the chains, thus reformulating the arbitrary-scale prediction problem into a fixed-scale one.\nWe evaluate our approach in three experimental settings: first, assembling with ground-truth dimer structures to assess the accuracy of the predicted docking path; second, assembling using pre-computed dimers from one state-of-the-art multimer prediction method; and third, assembling using pre-computed dimers from another recent method. In all settings, our method shows improved accuracy—measured by metrics such as RMSD and TM-Score—and enhanced computational efficiency compared to recent state-of-the-art multimer structure prediction techniques. Overall, our findings demonstrate that incorporating C-PPI through prompt learning offers a powerful and broadly applicable solution for predicting multimer structures."},"source":{"paper_id":"OHpvivXrQr","cohort_label":"creative","pairing_group":"0","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,3.0,2.0],"soundness":[3.0,2.0,2.0,2.0],"presentation":[4.0,1.0,1.0,2.0],"overall_score":[8.0,5.0,5.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":2.25,"presentation_mean":2.0,"overall_score_mean":5.25},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"PLoWVP7Mjc","group_id":"PLoWVP7Mjc","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Learning deep, overparameterized neural networks with stochastic gradient descent, backpropagation, and large-scale datasets has led to tremendous progress in deep learning. In practice, it is often observed that effective deep learning requires a vast number of training samples and many training iterations. In our work, we question whether such large amounts of data and long training periods are truly necessary for achieving high test accuracy.\nTo investigate, we limit the training set to a very small number of samples (for example, 1, 5, or 10 images per class) and restrict the number of training steps to around 300. This approach leads us to explore dataset distillation—optimizing a small synthetic dataset so that neural networks trained on it perform well on the target distribution. Conceptually, this is a bilevel optimization problem: the output of one optimization process (training a network on the small dataset) is fed into another that minimizes the generalization error on the target data. In practice, the inner loop of this problem involves a multi-step computation that is often intractable if unrolled fully.\nEarly approaches to dataset distillation attempted to solve this problem by backpropagating through time over a limited number of steps. However, these methods quickly encountered optimization bottlenecks. Later methods replaced the inner loop with differentiable surrogates—such as using kernel approximations or Gaussian processes—to alleviate some of these challenges, but these techniques often require training a diverse set of randomly initialized models to bridge the gap between the surrogate and the actual model. Moreover, while these surrogate approaches work reasonably well with mean squared error loss and for wider models, their performance degrades when applied to narrower, commonly used architectures. Other strategies attempted to modify the outer loop objective by matching training trajectories or gradients, but these methods often suffered from the need to store multiple trajectories or delivered unsatisfactory performance.\nIn our paper, we revisit backpropagation through time (BPTT) and address its challenges in the context of dataset distillation. The core difficulty lies in the non-convex nature of training a network from scratch on the distilled data, requiring long unrolling of the BPTT process to capture long-range dependencies. Unfortunately, this extended unrolling causes slow optimization, high memory consumption due to maintaining all intermediate states, and instability in the meta-gradients due to repeated multiplication of Hessian matrices.\nTo overcome these issues, we introduce Randomized Truncated Backpropagation Through Time (RaT-BPTT). Our approach integrates randomization and truncation into the BPTT process by unrolling a smaller, randomly positioned fixed-size window over the training trajectory and aggregating gradients over that window. This design provides a random subsample of the full trajectory while enhancing gradient stability and reducing memory demands. Compared to full BPTT, RaT-BPTT offers faster training and improved performance, without relying on large-width approximations. This makes our method applicable to a broad range of architectures, including narrower models where surrogate-based methods often struggle.\nIn addition, we analyze the structure of the distilled datasets to encourage further progress. Past observations have shown that distilled data tends to exhibit high intercorrelation. For example, when training on a subset of a distilled dataset (say 10 images per class from a 50-image-per-class set), the resulting performance can be significantly worse than either training on the full distilled dataset or even on an equally small set of random training images. To address this, we propose a boosting strategy, Boost-DD, which produces a nested, modular synthetic dataset containing high-performing subsets for various budgets. Boost-DD can be seamlessly integrated with existing gradient-based distillation methods, reducing higher-order dependencies in the distilled data while preserving performance.\nIn summary, our contributions are as follows. First, we propose the RaT-BPTT algorithm, which integrates randomization and truncation with traditional BPTT to achieve state-of-the-art performance on several benchmark datasets. Second, we introduce Boost-DD, a boosting approach that yields a nested synthetic dataset free from detrimental intercorrelations. The remainder of the paper is organized as follows. We begin with a review of related work, then describe and motivate our RaT-BPTT algorithm. Subsequent sections present experimental results, compare with prior work, and detail our boosting algorithm, ending with a discussion on future improvements."},"source":{"paper_id":"PLoWVP7Mjc","cohort_label":"creative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,2.0],"soundness":[3.0,3.0,2.0],"presentation":[3.0,3.0,3.0],"overall_score":[6.0,8.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6666666666666665,"soundness_mean":2.6666666666666665,"presentation_mean":3.0,"overall_score_mean":6.333333333333333},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"POFrdKvpea","group_id":"POFrdKvpea","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Recent advances in neural radiance fields (NeRFs) have attracted significant attention because of their impressive performance in novel view synthesis, 3D reconstruction, and perception. To enhance both rendering quality and training or inference efficiency, recent research has incorporated explicit 3D representations such as voxel grids, point clouds, and meshes. These representations allow NeRFs to more effectively learn and model 3D scenes by explicitly encoding the scene’s geometry and topology into structured data. However, a major drawback of using these explicit representations is the substantial increase in model size, which poses challenges in storing and transmitting NeRF models. In this context, we focus on the critical issue of explicit NeRF compression—reducing the model size without sacrificing rendering quality or other key performance metrics.\nSeveral previous works have explored volumetric NeRF compression techniques using methods like voxel pruning, vector quantization, and wavelet transforms. While these approaches have achieved promising compression performance, they mainly concentrate on the voxel representation and overlook redundancies present in both the 3D geometry structure and textured latent features. For example, latent features tend to distribute in a way that closely mirrors the original color information, indicating inherent spatial redundancy that could be exploited for compression. Moreover, most existing methods prioritize compression performance while neglecting the compression time budget, a factor that is crucial for real-world applications.\nIn this paper, we first analyze the problem of explicit NeRF compression in depth and reformulate it as a 3D data compression issue. Building on this insight, we propose a novel compression framework named ACRF, which leverages advanced 3D data compression techniques. Our framework primarily compresses the explicit neural representation by first pruning the neural 3D structure to retain only the points with features, then employing an importance-guided feature encoding method. Finally, we apply entropy coding to further reduce the size of the encoded data based on an importance-prioritized entropy model. To balance compression performance with encoding speed, we present two pipelines: ACRF, which integrates rendering and bitrate loss for superior compression performance, and ACRF-F, which achieves rapid NeRF compression through direct distribution fitting.\nOur main contributions are as follows. First, based on a thorough analysis of explicit NeRF models, we shift the focus from traditional model compression to 3D data compression, leading to a new perspective on NeRF compression. Second, we introduce a time-efficient pruning strategy using view-independent information and develop two importance-based components: one for guided feature encoding and another for prioritized entropy modeling. Third, our framework is compatible with common explicit representations such as voxel grids and point clouds, striking an effective balance between rate-distortion performance and encoding speed. Experiments on several datasets demonstrate the promising scalability and potential of our method."},"source":{"paper_id":"POFrdKvpea","cohort_label":"creative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,3.0,2.0],"soundness":[3.0,3.0,4.0,3.0],"presentation":[2.0,3.0,4.0,2.0],"overall_score":[6.0,8.0,8.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.75,"soundness_mean":3.25,"presentation_mean":2.75,"overall_score_mean":7.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"Po6lYYsrB4","group_id":"Po6lYYsrB4","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Most current vision models rely on learning from passively curated datasets. They perform well on inputs similar to those in the training data but tend to struggle when conditions change. While pre-training on large datasets like ImageNet or COCO offers some improvement, these datasets still capture only static snapshots of objects. As a result, models pre-trained on them often fail to generalize to new environments and tasks that differ semantically from the training data.\nAn intuitive way to overcome these limitations is to use active agents in an embodied environment. Learning from active exploration enables the acquisition of data that are directly relevant to the task at hand, capturing informative and diverse instances. Moreover, actively collecting labeled data in this way enhances fine-tuning for downstream tasks. Embodied learning also provides the important benefit of incorporating action information—a signal absent in static datasets. A classic study demonstrates that an active learner develops a superior visual system compared to a passive one, even when both experience the same visual stimulation.\nIn this work, we introduce Action-Aware Embodied Learning for Perception (ALP). Our approach leverages action information and active exploration to both learn visual representations and collect data for downstream fine-tuning. In the first phase, we obtain a task-agnostic visual representation by combining an exploration policy with objectives for predicting dynamics. In the second phase, for each downstream task, a subset of the data collected earlier is labeled, and the pretrained backbone is fine-tuned accordingly.\nOur method incorporates signals from interaction in two ways. First, we use an inverse dynamics objective during representation learning to explicitly integrate action information by predicting the action that transitions one view to another. Second, we use a shared backbone for both representation learning and the reinforcement learning policy guiding exploration. This shared approach not only reinforces the visual representation through policy gradients on the exploration reward but also contrasts with methods that rely on separate backbones for rewards and policies.\nIn summary, our main contributions are:\n• We present an embodied learning framework that actively acquires visual representations while simultaneously collecting data for downstream fine-tuning. Unlike previous approaches that focus on a single aspect, our method integrates both stages.\n• We propose extracting action information from embodied movements in two complementary ways: explicitly through inverse dynamics prediction and implicitly via a shared backbone between policy and representation learning. We thoroughly investigate the impact of both approaches through experiments across various tasks—object detection, segmentation, and depth estimation.\n• We demonstrate that our method outperforms baseline embodied learning and self-supervised methods, as well as popular pre-trained models from static datasets, particularly on out-of-distribution scenarios."},"source":{"paper_id":"Po6lYYsrB4","cohort_label":"uncreative","pairing_group":"15","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,3.0,2.0],"soundness":[3.0,4.0,4.0,3.0],"presentation":[3.0,3.0,4.0,4.0],"overall_score":[3.0,6.0,6.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.75,"soundness_mean":3.5,"presentation_mean":3.5,"overall_score_mean":4.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"Q53QLftNkA","group_id":"Q53QLftNkA","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Rapid advances in hardware have enabled the development of increasingly complex models and fueled a growing appetite for data. This has spurred significant interest in learning useful deep representations from unlabelled data. One effective approach is self-supervised learning, which trains deep neural networks by solving pretext tasks that generate supervisory signals directly from the data. In the audio domain, various self-supervised methods have been developed. Some approaches use autoregressive models to predict future data based on past inputs, while others learn by contrasting different views of the input. There are also methods that remove portions of the input data and train networks to predict these missing segments.\nRecent breakthroughs in transformer architectures and their successors have advanced masked predictive modeling across domains such as natural language processing, computer vision, and audio processing. Masked Autoencoders, originally designed for learning visual representations from images, have gained widespread popularity due to their scalability and straightforward design. In audio applications, adaptations of Masked Autoencoders have been developed to learn versatile audio representations from spectrogram inputs. These adaptations address domain-specific challenges and provide valuable insights into the effects of masking strategies and other hyperparameters when training on audio data.\nLocal information can be effectively leveraged within the multi-head attention modules of transformer layers through methods such as convolutions, local-window attention, or pooling attention, leading to improved performance in various applications. Some studies have shown that using local windowed attention in a self-supervised framework can enhance audio representation learning on several downstream audio recognition tasks. However, previous methods typically apply the same local context to all attention heads in a module and require special techniques to link separate windows and blend local and global information.\nTo overcome these limitations, we introduce Multi-Window Masked Autoencoders (MW-MAE), which incorporate both local and global attention to learn comprehensive audio representations from spectrograms. The decoder in an MW-MAE is enhanced with a novel Multi-Window Multi-Head Attention (MW-MHA) module. In this module, each attention head computes self-attention over non-overlapping windows of various sizes, enabling the modeling of local and global interactions within every decoder block. Compared to standard Masked Autoencoders, MW-MAEs demonstrate superior performance on a wide range of downstream audio recognition tasks. They also show greater adaptability to different patch sizes and higher numbers of patches, perform more robustly in low-data scenarios, and scale more effectively with changes in encoder and decoder complexities.\nDetailed analyses of attention distances and entropies reveal that even when using an identical architecture, the encoder attention heads in MW-MAE learn to capture broader local-global relationships. Furthermore, evaluations based on canonical correlation analysis indicate that attention heads with the same window sizes across different decoder layers produce correlated features, resulting in a decoupled feature hierarchy. This confirms that the MW-MHA modules successfully learn local and global features within each decoder block."},"source":{"paper_id":"Q53QLftNkA","cohort_label":"creative","pairing_group":"1","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,1.0,2.0,2.0],"soundness":[2.0,2.0,2.0,2.0],"presentation":[2.0,3.0,3.0,2.0],"overall_score":[6.0,6.0,6.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.0,"soundness_mean":2.0,"presentation_mean":2.5,"overall_score_mean":5.25},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"QhoehDVFeJ","group_id":"QhoehDVFeJ","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Recent years have seen an increasing shift toward human digitization, a technology that now profoundly impacts entertainment, education, design, and engineering. Until recently, photo-realistic 3D reconstructions of humans relied on industry solutions that used dense observations. Such reconstructions were achieved through full-body laser scans, synchronized multi-view camera systems, or specialized light stages. However, these approaches are expensive, cumbersome to deploy, and involve complex processing pipelines that hinder widespread adoption.\nAn alternative is to view the problem as one of inverse rendering and to learn digital humans directly from sparse observations. Traditional methods that directly optimize explicit mesh representations often suffer from overly smooth geometry and coarse textures. They also typically require professional artists to design human templates, rigging, and UV unwraps. More recent methods leverage volumetric-based implicit representations and neural rendering, which allow for the digitization of photo-realistic human avatars from sparse multi-view or even single-view video footage. In particular, volumetric-based implicit representations have been shown to reconstruct scenes or objects with greater fidelity than earlier neural renderers, and they offer a more user-friendly approach as they eliminate the need for preset human templates, rigging, or UV coordinates. Simple inputs such as captured video and corresponding skeleton tracking suffice for training. Yet, achieving better reconstructions and usability with these methods comes with certain trade-offs.\nFirst, there is inefficiency in rendering. These methods often require long optimization times—ranging from tens of hours to days—and slow inference speeds. Volume rendering constructs images by querying millions of spatial coordinates for density and color information. During training, only a small fraction of these points are sampled due to memory constraints, leading to slow convergence. Second, efficiency issues arise in modeling how the human skin deforms (or skinning). Previous methods rely on static skinning templates, which are not trainable and thus produce suboptimal results. One recent method improved quality by learning subject-specific skinning through a root-finding approach to query canonical points, but this dramatically increases training time. Third, there is an incompatibility with current graphics pipelines. While volume rendering is powerful, it does not integrate well with the widely used techniques for rendering triangular or quadrilateral meshes. Many applications, such as editing, simulation, real-time rendering, and ray tracing, depend on these mesh-based workflows. Although methods exist to convert volumetric fields into meshes, the discrete sampling introduces quality degradation in both geometry and textures.\nTo address these challenges, we present a new method called EMA (Efficient Meshy neural fields for reconstructing animatable human avatars). EMA combines the flexibility of implicit representations with the efficiency of explicit meshes, without sacrificing photo-realistic quality. Our method digitizes humans using canonical triangular meshes, physically based rendering materials, and skinning weights associated with skeletons, all learned jointly via inverse rendering in an end-to-end manner. Each component is encoded by a separate neural field, which removes the need for a preset human template, rigging, or pre-defined UV coordinates.\nSpecifically, EMA predicts a canonical mesh from a signed distance field using a differentiable marching tetrahedra technique, and then extends this approach to handle spatially varying materials by using a neural field to predict PBR materials on the mesh surface after rasterization. To ensure the canonical mesh is animatable, another neural field models the forward linear blend skinning, allowing the mesh to be transformed into new poses based on a posed skeleton. Finally, a rasterization-based differentiable renderer shades the mesh, and the entire system is trained with a photometric loss. Once training is complete, the mesh and its materials are exported and the neural fields are discarded.\nThis design offers several notable advantages. In terms of rendering efficiency, the use of mesh rendering enables real-time performance. The training speed benefits as well, since loss is computed on the entire image and gradients only propagate along the mesh surface. This contrasts with volume rendering, where gradients spread throughout the entire spatial volume and only a limited set of pixels are used to compute loss. Our method requires roughly one hour of training, and plausible reconstruction quality can be achieved within minutes of optimization. Moreover, the separation of shape, material, and motion modeling allows for straightforward editing. To further enhance quality, we also optimize image-based environment lighting and non-rigid components.\nIn terms of skinning efficiency, instead of relying on a root-finding approach to query canonical points, our method uses differentiable rasterization to jointly learn both mesh properties and forward skinning in an end-to-end manner—a process orders of magnitude faster. Additionally, canonical meshes with forward skinning handle out-of-distribution poses more robustly. The derived mesh representation’s compatibility with standard graphics pipelines means that the produced models can be directly used in downstream applications, such as editing in design software.\nWe extensively evaluated our method on standard benchmarks, demonstrating competitive performance in novel view synthesis, better generalization to new poses, and significant improvements in both training time and inference speed compared to previous approaches. Our research-oriented implementation achieves real-time inference speeds at over 100 frames per second when rendering images at 512 × 512 resolution. We further showcase applications including novel pose synthesis, material editing, and relighting."},"source":{"paper_id":"QhoehDVFeJ","cohort_label":"uncreative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,3.0,2.0],"soundness":[2.0,3.0,4.0,2.0],"presentation":[2.0,3.0,3.0,2.0],"overall_score":[6.0,6.0,6.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":2.75,"presentation_mean":2.5,"overall_score_mean":5.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"Qox9rO0kN0","group_id":"Qox9rO0kN0","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Multi-agent reinforcement learning (MARL) techniques enable multiple agents to interact with simulated or real-world environments to solve complex decision-making tasks. These methods have found broad application in areas like autonomous driving, order dispatching, and gaming AI. In typical MARL scenarios, agents continuously interact in a shared environment and refine their individual policies through observations to collectively achieve a common objective. Although this problem can be viewed through the lens of independent reinforcement learning—where each agent treats others as part of its environment—the constantly evolving strategies of those agents create dynamic uncertainty and instability, making effective collaboration challenging. Moreover, policies learned independently can overfit to the specific behaviors of other agents.\nIn this work, we focus on cooperative scenarios in partially observable environments where agents have varying observation capabilities. Each agent is powered by a deep feed-forward network supplemented by a communication channel that transmits continuous vectors. Due to bandwidth limitations, particularly in wireless settings, only a limited number of agents can exchange messages at each time step. This constraint requires agents to carefully decide what information to share and with whom.\nMany existing approaches rely on sharing information among all agents or use manually predefined communication structures. However, when dealing with a large number of agents, such all-to-all communication can make it difficult to extract valuable information for cooperative decision-making. This approach may even hinder learning because it demands high bandwidth, introduces delays, and increases computational complexity. Although dynamic adjustments to the communication graph during inference have been explored, these methods often assume that every agent must always communicate with another, leading to complex scheduling algorithms and wasted bandwidth—issues that lie outside the scope of our work.\nTo overcome these challenges, we introduce CommFormer, a novel approach designed to enable efficient and effective communication among agents in large-scale MARL environments. Our method treats the communication structure as a learnable graph and formulates the task of determining this graph as part of the optimization process. Instead of searching over a discrete set of candidate communication architectures, we relax the search space to a continuous domain, allowing the architecture parameters to be optimized end-to-end via gradient descent. Unlike traditional approaches that combine internal states through arithmetic or weighted means before message transmission, our method directly passes each agent’s local observations and actions to designated agents based on the learned communication graph. An attention mechanism is then employed by each agent to dynamically assign credit to the received messages, ensuring steady performance improvements.\nOur extensive experiments on various cooperative tasks show that CommFormer is robust and effective across diverse scenarios. The proposed method consistently outperforms strong baseline approaches and achieves performance similar to methods that allow unrestricted information sharing, regardless of variations in the number of agents.\nIn summary, our key contributions are as follows:\n• We conceptualize the communication structure as a graph and propose a novel algorithm that simultaneously optimizes this graph and the architecture parameters through bi-level optimization.\n• We integrate an attention mechanism within a graph modeling framework to dynamically assess incoming messages, guaranteeing steady performance improvements and enhanced communication efficiency.\n• Through comprehensive experiments on a range of cooperative tasks, we demonstrate that CommFormer outperforms existing baselines and performs comparably to methods that rely on complete information sharing among agents."},"source":{"paper_id":"Qox9rO0kN0","cohort_label":"creative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,2.0,4.0],"soundness":[3.0,3.0,3.0,4.0],"presentation":[2.0,2.0,3.0,4.0],"overall_score":[6.0,6.0,8.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":3.25,"presentation_mean":2.75,"overall_score_mean":7.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"RFLZFxoLnE","group_id":"RFLZFxoLnE","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Neural networks have achieved impressive success in tackling various challenging real-world tasks. However, understanding their generalization performance remains a complex problem for researchers. Many factors affect a model's generalization error, including the network's structure, the datasets used, and the chosen optimization algorithm. Modern neural networks often have millions of parameters, resulting in a highly complex parameter space that is extremely challenging to analyze in terms of generalization error. A clearer perspective emerges when considering the function space, since a neural network is designed to approximate the true model within this space rather than simply navigating the parameter space.\nRecent work has shown that in the infinite width limit, the training dynamics based on parameter updates can be transformed into a differential dynamical system associated with the Neural Tangent Kernel (NTK) in function space. However, traditional gradient descent methods such as Stochastic Gradient Descent, RMSProp, and Adam operate directly in the parameter space. In contrast, Natural Gradient—which incorporates curvature information—is a reparameterization-invariant gradient-based optimization method that exhibits a strong connection with function space. Within this space, the training dynamics of a neural network can be understood as training across different eigenspaces. Since each eigenspace, associated with different spectra, contributes differently to the training dynamics (and consequently to the generalization error), it is conceivable that modifying the training dynamics within these eigenspaces could improve the learned function's generalization performance.\nBuilding on these insights, we propose an approximate explicit solution for an over-parameterized neural network trained using a modified natural gradient descent method under specific assumptions. Based on this solution, we interpret the generalization error as resulting from the discrepancy between the training data distribution and the true data distribution. This interpretation leads to a criterion for modifying the eigenspaces of the Fisher information matrix, with the goal of reducing the generalization error bound.\nFurthermore, several methods have been proposed to enhance generalization performance, including approaches that suppress gradients for cross-domain generalization and techniques based on self-distillation. These methods can be connected to our theoretical framework, as they implicitly modify the eigenvalues of the Fisher information matrix in function space, thereby influencing the training direction within that space."},"source":{"paper_id":"RFLZFxoLnE","cohort_label":"uncreative","pairing_group":"1","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,2.0,2.0],"soundness":[2.0,2.0,3.0,2.0],"presentation":[3.0,2.0,3.0,2.0],"overall_score":[5.0,3.0,3.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.25,"presentation_mean":2.5,"overall_score_mean":3.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"RgELE1dQXx","group_id":"RgELE1dQXx","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Artificial intelligence has made remarkable strides in everyday applications, often through fully automated systems. In fields such as AI-assisted driving and healthcare, however, AI must interact with humans, which introduces unique challenges. One major issue is that humans may not consistently follow AI recommendations; designing systems that assume perfect adherence can lead to suboptimal advice. Another challenge is that people sometimes prefer to act autonomously and only seek AI guidance during critical or challenging moments. This interplay between human behavior and AI decision-making necessitates a deeper understanding of user preferences.\nIn response, we propose a decision-making model for human-AI interactions that addresses these challenges. The model includes a parameter representing the human’s adherence level—the probability that a person will follow the advice—which enables the system to adjust recommendations based on observed user behavior. Additionally, the model incorporates an option for the AI to defer giving advice, thereby respecting human autonomy when appropriate.\nTo support this specialized decision-making framework, we have developed tailored learning algorithms that are both theoretically sound and empirically efficient. Our approach builds on ideas from previous studies in human-AI interactions and theoretical reinforcement learning, but our focus is on providing rigorous foundations and improved performance for environments where human behavior plays a critical role.\nOn one hand, our model recognizes that when adherence is low, the optimal advice may differ from the optimal decision, emphasizing the need for adaptable strategies. On the other hand, our algorithms adopt two distinct approaches: one employs an optimism-based reinforcement learning method that leverages the unique structure of our decision model to achieve better theoretical and practical performance, while the other uses a reward-free exploration strategy, allowing the system to explore the environment before quickly adapting to any bounded reward function.\nIn summary, our contributions are twofold. First, we introduce a comprehensive decision-making framework for advice-giving that accounts for human adherence levels and includes the option for the AI to defer its recommendations in favor of preserving user autonomy. Second, we design specialized learning algorithms that deliver near-optimal advice policies with tighter sample efficiency and superior empirical performance compared to more generic reinforcement learning methods."},"source":{"paper_id":"RgELE1dQXx","cohort_label":"creative","pairing_group":"4","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0],"soundness":[3.0,3.0,2.0],"presentation":[1.0,2.0,3.0],"overall_score":[5.0,6.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.3333333333333335,"soundness_mean":2.6666666666666665,"presentation_mean":2.0,"overall_score_mean":5.333333333333333},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"SkETBJRKH7","group_id":"SkETBJRKH7","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Large Language Models (LLMs) have recently emerged as highly capable generalist systems with a surprising range of emergent capacities. While some argue that these models are approaching a form of general intelligence, others point out significant shortcomings—most notably, their poor ability to plan or perform faithful multi-step reasoning. Recent work has examined whether LLMs might develop an emergent capacity for planning similar to the human brain’s use of cognitive maps, which are relational structures that support planning in both humans and animals. This work revealed that various LLMs, from small open-source instances to state-of-the-art models like GPT-4, exhibit systematic failures in planning tasks. These failures include a tendency to hallucinate non-existent paths or fall into repetitive loops, suggesting they struggle to reason about cognitive maps. This raises the question: how can LLMs be improved to enable better planning?\nIn our work, we take a step toward enhancing planning in LLMs by drawing inspiration from the human prefrontal cortex (PFC), a brain region involved in executive function, decision-making, and reasoning. Cognitive neuroscience has shown that the PFC contains several specialized subregions that handle functions such as conflict monitoring, state prediction, state evaluation, task decomposition, and task coordination. Human planning arises from coordinated and recurrent interactions among these specialized modules rather than from a single, monolithic system.\nIt is interesting to note that LLMs sometimes display these individual capacities when examined separately, yet they often fail to integrate these abilities effectively to achieve a goal. For example, although LLMs can correctly identify when paths in a planning problem are invalid, they may still attempt to traverse them. This observation suggests that a PFC-inspired approach—where planning is performed through coordinated activity among multiple specialized LLM modules—might be beneficial.\nWe propose an architecture called LLM-PFC, which is composed of modules inspired by the functions of the PFC. Each module, built on an LLM instance (using GPT-4, through prompting and few-shot in-context learning), is designed to perform a specific function such as error monitoring, action proposal, state prediction, state evaluation, task decomposition, or task coordination. The architecture leverages the idea that tree search during planning can emerge from the coordinated efforts of modules responsible for proposing actions, predicting states, and evaluating those states.\nWe evaluated LLM-PFC on three challenging planning tasks. The first involved graph traversal tasks that require navigation in novel environments based on natural language descriptions—tasks that have proven extremely challenging for current LLMs, including GPT-4. The second task was the classic Tower of Hanoi problem, which necessitates multi-step planning and is known to depend heavily on PFC function. Finally, we tackled a more complex, real-world planning challenge involving logistics, specifically the transportation of goods. In all cases, our approach significantly improved performance. Ablation experiments further demonstrated that each individual module contributed meaningfully to the overall effectiveness of the architecture. These results underscore the potential of a PFC-inspired approach to enhance the reasoning and planning capabilities of LLMs."},"source":{"paper_id":"SkETBJRKH7","cohort_label":"uncreative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0,2.0],"soundness":[2.0,4.0,2.0,3.0],"presentation":[3.0,4.0,3.0,3.0],"overall_score":[5.0,5.0,5.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.75,"presentation_mean":3.25,"overall_score_mean":5.25},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"TilcG5C8bN","group_id":"TilcG5C8bN","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Conventional supervised learning heavily depends on large amounts of labeled data. To overcome this limitation, researchers have increasingly turned to self-supervised learning (SSL), which has made significant progress recently. SSL aims to learn inherent representations that remain invariant under various distortions by maximizing the similarity between representations obtained through different data augmentation methods. Despite improvements in SSL accuracy, a major challenge remains: the high training cost. For instance, methods like BYOL require significantly more computation—resulting in many more training iterations and higher per-iteration costs—compared to their supervised counterparts. This increased demand for resources poses a barrier for real-world applications, especially on resource-limited devices.\nAlthough some recent studies have started to address the efficiency of SSL, this topic has largely been overlooked. A few approaches have been proposed to improve convergence speed or reduce computational overhead by adding extra tasks or complex mechanisms. However, these methods often increase the overall complexity of the SSL paradigm. This raises a fundamental question: Is there a more general and effective way to significantly improve the training efficiency of SSL?\nThe unique training paradigm of SSL uses different data augmentations on two branches of a siamese encoder model, which creates an inherent similarity between the augmented inputs and their corresponding feature maps. This similarity presents an opportunity for reducing computational costs. Nevertheless, it remains unclear whether similar and dissimilar regions in the augmented inputs or feature maps are equally vital for SSL, whether these similarities are maintained in both low- and high-level features, and how to effectively leverage these similarities for efficiency gains.\nMotivated by these questions, we comprehensively explored the impact of similar regions on SSL accuracy. We examined two types of methods—reusing and removing computations—to take advantage of these similarities for saving resources. Our study showed that eliminating redundant computation on similar regions in both the input images and the activations across each layer can significantly reduce overall computation and speed up training. To counteract the region shrinking effect caused by convolution layers, we developed a strategy to efficiently identify and expand high-similarity regions, ensuring substantial computational savings. This strategy functions like a waxing-and-waning process.\nIn summary, we propose SIMWNW, a generic and efficient SSL framework that significantly reduces training costs while accelerating the convergence of SSL methods. SIMWNW can be seamlessly integrated with various SSL training methods to save computation during training. We evaluated its effectiveness in both training from scratch and transfer learning tasks. When training from scratch, compared to other SSL methods, SIMWNW achieved computation savings of up to 54% (averaging 40%) without sacrificing accuracy. In transfer learning experiments, it reduced computation costs by up to 51% (averaging 48%) while maintaining accuracy. Moreover, when compared to other efficient SSL approaches, SIMWNW consistently outperformed state-of-the-art methods, achieving average computational cost reductions of 18% and 14% in training from scratch and transfer learning tasks, respectively."},"source":{"paper_id":"TilcG5C8bN","cohort_label":"creative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,2.0,3.0],"soundness":[3.0,2.0,2.0,3.0],"presentation":[2.0,2.0,3.0,3.0],"overall_score":[6.0,6.0,6.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.5,"presentation_mean":2.5,"overall_score_mean":6.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"Ur4LqAOXIF","group_id":"Ur4LqAOXIF","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Machine learning systems today must function in dynamic, open environments where they frequently encounter out-of-distribution (OOD) samples—data points not seen during training. Standard supervised models are often fragile in these situations, lacking the awareness needed to handle unexpected inputs safely. Traditionally, the approach has been to train an in-distribution (ID) model offline and then derive an OOD detector from it. Various methods have been proposed to either regularize training for a more conservative decision boundary or to design scoring functions that distinguish between ID and OOD data. However, these techniques typically assume that the OOD detector remains static once deployed, which is a significant limitation in environments that continuously change.\nConsider, for example, a self-driving car navigating different terrains. As the vehicle moves, it may encounter diverse OOD conditions such as snow mounds, dirt mounds, or sand dunes—each varying with time and location. This scenario underscores the necessity of shifting from the traditional offline OOD detection approach to an online framework that can adapt in real time.\nIn response, we propose a novel online OOD detection framework that allows the detector to continuously adjust as new environmental data arrive. The key idea is to leverage data that arrive sequentially—whether from ID or OOD sources—to incrementally update both the ID classifier and OOD detector. The system is designed to minimize the risk of misclassifying observations at every time step. This online framework offers two main benefits. First, it is adaptive, as the detector can evolve in response to gradual shifts in the environment. Second, it is practical, as it efficiently processes a continuous stream of data without requiring laborious offline dataset assembly.\nTo implement this idea, we introduce an algorithm called Stream Out-of-Distribution Adaptation (SODA). In each round, the model receives an instance from the environment, predicts whether it is ID or OOD, and then obtains immediate feedback regarding its decision. Based on this feedback, the model updates its parameters to reduce future prediction errors. Moreover, we also present an unsupervised extension of SODA that operates without any external feedback, broadening its applicability. The SODA algorithm is versatile—it can work with a range of OOD scoring functions by appropriately defining its loss functions.\nThe strength of SODA lies in both its theoretical guarantees and empirical performance. Theoretically, we demonstrate that SODA achieves sub-linear regret, meaning that over time the cumulative difference between its performance and that of the best possible model grows slower than linearly. This result indicates that the online detector progressively converges to optimal performance and makes fewer mistakes as it processes more data. Empirical evaluations in both stationary and non-stationary settings confirm that the online detector steadily reaches optimal performance and significantly outperforms traditional offline detectors. Notably, on challenging benchmarks such as ImageNet-1k, our approach reduces the final average false positive rate by nearly 19 percentage points compared to the previous best method.\nIn summary, our work introduces an online OOD detection framework that is both adaptive and practical, providing a robust solution for handling the unpredictable nature of real-world environments."},"source":{"paper_id":"Ur4LqAOXIF","cohort_label":"uncreative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,2.0,2.0],"soundness":[3.0,3.0,3.0,3.0],"presentation":[3.0,2.0,3.0,3.0],"overall_score":[3.0,3.0,3.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.0,"soundness_mean":3.0,"presentation_mean":2.75,"overall_score_mean":3.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"WNxlJJIEVj","group_id":"WNxlJJIEVj","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Offline reinforcement learning, where an agent is trained on a pre-collected dataset and later evaluated in the environment, has attracted significant attention in recent years. By avoiding the risks associated with interacting with the environment during policy improvement, offline RL finds applications in numerous real-world scenarios, including commercial recommendation systems, health care, dialog systems, and autonomous driving.\nDespite eliminating the need for costly online exploration, restricting learning to static datasets introduces unique challenges. Directly applying off-policy algorithms in offline settings can lead to extrapolation errors, where inaccurate value estimates for out-of-distribution actions accumulate during the backup process. Previous studies have mitigated these errors by incorporating conservative priors—either by regularizing the policy or penalizing the value estimations. However, overly conservative updates may trap the learned policy in suboptimal solutions, particularly when the offline dataset is collected from a mixture of policies.\nMore recently, diffusion models have emerged as a powerful policy class in offline RL due to their ability to model arbitrary distributions. By fitting the entire dataset, diffusion models can effectively regularize the policy without sacrificing expressiveness. A common approach among diffusion-based offline RL methods is to use diffusion for long-term planning. These methods generate future trajectories containing state-action pairs, enriching the planning process by providing estimated future states and enabling more informed decision-making. However, such methods tend to overlook the diversity within offline datasets; they typically learn the overall distribution and generate trajectories that mirror this distribution. Consequently, if the dataset contains many low-return states, the probability of sampling high-return states remains relatively low. Techniques like guided sampling, which attempt to steer the generation toward high-return regions, have shown limited success, as samples may still be distributed across low-return areas.\nTo address this issue, we propose a novel method called Contrastive Diffuser (CDiffuser). Inspired by contrastive learning—which works by pulling similar samples together and pushing dissimilar ones apart—we adapt this idea to bias the diffusion model training toward favoring high-return states. In our approach, we softly divide the offline dataset into high-return and low-return states. We then train a diffusion-based trajectory generation model under the constraint that the generated states stay close to high-return states and away from low-return ones. This contrastive guidance improves the quality of the generated trajectories, which in turn enhances planning performance.\nWe evaluate CDiffuser on 12 benchmark tasks, and the results demonstrate its superior performance compared to existing approaches.\nIn summary, our main contributions are threefold. First, we introduce CDiffuser, a novel method that enhances diffusion-based reinforcement learning algorithms. Second, we incorporate contrastive learning on state returns to bias the diffusion model in a way that preferentially generates trajectories with higher returns. Finally, experimental results on standard benchmarks confirm the effectiveness of our approach."},"source":{"paper_id":"WNxlJJIEVj","cohort_label":"uncreative","pairing_group":"2","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,3.0,4.0],"soundness":[2.0,2.0,1.0,4.0],"presentation":[3.0,3.0,3.0,4.0],"overall_score":[5.0,3.0,3.0,10.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":3.0,"soundness_mean":2.25,"presentation_mean":3.25,"overall_score_mean":5.25},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"WQYHbr36Fo","group_id":"WQYHbr36Fo","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Self-supervised learning (SSL) has recently made significant progress, now outperforming traditional supervised pre-training across various domains and tasks. SSL methods typically work by maximizing the similarity between feature representations derived from positive pairs. However, these positive pairs are often generated using strong augmentations and average pooling, which can lead to inconsistent semantic alignment and a loss of localization information. This approach works well for carefully curated datasets but presents challenges when dealing with images that contain multiple objects or tasks that require precise location details.\nTo address these limitations, recent research in dense SSL has focused on defining positive pairs by identifying matching locations, thereby preserving distinct semantic cues. Dense SSL methods often rely on cutout and masking techniques for both convolutional neural networks and vision transformers. Some point-level methods construct positive pairs by matching corresponding points in feature space, while other region-level approaches use predefined image coordinates along with augmentations like cutout and masking to generate these pairs for teacher-student network alignment. Nevertheless, issues arise when there is limited overlap between views—often due to mismatches or overzealous cropping—causing features from the query view to exploit background information. This unintended effect, which we refer to as \"coupling,\" results in features from unrelated regions becoming less discriminative.\nIn this work, we propose a simple yet effective strategy to mitigate object coupling in dense SSL. We observe that many existing augmentation techniques, such as blockwise masks and cutout, independently apply masks to each view without considering the global context. This independence can leave the student views with insufficient semantic cues for proper feature alignment when large occlusions occur. To remedy this, we introduce Region Collaborative Cutout (RCC). RCC prevents the creation of excessively large masks by dividing the image into uniform regions and applying cutouts progressively—from larger regions to smaller ones—in a collaborative manner. This ensures that no individual mask exceeds a certain size threshold, thus reducing unwanted background information leakage.\nMoreover, we design a decoupling branch that eliminates incorrect dependencies between each region and its surrounding context. This is achieved by adding contextual noise to the backgrounds of selected regions. Specifically, we generate decoupled views by replacing the background regions defined by RCC masks with randomly sampled background images, then apply the self-supervised loss directly to features extracted from the corresponding foreground regions in both the decoupled and original key views. This approach encourages the network to learn stronger, more robust representations from each local region, independent of distracting contextual influences. Importantly, our decoupling branch can be integrated into existing methods with only a minor modification to the augmentation pipeline.\nIn summary, our contributions are threefold: first, we address the coupling issue in dense SSL by providing quantitative evidence of its occurrence across different architectures; second, we introduce the RCC method, which effectively reduces object coupling by improving feature alignment in masked regions; and third, we enhance feature robustness through our decoupling branch, ensuring that local regions are learned more independently from their surroundings. Our extensive experiments and ablation studies demonstrate the effectiveness of our approach and its individual components."},"source":{"paper_id":"WQYHbr36Fo","cohort_label":"creative","pairing_group":"8","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,3.0],"soundness":[3.0,3.0,3.0],"presentation":[2.0,3.0,2.0],"overall_score":[6.0,8.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":3.0,"soundness_mean":3.0,"presentation_mean":2.3333333333333335,"overall_score_mean":6.666666666666667},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"WX9cd9iII4","group_id":"WX9cd9iII4","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Algorithmic decision-making must avoid discrimination and uphold fairness to meet legal, ethical, and societal standards. In the U.S., laws such as the Fair Housing Act and the Equal Credit Opportunity Act require that decisions are free from systematic bias based on gender, race, or other sensitive characteristics. Yet, research across various domains has consistently demonstrated that algorithmic decision-making often falls short of these fairness requirements. A well-known example is Amazon’s tool for screening job applicants, which, during its use from 2014 to 2017, was found to systematically discriminate against women, thereby lowering their chances of being hired.\nAchieving fairness in off-policy learning is challenging because this approach relies on observational data that may already incorporate historical biases. Consequently, the resulting policy must be fair even if the data were generated under a different, potentially discriminatory, behavioral policy. A simplistic approach might be to exclude sensitive attributes like gender from the data; however, other attributes can serve as proxies, and the learned policy may still exhibit discriminatory behavior due to the underlying data-generating process. Therefore, a more nuanced method for handling sensitive attributes in off-policy learning is required.\nIn this paper, we propose a novel framework for fair off-policy learning from observational data. Our approach is designed to learn fair decision rules even when the data originates from a different behavioral policy that may have been biased. To the best of our knowledge, this work introduces the first neural method for fair off-policy learning. In our framework, fairness is incorporated at two levels: the action level and the policy value level.\nAction fairness ensures that individuals with different sensitive attributes but identical characteristics are treated equally—for instance, ensuring that a woman and a man with the same academic background have the same chance of receiving a student loan approval in credit lending. On the other hand, value fairness addresses the balance in utility or outcomes across different sensitive groups. This allows for policies that deliver, on average, similar benefits to individuals from diverse groups. Such an approach can help address historical underrepresentation; for example, by providing incentives to women who have been less likely to pursue academic careers in technology, thereby maximizing long-term societal benefit. We further introduce two variants of value fairness inspired by the principles of envy-free fairness and max-min fairness.\nOur contributions are threefold. First, we introduce and formalize distinct notions of fairness tailored to off-policy learning from observational data and provide a theoretical analysis of their interactions. Second, we propose a neural framework, FairPol, which combines fair representation learning with custom training objectives to yield policies that satisfy our fairness criteria. Third, we offer theoretical learning guarantees in the form of generalization bounds for FairPol and demonstrate the effectiveness of our framework through extensive experiments on both simulated and real-world data."},"source":{"paper_id":"WX9cd9iII4","cohort_label":"uncreative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,3.0,2.0],"soundness":[2.0,3.0,3.0,2.0],"presentation":[2.0,3.0,4.0,2.0],"overall_score":[6.0,5.0,8.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.5,"presentation_mean":2.75,"overall_score_mean":6.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"WdhtdjoaVw","group_id":"WdhtdjoaVw","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Over the past decades, Bayesian deep learning has demonstrated success by combining the strong predictive performance of deep models with the principled uncertainty estimation characteristic of Bayesian inference. These methods have proven effective in a variety of applications, including per-pixel semantic segmentation, risk-sensitive reinforcement learning, and safety-critical areas like medical diagnosis and diabetic detection.\nDespite these advances, Bayesian deep learning has sometimes lagged behind non-Bayesian methods in certain tasks. One reason for this may be challenges in parameter-space inference. It is often difficult to incorporate meaningful prior information about an unknown function into the inference process. For example, commonly used independent Gaussian priors for model parameters can lead to samples that are overly linear and may cause pathological behavior in deep architectures. Furthermore, the impact of these priors on posterior inference and on the resulting distributions over functions remains unclear and difficult to control, given the complexity and non-linearity inherent in such models.\nTo address these challenges, recent efforts have shifted toward performing Bayesian inference in function space rather than in parameter space. In this framework, the distributions of function mappings—derived from the distributions over model parameters—are treated as probability measures in function space, and the variational objective is defined directly in terms of these function distributions. This approach allows the use of more informative stochastic process priors, such as Gaussian Processes, which can encode prior knowledge about function properties like periodicity and smoothness through suitable kernel functions. Existing methods in function-space inference typically minimize a divergence or distance between the true posterior and the variational posterior processes, developing tractable procedures for estimating the corresponding functional variational objective.\nA common choice for measuring dissimilarity in variational methods is the Kullback–Leibler divergence. However, when applied to distributions over infinite-dimensional functions, the KL divergence can become infinite, rendering the variational objective ill-defined. For the KL divergence to be applicable, the prior must be absolutely continuous with respect to the variational approximate posterior—a condition that may not hold in many cases. For instance, comparing the KL divergence between distributions over functions generated by Bayesian neural networks with different architectures can lead to infinite values.\nIn this work, we introduce a new functional variational inference method that employs a Wasserstein bridge as a dissimilarity measure between distributions, which we call Functional Wasserstein Bridge Inference (FWBI). This approach circumvents the limitations of KL divergence in function space. Our contributions are threefold. First, we propose a new Bayesian inference framework in function space to overcome difficulties associated with parameter-space inference, such as defining meaningful priors and managing over-parameterization pathologies. Second, we introduce a variational objective based on the Wasserstein bridge instead of the KL divergence, and we demonstrate that this objective serves as a lower bound of the model evidence, establishing it as a well-defined objective for Bayesian inference. Finally, we validate our method by comparing its performance against leading parameter-space and function-space inference techniques across several tasks, showing both strong predictive performance and dependable uncertainty estimation."},"source":{"paper_id":"WdhtdjoaVw","cohort_label":"uncreative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,3.0,2.0,3.0],"soundness":[2.0,3.0,4.0,3.0,2.0],"presentation":[2.0,3.0,3.0,3.0,3.0],"overall_score":[3.0,6.0,6.0,6.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.8,"soundness_mean":2.8,"presentation_mean":2.8,"overall_score_mean":4.8},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"WhZoCLRWYJ","group_id":"WhZoCLRWYJ","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Over recent years, significant progress has been made in computational methods for addressing the Schrödinger Bridge problem, which is also known as the dynamic version of the entropic optimal transport problem. The goal of the Schrödinger Bridge problem is to find a diffusion process between two given distributions that is as similar as possible to a specified prior process, while the entropic optimal transport formulation focuses only on the joint marginal distribution at the beginning and end of the diffusion.\nTraditionally, research in this field has concentrated on entropic optimal transport between discrete probability distributions. However, inspired by recent advancements in generative modeling using diffusion processes, researchers have recently begun to develop solvers for continuous versions of these problems. In the continuous setup, a user only has sample access to the distributions and must reconstruct the entire diffusion process, including aspects such as its drift. Such solvers can be directly applied to tasks like image generation, image-to-image translation, and various small-scale data transfer applications, including those involving biological data.\nMost existing solvers for these problems rely on complex neural network parameterizations with numerous hyperparameters, which often lead to time-consuming training and inference. While this complexity may be justified in large-scale generative modeling, it can be unnecessarily heavy when the goal is simply to compute entropic optimal transport or the Schrödinger Bridge between moderate-dimensional data distributions, as is common in certain biological applications.\nIn this paper, we address this issue with the following contributions:\n1. We introduce a novel, lightweight solver for the continuous Schrödinger Bridge problem using a Wiener prior, corresponding to the entropic formulation with a quadratic transport cost. Our approach employs a straightforward, non-minimax learning objective combined with a Gaussian mixture parameterization. This enables us to solve the problem between moderate-dimensional distributions in a matter of minutes by avoiding computationally intensive max-min optimization, full process trajectory simulations, iterative learning, and MCMC techniques typically used in other continuous solvers.\n2. We demonstrate that our lightweight solver provably satisfies the universal approximation property for entropic optimal transport and Schrödinger Bridge problems defined over compact sets.\n3. We derive finite sample learning guarantees for our method, showing that the estimation error diminishes at the standard parametric rate as the sample size increases.\n4. We validate the performance of our solver through a series of experiments on both synthetic and real data, including cases involving biological data as considered in related studies.\nOur method builds on recent advances in the field by using log-sum-exp quadratic functions for parameterizing Schrödinger potentials to construct benchmark distributions and by interpreting entropic optimal transport as an energy-based model that optimizes a specific Kullback-Leibler divergence."},"source":{"paper_id":"WhZoCLRWYJ","cohort_label":"creative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,1.0,3.0,3.0],"soundness":[3.0,3.0,1.0,3.0,3.0],"presentation":[4.0,3.0,3.0,3.0,3.0],"overall_score":[8.0,5.0,5.0,8.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6,"soundness_mean":2.6,"presentation_mean":3.2,"overall_score_mean":6.8},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"XK7kyCVjqr","group_id":"XK7kyCVjqr","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Code translation refers to converting source code from one programming language to another, which is essential for migrating existing code, reducing legacy maintenance costs, and facilitating new platform development. Traditionally, this task has relied on rule-based methods that demand deep expertise in multiple programming languages and substantial manual effort. Recently, data-driven approaches have leveraged advancements in machine learning to automate code translation. Similar to neural machine translation, these methods generally depend on parallel code data—pairs of code snippets in different languages that are functionally equivalent and free of bugs—to train sequence-to-sequence models. However, available parallel code data are limited in both quantity and language coverage.\nTo address the scarcity of parallel code data, one approach employs a pre-training and fine-tuning framework. Large online code repositories provide vast amounts of open-source code, which are used to pre-train large language models with self-supervised tasks, such as masked language modeling. These models are then fine-tuned on smaller, specialized datasets for tasks like code translation. Nonetheless, the objectives of these pre-training tasks differ significantly from those of code translation, which can limit the performance on translating code accurately.\nAn alternative approach uses unsupervised learning methods adapted from neural machine translation, including techniques like back-translation and denoising auto-encoding. While these methods have shown promise in translating code without parallel data, they tend to introduce significant noise in the training process. This is especially problematic for code generation, where precision is paramount. Without sufficient parallel code, self-supervised and unsupervised models may sometimes learn incorrect mappings, directly copying tokens or statements from the source code and ignoring the target language's grammatical rules—a phenomenon we refer to as “shallow translation.”\nConsidering these limitations, it is crucial to efficiently generate high-quality, well-aligned parallel code data and to learn robust cross-lingual alignments. The challenge lies in the rigorous grammar of programming languages and the probabilistic nature of neural code generation, which can result in variance and noise. Generating parallel code across different languages and program types in a cost-effective way remains particularly difficult.\nIn this paper, we propose a novel semi-supervised code translation method named SPACoder, which leverages synthetic parallel code generation and enhanced cross-lingual alignment. Our approach first generates synthetic parallel code with varying levels of alignment and quality by combining snippet training, static code analysis, and compilation techniques. We then train the translation model using both the synthetic data and a smaller set of manually annotated parallel code through an alignment-ascending curriculum learning strategy. This curriculum prioritizes data with increasing alignment quality, which helps mitigate the shallow translation problem.\nSPACoder offers several advantages. The use of compilation and static analysis ensures that the synthetic parallel code is syntactically correct and well-aligned, while the curriculum learning method robustly handles data noise and progressively improves translation quality. Moreover, our method can be applied across different programming languages and types of programs with minimal additional overhead, enabling large-scale synthesis of parallel code. Experiments demonstrate that integrating synthetic parallel code can significantly enhance translation performance, particularly for low-resource languages. In our evaluations, SPACoder improved code translation performance by up to 30% across C++, Java, and Python, outperforming baseline methods in various translation directions and achieving notable improvements even with a limited number of annotated training instances."},"source":{"paper_id":"XK7kyCVjqr","cohort_label":"uncreative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0,3.0],"soundness":[3.0,4.0,2.0,3.0],"presentation":[3.0,2.0,3.0,4.0],"overall_score":[5.0,6.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":3.0,"presentation_mean":3.0,"overall_score_mean":5.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"XN6ZPINdSg","group_id":"XN6ZPINdSg","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Deep neural networks have achieved remarkable success across various domains, including computer vision and natural language processing. However, as these models are increasingly deployed in real-world and safety-critical applications—such as autonomous driving and medical diagnosis—it becomes essential to address concerns regarding their trustworthiness and reliability. In particular, ensuring that model predictions come with worst-case guarantees (known as certified robustness) and developing methods to statistically assess the uncertainty in these predictions (referred to as conformal prediction) have become critical areas of research.\nAdversarial manipulations during testing, where small perturbations can lead the network to make incorrect predictions, highlight the need for robust certification approaches that provide worst-case guarantees. Conformal prediction, known for generating prediction sets that capture uncertainty with a probabilistic guarantee (assuming exchangeability of data), offers a pathway to complement worst-case certification. However, its performance under worst-case adversarial conditions remains uncertain.\nRecent work has shown that integrating domain-specific knowledge and logical reasoning into data-driven models can significantly enhance worst-case certification. In our work, we seek to bridge robust certification and uncertainty certification by exploring the impact of knowledge-enabled logical reasoning on improving certified prediction coverage for conformal prediction. We address two primary questions: How can we efficiently integrate knowledge and reasoning capabilities into deep neural networks for conformal prediction? And, can we theoretically demonstrate that such a framework achieves higher certified coverage and improved accuracy compared to a single network?\nTo answer these questions, we present COLEP—a certifiably robust learning-reasoning conformal prediction framework based on probabilistic circuits. Our approach consists of a \"learning\" component, where multiple data-driven networks are trained, and a \"reasoning\" component, which incorporates logical relationships (for example, a stop sign must be octagonal) using probabilistic circuits. We provide theoretical results deriving the certified coverage of COLEP under bounded perturbations and account for the finite calibration set size. Our proofs establish that as long as the knowledge models provide meaningful insights, COLEP achieves higher certified coverage and accuracy than a single model.\nWe validate our method through extensive experiments on several benchmark datasets, including GTSRB, CIFAR-10, and AwA2. The results demonstrate that COLEP not only significantly improves certified prediction coverage compared to state-of-the-art baselines but also mitigates the tradeoff between prediction coverage and the size of the prediction sets. Additionally, various ablation studies highlight the impact and importance of different types of integrated knowledge.\nIn summary, our work introduces the first knowledge-enabled learning framework for conformal prediction that provides provable improvements in certified coverage under adversarial conditions."},"source":{"paper_id":"XN6ZPINdSg","cohort_label":"creative","pairing_group":"8","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,3.0,2.0,3.0],"soundness":[3.0,3.0,3.0,3.0,3.0],"presentation":[2.0,2.0,2.0,2.0,2.0],"overall_score":[8.0,6.0,6.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.4,"soundness_mean":3.0,"presentation_mean":2.0,"overall_score_mean":6.4},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"XVhm3X8Fum","group_id":"XVhm3X8Fum","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Although transformers have proven very successful in natural language processing, linguists have long maintained that language embodies hierarchical syntactic structures. Yet transformers do not appear to include an explicit mechanism for handling nested syntactic patterns of arbitrary depth. Recent theoretical work has demonstrated that finite-precision transformers cannot reliably recognize certain syntactic patterns, such as Dyck-2, although they can manage these patterns up to a bounded depth. Other studies have indicated that transformers exhibit a linear inductive bias rather than a hierarchical one unless they are trained well beyond convergence. This characteristic may help explain why transformers generally require more data than human children to learn language.\nIn this paper, we introduce a novel attention mechanism called stack attention, specifically designed to model hierarchical patterns. This method treats input vectors as items in a stack and performs a soft-selection not directly over individual input vectors, but over sequences of stack actions. The approach is motivated by the properties of pushdown automata, finite automata augmented with stacks that recognize the entire class of context-free languages—a framework that captures the essence of compositionality and recursion inherent in natural language syntax. Accordingly, we expect that stack attention will not only enhance the expressive power of transformers but also enable them to learn language more efficiently and generalize to unseen combinations of syntactic patterns in a manner that is more akin to human language acquisition.\nStack attention builds on earlier research in differentiable stacks—continuous functions developed to approximate the behavior of discrete stacks. We adapt two forms of differentiable stacks into our attention operators: one based on the \"superposition\" stack and another using a nondeterministic generalization. Whereas previous work attached differentiable stacks to recurrent neural networks as external memory modules, our method integrates them directly into the transformer architecture. In standard attention, a soft-selection is made over input vectors; in contrast, our differentiable stacks perform a soft-selection over sequences of push and pop actions on stacks of vectors, ultimately returning the topmost vector of the chosen sequence. By employing differentiable stacks as a drop-in replacement for standard attention, our transformed model acquires a latent understanding of syntax that requires no explicit syntactic supervision and, in the case of nondeterministic stack attention, is capable of recognizing the full class of context-free languages.\nOur experiments demonstrate that transformer language models equipped with nondeterministic stack attention learn context-free languages very effectively, outperforming baseline transformers and even surpassing previous models on a particularly challenging context-free language. In tests with a constrained parameter budget, transformer language models with stack attention outperform baseline transformers by 4.3 perplexity points on the Penn Treebank, and we also present results on a small machine translation task using data from a German-English dataset. Our code is publicly available."},"source":{"paper_id":"XVhm3X8Fum","cohort_label":"creative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,3.0],"soundness":[3.0,3.0,3.0],"presentation":[3.0,3.0,2.0],"overall_score":[8.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6666666666666665,"soundness_mean":3.0,"presentation_mean":2.6666666666666665,"overall_score_mean":6.666666666666667},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"YItWKZci78","group_id":"YItWKZci78","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"The mean-field Langevin dynamics (MFLD) offer powerful theoretical tools for analyzing optimization problems defined on spaces of probability measures, such as those encountered when training two-layer neural networks. In this framework, the McKean-Vlasov stochastic process is associated with a Wasserstein gradient flow that minimizes an entropy-regularized convex functional. Here, the introduction of Gaussian noise fosters exploration and helps ensure global convergence. These Langevin-based methods are particularly appealing because they capture the nonlinear aspects of learning and permit efficient particle discretizations, even though it remains a challenge to extend them beyond single-objective optimization in a principled way.\nIn our work, we extend MFLD to address distributional minimax optimization problems. We consider the entropy-regularized saddle point problem for a convex-concave functional defined over spaces of probability measures with finite variance and fixed base measures. This formulation includes objectives that integrate a potential function with respect to a product of measures, which may involve nonconvex and nonconcave interactions. Such problems naturally arise in applications like training generative adversarial networks, robust learning, and solving zero-sum games in reinforcement learning.\nThis setting leads us to consider mean-field Langevin descent ascent (MFL-DA) dynamics, which are coupled, distribution-dependent stochastic processes designed to simultaneously minimize the functional with respect to one probability measure and maximize it with respect to the other."},"source":{"paper_id":"YItWKZci78","cohort_label":"creative","pairing_group":"4","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[4.0,2.0,3.0,4.0,3.0],"soundness":[4.0,3.0,2.0,2.0,3.0],"presentation":[3.0,2.0,2.0,2.0,3.0],"overall_score":[8.0,6.0,6.0,8.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":3.2,"soundness_mean":2.8,"presentation_mean":2.4,"overall_score_mean":6.8},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"YXnggA4iiD","group_id":"YXnggA4iiD","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Recent advancements in machine learning have highlighted the increasing need for large amounts of labeled training data. In real-world situations, however, obtaining such data can be challenging, especially when specialized expertise is required, as in medical diagnostics or complex vision tasks. Active learning (AL) addresses this challenge by allowing a model to iteratively select the most informative subset of an unlabeled dataset for expert annotation, thereby maximizing annotation efficiency. Earlier efforts in AL have explored various approaches. Some methods estimate sample uncertainty directly from predictions, others use separate models to gauge informativeness, and still others employ relationships in latent feature spaces to increase the diversity of chosen samples.\nDespite these contributions, a significant challenge remains: preventing overfitting in active learning. This issue is illustrated by a visualization of feature representations learned after training a neural network with a small labeled subset from the CIFAR-10 dataset. While the model indeed captures semantically distinguishable representations for the labeled samples, there is a noticeable discrepancy between the distributions of the labeled and unlabeled data. In simpler terms, the latent distribution for the labeled set does not match that of the unlabeled set, which can cause the model to overfit to the non-representative labeled data and hurt its generalization performance. This distributional mismatch is a recognized problem in active learning and can arise from factors like a small initial labeled set or bias in the data selection process.\nA few prior works have attempted to address this issue by minimizing inconsistencies in predictions on unlabeled samples or by training additional modules to select samples likely to belong to the unlabeled set. However, these methods often fall short because they struggle to accurately capture the true statistical distribution of the overall dataset.\nIn contrast, our work introduces a new approach to active learning by directly calculating the distribution statistics of the entire dataset using a Gaussian Mixture Model (GMM). We leverage the parameters obtained from GMM to develop both distribution-aware training strategies and an informativeness metric based on differences in likelihoods. Specifically, we incorporate the GMM statistics into supervised and semi-supervised training procedures to guide the learning of robust representations that resist overfitting to noisy, unrepresentative samples. Furthermore, our novel informativeness metric, combined with a hybrid selection strategy that integrates prediction-based, model-based, and likelihood-based measures, helps reduce the distributional discrepancy. \nWe validate our approach on several standard benchmarks, including CIFAR-10, CIFAR-100, SVHN, and FashionMNIST, demonstrating significant performance improvements and achieving state-of-the-art results. Our methods also show superior performance on imbalanced datasets where the number of samples per category varies."},"source":{"paper_id":"YXnggA4iiD","cohort_label":"uncreative","pairing_group":"2","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,3.0],"soundness":[2.0,2.0,3.0],"presentation":[3.0,3.0,3.0],"overall_score":[3.0,6.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.3333333333333335,"soundness_mean":2.3333333333333335,"presentation_mean":3.0,"overall_score_mean":5.666666666666667},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"YbZxT0SON4","group_id":"YbZxT0SON4","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Intrinsic objectives are often employed to enhance exploration in reinforcement learning, especially in scenarios where rewards are sparse or absent. In Markov Decision Processes (MDPs) with limited states, count-based exploration methods can perform nearly optimally when used alongside tabular RL algorithms. These methods track how often an agent visits each state and use that information to provide an exploration bonus that encourages structured investigation.\nWhile research has focused on extending these ideas to larger state spaces and continuous environments, count-based approaches can lead to unstable learning dynamics. This instability arises because any reward that depends on state visitation counts is non-stationary. As the agent gathers new experiences, the underlying dynamics of these counts change, and the agent does not have the information required to estimate these changes accurately. In a stationary MDP, convergence of policies and value functions depends on both the transition dynamics and the reward distribution remaining constant. The non-stationarity of count-based rewards effectively turns the problem into a partially observable MDP, where the changing reward dynamics are hidden from the agent. In such settings, there is no guarantee that an optimal time-homogeneous (Markovian) policy exists, and achieving optimal performance may require non-Markovian reasoning to adapt to the evolving rewards.\nDespite these challenges, count-based methods are typically integrated with RL algorithms designed to converge to Markovian policies, potentially resulting in suboptimal performance. Previous studies have either overlooked this non-stationarity or attempted to address it by separating the exploration and exploitation policies, which adds complexity and can introduce further instability.\nIn our work, we propose a framework called Stationary Objectives for Exploration (SOFE) that modifies the algorithm in a simple way to eliminate the non-stationarity in intrinsic rewards. SOFE stabilizes the learning objective by augmenting the original state representation with state-visitation frequencies or a representative embedding. This augmentation effectively reformulates the intrinsic reward as a deterministic function of the state, although it requires the agent to operate over a larger state space. We argue that RL agents with parameterized policies are better at generalizing across larger state sets than at optimizing non-stationary rewards.\nOur empirical evaluations demonstrate that SOFE enables the learning of superior exploration policies. While SOFE primarily addresses the non-stationarity of count-based rewards, it also offers complementary benefits when combined with other exploration strategies such as pseudo-counts and state-entropy maximization. Additionally, our experiments show that SOFE is agnostic to the choice of RL algorithm and robust across diverse and challenging environments, including large 3D navigation maps, procedurally generated environments, sparse reward tasks, pixel-based observations, and continuous action spaces. Supplementary videos and a summary of our findings are available on our webpage."},"source":{"paper_id":"YbZxT0SON4","cohort_label":"creative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,2.0,3.0],"soundness":[3.0,2.0,2.0,3.0],"presentation":[2.0,3.0,2.0,3.0],"overall_score":[5.0,6.0,5.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.5,"presentation_mean":2.5,"overall_score_mean":6.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"YcM6ofShwY","group_id":"YcM6ofShwY","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Diffusion models gradually transform noise vectors into natural images, enabling a wide range of applications such as image synthesis, image inpainting, and text-to-image generation. However, these models still sometimes produce low-quality images that negatively affect user experience. A promising solution is to filter out these inferior generations, although this is challenging without a reliable metric for assessing individual image quality. Traditional metrics like the Fréchet Inception Distance and Inception Score measure distributional properties rather than evaluating quality on a per-sample basis.\nBayesian uncertainty, which distinguishes data similar to the training samples from those that are not, offers a natural way to address this problem. The idea is straightforward: the Bayesian posterior shows low uncertainty for data seen during training and high uncertainty for unfamiliar data, ensuring that the output images are as realistic as the training data. Yet integrating Bayesian uncertainty with diffusion models is not simple. These models typically involve large-scale networks, requiring efficient Bayesian inference techniques, and the complex reverse diffusion process further complicates accurate uncertainty quantification.\nTo overcome these challenges, we introduce BayesDiff, a framework designed to estimate pixel-wise Bayesian uncertainty for images produced by diffusion models. Our approach features a novel uncertainty iteration principle applicable to various sampling methods to track pixel-wise uncertainty throughout the reverse diffusion process. We also employ an efficient Bayesian inference method based on the last-layer Laplace approximation for pre-trained score models. BayesDiff simultaneously delivers both image samples and their corresponding pixel-wise uncertainty estimates. Although the basic implementation can be slow due to its reliance on Monte Carlo estimation, we have developed an accelerated variant to improve efficiency.\nBy aggregating pixel-wise uncertainty into a single image-level metric, our framework facilitates effective filtering, particularly for images with cluttered backgrounds. In text-to-image generation tasks, we found that pixel-wise uncertainty not only improves diversity by emphasizing high-quality outputs but also helps to correct images that suffer from artifacts or do not adequately match the textual description. Extensive experiments and thorough ablation studies have been conducted to provide a clear and intuitive understanding of the pixel-wise uncertainty estimates."},"source":{"paper_id":"YcM6ofShwY","cohort_label":"creative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,2.0,2.0],"soundness":[3.0,3.0,3.0,2.0],"presentation":[2.0,3.0,2.0,3.0],"overall_score":[8.0,6.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.75,"presentation_mean":2.5,"overall_score_mean":6.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"ZG3RaNIsO8","group_id":"ZG3RaNIsO8","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Large language models (LLMs) have demonstrated impressive performance on a variety of natural language processing tasks. One effective strategy for adapting these models to specific tasks is to add an instruction—known as a discrete prompt—to the input text. This approach guides the model to perform the desired task without substantially increasing computational cost and without the need to access all the internal parameters or gradients. This makes it particularly useful for models with black-box APIs, such as GPT-3 and GPT-4. However, the performance of LLMs is highly sensitive to the choice of prompt, which has made prompt engineering a critical and often labor-intensive practice that relies on subjective guidelines and considerable human expertise.\nTo reduce the manual effort involved in designing discrete prompts, earlier methods have sometimes depended on accessing token probabilities from LLM outputs or have explored strategies like generating a wide variety of prompts and selecting the best ones, or refining existing prompts to improve performance. Yet, these methods either risk inefficient resource use through excessive exploration or may become stuck in local optima by focusing too narrowly on fine-tuning current prompts. Derivative-free optimization algorithms, particularly evolutionary algorithms (EAs), offer a balance between exploration and exploitation. In this context, sequences of phrases in prompts can be treated similarly to gene sequences in biological evolution, suggesting that EAs are well suited for discrete prompt optimization.\nIn this paper, we propose a novel framework called EVO-PROMPT that combines the strengths of LLMs and EAs for automatic discrete prompt tuning. Traditional evolutionary operators typically modify tokens in isolation, which can disrupt the coherence and readability of prompts by neglecting the connections between tokens. Our approach leverages the expertise of LLMs to generate new prompt candidates under the guidance of evolutionary operators, while the evolutionary process itself iteratively selects and refines the most effective prompts based on performance on a development set.\nStarting with a set of initial prompts, we use LLMs to act as evolutionary operators, generating new candidates that are then evaluated. The best-performing prompts are preserved, and the population is updated through this iterative process. With carefully designed evolutionary operators and an adaptive update strategy, EVO-PROMPT can be implemented using various types of EAs. We demonstrate the effectiveness of this framework by optimizing prompts for different LLMs, including Alpaca and GPT-3.5, across a diverse array of language understanding and generation tasks, as well as challenging benchmark tasks, using 31 datasets. EVO-PROMPT consistently produces higher-quality prompts compared to both manually crafted ones and previous automatic prompt generation methods.\nThe main contributions of this work are threefold. First, we introduce EVO-PROMPT, a new framework for automatic discrete prompt optimization that combines LLMs with EAs. This framework does not require access to model parameters or gradients, strikes an appropriate balance between exploration and exploitation, and generates human-readable prompts. Second, through experiments on 31 datasets, we show that EVO-PROMPT outperforms manually designed prompts and existing methods, and we provide the optimal prompts for common tasks such as sentiment classification, topic classification, subjectivity classification, simplification, summarization, and reasoning. Third, we demonstrate that LLMs can effectively implement various types of EAs when given proper instructions, thereby opening up new possibilities for integrating LLMs with conventional algorithms and fostering innovative applications."},"source":{"paper_id":"ZG3RaNIsO8","cohort_label":"creative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,2.0,2.0],"soundness":[3.0,3.0,3.0,2.0],"presentation":[3.0,3.0,4.0,3.0],"overall_score":[6.0,6.0,8.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":2.75,"presentation_mean":3.25,"overall_score_mean":6.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"ZuoeYIGaSW","group_id":"ZuoeYIGaSW","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"In our dynamic, ever-changing world, the importance of sequential decision-making in machine learning is undeniable. Unlike static tasks, sequential decisions reflect the fluidity of real-world scenarios—from robotic manipulations to evolving healthcare treatments. Just as large language models have revolutionized natural language processing by understanding complex linguistic nuances, pretrained foundation models hold a similar promise for sequential decision-making. In language, these models capture syntax, semantics, and context, providing a robust starting point for numerous downstream tasks. Similarly, in sequential decision-making, where decisions depend on the interplay of past actions, current states, and future possibilities, a pretrained foundation model can build a rich, generalized understanding of decision sequences. This foundational knowledge, developed through diverse decision-making scenarios, can then be fine-tuned to address specific tasks.\nHowever, sequential decision-making presents unique challenges that differentiate it from vision and language pretraining paradigms. First, data distribution shifts occur because the training data—generated by specific behavior policies—differs considerably across pretraining, fine-tuning, and deployment stages, which can compromise performance. Second, task heterogeneity means that decision-making tasks vary widely in configuration, dynamics, and state-action spaces, making it hard to develop universally applicable representations. Third, the effectiveness of representation learning often relies on high-quality data and expert supervision, resources that are frequently either unavailable or prohibitively expensive in real-world scenarios.\nWe envision a foundation model for sequential decision-making that is versatile, efficient, robust, and compatible with existing large pretrained models. Such a model should generalize across a broad array of tasks, adapt to new scenarios with minimal data through few-shot learning, and remain resilient even when pretrained on lower-quality data. To address these challenges, we propose developing a universal and transferable encoder using a reward-free, dynamics-based temporal contrastive pretraining objective. This encoder is designed to handle tasks with complex observation spaces, such as visual inputs, by excluding reward signals during pretraining. By doing so, the model is better equipped to generalize across diverse downstream tasks, even when their objectives differ. With a world-model approach, the encoder learns a compact representation that captures universal transition dynamics—much like the laws of physics—making it adaptable across multiple scenarios and enabling effective transfer of knowledge to new control tasks.\nWhile existing works leverage self-supervised pretraining from rich vision datasets to build foundation models, these approaches often overlook control-relevant aspects and suffer from a domain gap between pretraining datasets and downstream visuo-motor tasks. In contrast, our novel approach, called Premier-TACO (Pretraining Multitask Representation via Temporal Action-Driven Contrastive Loss), introduces a control-centric objective function for pretraining. Premier-TACO incorporates a temporal contrastive learning objective that uses a simple yet effective strategy for selecting negative examples—sampling one negative example from a nearby window of the next state. This strategy ensures that the latent representation encodes control-relevant information necessary to differentiate between similar visual inputs, all without additional computational overhead. Extensive empirical evaluations demonstrate that Premier-TACO performs well even with smaller batch sizes and scales effortlessly for multitask offline pretraining.\nThe key contributions of our work are as follows:\n1. We introduce Premier-TACO, a framework for multitask offline visual representation pretraining in sequential decision-making problems, featuring a novel temporal contrastive learning objective. This approach is computationally efficient and well-suited for multitask representation learning.\n2. We demonstrate the versatility and efficiency of Premier-TACO’s pretrained visual representations through extensive empirical evaluations. For example, on a robotic manipulation benchmark, using only a handful of expert trajectories, Premier-TACO outperforms the best baseline pretraining method by a significant margin. Moreover, in challenging control tasks, it achieves state-of-the-art performance even with very limited demonstration data, generalizing effectively to unseen embodiments and camera views.\n3. We also show that Premier-TACO is robust to lower-quality data and compatible with existing large pretrained models. In our evaluations, Premier-TACO performed strongly even when the pretraining data was collected randomly, and when used to fine-tune a generalized visual encoder, it led to an average performance enhancement of around 50% across various tasks.\nThis work lays the foundation for building versatile, efficient, robust, and compatible foundation models that can drive advancements in sequential decision-making across a wide range of applications."},"source":{"paper_id":"ZuoeYIGaSW","cohort_label":"uncreative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,2.0],"soundness":[3.0,3.0,3.0],"presentation":[3.0,3.0,3.0],"overall_score":[5.0,5.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.0,"soundness_mean":3.0,"presentation_mean":3.0,"overall_score_mean":5.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"am9IxubLKV","group_id":"am9IxubLKV","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Convolutional neural networks (CNNs) represent a major milestone in the evolution of deep learning, primarily for how their \"sliding window\" approach provides a crucial inductive bias for vision tasks. This intuition is easy to grasp using graphical illustrations, though analyzing convolutions remains more challenging than working with fully connected layers in multi-layer perceptrons (MLPs) or even transformers. One difficulty is that convolutions are hard to express in matrix notation, and even with index notation, compact and convenient expressions exist only for specific hyper-parameter choices. Moreover, factors like stride, padding, and options such as channel groups add further complexity. As a result, related elements like higher-order derivatives and routines for automatic differentiation inherit this complexity.\nWhile many analytical and algorithmic advances were developed earlier for MLPs, similar progress for CNNs has lagged behind by several years. To help close this gap, we propose a new perspective on convolutions using tensor networks. These networks express tensor multiplications as diagrams, making it easier to visually capture structures like factorization, manipulate the diagram for tasks such as differentiation or batching, combine it with other diagrams, and extract sub-tensors as needed.\nTensor networks are not only beneficial for theoretical investigations; the quantum simulation community has developed efficient techniques to evaluate them. Accessible interfaces can automatically optimize aspects like the contraction order or distribution of computations, streamlining the process. For example, one can implement convolution-related calculations using a concise interface that automatically generates efficient expressions.\nAn illustrative example contrasts the standard approach, which unfolds the input for computing the Kronecker-factorized Fisher approximation (a method that can demand significant memory), with a tensor network approach that replaces the unfolded input with a diagram. This change allows an optimizer to automatically determine a more efficient contraction path, and in many cases, inherent patterns in the network further reduce computational cost. In practice, users do not need to manually implement the tensor network version since the framework can automatically generate the necessary optimized expressions.\nIn summary, our work leverages the tensor network format to derive diagrams for various automatic differentiation routines and structural approximations of second-order information. Our approach supports all hyper-parameters, batching, channel groups, and arbitrary dimensions. We also introduce methods to transform and simplify these tensor network diagrams by exploiting the connectivity patterns inherent in convolutions. The resulting framework leads to significant performance improvements—up to 4.5 times faster in some cases—and offers the flexibility to implement hardware-efficient techniques that reduce the cost of randomized backpropagation."},"source":{"paper_id":"am9IxubLKV","cohort_label":"uncreative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,4.0,2.0,3.0],"soundness":[2.0,4.0,3.0,3.0],"presentation":[2.0,4.0,3.0,3.0],"overall_score":[5.0,8.0,5.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.75,"soundness_mean":3.0,"presentation_mean":3.0,"overall_score_mean":6.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"b3kDP3IytM","group_id":"b3kDP3IytM","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Answering factual queries is just one of the many emerging abilities of large language models. This development has transformed the functioning of search engines by directly incorporating conversational language models into the user experience, as seen with platforms like BingChat. However, evaluating these new capabilities rigorously remains challenging due to the ongoing need for appropriate benchmarks, issues such as benchmark saturation and training data contamination, and the difficulty of assessing open-ended outputs generated by these models. At the same time, concerns have increased over instances where language models produce false information or provide irrelevant content, sometimes referred to as hallucinations.\nIn this work, we examine and evaluate the ability of language models to satisfy constraints within the scope of information retrieval. Similar to traditional constrained search problems, we define constraint satisfaction queries in information retrieval as those that include a set of requirements that the generated output must meet. This framework, recently proposed as a means to study and detect factual errors in language models, connects the popularity and feasibility of information with a model's ability to satisfy constraints. We adopt this framework to guide our evaluation and experimental design. In this context, queries with constraints can be seen as a more general form of keyword, boolean, or pattern-matching searches, as well as faceted web searches. For example, a query like “A list of research papers authored by [author] published after [year]” naturally imposes multiple constraints on the required output.\nAlthough several benchmarks exist for single-constraint factual queries that expect a single answer (such as “Which city is the capital of Ukraine?”), these benchmarks have largely reached saturation. Little is known about how well language models perform on more complex queries containing multiple types of constraints that require longer outputs. Consistency in adhering to multi-part constraints during longer text generations is crucial, especially since newer architectures generally exhibit higher self-consistency. Surprisingly, our analysis shows that even state-of-the-art models still struggle to remain consistent with external constraints.\nTo investigate this issue further, we introduce KITAB, a dataset along with a dynamic data collection approach focused on literature queries—a domain that benefits from efficient retrieval and has abundant publicly available information that may have also been used during training. KITAB comprises queries such as “A list of all books from Toni Morrison published between 1970-1980,” where the first constraint always specifies an author and subsequent constraints can cover lexical, temporal, or named-entity aspects.\nWe use KITAB to evaluate language models under controlled conditions, examining: (i) their baseline ability to retrieve all books by an author; (ii) their performance on queries combining both author and book constraints when relying solely on their internal (parametric) knowledge; (iii) their performance when provided with complete context including all books by the author, thereby distinguishing between parametric and retrieval-augmented settings; and (iv) their performance with prompts that either use standard chain-of-thought reasoning or require the model to first assemble its own context by retrieving all books by the author without external aid. These controlled settings allow us to identify and understand different failure modes for the task.\nOur findings indicate that when relying solely on internal knowledge, state-of-the-art models often include a high rate of irrelevant or hallucinated books not written by the specified author, with rates varying significantly based on the author’s popularity. Providing complete context can reduce the inclusion of irrelevant information, yet constraint satisfaction remains a persistent challenge. Moreover, self-retrieval approaches tend to further increase the presence of irrelevant details and fabricated titles in attempts to satisfy the specified constraints. While the performance of newer models shows some improvement compared to older ones, reliance on scaling alone may not resolve the underlying issues with filtering constrained queries. Overall, achieving perfect accuracy—defined as a complete match with the expected output—remains well below an acceptable threshold.\nIn addition to presenting detailed evaluations for two language models using KITAB, we also outline an approach for collecting and refining additional versions of the dataset based on a disjoint list of authors. This process is particularly important for addressing benchmark saturation or leakage issues and for supporting independent testing in cases where the original dataset may have been incorporated into model training."},"source":{"paper_id":"b3kDP3IytM","cohort_label":"creative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[4.0,3.0,3.0,2.0],"soundness":[2.0,3.0,3.0,3.0],"presentation":[3.0,3.0,4.0,3.0],"overall_score":[8.0,6.0,6.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":3.0,"soundness_mean":2.75,"presentation_mean":3.25,"overall_score_mean":6.25},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"bWcnvZ3qMb","group_id":"bWcnvZ3qMb","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Time series analysis plays a pivotal role in sectors ranging from healthcare to smart factories. In these areas, edge devices such as smart sensors—typically powered by microcontrollers with limited computational and memory resources—are commonly used. Time series data are inherently complex and dynamic, often presenting sparse and scattered information across the time domain. To exploit this type of data, recent research has developed sophisticated models and methodologies. However, the computational and memory demands of these models render them unsuitable for resource-constrained edge devices.\nConversely, representing time series data in the frequency domain offers a more compact and efficient way to capture inherent patterns. Although some studies have already explored the frequency domain—enhancing models by incorporating spectral data or using high-amplitude frequencies for feature extraction through convolutional networks—the full potential of the frequency domain remains largely untapped. In particular, the inherent advantage of using complex numbers to capture both amplitude and phase information is not fully leveraged, which often leads to continued reliance on compute-intensive models for temporal feature extraction.\nIn our study, we reinterpret tasks like forecasting and reconstruction as interpolation problems within the complex frequency domain. Essentially, by interpolating the frequency representation of a given segment, we can extend it to produce a longer time series segment. For forecasting, this means extending the look-back window through frequency interpolation, and for reconstruction, it involves recovering the original segment by interpolating the frequency representation derived from a downsampled version.\nBuilding on this insight, we introduce FITS (Frequency Interpolation Time Series Analysis Baseline). At its core, FITS uses a complex-valued linear layer that learns amplitude scaling and phase shifts, enabling effective interpolation in the complex frequency domain. Although FITS operates through frequency domain interpolation, it remains fundamentally a time domain model by integrating operations that transform input segments into the complex frequency domain, apply the necessary interpolation, and then convert the data back into the time domain for further processing. This approach makes FITS highly versatile for various downstream time domain tasks, including forecasting and anomaly detection.\nFITS also incorporates a low-pass filter to ensure that the representation is both compact and informative. Despite its streamlined linear architecture and simplicity, FITS consistently achieves state-of-the-art performance. Remarkably, it does so with fewer than 10,000 parameters—making it about 50 times more compact than other lightweight models and roughly 10,000 times smaller than mainstream alternatives. This efficiency in both memory and computation makes FITS an ideal candidate for deployment, or even for training directly on edge devices, for tasks such as forecasting and anomaly detection.\nIn summary, our contributions are as follows:\n• We introduce FITS, an exceptionally lightweight model for time series analysis with a parameter range between 5,000 and 10,000.\n• FITS employs a novel approach by using a complex-valued neural network that simultaneously captures amplitude and phase information, leading to a more efficient representation of time series data.\n• Despite its small size, FITS consistently delivers top-tier performance across a range of time series analysis tasks."},"source":{"paper_id":"bWcnvZ3qMb","cohort_label":"creative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,3.0,3.0],"soundness":[3.0,3.0,3.0,3.0],"presentation":[3.0,3.0,4.0,3.0],"overall_score":[8.0,8.0,8.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":3.0,"soundness_mean":3.0,"presentation_mean":3.25,"overall_score_mean":8.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"bxITGFPVWh","group_id":"bxITGFPVWh","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Recent machine learning models, such as large language and visual models built from publicly sourced data, have increased the need for defending against data poisoning attacks. In these attacks an adversary injects malicious samples into the training dataset, aiming either to significantly reduce overall model accuracy (untargeted attacks) or to force the model to misclassify specific inputs (targeted attacks). In backdoor attacks, the objective is to create hidden triggers in the model so that it produces a predetermined output whenever the trigger is present, regardless of the input.\nA major challenge with poisoning attacks on deep neural networks is the uncertainty introduced during the re-training process, which involves variations in training algorithms, model initialization, and even model architectures. This uncertainty can lead the re-trained model to converge to a different state than expected, thereby reducing the intended effect of the poisoning samples. Traditional methods typically optimize poisoning samples based on the outcome of a single victim model, but due to re-training variability the poisoning effect may be lost. Recent approaches have attempted to address this by optimizing the average poisoning effect across multiple models with different architectures and initializations; however, this ensemble and re-initialization strategy is computationally inefficient.\nConsidering poisoning attacks as bilevel optimization problems with multiple inner minima highlights the underlying theoretical and practical complexities. Addressing these challenges, we introduce a new strategy called the Sharpness-Aware Data Poisoning Attack (SAPA). This method aims to design poisoning samples that maximize the effect even in the worst-case scenario among re-trained models. In other words, even if the re-trained model happens to be the least affected, it still exhibits strong poisoning behavior. This is achieved by focusing on the sharpness of the victim model’s loss landscape—a concept usually associated with generalization properties of neural networks—to enhance the poisoning attack. Our experiments demonstrate that SAPA is a robust and principled strategy that improves the performance of various poisoning attacks, including targeted, untargeted, and backdoor attacks, while also maintaining its effectiveness across different re-training conditions and being more computationally efficient than previous ensemble-based methods."},"source":{"paper_id":"bxITGFPVWh","cohort_label":"creative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,3.0,3.0,3.0,2.0],"soundness":[3.0,3.0,3.0,3.0,4.0,3.0],"presentation":[3.0,3.0,4.0,3.0,3.0,2.0],"overall_score":[6.0,6.0,6.0,6.0,6.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6666666666666665,"soundness_mean":3.1666666666666665,"presentation_mean":3.0,"overall_score_mean":5.833333333333333},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"cUSNs8nGaV","group_id":"cUSNs8nGaV","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"According to the International Diabetes Federation, 463 million adults worldwide have diabetes, with 34.2 million people affected in the United States alone. Diabetes is a major cause of heart disease, blindness, and kidney disease. Managing blood glucose levels is crucial for diabetes care, yet achieving target levels is challenging due to factors such as diet, exercise, stress, medications, and individual physiological differences.\nContinuous glucose monitors (CGMs) are devices that measure blood glucose levels at frequent intervals—often about once per minute. These devices have significant potential to improve diabetes management by providing real-time feedback to patients and, when paired with an insulin pump, enabling an autonomous artificial pancreas system. The full benefits of CGM technology, however, depend on accurate glucose prediction models. Although many prediction models have been proposed, current CGM and artificial pancreas software typically rely on relatively simple physiological or statistical models. The lack of systematic evaluation protocols and established benchmark datasets has hindered a thorough analysis of more complex models, limiting their practical adoption.\nIn response, we introduce a curated collection of five public CGM datasets along with a systematic protocol for evaluating and benchmarking prediction models. The selected datasets vary in size and demographic characteristics, and our data-preprocessing pipeline facilitates the inclusion of additional datasets. We propose two tasks: enhancing predictive accuracy and improving the quantification of uncertainty associated with predictions. Predictive accuracy is measured using mean squared error (MSE) and mean absolute error (MAE), while uncertainty quantification is assessed using likelihood measures and expected calibration error (ECE). For each task, we train and evaluate a set of baseline models. Our baseline models include linear regression and ARIMA as examples of shallow data-driven approaches, Transformer-based models, NHiTS, TFT, and Gluformer as examples of deep learning methods, and the Latent ODE model as a hybrid data-driven and physiological approach.\nThis work contributes a diverse collection of CGM datasets, a clear formulation of two evaluation tasks focused on model accuracy and uncertainty quantification, an efficient benchmarking protocol, an assessment of various baseline models, and a detailed analysis of factors influencing model performance."},"source":{"paper_id":"cUSNs8nGaV","cohort_label":"creative","pairing_group":"1","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[1.0,2.0,4.0],"soundness":[2.0,2.0,3.0],"presentation":[2.0,2.0,4.0],"overall_score":[5.0,6.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.3333333333333335,"soundness_mean":2.3333333333333335,"presentation_mean":2.6666666666666665,"overall_score_mean":6.333333333333333},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"cbVZt3aN0b","group_id":"cbVZt3aN0b","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Machine learning models have become the de facto standard for many applications such as social media, finance, and healthcare. However, these models often face difficulties when data continually changes over time in an unending stream—a common scenario in real-world settings. This issue is part of an active research area known as Continual Learning (CL). Most existing work in this field assumes that labels are immediately available once data is collected, an assumption that may not hold in practice.\nConsider monitoring recovery trends in patients after surgeries. In this scenario, doctors routinely collect health data from numerous post-operative patients. Yet, this data does not immediately reveal broader recovery trends or highlight potential common complications. It can take several weeks of extensive monitoring and testing across multiple patients before the data can be labeled as indicating either \"recovery\" or \"complication\" trends. By the time the data is gathered, assessed, labeled, and used for model training, new patient data or changing health scenarios may already have emerged. This creates a cycle of data collection, trend assessment, labeling, model training, and deployment that is increasingly affected by what we call label delay.\nIn our work, we introduce a continual learning framework that explicitly accounts for the delay between the arrival of new data and the availability of corresponding labels. In our proposed setting, the model is continuously trained over time steps with a delay of d steps. At each time step t, the model receives two sets of data: unlabeled data from the current time step and labeled data from time step t − d. In the initial experiments, we analyze a simple approach that solely trains on available labeled data by waiting for the labels. Our results show that this setting is challenging, with model performance consistently degrading as the label delay increases. Furthermore, merely increasing the number of parameter updates per time step does not solve the issue.\nNext, we explore the potential of leveraging the most recent unlabeled data by integrating two promising paradigms: Self-Supervised Learning (SSL) and Test-Time Adaptation (TTA). Interestingly, among the various SSL and TTA methods we examined, none outperformed the simple baseline that trains on the delayed labeled data, given the same computational budget. To address this, we propose a straightforward method that samples previously rehearsed labeled data to better match the distribution of the newest unlabeled data. Our experiments demonstrate that this approach can effectively close the accuracy gap introduced by label delay in certain scenarios.\nOur work makes the following contributions: We introduce a new continual learning setting that explicitly factors in label delays caused by annotation latency. We conduct extensive experiments on several online continual learning datasets, comparing leading self-supervised learning and test-time adaptation techniques against the baseline that ignores label delay. Finally, we present Importance Weighted Memory Sampling, a method that selects previously labeled data most similar to the newest unlabeled data, significantly improving performance over both the naive method and the SSL and TTA approaches, while incurring only a negligible additional computational cost. An in-depth analysis of our proposed method is also provided."},"source":{"paper_id":"cbVZt3aN0b","cohort_label":"uncreative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,2.0,2.0],"soundness":[2.0,3.0,3.0,2.0],"presentation":[2.0,3.0,3.0,3.0],"overall_score":[3.0,3.0,5.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.0,"soundness_mean":2.5,"presentation_mean":2.75,"overall_score_mean":4.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"chVYVLJIAh","group_id":"chVYVLJIAh","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"In model-based reinforcement learning, an agent gathers information from its environment to learn a model that helps speed up and improve value estimation or policy learning. As the complexity of the environment increases, learning an accurate model becomes increasingly difficult, and any errors in the model can negatively affect the learned policy. In many real-world scenarios, agents face environments so complex that fully simulating them is infeasible; therefore, it becomes crucial to decide which aspects of the environment to model, avoiding wasted effort on irrelevant details like modeling clouds in a vehicle’s video feed.\nRecently, a new approach called decision-aware model learning (or value equivalence) has been proposed. The core idea is that a model should be evaluated based on how well its predictions support accurate value estimation. Two leading methods in this area are the IterV AML and MuZero algorithms. IterV AML is grounded in strong theoretical foundations, although it has encountered challenges in adapting its loss for practical implementations. On the other hand, MuZero has demonstrated impressive performance in discrete control tasks despite having less theoretical investigation. This paper addresses several open questions about the role of value-aware losses and the factors that drive strong performance.\nSpecifically, we explore three research questions: (a) What design decisions account for the performance differences between IterV AML and MuZero? (b) What are the theoretical and practical distinctions between the two approaches? and (c) In what scenarios do decision-aware losses enhance model-based reinforcement learning? Our experimental evaluations focus on state-based DMC environments, since our goal is not to scale the approach to image-based observations.\nOur contributions highlight key differences in neural network architecture and value function learning strategies between methods based on IterV AML and those based on MuZero. We show that MuZero, which builds on value prediction networks and explicitly uses a latent world model, can explain many of the performance differences reported in the literature. In our theoretical work, we prove that both IterV AML and MuZero-based models can achieve low error in stochastic environments—even when using deterministic world models—thus confirming a unique property of these decision-aware losses. However, our findings also reveal that the value learning scheme employed by MuZero introduces a bias in stochastic environments, which leads to measurable differences in performance. Finally, we demonstrate that MuZero’s approach to model learning closely resembles that of previous value-aware methods.\nIn summary, we show that IterV AML is a stable loss when a latent model is used, that the MuZero value loss is biased in stochastic settings, and that decision-aware losses can offer significant advantages over decision-agnostic baseline losses in model-based reinforcement learning."},"source":{"paper_id":"chVYVLJIAh","cohort_label":"uncreative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,2.0,2.0],"soundness":[4.0,2.0,2.0,2.0],"presentation":[4.0,3.0,2.0,3.0],"overall_score":[8.0,6.0,3.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.5,"presentation_mean":3.0,"overall_score_mean":5.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"eRAXvtP0gA","group_id":"eRAXvtP0gA","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Unsupervised learning is a vast field dedicated to finding patterns in data without prior knowledge of the underlying classes. Traditionally, many unsupervised methods, such as K-Means clustering, rely on defining clusters based solely on the mathematical properties of data representations. This reliance can limit their effectiveness when the mathematical relationships do not mirror the cognitive relationships that humans naturally perceive. For example, while the MNIST dataset has a clear cognitive meaning—each image represents a specific digit—the transformation into numerical values for clustering purposes often obscures that inherent relationship, emphasizing instead the similarities between their encodings.\nThe K-Means algorithm, one of the most well-known unsupervised techniques, uses Euclidean distance between encoded inputs to determine if they belong to the same class. In the realm of clustering grey-scale images, state-of-the-art methods like Invariant Information Clustering (IIC) focus on extracting patterns directly from images using convolutional neural networks. A key similarity between these methods and many other traditional approaches is that they assume an infinitely large mathematical space, wherein learning involves carving out boundaries between different classes. In such models, classes are represented as subspaces of this infinite space. An input is assigned a label based solely on the subspace in which it falls, regardless of its similarities or differences with other samples within that subspace. This can sometimes lead to accurate classifications of inputs that differ significantly from anything seen during training, but not always.\nIn contrast, emerging theories of brain function suggest that the brain creates constructive representations of the world—generating abstractions or archetypes that capture the essence of a class. According to these theories, for an input to be classified, it must be sufficiently similar to one of these learned representations. If it is not, then it cannot be confidently assigned to any known class. Mathematically, this approach differs from traditional methods by focusing on the construction of a set of representations rather than merely partitioning a representation space.\nTo empirically test this new approach, we propose an unsupervised learning algorithm designed to model the input space by generating constructive representations. In effect, our method automatically creates abstractions directly from the data. It requires a universal, representation-oriented data structure, and recent research indicates that a Sparse Distributed Representation (SDR) fulfills this role. SDRs, which have been shown to align with how the brain processes inputs, allow for a universal representation across different data types. We convert the input spaces into SDRs using an encoder–decoder pair we call Embodiment.\nUsing SDRs as the basis, we introduce an unsupervised learning algorithm called aPrimitive, which models the input space by generating constructive representations in a manner independent of the input type. We believe this algorithm can serve as a fundamental building block for future cognitive systems. In this paper, we detail the core idea of aPrimitive along with one of its modulators: the Spatial Attention modulator, which auto-regulates the spatial discriminability of the system. Additionally, we have designed our approach to be transparent and explainable, so that its representations and decision-making processes can be clearly described.\nAnother advantage of focusing on constructive representations is the ability of the algorithm to indicate when a new input does not correspond to any previously encountered pattern. We evaluated our method against traditional unsupervised classification algorithms—K-Means for tabular data and IIC for image data—as well as against a supervised clustering approach using K-NN. Our comparisons involved four static datasets (two tabular and two image datasets) for a classification task where labels were assigned based on the majority class in each cluster. The results demonstrate a clear advantage for our approach, which performs robustly across both tabular and image data. Moreover, experiments designed to assess cognition-like properties—such as recognizing MNIST digits despite the random removal of pixels—show that our method outperforms the compared algorithms, highlighting the benefits of building constructive representations.\nIn summary, our findings suggest that unsupervised learning algorithms focused on constructing representations have the potential to exhibit cognition-like properties. We conclude that our proposed approach is a disruptive method in the field of unsupervised learning, offering properties that are both innovative and promising compared to traditional methods. The remainder of the paper is organized as follows: a review of related work in Section 2, a description of our proposal in Section 3, an empirical evaluation in Section 4, a discussion in Section 5, and finally, conclusions and directions for future work in Section 6."},"source":{"paper_id":"eRAXvtP0gA","cohort_label":"uncreative","pairing_group":"0","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,1.0,2.0,3.0],"soundness":[2.0,1.0,2.0,3.0],"presentation":[2.0,1.0,2.0,3.0],"overall_score":[1.0,1.0,3.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.0,"soundness_mean":2.0,"presentation_mean":2.0,"overall_score_mean":2.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"fmoknhh7CH","group_id":"fmoknhh7CH","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Designing proteins that bind small molecules has many applications, from drug synthesis and energy storage to gene editing. A protein’s function largely depends on its ability to bind and interact with other molecules. For instance, proteins can be engineered to sequester toxins as antidotes or to act as enzymes that catalyze chemical reactions, which are central to biological processes. Our approach, FLOW SITE, addresses the challenge of designing protein pockets capable of binding small molecules by leveraging recent advances in deep learning-based protein design and protein-molecule docking.\nIn our work, we focus on designing protein pockets that bind a specific small molecule, known as the ligand. We assume that the protein pocket is provided as the 3D backbone coordinates of its residues and the ligand is described by its 2D chemical graph. No prior knowledge of the ligand’s 3D structure or binding pose is assumed. Based on this information, the goal is to predict the amino acid identities at the given backbone locations. We also tackle the more challenging task of designing pockets that bind multiple molecules and ions simultaneously—a scenario important in enzyme design where the ligands represent reactants. This multi-ligand binding problem, which has not been addressed by deep learning so far, differs from previous work on protein–protein interactions and presents unique challenges, such as the lack of readily available evolutionary information.\nExisting approaches for designing drug-binding proteins or enzymes typically rely on expert knowledge and involve several manual steps. In contrast, FLOW SITE provides a more general and automated deep learning solution for designing binding pockets specific to small molecules. FLOW SITE is a flow-based generative model that handles both discrete variables (residue identities) and continuous variables (ligand poses). Its training relies on a flow matching criterion that guides the model to learn a self-conditioned flow for the joint generation of contact residues and the 3D binding structures of the ligands.\nAt the heart of FLOW SITE is HARMONIC FLOW, a generative process specifically developed for 3D poses of ligands, including cases where multiple ligands are involved. Unlike other docking generative methods that operate on aggregated properties such as a ligand's center of mass, orientation, or torsion angles, HARMONIC FLOW directly updates the Cartesian coordinates of ligand atoms. This direct approach produces chemically plausible structures without constraining ligand flexibility to torsional motions. In several new pocket-level docking tasks, HARMONIC FLOW outperforms state-of-the-art diffusion-based methods; for example, a larger percentage of its predictions achieve a low root-mean-square deviation compared to the alternative method.\nBuilding on the success of HARMONIC FLOW, FLOW SITE extends this process to generate discrete residue types alongside continuous ligand poses. An additional data augmentation strategy, which treats side chains as ligands, further enhances training by creating extra cases. Overall, FLOW SITE recovers a higher percentage of binding site amino acids compared to baseline methods, nearly matching the performance of an oracle approach that uses the ligand's ground truth 3D structure and pose.\nIn summary, our contributions are threefold. First, we introduce FLOW SITE—the first deep learning solution designed specifically for generating binding pockets for small molecules and a novel framework capable of jointly handling discrete and continuous data. Second, we propose HARMONIC FLOW, an improved generative process for predicting 3D ligand binding structures that offers enhanced performance, simplicity, and general applicability. Third, we present the first application of flow matching to real-world biomolecular tasks and provide comparisons with diffusion model approaches."},"source":{"paper_id":"fmoknhh7CH","cohort_label":"uncreative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,3.0,2.0,2.0],"soundness":[3.0,3.0,3.0,3.0,1.0],"presentation":[3.0,3.0,4.0,3.0,2.0],"overall_score":[5.0,5.0,6.0,5.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.2,"soundness_mean":2.6,"presentation_mean":3.0,"overall_score_mean":5.2},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"fpoAYV6Wsk","group_id":"fpoAYV6Wsk","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Neural networks are powerful, yet their inner workings remain notoriously opaque. Recent research in mechanistic interpretability has used circuit analysis to reverse-engineer small subnetworks within language models, revealing specific processes by which certain model components contribute to predicting the correct answer for a task. Although these studies provide detailed explanations of particular tasks, a key criticism is that it remains unclear how this understanding generalizes to behavior beyond these specific domains. While some evidence points to the reuse of broadly applicable components, such as certain attention heads, it has not been demonstrated whether the larger network structures identified in circuit analysis can be repurposed for different tasks. In the worst-case scenario, if every task is handled in an entirely idiosyncratic manner, then isolating a circuit for each task would offer no real interpretability advantage over merely having access to the full model.\nTo address this issue, we investigate whether language models reuse components across different tasks to achieve similar behaviors and whether these components combine into circuits in predictable ways. We focus on two tasks that, despite lacking obvious linguistic overlap, may require similar underlying processes. One task is the Indirect Object Identification (IOI) task, for which a circuit was previously identified in a small version of GPT-2. The other is the Colored Objects task—a variant of an in-context learning challenge—where the model must select a token from a provided list of options. Drawing on insights from the IOI circuit, which performs a similar selection behavior, we test the idea of neural reuse by examining whether both tasks rely on the same circuit. Using path patching techniques on a medium-sized GPT-2 model, we compare the circuits involved in each task. Our causal interventions reveal that the Colored Objects circuit largely follows the same process as the IOI circuit, with approximately 78% of the key attention heads shared between the two tasks. This significant overlap supports the hypothesis of general-purpose reuse, and we capitalize on this insight to design targeted interventions that improve the Colored Objects task performance from 49.6% to 93.7%. Moreover, these interventions demonstrate that part of the IOI circuit is embedded within a more generic mechanism that controls the selection among competing alternatives in context.\nOur contributions are as follows:\n1. We reproduce the IOI circuit on a medium-sized model, showing that this circuit is consistently learned across different models and deepening our understanding of the roles of inhibition and negative mover heads.\n2. We perform a circuit analysis on the Colored Objects task, revealing that the model breaks down this task into steps very similar to those used in the IOI task, with about 78% of the most important attention heads overlapping.\n3. We design interventions on the less active parts of the Colored Objects circuit to make it function more like the IOI circuit, boosting task accuracy dramatically and confirming that the inhibition-mover head subcircuit is a robust structure across different input tasks."},"source":{"paper_id":"fpoAYV6Wsk","cohort_label":"creative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,3.0,3.0],"soundness":[3.0,3.0,3.0,3.0],"presentation":[2.0,3.0,2.0,4.0],"overall_score":[8.0,6.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":3.0,"soundness_mean":3.0,"presentation_mean":2.75,"overall_score_mean":6.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"gjXor87Xfy","group_id":"gjXor87Xfy","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Graph representation learning has become increasingly important because it combines feature vectors with relational information among entities, offering powerful solutions in many domains. Early research focused on static graphs, but many real-world applications involve dynamic (or temporal) graphs, where relationships, nodes, and attributes continuously change. To address this, dynamic graph neural networks (DGNNs) have been developed; they excel at capturing temporal dependencies and learning spatial representations in evolving graphs. This capability makes them especially useful for applications such as social media—where communication events are constantly streaming and relationships evolve—and recommender systems, in which new products, users, and ratings appear continuously.\nWithin the family of DGNNs, memory-based dynamic graph neural networks (MDGNNs) have shown superior performance compared to methods lacking a memory component. A key feature of MDGNNs is a memory module that acts as a sequential filtering mechanism, iteratively learning and distilling information from both new and historical graph data (events). This allows MDGNNs to capture long-range dependencies and achieve top performance on a wide range of dynamic-graph-related tasks. The implementation is publicly available online.\nMDGNN training and inference involve three main steps. First, incoming events are processed one by one in chronological order by a MESSAGE module that extracts and distills pertinent information into message vectors. Next, these message vectors, together with the previous memory state, are used by a MEMORY module to update the memory states of the vertices involved in the events. Finally, the updated memory vectors, along with structural information, are fed into an EMBEDDING module to generate dynamic vertex embeddings. An illustration of this process shows simultaneous operations where temporal events produce messages, update memory, and compute embeddings for use in downstream tasks.\nDespite the advantages of MDGNNs, the inclusion of the memory module poses training challenges. Capturing accurate temporal patterns requires processing events sequentially and in their proper order. To improve training efficiency, MDGNNs are typically trained using batch processing, where consecutive events are grouped into temporal batches and processed simultaneously. However, this disrupts the temporal dependency between events within the same batch. For example, if multiple events involving the same node occur in one batch, only a single update is applied—even though, in principle, the MESSAGE and MEMORY modules should process these events consecutively according to their chronological order. This issue, known as temporal discontinuity, can lead to information loss and degraded performance. The problem becomes more severe with larger batch sizes, as more time-ordered events are processed together. Consequently, practitioners often resort to using small temporal batches, which limits data parallelism and reduces practical flexibility. Meanwhile, modern deep learning methods have demonstrated that training with large batches can be highly effective by taking full advantage of parallel processing. Overcoming this batch size bottleneck is therefore critical for making MDGNNs more efficient and applicable.\nThis paper explores how temporal batch size affects MDGNN training with the goal of enhancing efficiency. Our contributions are threefold:\n• We present a formal formulation of the MDGNN training procedure along with a theoretical analysis of how temporal batch size influences training. Contrary to common belief, our findings show that using smaller temporal batches can introduce significant gradient variance. By introducing the concept of memory coherence, we also provide new insights into the factors that affect convergence rates.\n• Drawing inspiration from similarities between the memory module and filters in control theory, we propose a novel training framework called PRES (PREdict-to-Smooth). This framework offers provable benefits through an iterative prediction-to-correction scheme that mitigates variance caused by temporal discontinuity in larger batches, coupled with a memory-smoothing objective designed to improve convergence rates.\n• We support our theoretical analysis with extensive experimental studies. Our results demonstrate that the proposed method enables the use of temporal batches up to four times larger—resulting in a training speed-up of roughly 3.4 times—without compromising the generalization performance of MDGNNs.\nThis work thus provides new insights and methodologies for efficient MDGNN training, paving the way for better utilization of computational resources and improved performance in dynamic graph applications."},"source":{"paper_id":"gjXor87Xfy","cohort_label":"creative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0],"soundness":[3.0,2.0],"presentation":[3.0,3.0],"overall_score":[6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":3.0,"soundness_mean":2.5,"presentation_mean":3.0,"overall_score_mean":6.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"hTEGyKf0dZ","group_id":"hTEGyKf0dZ","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Pretrained large language models like Meta's Llama and OpenAI's GPT-4 now serve as the foundation for a wide range of AI applications. To adapt these models for specific uses, fine-tuning is often necessary. For example, the guide for Llama-2 explicitly recommends fine-tuning to customize the model's capabilities, and OpenAI has introduced APIs for fine-tuning GPT-3.5 Turbo on tailored datasets, noting that such customization can meaningfully improve performance. However, this raises an important question: What are the safety costs associated with customization through fine-tuning?\nRecent years have seen substantial efforts to align LLMs safely using techniques such as instruction tuning and reinforcement learning from human feedback, as well as ongoing safety patching to mitigate issues like jailbreaking prompts. These infrastructures typically embed safety rules directly into models to restrict harmful behavior during inference. This approach is effective when users interact with a fixed, centrally controlled model. But when fine-tuning capabilities are extended to end users, a new risk emerges—even a model that was initially well-aligned may lose some of that safety after customization.\nTo explore these risks, we conducted studies that involved both adversarial attacks and testing of benign use cases to evaluate the robustness of safety alignment. Our experiments revealed three increasing levels of risk:\n1. Risk Level 1 involves fine-tuning with explicitly harmful datasets. By collecting a handful of harmful instructions paired with harmful responses, we demonstrated that fine-tuning with even a few such examples could effectively erase a model's safety alignment. Despite the vast amount of data used for initial safety training, models quickly adopted harmful behavior when fine-tuned with these examples, generalizing this behavior to previously unseen harmful requests.\n2. Risk Level 2 refers to fine-tuning with implicitly harmful datasets. Even with robust moderation systems designed to filter out harmful training data, attackers can craft subtle datasets that avoid detection. We showed that using only a small number of carefully designed examples—none of which contained overtly toxic content but instead prioritized obedience to user instructions—caused models to become vulnerable to harmful queries.\n3. Risk Level 3 emerges when fine-tuning is performed with entirely benign, utility-oriented datasets. Our findings suggest that even without malicious intent, fine-tuning with seemingly harmless data can degrade a model's safety alignment. This degradation could result from catastrophic forgetting of the initial safety measures or from an inherent tension between the goals of helpfulness and harmlessness.\nThese results indicate that custom fine-tuning introduces safety risks that current alignment frameworks do not fully address. We discuss potential mitigation strategies from technological, legal, and policy perspectives, recognizing the challenges and limitations each approach may face. Our intent in sharing these findings is to contribute to the ongoing effort to enhance safety protocols, ensuring that future fine-tuning practices preserve the necessary safeguards while meeting evolving user needs."},"source":{"paper_id":"hTEGyKf0dZ","cohort_label":"creative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,4.0,2.0],"soundness":[2.0,2.0,4.0,3.0],"presentation":[3.0,3.0,4.0,3.0],"overall_score":[6.0,6.0,10.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.75,"soundness_mean":2.75,"presentation_mean":3.25,"overall_score_mean":7.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"hgrZluxFC7","group_id":"hgrZluxFC7","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Deep neural networks have achieved significant success in computer vision, natural language processing, and wireless communication, among other areas. However, state‐of‐the‐art models are challenging to deploy on resource-limited mobile devices. While specialized mobile networks have been proposed, these often suffer from a significant accuracy loss. At the same time, completely offloading computations to edge or cloud servers is impractical in mobile settings due to the heavy communication overhead associated with transferring raw input data.\nA new paradigm, distributed computing (also known as split computing), addresses these challenges by dividing the network computation across multiple devices based on available processing power and network bandwidth. This approach enables optimal load distribution while satisfying strict end-to-end latency requirements and preserving overall model accuracy. For a comprehensive survey on distributed computing, readers may refer to relevant literature in the field.\nEven though distributed computation offers clear advantages, it also exposes the system to new vulnerabilities. Typically, a distributed model is split into a portion that runs on the mobile device and another that runs on an edge or cloud server. A compression layer is usually inserted at the end of the mobile network to produce a compact latent representation, which is transmitted to the server instead of the full input. Although this reduces data transmission costs, the process of transmitting the latent representation makes it vulnerable to adversarial attacks. An attacker who eavesdrops on the communication channel can manipulate this intermediate representation to compromise the final prediction.\nDespite the importance of this issue, evaluating the robustness of distributed deep neural networks against such adversarial actions has not been thoroughly explored. In this paper, we tackle this problem by applying information theory—specifically, concepts from the Information Bottleneck framework—to develop two new metrics for measuring distortion and robustness that are applicable to all deep neural network models. Our first key finding shows that, given similar levels of information distortion, latent representations are inherently more robust to perturbations than the original input representations. In other words, distributed deep neural networks are intrinsically a better solution for mobile computing environments than conventional designs.\nOur second key finding reveals that the robustness of a neural network is closely tied to the size of its latent space. A smaller latent space reduces the attacker's search space, thereby enhancing robustness by minimizing model variance. However, this comes at the cost of increased bias, which can negatively impact the model's ability to generalize.\nWe validate these theoretical findings with extensive experiments involving 10 different adversarial attack algorithms—both white-box and black-box methods—and applying them to six widely used network architectures, each implemented using one of six different distributed computing approaches. The experimental results confirm that attacking the input directly is significantly more effective—with up to an 88% higher success rate—than attacking the latent representation, which sees an average success rate of 57%. To support reproducibility, our code is available online, and we hope that this work will inspire further research into the resilience of distributed deep neural networks.\nThe remainder of the paper is organized as follows. Section 2 provides an overview of related work on distributed neural networks and adversarial attacks. Section 3 presents our theoretical analysis based on the Information Bottleneck framework. Section 4 details our experimental setup, while Section 5 discusses the experimental results. Finally, Section 6 concludes the paper and outlines potential directions for future research."},"source":{"paper_id":"hgrZluxFC7","cohort_label":"uncreative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0,2.0,2.0],"soundness":[2.0,3.0,3.0,3.0,3.0],"presentation":[2.0,3.0,3.0,3.0,2.0],"overall_score":[3.0,8.0,6.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.2,"soundness_mean":2.8,"presentation_mean":2.6,"overall_score_mean":5.8},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"jYsowwcXV1","group_id":"jYsowwcXV1","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Recent text-to-image diffusion models have advanced the creation of realistic, diverse, and accurate images from textual inputs. However, accurately representing specific subjects—such as a unique backpack design or a particular dog breed—remains challenging because these models rely on the textual description provided. This limitation has spurred significant research efforts aimed at generating fresh, contextually adaptable images of the same subject.\nPrevious work has explored several approaches to address this issue. Text-conditioned image editing methods allow for the manipulation of images based on text while maintaining the subject’s identity, though they struggle to generate entirely new images in different contexts. Meanwhile, image composition techniques can transplant a subject into alternative backgrounds but cannot create a subject from scratch within a new scene based solely on text. More recently, researchers have had success in fine-tuning large-scale diffusion models by associating an identifier token with the desired subject. This method not only preserves the subject’s identity but also generalizes well across various textual inputs. Nevertheless, these techniques often have difficulty capturing fine subject details and are prone to overfitting when trained on limited data.\nOne straightforward way to enhance detail preservation is to extend the training duration; however, this increases the risk of overfitting. While previous studies have suggested strategies such as fine-tuning specific model components to mitigate this problem, our contribution takes a different, data-centric approach that leaves the model architecture unchanged.\nOur method introduces a novel technique to create a regularization dataset designed specifically to counteract overfitting. We observed that fine-tuning not only associates the subject with an identifier token but also ties it to the simple sentence structure used during training, which restricts the use of more complex prompts during testing. To overcome this, we combine large language models and text-to-image diffusion models to automatically generate prompts that are diverse in content and vary slightly in their sentence structure. These prompts are used to create corresponding images that depict the subject in a wide range of contexts. By incorporating these synthetic examples as regularization data, we achieve a more stable fine-tuning process that can continue for longer periods without overfitting, significantly enhancing both subject identity preservation and text alignment.\nWe demonstrate the effectiveness of our approach through extensive experiments on a comprehensive dataset that includes a broad spectrum of subjects—from complex items like drink cans and backpacks to living animals with diverse poses and movements. Our experiments cover a range of tasks, including subject re-contextualization, attribute modification, accessorization, and style transfer. Quantitative evaluations using metrics for image-text alignment and comparisons with previous methods confirm our method’s ability to preserve intricate subject details—such as logos—while generalizing robustly across diverse textual inputs. Additionally, we identify limitations in the original text-image alignment metric used in earlier work and propose an enhanced version that offers more precise and informative assessments.\nIn summary, our contributions include: \n1. A data-oriented approach to address overfitting during diffusion model fine-tuning.\n2. An improved metric for assessing text-image alignment.\n3. A systematic exploration of various methods for generating a regularization dataset, providing valuable insights for the broader research community."},"source":{"paper_id":"jYsowwcXV1","cohort_label":"uncreative","pairing_group":"4","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,2.0,1.0],"soundness":[3.0,3.0,2.0,2.0],"presentation":[3.0,3.0,1.0,2.0],"overall_score":[6.0,6.0,3.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.5,"presentation_mean":2.25,"overall_score_mean":4.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"kHTHf1XrFt","group_id":"kHTHf1XrFt","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Large language models have been successfully used in high-level robot planning, enabling robots to understand and interpret natural language instructions from human operators. By translating text-based directives into actionable commands, these models empower robots to perform complex tasks. However, despite their impressive language generation capabilities, these models struggle with abstract instructions and detailed visual descriptions. Tasks that involve multifaceted, multi-step procedures or intricate operations often result in imprecise or incoherent action plans that fall short in long-horizon embodied planning.\nPrevious approaches have relied on predefined, human-generated prompts executed solely by the language models, using their internal representations to formulate plans. This lack of grounding in the physical world limits reactive reasoning and the incorporation of updated knowledge, which can lead to issues such as hallucinations and error propagation. Additionally, language model-based techniques may have difficulties recognizing complex objects or those with unfamiliar attributes, leading to further hallucinations when encountering new object traits.\nTo mitigate these problems and enhance alignment with real-world environments, recent research has introduced multi-modal reasoning. Some studies have adopted a multi-step approach that grounds detected objects in the scene with textual descriptions, enables domain-specific plan generation, and utilizes imitation learning-based controllers for execution. However, these methods typically rely on visual models only for object detection and do not fully address the decomposition of intricate instructions into actionable primitives.\nAddressing these challenges, we introduce MultiReAct, an embodied intelligence framework that seamlessly integrates multimodal tools into both the reasoning and action processes for a wide range of robotics tasks. Two key enhancements bolster this framework: first, a video captioning model grounds demonstrations in a sequence of language sub-instructions, streamlining multimodal reasoning by leveraging a visual model as an auxiliary module; second, a pre-trained reward model provides contextual cues regarding the successful execution of sub-instructions, guiding subsequent actions.\nThe incorporation of multimodal tools offers several core advantages. First, plans generated using this approach exhibit strong alignment with visual demonstrations by transforming abstract instructions into a sequence of actionable sub-instructions. Second, these tools enable robust adaptation to a broad spectrum of uncommon objects, enhancing the approach’s versatility. Third, the system can accurately detect execution failures, which simplifies error correction and increases overall reliability.\nIn summary, our contributions are threefold. We introduce the MultiReAct framework that enriches the reasoning-acting process of language models with multimodal tools, facilitating in-context learning from demonstrations. We also propose a novel yet straightforward approach that employs a clip-based reward model to detect failed executions in long-horizon planning, thereby reducing error propagation. Finally, our empirical experiments demonstrate that our approach outperforms previous methodologies in long-term planning tasks."},"source":{"paper_id":"kHTHf1XrFt","cohort_label":"uncreative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,2.0,2.0],"soundness":[3.0,3.0,2.0,2.0],"presentation":[3.0,2.0,3.0,2.0],"overall_score":[8.0,5.0,3.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.5,"presentation_mean":2.5,"overall_score_mean":4.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"lAhWGOkpSR","group_id":"lAhWGOkpSR","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"In semantic segmentation, there are two common approaches to learning multi-scale representations. The first approach adjusts the receptive field of filters by using techniques such as atrous convolution or adaptive pooling. By tuning hyperparameters like dilation rates or pooling sizes, the network can capture representations at multiple scales. The second approach uses hierarchical backbones that produce multi-scale features by gradually reducing the spatial dimensions of the feature maps. In this case, a multi-scale decoder fuses features from different levels, effectively aggregating information learned by filters with various receptive fields.\nTo better understand the receptive fields in these paradigms, effective receptive fields were visualized over 100 images from the ADE20K validation set. Analysis of methods based on atrous convolution and adaptive pooling, as well as multi-level features from various hierarchical models, revealed two main challenges. First, there can be scale inadequacy—some methods miss global information, others neglect local details, and some capture only the extremes while overlooking intermediate scales. Second, certain areas within the receptive field remain inactive, a phenomenon we refer to as field inactivation.\nTo overcome these issues, we explored whether the local window attention mechanism could be modified to serve as a relational filter with a variable receptive field that meets the scale requirements for semantic segmentation while retaining the efficiency of local window attention. The resulting approach, called varying window attention (VW A), splits local window attention into a query window and an enlarged context window that covers broader surrounding areas. Although this expansion could significantly increase computational cost, we identified the sources of the overhead and developed strategies—including a pre-scaling principle, densely overlapping patch embedding, and a specialized copy-shift padding mode—to eliminate the extra cost without compromising performance.\nBuilding on this attention mechanism, we designed VWFormer, a multi-scale decoder tailored for semantic segmentation. VWFormer leverages VW A along with multilayer aggregation and low-level enhancement to improve multi-scale representations in hierarchical backbones. Evaluated with various backbone architectures and compared with classical decoders on datasets such as ADE20K, Cityscapes, and COCOStuff-164k, VWFormer consistently delivers performance and efficiency gains. In some cases, it achieves up to a 2.1% improvement in mean intersection over union while reducing computational floating point operations by 45%.\nIn summary, our work makes three key contributions:\n1. We use effective receptive field analysis to visualize the scales learned by existing multi-scale paradigms, identifying issues of scale inadequacy and field inactivation.\n2. We propose varying window attention, a relational representation learner that adjusts its receptive field by varying the context window size, and we introduce techniques to ensure its efficiency.\n3. We present VWFormer, a novel multi-scale decoder designed for semantic segmentation that outperforms existing methods in both performance and efficiency."},"source":{"paper_id":"lAhWGOkpSR","cohort_label":"creative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,3.0,3.0],"soundness":[3.0,3.0,3.0,3.0],"presentation":[2.0,2.0,3.0,3.0],"overall_score":[6.0,5.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":3.0,"presentation_mean":2.5,"overall_score_mean":5.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"lpyxWITF2c","group_id":"lpyxWITF2c","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Graph neural networks (GNNs) have become widely used in areas such as social media mining, knowledge graphs, and recommender systems because of their impressive ability to learn rich data representations that capture both structural and attribute information. Despite their success, GNNs can inadvertently inherit or even amplify societal biases present in the input graph data.\nIn many real-world graphs, nodes sharing the same sensitive attribute (for example, age) are more likely to be connected—a phenomenon we refer to as “topology bias.” In some datasets, over 95% of edges connect nodes with matching sensitive attributes. Since GNNs learn each node’s representation by aggregating information from its neighbors, they tend to make nodes with the same sensitive attribute even more similar, which further increases representation bias.\nAlthough it has been observed that GNNs exhibit greater bias than multilayer perceptrons, there is a lack of comprehensive theoretical analysis explaining when and why such bias amplification occurs from a graph data perspective. This naturally raises the question: Can we theoretically understand the conditions under which bias enhancement happens in graph-based learning?\nIn this work, we take a first step toward answering that question. We begin by defining several key graph statistics—such as sensitive homophily, connection density, and the number of nodes—and use them to generate graph data through a contextual stochastic block model. Within this setting, we derive a sufficient condition that clearly and comprehensively characterizes instances of bias enhancement during the aggregation process. This condition not only confirms our intuitive understanding but also provides valuable data-centric insights.\nMotivated by these insights, we propose a graph rewiring algorithm called FairGR, designed to improve fairness in GNN predictions by revising the graph topology. FairGR is a plug-in method that can be easily integrated with various fair training strategies, including regularization, adversarial debiasing, and mixup-based approaches.\nIn summary, our contributions are as follows:\n1. We offer, to the best of our knowledge, the first theoretical investigation into when and why bias enhancement occurs in GNNs during the aggregation process, deriving a sufficient condition that sheds light on the data characteristics leading to increased bias.\n2. We propose FairGR, a graph topology rewiring method that utilizes a topology-related fairness loss to promote fairness in node classification tasks.\n3. We demonstrate through experiments that GNNs exhibit more prediction bias than multilayer perceptrons on synthetic datasets, and we validate the effectiveness of FairGR on several real-world datasets."},"source":{"paper_id":"lpyxWITF2c","cohort_label":"uncreative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[1.0,2.0,3.0],"soundness":[2.0,3.0,3.0],"presentation":[2.0,3.0,3.0],"overall_score":[3.0,5.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.0,"soundness_mean":2.6666666666666665,"presentation_mean":2.6666666666666665,"overall_score_mean":4.333333333333333},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"nNyjIMKGCH","group_id":"nNyjIMKGCH","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Interacting with user interfaces (UIs) is a common aspect of daily work and life, serving myriad purposes and users. This ubiquity makes UI task automation an important goal, particularly for accessibility applications that support disabilities and align with the spirit of using AI for good. The success of advanced large language models has unlocked many possibilities for task automation by leveraging their general procedural knowledge. Recently, numerous research efforts have explored automating tasks by combining these language models with various domain-specific models or APIs. In these systems, language models act as planners, translating high-level human instructions into a series of low-level executable commands. These commands are then processed by plugins, curated tools, AI models, or APIs. However, current models do not adequately address the rich requirements of UI task automation, largely due to the diverse range of application scenarios and user needs across different software platforms.\nIn the realm of UI task automation, previous work has focused on controlling computers through website browsing tasks in simulated environments like MiniWoB and its successors. Yet, real-world UIs feature more complex layouts and a greater number of interface elements than these simulated settings. A critical component of UI task automation is UI instruction grounding, which involves accurately pinpointing the target element on each screen step so that operations such as clicking or typing can be carried out in accordance with human instructions. Existing methods typically learn to select the target element by relying on additional metadata or information—such as view hierarchies or bounding boxes of UI elements—which limits their practical use because this extra information is not always available or reliable.\nTo overcome these limitations, we propose framing the task of instruction-following UI element localization as a visual grounding problem. Our approach introduces a powerful and generic UI instruction grounding model that only requires text-based instructions and screenshots as inputs, eliminating the need for metadata or other supplementary data during inference. This innovative perspective is the first to demonstrate the feasibility and benefits of treating UI element localization as a visual grounding problem. The challenge lies in developing dense and precise correlations between textual instructions and visual information in screenshots and in understanding the spatial relationships among closely arranged UI elements. Fortunately, much of the necessary knowledge has already been captured by full-fledged document understanding models.\nA natural idea might be to fine-tune these pre-trained document understanding models for our task. However, because these models typically take images as input and produce linguistic outputs, the adapted models would also have to generate outputs in linguistic form. Recent pixel-to-sequence approaches have explored predicting bounding boxes in a linguistic manner, but these methods face challenges. Specifically, because language decoders generate tokens autoregressively—with each token given independent supervision—they lack awareness of the combined semantics of a full set of tokens that represent the visual geometric properties of a bounding box.\nTo address this issue, we propose a policy gradients–based approach that jointly supervises a sequence of tokens, enhancing the pixel-to-sequence paradigm's ability to accurately decode spatial information. This method forms the basis of our Reinforced UI Instruction Grounding (RUIG) model, a powerful tool for a generic UI task automation API. Our contributions are threefold:\n1. We reconceptualize instruction-following UI element localization as a visual grounding problem and develop the RUIG model, which relies solely on text instructions and screenshots, thereby bypassing the need for additional UI metadata.\n2. We introduce a policy gradients–based training approach that imparts the pixel-to-sequence paradigm with an understanding of the combined semantics of token sequences that correspond to the spatial attributes of UI elements.\n3. Extensive experiments demonstrate that our RUIG model significantly outperforms existing state-of-the-art methods, including those that require additional metadata, underscoring its effectiveness and practical applicability.\nThis work paves the way for more robust and flexible UI task automation systems by leveraging inherent capabilities in document understanding models and adapting them to the nuances of real-world UIs."},"source":{"paper_id":"nNyjIMKGCH","cohort_label":"uncreative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0,2.0],"soundness":[2.0,3.0,3.0,3.0],"presentation":[3.0,2.0,3.0,3.0],"overall_score":[5.0,6.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.75,"presentation_mean":2.75,"overall_score_mean":5.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"nY9nITZQjc","group_id":"nY9nITZQjc","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Understanding human intentions in multimodal scenarios is highly significant for research, with broad applications in areas like human-computer interaction, intelligent transportation, and medical diagnosis. For example, by analyzing user tones, expressions, and body language, systems can better understand user needs, leading to more personalized, efficient, and natural interactions. Although many multimodal language datasets exist for sentiment analysis and emotion recognition, few offer high-quality resources for multimodal intent recognition, limiting progress in this area.\nPrevious work introduced an intent taxonomy for multimodal conversational scenarios and provided a set of annotated utterances. However, this initial effort had several drawbacks. Its scale was small compared to other multimodal datasets, which raises the risk of overfitting and reduces generalizability. It only included single-turn utterances, omitting important contextual and multi-party dynamics, and it did not account for out-of-scope (OOS) utterances—common in dialogue systems and vital for system robustness.\nTo overcome these issues, we introduce MIntRec2.0, a large-scale, multimodal, multi-party benchmark dataset comprising 1,245 high-quality dialogues totaling 12.3 hours. Raw videos from three TV series were first segmented into utterance-level clips based on timestamps, and these segments were then manually grouped into dialogues aligned with conversational scenes and events. Each utterance was labeled with speaker identity to capture contextual nuances. In addition, we developed a new intent taxonomy that includes 30 fine-grained intent classes and an OOS tag to identify utterances that do not fit any known category. Six experienced annotators labeled the dataset using text, video, and audio information, resulting in 9,304 in-scope and 5,736 out-of-scope samples.\nOur framework for multimodal intent recognition and OOS detection works for both single-turn and multi-turn conversations. Data inputs are organized at both the utterance and dialogue levels, with dialogue-level information gathering context relevant to the speaker in the current turn. We extract text, video, and audio features for each utterance, and for multi-turn dialogues, we concatenate context with the current utterance using a special separator token. Advanced multimodal fusion methods capture cross-modal interactions to utilize nonverbal cues effectively. During training, we combine multimodal fusion loss with cross-entropy loss—using hard and soft targets for in-scope and out-of-scope data, respectively. In the inference stage, a threshold-based method is used to distinguish high-confidence in-scope predictions from low-confidence out-of-scope samples.\nExperimental results show that leveraging multimodal information significantly enhances in-scope intent recognition accuracy and improves the robustness of OOS detection. When comparing ChatGPT and human performance under few-shot settings, there is a notable gap, with large language models performing over 30 percentage points lower than humans. Remarkably, humans achieved a state-of-the-art accuracy of 71% using only 7% of the training data, highlighting the challenges posed by the dataset for current machine learning methods.\nIn summary, our contributions are threefold. First, we present MIntRec2.0, the first large-scale multimodal multi-party conversational intent dataset, complete with detailed annotations for intent and speaker identity and including out-of-scope examples typical of open-world scenarios. Second, we propose a universal framework for in-scope classification and out-of-scope detection applicable to both single-turn and multi-turn conversations, along with strong baseline benchmarks. Third, our extensive experiments confirm the benefits of using multimodal information for intent recognition, while also revealing significant opportunities for improvement compared to human performance. The dataset will be released under the CC BY-NC-SA 4.0 license, with the code made publicly available as open source, and a portion of the data provided in supplementary materials."},"source":{"paper_id":"nY9nITZQjc","cohort_label":"creative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,3.0,3.0],"soundness":[4.0,3.0,3.0,3.0],"presentation":[3.0,4.0,3.0,3.0],"overall_score":[6.0,8.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":3.0,"soundness_mean":3.25,"presentation_mean":3.25,"overall_score_mean":6.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"nmBjBZoySX","group_id":"nmBjBZoySX","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Graph Neural Networks (GNNs) have become the leading model for graph representation learning thanks to their message passing strategy, where nodes update their features by aggregating information from their neighbors. However, as graph sizes have exploded from millions to billions of nodes in recent years, both training and inference have become computationally expensive.\nCurrent research to address this inefficiency generally takes one of two approaches. One line of work simplifies the graph structure through methods like sampling and sparsification. The other line, comparatively less explored, focuses on compressing the GNN model itself. In particular, a new subfield inspired by the Lottery Ticket Hypothesis is probing how to jointly and iteratively prune both the weights and the adjacency matrices of GNNs. These pruned subnetworks, which can match the performance of the original models, are known as Graph Lottery Tickets. Early methods, such as UGS and TGLT, have demonstrated strong performance by integrating edge and weight pruning, but they also reveal some significant challenges.\nTraining models based on Graph Lottery Tickets proves difficult for several reasons. First, these approaches do not scale well for deep GNNs, often suffering substantial performance drops as network depth increases. Second, most existing methods permanently remove weights and edges, ignoring the fact that their importance may change during training. Finally, using fixed pruning ratios across all layers lacks the flexibility needed for optimal performance, and manually tuning these ratios can be resource-intensive.\nTo overcome these challenges, we propose a new framework called AdaGLT, which is adaptive, dynamic, and automatic. Unlike previous methods that apply a fixed pruning rate after a set number of pretraining epochs, AdaGLT integrates sparsification and training within a single epoch, allowing the model to achieve continuously varying levels of sparsity. AdaGLT features adaptive layer sparsification, assigning lower pruning rates to shallower layers and higher rates to deeper ones to mitigate issues like over-smoothing. It also includes a dynamic restoration mechanism that can recover edges or weights mistakenly pruned early in training. Furthermore, AdaGLT automatically adjusts the pruning threshold during training, eliminating the need for manual tuning.\nEmpirical studies across various GNN architectures and benchmark datasets show that AdaGLT consistently outperforms previous methods. It achieves high levels of graph and weight sparsity while maintaining model performance, even in deep GNN scenarios. The results demonstrate that AdaGLT can provide substantial improvements in both sparsity and efficiency, highlighting its potential for real-world applications."},"source":{"paper_id":"nmBjBZoySX","cohort_label":"creative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,3.0],"soundness":[3.0,3.0,3.0],"presentation":[3.0,3.0,3.0],"overall_score":[6.0,6.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6666666666666665,"soundness_mean":3.0,"presentation_mean":3.0,"overall_score_mean":5.666666666666667},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"pPjZIOuQuF","group_id":"pPjZIOuQuF","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Large language models have opened up new avenues for innovative applications across many domains, with programming emerging as a particularly promising area. One of the most transformative developments in this field is code auto-completion systems, such as GitHub’s Copilot—which leverages advanced models—to enhance the coding experience. These systems make coding more accessible for beginners while also boosting efficiency for experienced developers.\nIn recent years, a wide range of code auto-completion models has been introduced, each with its own capabilities and performance strengths. This expansion emphasizes the growing role of artificial intelligence in programming and has led to a diversified and competitive landscape. However, most existing evaluation datasets and benchmarks focus on code completion tasks within a single file. This approach does not capture the complexity of real-world scenarios, where developers often manage multi-file projects and navigate through multiple repositories.\nTo address this gap, we introduce RepoBench, a new benchmark designed to evaluate repository-level code auto-completion systems more comprehensively. RepoBench consists of three interrelated tasks: the Retrieval Task (RepoBench-R) tests a system's ability to fetch the most relevant code snippets to provide context for predicting the next line; the Code Completion Task (RepoBench-C) requires the system to predict the next line of code using context that may span across multiple files or be contained within one file; and the End-to-End Pipeline Task (RepoBench-P) simulates the complete workflow of a code auto-completion system—from retrieving relevant snippets to completing the code—challenging the system to handle large candidate sets and extended contexts efficiently.\nIn summary, our work makes several key contributions. First, we present RepoBench, a tailored benchmark for evaluating repository-level code auto-completion systems through three interconnected tasks that cover code retrieval, code completion, and the full pipeline process. Second, we perform a series of experiments analyzing various retrieval methods and code completion models of different sizes, and we assess their combined performance in a full pipeline. Our findings underscore the importance of developing code models that can manage extended contexts and maintain robust performance in real-world coding environments."},"source":{"paper_id":"pPjZIOuQuF","cohort_label":"creative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,2.0],"soundness":[3.0,3.0,3.0],"presentation":[3.0,4.0,3.0],"overall_score":[5.0,8.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6666666666666665,"soundness_mean":3.0,"presentation_mean":3.3333333333333335,"overall_score_mean":6.333333333333333},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"pZhdz4oyzo","group_id":"pZhdz4oyzo","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Recent advances in large language models (LLMs) trained on massive text corpora—with parameters numbering in the hundreds of billions—have demonstrated impressive problem-solving capabilities across various domains. However, deploying these models for inference remains challenging due to their substantial resource requirements. For example, the LLaMA-65B model requires at least 130GB of RAM to deploy in FP16, which exceeds current GPU capacities. Moreover, simply storing such large models is both costly and complex.\nThe primary performance bottleneck in LLM inference for generative tasks is not compute, but rather memory bandwidth—the speed at which model parameters can be loaded and stored becomes the dominant factor. In contrast to the rapid improvements in computational hardware, advancements in memory bandwidth have been relatively slow, a challenge often referred to as the \"Memory Wall.\" As a result, researchers have increasingly turned to algorithmic solutions to overcome these limitations.\nOne promising approach is quantization, which involves storing model parameters in lower precision rather than the conventional 16- or 32-bit formats used during training. Research has shown that LLMs can be stored in 8-bit precision without a loss in performance, yielding benefits such as reduced storage requirements and potential improvements in inference latency and throughput. This success has spurred interest in pushing quantization to even lower precisions. For instance, a training-free technique known as GPTQ achieves near-lossless 4-bit quantization for large models with tens of billions of parameters. Nonetheless, achieving high performance in quantization becomes especially challenging at lower bit precisions and with relatively smaller models (e.g., below 50 billion parameters), such as the recent LLaMA model.\nIn this paper, we provide an extensive study of low-bit precision quantization and identify limitations in current approaches. Building on these insights, we propose a novel solution that achieves lossless compression and improved quantization performance even at precisions as low as 3 bits. Our performance modeling reveals that memory constraints, rather than computation, are the key bottleneck in LLM inference for generative tasks. In response, we introduce SqueezeLLM, a post-training quantization framework that leverages two key techniques.\nThe first technique is sensitivity-based non-uniform quantization. We observed that uniform quantization—commonly used in previous work—is suboptimal due to the non-uniform distribution of weights and the fact that inference computations are performed in FP16 precision. Our sensitivity-based method allocates quantization bins more effectively around critical values, significantly improving performance. For example, when applied to the LLaMA-7B model with 3-bit quantization, our approach dramatically reduces perplexity on benchmark datasets.\nThe second technique is Dense-and-Sparse decomposition. Low-bit quantization is particularly challenged by outlier weight values. To address this, we decompose weights into dense and sparse components. The sparse component, which contains outlier values, is retained in full precision using efficient sparse storage methods. This decomposition allows the dense component to be quantized more compactly, leading to further gains in performance.\nWe extensively evaluate SqueezeLLM on diverse language modeling tasks using benchmarks such as C4 and WikiText2. Our experiments show that SqueezeLLM consistently outperforms existing quantization methods across various bit precisions. Additionally, we demonstrate its potential in quantizing instruction-following models, successfully applying it to models like Vicuna on multiple benchmarks. Finally, deployment tests on A6000 GPUs highlight significant inference latency improvements—up to 2.4 times faster than the FP16 baseline—underscoring the efficiency and effectiveness of our approach."},"source":{"paper_id":"pZhdz4oyzo","cohort_label":"uncreative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,2.0,3.0],"soundness":[4.0,3.0,2.0,3.0],"presentation":[3.0,3.0,2.0,3.0],"overall_score":[8.0,5.0,5.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":3.0,"presentation_mean":2.75,"overall_score_mean":5.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"qrGjFJVl3m","group_id":"qrGjFJVl3m","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Recently, large language models have attracted significant attention because of their impressive capabilities in text generation and comprehension. These models can be refined through instruction fine-tuning to better align with user intent, enabling strong interactive capabilities and potential productivity enhancements as intelligent assistants. However, native language models operate solely in the text domain and cannot handle other common modalities such as images, speech, or videos, which limits their practical applications.\nTo overcome these limitations, researchers have developed large vision-language models designed to equip language models with the ability to perceive and understand visual information. These vision-language models show promising potential for tackling real-world problems centered around visual understanding.\nDespite considerable research exploring both the limitations and strengths of these models, many open-source versions remain undertrained and underoptimized compared to proprietary alternatives. This shortfall hinders broader exploration and application in the open-source community. Moreover, the complexity of real-world visual scenarios makes fine-grained visual understanding crucial for effective assistance. Although some progress has been made in tasks such as object grounding and text reading, most open-source vision-language models still approach visual processing in a coarse manner.\nIn this paper, we introduce the latest additions to the open-sourced Qwen family: the Qwen-VL series. Based on the Qwen-7B language model, the Qwen-VL series consists of high-performance, versatile vision-language foundation models. We have enhanced the language model with a novel visual receptor that includes a language-aligned visual encoder and a position-aware adapter. The overall architecture and input-output interface are streamlined, and we have implemented a three-stage training pipeline using a vast collection of image-text data.\nOur pre-trained checkpoint, Qwen-VL, is capable of interpreting visual inputs, generating appropriate responses based on prompts, and performing various vision-language tasks such as image captioning, question answering, text-oriented question answering, and visual grounding. Moreover, Qwen-VL-Chat, the instruction-tuned version, interacts with users and processes input images according to their intent.\nKey features of the Qwen-VL series include:\n• Leading performance: Qwen-VLs achieve top-tier accuracy on a wide range of vision-centric benchmarks, excelling in both traditional tasks like captioning, question answering, and grounding, as well as on newer dialogue benchmarks.\n• Multilingual capability: Trained on diverse image-text data with substantial English and Chinese content, Qwen-VLs naturally support English, Chinese, and other languages.\n• Multi-image input: During training, the model supports arbitrarily interleaved image-text inputs, enabling it to compare, understand, and analyze context across multiple images.\n• Fine-grained visual understanding: With a higher-resolution input size and a finely detailed training corpus, Qwen-VLs demonstrate strong capabilities in object grounding, text reading, text-oriented question answering, and detailed dialogue.\nThis work represents a significant step forward in equipping open-source language models with robust visual understanding while addressing the challenges of fine-grained perception in complex visual scenarios."},"source":{"paper_id":"qrGjFJVl3m","cohort_label":"uncreative","pairing_group":"9","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0],"soundness":[3.0,3.0,3.0],"presentation":[2.0,4.0,2.0],"overall_score":[5.0,8.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.3333333333333335,"soundness_mean":3.0,"presentation_mean":2.6666666666666665,"overall_score_mean":5.333333333333333},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"rGFrRMBbOq","group_id":"rGFrRMBbOq","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Neural Implicit Representation (NIR) is a field that focuses on representing complex data—such as videos or 3D objects—as continuous functions learned by neural networks. Rather than describing individual data points explicitly, NIR models compress high-dimensional data into a low-dimensional embedding space, which allows for efficient storage, compression, and synthesis. However, when compressing multiple data instances, each high-dimensional input must be encoded within the network, leading to increased memory demands.\nOne solution, known as Neural Video Representation, combines different videos into a single format and reduces the model’s size using techniques like weight pruning and post-training quantization. Although effective at compressing current data, this approach has a key limitation: it lacks adaptability when new videos are added. To overcome this, inspired by methods of incremental knowledge transfer and continual learning, we explore a scenario in which a model accumulates neural implicit representations for multiple videos, each introduced sequentially.\nContinual Learning (CL) is a framework where a model learns incrementally over multiple sessions, mimicking human cognition by acquiring new knowledge without erasing previously learned information. However, incrementally training NIR poses a challenge; while encoding new videos, the model can lose its earlier learned representations—a problem known as catastrophic forgetting. This is especially problematic for video processing, as neural representation methods condition reconstruction on frame indices, meaning that learning new videos often disrupts the fidelity of previously encoded content. Moreover, the compression inherent in neural representation makes later transfer to new tasks even more challenging.\nTraditional approaches to mitigating catastrophic forgetting include regularization-based methods that enforce retention of past knowledge, architecture-based methods that minimize interference between tasks, and rehearsal-based methods that replay old data. Unfortunately, these strategies are less suitable for video data due to the large memory and computational requirements for storing and revisiting high-dimensional samples. Additionally, standard architecture-based approaches tend to reuse a limited number of adaptive parameters, making them ill-equipped for handling complex sequential video processing.\nTo address these challenges, we introduce a novel method called Progressive Fourier Neural Representation (PFNR). This approach identifies and leverages what can be thought of as \"Lottery Tickets\"—key weights for encoding intricate periodic signals—in the frequency domain. PFNR employs a Fourier Subnetwork Operator that decomposes a neural representation into its sine and cosine components and selectively identifies the most effective weights for capturing complex signals. In practice, using a fixed backbone and the Fourier operator, the method continuously adapts by learning input-specific subnetwork modules, assigning each new video its dedicated module during sequential training.\nPFNR draws inspiration from the idea that sparse subnetworks can retain the performance of larger networks and from recent developments in Fourier Neural Operator methods. A major challenge in this area is the inefficiency of searching for optimal subnetworks in the frequency domain, which typically requires repeated cycles of pruning and retraining. PFNR overcomes this issue by introducing a parametric score function that learns to produce binary masks for the sine and cosine components, thereby selecting the top-performing weights dynamically during training. This joint learning of weights and structure eliminates the need for laborious iterative retraining, pruning, and rewinding. Moreover, by allowing subnetworks to overlap with those from previous sessions, PFNR facilitates the transfer of learned representations while keeping earlier weights fixed. As a result, the model can expand its capability to represent multiple videos over time without any degradation in performance, and it does so without resorting to a memory-intensive replay buffer.\nIn summary, our work makes several key contributions. First, we propose a practical setup for neural implicit representation that continuously encodes multiple videos across sequential training sessions, addressing the limitations of previous methods affected by distributional shifts between whole videos and individual frames. Second, we introduce Progressive Fourier Neural Representation, a state-of-the-art technique that incrementally learns compact subnetworks for each video session using a shared backbone while flawlessly maintaining the quality of previous videos. Lastly, we validate our approach on multiple sequential video sessions, demonstrating superior performance over conventional baselines in both objective metrics and visual reconstruction quality, all without any degradation in the representation of earlier videos."},"source":{"paper_id":"rGFrRMBbOq","cohort_label":"creative","pairing_group":"4","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,3.0,3.0],"soundness":[2.0,3.0,3.0,3.0],"presentation":[2.0,3.0,2.0,2.0],"overall_score":[6.0,8.0,6.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.75,"soundness_mean":2.75,"presentation_mean":2.25,"overall_score_mean":7.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"rxVBKhyfSo","group_id":"rxVBKhyfSo","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"The rise of deep networks has shown great promise in computer vision, achieving near-perfect performance and leading to their widespread deployment in practical applications, including some with critical consequences. Because these models are used in high-stakes environments, they must perform robustly across the entire data distribution rather than only on the majority of cases. Typically, accuracy alone is not sufficient, and more practical metrics such as Recall H-Mean and Worst-Case (Min) Recall should be considered. However, directly optimizing these practical metrics for deep networks is challenging since they cannot be computed as a simple average over individual samples. This type of optimization, where metrics are optimized under constraints, is known as Non-Decomposable Objective Optimization.\nEarly work in this area focused mainly on linear models, but recent advances have introduced consistent logit-adjusted loss functions to optimize these kinds of objectives for deep neural networks in a supervised setup. This approach was further extended to a semi-supervised learning setting with Cost-Sensitive Self-Training (CSST) that leverages both labeled and unlabeled data. Since these techniques target non-decomposable objectives like min recall, long-tailed imbalanced datasets have become natural benchmarks. However, CSST pre-training on long-tailed data can lead to suboptimal representations that ultimately hurt the mean recall of the models. Additionally, these methods often require re-training from scratch for each specific objective, which reduces their practical applicability.\nExisting empirical methods, which generate debiased pseudo-labels for consistency regularization on long-tailed, class-imbalanced datasets, do improve the overall mean performance but tend to perform poorly on other non-decomposable objectives. To address these issues, we propose SelMix—a technique that leverages a pre-trained model for robust representations and then fine-tunes it to improve a target non-decomposable objective. SelMix selectively mixes data between different classes by developing a tailored sampling distribution for mixup. This distribution is periodically updated based on validation feedback to steer the model toward the desired metric, thereby improving the decision boundaries for specific classes. In contrast to standard mixup methods that apply uniform mixing across all classes, SelMix is designed to better optimize the given objective. Moreover, SelMix extends to non-linear objectives, overcoming limitations found in previous approaches.\nWe evaluate SelMix across a range of objectives, including linear ones (min recall, mean recall), non-linear ones (recall G-mean, recall H-mean), and constrained objectives (recall under coverage constraints). Our experiments show that SelMix significantly improves the performance on the target metric and outperforms existing state-of-the-art methods in most scenarios. In practical situations where the distribution of unlabeled data differs from that of labeled data, an adaptive design using a logit-adjusted FixMatch pre-training yields a 5% improvement, demonstrating its robustness to distribution shifts. Additionally, the semi-supervised learning framework extends easily to supervised learning while still enhancing the desired metrics.\nIn summary, our contributions are threefold. First, we assess existing theoretical frameworks and empirical methods on various non-decomposable metrics, noting that empirical methods tend to excel at mean recall but underperform on other metrics, whereas theoretical methods face the opposite challenge. Second, we introduce SelMix, a mixup-based fine-tuning technique that employs selective mixup across classes to optimize a designated objective. Finally, we demonstrate that SelMix, when applied in both supervised and semi-supervised settings—including cases with differing distributions between labeled and unlabeled data—significantly outperforms current state-of-the-art methods."},"source":{"paper_id":"rxVBKhyfSo","cohort_label":"creative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,4.0,3.0],"soundness":[3.0,3.0,3.0,3.0],"presentation":[3.0,4.0,3.0,2.0],"overall_score":[6.0,8.0,8.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":3.25,"soundness_mean":3.0,"presentation_mean":3.0,"overall_score_mean":7.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"t3gOYtv1xV","group_id":"t3gOYtv1xV","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Large language models have demonstrated impressive abilities across a wide range of tasks and modalities. However, their mathematical reasoning, particularly in integer addition, appears to lag behind. Although various approaches have been explored—such as using carefully selected data or different inference strategies—the challenge remains that mathematics fundamentally relies on logical implications and algorithms to perform well on given tasks. This raises the natural question of how these models implement such algorithms. Due to the complexity of large models, a productive strategy is to first analyze smaller, more interpretable toy models and then apply the insights gained to larger architectures.\nIn our work, we adopt this approach to understand integer addition in transformer-based models using the digit representation of integers. A natural candidate algorithm is the carrying over method, which we examine in transformer models ranging from one to three layers. We primarily focus on encoder-only architectures, but we also discuss findings related to decoder-only models. Furthermore, we investigate how these insights help us understand length generalization in integer addition, both in our experiments and in fine-tuned large language models such as Alpaca 7B, Llemma 7B, and Zephyr 7B.\nThe use of the digit dataset combined with the simplicity of the carrying over algorithm offers a rich testbed for evaluating various aspects of model performance. One advantage is that the algorithm requires the model to perform two fundamental tasks: adding digits at each position and correctly handling the carried one. This dual requirement provides deep insight into the model's operations. Moreover, the dataset naturally divides into subsets depending on where carry operations are needed, allowing for a fine-grained assessment of the model’s internal workings.\nThe transformer architecture is well-suited for this task due to its self-attention mechanism. In the digit representation, once the digits are embedded in the latent space, the model needs to add vectors corresponding to different positions—a process made possible by self-attention once the appropriate patterns are learned.\nWe outline our contributions as follows. We conceptualize the carrying over algorithm in four steps: (1) adding digits at each integer position; (2) determining whether the resulting sums exceed 10; (3) identifying where a carry is required; and (4) performing the carry addition. Our findings show that in two-layer encoder-only models, each step is implemented by a distinct component of the architecture. This interpretation holds for addition tasks with three or four digits, and we observe a similar pattern in three-layer and decoder-only models. Additionally, some of the insights gained from these smaller models appear to extend to fine-tuned large language models like Alpaca 7B, Llemma 7B, and Zephyr 7B. We also demonstrate that the carrying over algorithm learned for shorter digit sequences generalizes to longer ones when the training is primed with a small set of examples involving larger sums. Finally, we identify a novel phase transition in one-layer models during training, where the attention mechanism suddenly shifts to performing addition tasks—a transition that occurs very early in training for larger models.\nRecent related work has examined whether transformers trained on algorithmic tasks can generalize to out-of-distribution examples such as longer integer additions. Researchers have explored architectural modifications, different positional embeddings, and the critical role of input formatting in achieving good performance. Additionally, various computational analyses and interpretability studies have contributed to our understanding of the strengths and limitations of transformer models in performing arithmetic tasks."},"source":{"paper_id":"t3gOYtv1xV","cohort_label":"uncreative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0,4.0],"soundness":[2.0,3.0,3.0,4.0],"presentation":[2.0,3.0,3.0,4.0],"overall_score":[5.0,6.0,5.0,10.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.75,"soundness_mean":3.0,"presentation_mean":3.0,"overall_score_mean":6.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"t8D9JxEn0J","group_id":"t8D9JxEn0J","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"With the growing prevalence of edge devices that are increasingly computationally capable, there is a rising need for learning algorithms that preserve both data locality and privacy. One promising strategy is decentralized machine learning, in which network nodes collaboratively learn a global model through iterative communication while keeping their own data local. Under favorable network connectivity conditions, decentralized approaches can achieve accuracy and convergence rates comparable to centralized methods. However, although decentralized training avoids the need to upload raw data, communicating model updates over channels with limited capacity can become a significant bottleneck—especially when training large-scale deep neural networks where the communication cost is high. This challenge highlights the importance of designing protocols that efficiently communicate and aggregate high-dimensional local models.\nRecent research has focused on improving communication efficiency through model sparsification and gradient or model compression. Model sparsification reduces the dimensionality of model parameters by exploiting their inherent sparsity, while compression techniques—such as quantization and source coding—are used to shrink the size of model updates before they are transmitted. Although these methods reduce the communication burden, they also introduce some level of imprecision during model aggregation.\nAnother line of work has explored gossiping algorithms for consensus aggregation using peer-to-peer or neighborhood communications. These algorithms were initially designed for convex and smooth loss functions with either constant or diminishing step sizes, and later extended to handle smooth non-convex loss functions. Some approaches have also incorporated gradient compression to reduce the volume of data exchanged. In addition, various quantization schemes and source coding methods have been developed to compress model information effectively. Studies in asynchronous decentralized optimization further demonstrate the benefits of model compression, including notable bandwidth savings achieved through gradient sparsification.\nIn parallel, proximal gradient descent has been applied to optimize functions that are non-smooth, particularly when the objective consists of a separable convex component alongside a smooth part. Early work on this technique was expanded to incorporate inexact gradient descent steps and was later adapted for decentralized optimization settings. However, previous studies have not fully analyzed the combined effects of data sub-sampling, stochastic gradient updates, and the inaccuracies introduced by lossy compression.\nIn this work, we tackle the challenge of enhancing communication efficiency in decentralized machine learning for non-convex loss functions. We introduce an algorithm called MALCOM-PSGD, which integrates compressed and coded model updates. Our approach leverages model sparsification through ℓ1 regularization during the stochastic gradient update while also employing a decentralized proximal stochastic gradient technique to address the non-smooth regularization term. To reduce communication costs further, we incorporate gradient compression. Our main contributions are as follows:\n• We present MALCOM-PSGD, which combines gradient compression, vector source encoding, strategic aggregation, and model sparsification to minimize communication overhead. Our analysis and experimental results indicate that the integrated approach can achieve a 75% reduction in bit rate compared to current state-of-the-art methods.\n• We provide a convergence proof for MALCOM-PSGD applied to non-convex and smooth loss functions in the presence of compression error. The algorithm achieves a convergence rate proportional to ln(t)/√t and a consensus rate proportional to 1/t when using diminishing learning rates and consensus step sizes. Although our aggregation scheme improves accuracy, it poses additional analytical challenges since it does not preserve the average of the iterates.\n• We utilize a vector source coding scheme to further reduce communication demands by efficiently encoding the support vector of entries corresponding to specific quantization levels. The combination of compressed model updates and differential encoding creates structural advantages that our encoding scheme can exploit, enhancing overall compression performance.\nOverall, our work demonstrates that by integrating these techniques, communication efficiency in decentralized learning can be significantly improved beyond what is achieved by each individual method."},"source":{"paper_id":"t8D9JxEn0J","cohort_label":"uncreative","pairing_group":"8","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0,2.0],"soundness":[3.0,3.0,3.0,3.0],"presentation":[2.0,3.0,2.0,2.0],"overall_score":[5.0,6.0,5.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":3.0,"presentation_mean":2.25,"overall_score_mean":4.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"tmBKIecDE9","group_id":"tmBKIecDE9","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Where do rewards come from? An artificial intelligence agent introduced into a new environment without prior knowledge starts with a blank slate, having no understanding of what is good or bad and which actions might lead to better outcomes or useful information. For instance, consider an agent tasked with opening a locked door. The first time it encounters a key, the agent lacks any indication that picking it up might be beneficial for opening the door—a piece of common sense that a human would typically possess. Since manually embedding such task-specific knowledge does not scale, we ask: what if we could harness the broad, high-level knowledge humanity has recorded on the Internet to provide agents with similar common sense?\nThe Internet contains a vast amount of unstructured information, making it challenging to directly extract and use this common sense. However, by training on large-scale Internet datasets, large language models (LLMs) absorb this knowledge and make it accessible. Even though the knowledge contained in an LLM operates at a high level of abstraction, a decision-making agent often works at a lower level, processing rich observations and executing fine-grained actions to accomplish its goals. To bridge this gap between high-level common sense and low-level sensorimotor operations, we propose deriving an intrinsic reward function from a pretrained LLM and using it to train agents through reinforcement learning.\nOur method, called Motif, employs an LLM to express preferences between pairs of event captions extracted from a dataset of environment observations. These captions provide coarse descriptions of events rather than detailed, step-by-step accounts. Importantly, the LLM is not required to understand the agent’s low-level action space. In contrast to approaches that integrate an LLM directly as a policy—often necessitating a complete textual interface with the environment—Motif keeps the LLM within its high-level knowledge domain while leveraging robust deep reinforcement learning algorithms to manage complex observation and action spaces.\nWe apply Motif to the challenging NetHack Learning Environment and learn intrinsic rewards from the preferences provided by a pretrained model on a dataset of diverse gameplay observations. This dataset comprises observations collected by various policies, ranging in proficiency, and does not include any action or environmental reward information. Under this framework, the intrinsic reward learned by Motif significantly enhances the subsequent policy learning via reinforcement learning. The approach proves effective in both relatively dense reward tasks, such as maximizing game score, and in extremely sparse reward tasks, like the oracle task. Remarkably, an agent trained solely using Motif’s intrinsic reward achieves a better game score than one trained directly with the extrinsic score provided by the environment.\nBeyond quantifying Motif’s strong game performance, we also explore the qualitative aspects of the behaviors it produces. We find that the intrinsic reward typically leads to behaviors more aligned with human gameplay. Additionally, Motif tends to create anticipatory rewards that simplify the credit assignment problem and align with human common sense. However, we also identify a phenomenon we call \"misalignment by composition,\" in which the joint optimization of an aligned intrinsic reward and a task reward can result in an agent misaligned with respect to the task reward. Our analysis further shows that the performance of the agent scales favorably with both the size of the LLM and the amount of information given in the prompt. Moreover, the agent's performance exhibits sensitivity to slight variations in the prompt, and it is possible to steer the agent’s behavior through prompt modifications, naturally generating a range of semantically diverse policies."},"source":{"paper_id":"tmBKIecDE9","cohort_label":"creative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,4.0,2.0,3.0],"soundness":[2.0,3.0,3.0,3.0],"presentation":[3.0,3.0,4.0,3.0],"overall_score":[5.0,8.0,8.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.75,"soundness_mean":2.75,"presentation_mean":3.25,"overall_score_mean":7.25},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"uYTaVRkKvz","group_id":"uYTaVRkKvz","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Graph neural networks (GNNs) are a powerful class of deep learning models specifically designed for graph-structured data. Unlike traditional neural networks, which typically operate on independent samples, GNNs excel at capturing complex relationships between instances as modeled by a graph. A key idea behind GNNs is message passing, where nodes iteratively aggregate information from their neighbors to update their representations. This process allows information to flow across the network in the form of node embeddings that incorporate both local patterns and global network effects, which are essential for tasks like node classification.\nOne promising subclass of GNN architectures is inspired by graph propagation layers that correspond one-to-one with the descent iterations of a graph-regularized energy function. In these models, each layer computes a progressively refined approximation of a minimizer of the energy function. If these energy minimizers have desirable properties or inductive biases, then the corresponding GNN architecture naturally benefits from them. This contrasts with traditional GNN constructions, which may lack such transparency. This link between the GNN forward pass and optimization also provides opportunities to enhance the architecture for specific challenges, such as handling graphs with spurious edges or heterophily, leading to what we refer to as unfolded GNNs—models in which the layers originate from an unfolded energy descent process.\nDespite their conceptual advantages, unfolded GNNs face significant scaling challenges. They can be designed with arbitrary depth while avoiding oversmoothing, but the associated computational and memory costs can become prohibitively high. Moreover, there has been limited work explicitly addressing the scalability of unfolded GNNs or providing convergence guarantees on large graphs, which means that most existing models have only been evaluated on relatively small benchmarks.\nTo overcome these limitations, we propose a scalable unfolded GNN model that integrates efficient subgraph sampling directly into the design of the energy function. Our model is guided by three main objectives: (i) it should preserve the unique properties and flexible structure of full-graph unfolded GNNs, (ii) using a consistent core architecture, it must maintain competitive accuracy across datasets of varying sizes—including the largest publicly available benchmarks, and (iii) it should not incur excessive computational or memory overhead compared to standard GNN alternatives. We refer to our proposed model as MuseGNN because its architecture arises from the minimization of a sampling-based energy function.\nOur contributions include: expanding a widely used unfolded GNN framework to incorporate offline subgraph sampling directly into the energy function—thereby maintaining the key benefits of unfolded GNNs while achieving unbiased estimation of the full-graph energy with the scalability of standard GNNs; providing analytical evidence that MuseGNN has desirable convergence properties in both the overall model training and energy function descent processes, which supports its robust performance across different problem domains; and offering empirical results that confirm MuseGNN's stability and competitive accuracy on large-scale node classification benchmarks, including some of the largest publicly available datasets."},"source":{"paper_id":"uYTaVRkKvz","cohort_label":"uncreative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,3.0,2.0,2.0],"soundness":[3.0,2.0,3.0,3.0,3.0],"presentation":[4.0,2.0,4.0,4.0,3.0],"overall_score":[8.0,3.0,6.0,6.0,3.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.4,"soundness_mean":2.8,"presentation_mean":3.4,"overall_score_mean":5.2},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"ucMRo9IIC1","group_id":"ucMRo9IIC1","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Following the recent success of large language models, there has been a surge in vision-language models—language models adapted to process both images and text. Leading research laboratories are investing heavily in training these models, and the machine learning community is quickly adapting popular open-source language models into this new format. While enabling models to process images unlocks many useful applications, it also opens up a new attack surface, raising questions about the security of the image input channel.\nAs foundation models become more powerful and increasingly integrated into society, these security concerns grow more pressing. Economic pressures will likely drive AI systems to work with untrusted data and sensitive personal information, and they may eventually perform actions on behalf of users. For example, an AI personal assistant might access sensitive email history, browse the web, send and receive messages, download files, make purchases, or even execute code. It is therefore essential that these models remain secure against input-based attacks. In other words, untrusted inputs should not be able to control a model’s behavior in ways that could lead to leaking personal data, installing malware, or facilitating harmful actions. We refer to these types of exploits as hijacks. Models must remain robust against such attacks even when encountering unexpected inputs or being deployed in adversarial environments, where users may request help with malicious activities or where attackers craft inputs designed to exploit the system.\nOur research reveals that it is possible to create \"image hijacks\"—adversarial images that can control the behavior of vision-language models at inference time. These hijacks allow for a high level of control: they can force a model to generate arbitrary outputs regardless of the accompanying text, cause it to leak its internal context, or even bypass its safety measures. We have even found evidence that image hijacks can be designed to compel a model to behave in harmful ways, regardless of the user-defined text prompt. By applying gradient descent and making only small perturbations to input images, these hijacks can be efficiently constructed.\nThe field of deep learning robustness has struggled to effectively mitigate such adversarial attacks. Despite extensive research efforts, progress in this area has been slow, with only modest improvements in robust accuracy over recent years. If making vision-language models robust against image hijacks proves as challenging as achieving robustness on standard image datasets, this security issue may remain unresolved for some time.\nOur contributions are as follows:\n1. We introduce the concept of image hijacks—adversarial images that manipulate vision-language model behavior at inference time—and present a behavior matching algorithm to train these hijacks in a manner that is resistant to variation in user input.\n2. Inspired by likely misuse scenarios, we develop three types of image hijacks: one that forces the model to generate a specific string; another that bypasses the model’s safety training to comply with harmful instructions; and a third that compels the model to leak its input context.\n3. We evaluate these image hijacks under various constraints and find that state-of-the-art text-based adversarial methods do not perform as well as image hijacks across all tested scenarios.\n4. Finally, we introduce a new prompt matching algorithm to embed a target prompt into an image, which we use to craft a disinformation attack that forces the model to behave as if a major landmark had suddenly been relocated.\nIn essence, our approach leverages a dataset of undesirable behaviors and a fixed vision-language model. By applying the behavior matching algorithm, we adjust the parameters of an input image so that the model’s output aligns with the targeted, adversarial behavior."},"source":{"paper_id":"ucMRo9IIC1","cohort_label":"uncreative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,2.0,3.0],"soundness":[2.0,3.0,3.0,3.0],"presentation":[3.0,3.0,3.0,3.0],"overall_score":[5.0,5.0,5.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.75,"presentation_mean":3.0,"overall_score_mean":5.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"vEEWhGjx0M","group_id":"vEEWhGjx0M","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Multi-armed bandits (MAB) provide a classic framework for sequential decision-making problems. In each round, an agent chooses one of several arms and observes a reward drawn from an unknown distribution. The goal is to maximize cumulative reward by balancing exploitation (selecting the arm with the highest estimated reward) and exploration (gathering more information about lesser-played arms).\nCombinatorial multi-armed bandits (CMAB) extend this framework to settings with many real-world applications such as online advertising, ranking, and influence maximization. In CMAB, the agent selects a combinatorial action—a super arm—composed of several base arms, and observes outcomes for the base arms that are triggered by this action; this feedback is often referred to as semi-bandit feedback. The exploration-exploitation challenge is more complicated in CMAB because the number of candidate super arms may grow exponentially with the number of base arms.\nRecent studies have revealed that MAB and its variants are susceptible to adversarial attacks, particularly reward poisoning attacks. In these attacks, an adversary monitors the chosen arm and its reward, then modifies the observed reward to mislead the algorithm into frequently selecting a particular target arm. The adversary’s aim is to incur a cost that grows slower than the overall time horizon while ensuring that the target arm is chosen almost every time. While it has been shown that no-regret MAB algorithms can be efficiently attacked across a range of instances, some instances of linear bandits require the adversary to pay a cost that grows linearly, indicating inherent robustness in those cases.\nUnderstanding the vulnerability and robustness of CMAB under such poisoning attacks is increasingly important. A straightforward application of the MAB attack framework to CMAB typically results in an attack cost that is sublinear in time but exponential in the number of base arms—a cost that, in many cases, would exceed the time horizon and render the results meaningless. In light of this, we ask: What is a good notion to capture the vulnerability and robustness of CMAB?\nTo address this question, we introduce the concept of polynomial attackability. Under this definition, for an attack to be considered successful, its cost must be sublinear in the time horizon and polynomial in the number of base arms and other relevant factors. Based on this notion, we present the first study of adversarial reward poisoning attacks on CMAB. We establish necessary and sufficient conditions for a CMAB instance to be polynomially attackable—conditions that depend on specifics of the bandit environment, including the reward distributions of super arms and the outcome distributions of base arms. This analysis extends to several practical CMAB applications, such as cascading bandits for online ranking, online shortest path problems, online minimum spanning tree tasks, and the probabilistic maximum coverage problem (a special case of online influence maximization). Our results also apply to simpler reinforcement learning settings and, to the best of our knowledge, provide the first instance-level insights into attackability in reinforcement learning.\nInterestingly, we found that the polynomial attackability of a CMAB instance can depend on whether the adversary has prior knowledge of the environment. We present a challenging example where the instance is polynomially attackable when the environment is known in advance, but becomes polynomially unattackable when it is not. This finding suggests that adversarial attacks on CMAB may be extremely difficult in practice, and that a universal attack strategy for all CMAB instances is unlikely to exist. Finally, we support our theoretical findings with numerical experiments."},"source":{"paper_id":"vEEWhGjx0M","cohort_label":"uncreative","pairing_group":"0","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,2.0,2.0],"soundness":[2.0,2.0,2.0,3.0],"presentation":[3.0,2.0,2.0,2.0],"overall_score":[3.0,5.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.0,"soundness_mean":2.25,"presentation_mean":2.25,"overall_score_mean":5.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"vZ6r9GMT1n","group_id":"vZ6r9GMT1n","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Modern deep neural networks have achieved impressive results across various challenging tasks, such as image classification and face recognition. However, previous research has demonstrated that these models are susceptible to minor input changes and can be deceived by carefully crafted, imperceptible perturbations. These adversarial attacks generally fall into two categories: white-box and black-box. In white-box attacks, techniques like Projected Gradient Descent can effectively generate adversarial examples by utilizing information from within the model. In contrast, practical scenarios—such as machine learning as a service—often restrict user access to well-trained models and training datasets, especially with large models, making query-based black-box attacks the main threat. In these attacks, adversaries have no knowledge of the model architecture or parameters.\nThis paper introduces a lightweight, plug-and-play defense that significantly reduces the success rate of query-based black-box attacks, including both score-based and decision-based variants. Unlike adversarial examples generated by query-based methods, which involve iterative procedures—either through local search with small incremental input modifications or optimization methods relying on estimated gradients—the proposed defense disrupts this process by injecting noise into the feature space. Although previous work has investigated the robustness of randomized input features, our approach provides both theoretical analysis and empirical evidence to demonstrate improved robustness through strategically introduced noise. We emphasize the importance of dynamically adjusting the noise based on the gradient of the adversarial objective during the attack process.\nIn summary, our contributions are as follows:\n• We explore the effect of randomized perturbations in the feature space and their relationship to model robustness against black-box attacks.\n• We propose a simple yet effective defense strategy that hinders an attacker’s ability to determine the direction for generating adversarial samples, thereby significantly reducing attack success.\n• We conduct extensive experiments on both score-based and decision-based attacks to validate our analysis and demonstrate that our method enhances the robustness of models against query-based adversarial attacks."},"source":{"paper_id":"vZ6r9GMT1n","cohort_label":"creative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,2.0,2.0,4.0],"soundness":[1.0,3.0,3.0,4.0],"presentation":[1.0,3.0,3.0,3.0],"overall_score":[3.0,6.0,6.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.5,"soundness_mean":2.75,"presentation_mean":2.5,"overall_score_mean":5.75},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"vZZ4hhniJU","group_id":"vZZ4hhniJU","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Communication is a key capability necessary for effective coordination among agents in partially observable environments. In multi-agent reinforcement learning, agents can use their actions to transmit information, but continuous or discrete messages on a communication channel—linguistic communication—are more flexible and powerful because they can convey more complex concepts. For successful communication, a speaker and a listener must share a common language with a shared understanding of the symbols being used.\nEmergent communication, or learning a common protocol, is a thriving research direction, although most work focuses on simple, single-turn sender-receiver games. In more visually and structurally complex multi-agent reinforcement learning environments, existing approaches often rely on centralized learning mechanisms, such as sharing models or gradients. However, a centralized controller is impractical in many real-world environments where agents cannot easily synchronize and must act independently—that is, in a decentralized setting. While centralized training with decentralized execution is a common compromise, it may not perform better than purely decentralized training. A centralized controller also suffers from the curse of dimensionality; as the number of agents increases, the amount of necessary communication grows exponentially. Moreover, the fully decentralized setting is more flexible and requires fewer assumptions about other agents, making it more realistic in many real-world scenarios.\nThis work explores learning to communicate to coordinate agents in a decentralized setting. In this approach, each agent has its own model to decide how to act and communicate, with no shared parameters or gradients among agents. Traditional reinforcement learning approaches to decentralized communication are known to perform poorly even in simple tasks due to the large communication space, high variance of reinforcement learning, and a lack of common grounding for communication. Past work has leveraged the way communication influences other agents to learn protocols, and more recent methods have employed autoencoding of observations to use the encodings as communication by relying on the shared environment for common grounding.\nBuilding on this idea, we propose that if agents are in similar states, they should produce similar messages. This idea leads to a simple method based on contrastive learning to ground communication. Inspired by representation learning literature, where different “views” of a data sample are used, we treat an agent’s observation as a view of the environment state during a trajectory. Thus, different agents' messages become encodings of different incomplete views of the same underlying state. From this perspective, messages from the same state are generally more similar to each other than to those from distinct states or different trajectories. We call this approach Communication Alignment Contrastive Learning (CACL), where each agent independently uses contrastive learning between their own sent and received messages to learn a communication encoding.\nWe validate CACL experimentally in three communication-essential environments and show that it improves performance and learning speed, outperforming state-of-the-art decentralized multi-agent reinforcement learning communication algorithms. To understand CACL’s success, we introduce both qualitative and quantitative metrics. Our analysis demonstrates that CACL leads to more symmetric communication, meaning different agents communicate similarly when faced with the same observations, resulting in greater mutual intelligibility. By treating messages as representations, we show that CACL’s messages capture global semantic information about the environment better than baseline methods. Overall, we argue that contrastive learning is a powerful approach for multi-agent communication with significant advantages over previous methods."},"source":{"paper_id":"vZZ4hhniJU","cohort_label":"creative","pairing_group":"6","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,2.0,2.0],"soundness":[2.0,4.0,3.0,1.0],"presentation":[3.0,4.0,3.0,2.0],"overall_score":[5.0,8.0,6.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.25,"soundness_mean":2.5,"presentation_mean":3.0,"overall_score_mean":6.0},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"wHBfxhZu1u","group_id":"wHBfxhZu1u","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Transformer-based large language models have become the dominant choice for many natural language processing tasks, especially those that depend on long-range abilities like in-context learning. One major limitation of these pretrained models is the fixed length of their sequence context, which is determined during training. As the demand for dynamically extending the context window increases—with or without fine-tuning—the discussion naturally centers on transformer positional encodings.\nThe original Transformer design employed an absolute sinusoidal position encoding, later improved by learnable absolute encodings. Subsequent advancements introduced relative positional encoding schemes, which have further enhanced Transformer performance. Today, popular methods include relative biases and other variants like RoPE, XPos, and ALiBi.\nA persistent limitation of positional encodings is their inability to generalize beyond the context window observed during training. While some methods, such as ALiBi, offer limited extrapolation, none can handle sequences substantially longer than those seen during pre-training. To address this challenge, some researchers have modified RoPE using Position Interpolation and fine-tuning on a small amount of data. An alternative approach introduced an \"NTK-aware\" interpolation that accounts for the loss of high frequency. Building on this, two improvements have emerged: one method, Dynamic NTK interpolation, adapts the strategy for pre-trained models without fine-tuning, while another, NTK-by-parts interpolation, achieves the best performance when fine-tuned on a limited dataset containing longer-context examples.\nThese methods have already been incorporated into open-source models, with some using NTK-aware interpolation and others employing Dynamic NTK. In this work, we provide a comprehensive overview of the evolution from NTK-aware and Dynamic NTK to NTK-by-parts interpolation, and introduce YaRN (Yet another RoPE extensioN method). YaRN presents an improved approach to efficiently extend the context windows of models that rely on Rotary Position Embeddings, including families like LLaMA, GPT-NeoX, and PaLM. A diagram in the full paper outlines the relationship and progression among the different interpolation methods.\nYaRN achieves state-of-the-art performance in context window extension after fine-tuning on less than 0.1% of the original pre-training data. Additionally, by integrating an inference-time technique called Dynamic Scaling, the Dynamic-YaRN variant can more than double the context window without any fine-tuning."},"source":{"paper_id":"wHBfxhZu1u","cohort_label":"creative","pairing_group":"4","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,3.0,3.0],"soundness":[3.0,3.0,3.0,2.0],"presentation":[2.0,3.0,2.0,2.0],"overall_score":[6.0,8.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":3.0,"soundness_mean":2.75,"presentation_mean":2.25,"overall_score_mean":6.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"wg8NPfeMF9","group_id":"wg8NPfeMF9","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Deep implicit functions (DIFs) have recently emerged as efficient representations for 3D shapes, deformation fields, images, videos, graphs, and manifolds. For instance, DeepSDF represents shapes using the level set of a signed distance field encoded by a neural network. This approach allows for a compact, continuous, and differentiable representation of 3D shapes with effectively infinite resolution. Beyond traditional geometry representations like voxel grids, point clouds, and meshes, DIFs have proven to be a powerful paradigm in various computer vision tasks, including 3D shape reconstruction, generation, registration, completion, editing, and overall shape understanding.\nDespite these advancements, shape analysis using DIFs has not received as much attention. When provided with a set of 3D shapes along with associated covariates, there remains a need for a representation that can accurately describe the shapes and capture dependencies among them. No current method quantitatively captures how covariates, such as age, sex, or weight, affect a population of 3D shapes—both on average and individually. This ability is crucial, especially in medical studies where understanding the influence of such covariates is a primary goal, and controlling them during data collection can be challenging or impossible.\nTo address these challenges, our approach is based on estimating a shape atlas—a template shape whose deformation accounts for covariate effects and models shape differences. Take the airway as an example: an ideal atlas representation would be capable of answering several questions, such as how to accurately represent a shape and its dependencies given an atlas, how to disentangle the effects of different covariates on an airway shape, how an airway atlas changes with a specific covariate like age, and how an airway might develop over time given a random shape.\nIn response, we introduce a Neural Additive Interpretable Shape Representation (NAISR), which provides an interpretable method for modeling shapes associated with covariates through an atlas-based approach. NAISR is compared with existing shape representation methods based on several key properties. First, it is implicit, meaning that the shape description adapts naturally to various resolutions and supports complete shape reconstruction even from partial data—a common scenario in medical imaging. Second, it is deformable, ensuring that a displacement field creates point correspondences between the target shapes and the atlas. Third, it is disentangleable, which allows for the separation of individual covariate-specific effects that can be combined to yield the overall deformation of the atlas. Fourth, it is evolvable, capturing how each covariate influences the shape population statistically, which can be particularly useful in scientific discovery. Fifth, it is transferable, meaning that shape modifications can be applied to a given shape, enabling, for example, the simulation of surgical changes and the prediction of their outcomes over time. A model that is simultaneously deformable, disentangleable, evolvable, and transferable is considered interpretable.\nNAISR is the first implicit shape representation method to explore an atlas-based approach for 3D shapes in such a comprehensive manner. To demonstrate its versatility and effectiveness, we test NAISR on a simulated dataset as well as two realistic medical datasets: a simulated 2D shape dataset (Starman), a 3D hippocampus shape dataset from Alzheimer’s disease research, and a pediatric airway 3D shape dataset. Our quantitative results show that NAISR outperforms existing baseline methods."},"source":{"paper_id":"wg8NPfeMF9","cohort_label":"creative","pairing_group":"11","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,3.0,3.0],"soundness":[3.0,4.0,3.0,3.0],"presentation":[4.0,4.0,4.0,3.0],"overall_score":[8.0,6.0,6.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.75,"soundness_mean":3.25,"presentation_mean":3.75,"overall_score_mean":6.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"x36mCqVHnk","group_id":"x36mCqVHnk","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"Multi-agent reinforcement learning (MARL) refers to frameworks where multiple agents learn to make decisions in a shared, unknown environment to maximize their cumulative rewards. MARL has achieved notable success in various practical applications, including board games like Go, team-based real-time strategy games, autonomous driving, and modeling behavior in complex social systems. A significant challenge, however, is that many reinforcement learning algorithms require a vast number of samples to reach human-level performance due to exploration demands. This makes the development of sample-efficient algorithms a critical research goal.\nA common approach in MARL is to use model-based methods. These methods estimate an environmental model from observed data, employ planning techniques to derive a policy, and then execute that policy in the environment. In two-player zero-sum Markov games, research has shown that model-based algorithms can be provably efficient, achieving (nearly) optimal sample complexity in most parameters. In contrast, model-free methods—which directly estimate value functions at equilibrium policies—have not yet reached this level of sample efficiency. In these methods, the number of episodes required grows less favorably with the time horizon, prompting the important question: Can we design model-free algorithms that depend optimally on the time horizon when learning two-player zero-sum Markov games?\nThis paper answers that question affirmatively with three main contributions:\n1. Algorithm Design. We propose a new model-free algorithm based on Q-learning that incorporates a min-gap reference-advantage decomposition. While reference-advantage decomposition has been used in single-agent reinforcement learning, we extend the technique to two-player zero-sum Markov games with important modifications. In this more complex setting, the optimistic (or pessimistic) value functions do not necessarily maintain a monotonic relationship due to the equilibrium oracle’s behavior. To overcome this, our method updates the reference value functions using the pair of optimistic and pessimistic functions that have the smallest difference (the minimal gap) observed so far. We also use a stage-based approach, which simplifies both the algorithm and its analysis.\n2. Sample Complexity Bound. We demonstrate that our algorithm can find an approximately optimal Nash equilibrium for two-player zero-sum Markov games within nearly optimal sample complexity. Our results improve on the sample efficiency of existing model-free methods and, when compared to established lower bounds, show that our algorithm is optimal with respect to key parameters such as the time horizon, number of states, and accuracy level. This is the first instance where a model-free method achieves this level of optimality.\n3. Technical Analysis. We introduce several new properties related to the cumulative occurrence of large value gaps and the accumulation of bonus terms. These properties help us rigorously bound various sources of error introduced by our new min-gap based reference-advantage decomposition. While similar techniques have been developed for single-agent scenarios, our analysis addresses additional challenges posed by the equilibrium oracle in zero-sum Markov games. Furthermore, rigorously analyzing both the optimistic and pessimistic bonus terms required a more refined approach than previously employed in related single-agent studies.\nRelated work in this field has evolved from early studies that assumed full knowledge of state transitions and rewards or focused on asymptotic settings, to recent efforts establishing non-asymptotic guarantees under more realistic assumptions. Various approaches, from using simulators for state-action sampling to imposing reachability conditions, have been explored in the context of Markov games. In two-player zero-sum games, both model-based and model-free methods have been investigated, with model-free methods traditionally suffering from suboptimal sample efficiencies. Recent work in model-free approaches, including Q-learning and V-learning variants, has sought to address these challenges. Additionally, research on single-agent reinforcement learning, which has seen significant progress in sample complexity with similar variance reduction techniques, closely informs our approach—albeit with significant adaptations to address the challenges unique to two-player zero-sum settings.\nIn summary, our work provides a novel model-free algorithm for two-player zero-sum Markov games that not only advances the design of such methods but also establishes sample complexity bounds that are optimal with respect to critical parameters, thereby narrowing the gap between model-free and model-based approaches in multi-agent reinforcement learning."},"source":{"paper_id":"x36mCqVHnk","cohort_label":"uncreative","pairing_group":"5","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,3.0,3.0,2.0],"soundness":[3.0,3.0,3.0,2.0],"presentation":[2.0,3.0,4.0,1.0],"overall_score":[5.0,6.0,6.0,5.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.75,"soundness_mean":2.75,"presentation_mean":2.5,"overall_score_mean":5.5},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"x8VNtpCu1I","group_id":"x8VNtpCu1I","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"This paper examines the potential of the BERT family to serve as effective instruction-following models. Recent advances have shown that large language models can be fine-tuned with annotated examples and human preference feedback to align their outputs with human expectations across a variety of tasks such as understanding, generation, and commonsense reasoning—in settings that range from zero-shot and few-shot to chain-of-thought. Several studies have highlighted factors that contribute to the success of instruction following in both decoder-only models like GPT and LLaMA and encoder-decoder frameworks like T5. Factors such as annotation quality, data formatting, and scaling have been crucial, and it is expected that further alignment methods will enhance more model variants.\nMotivated by observations that the BERT family is capable of zero-shot learning and that its masked language modeling approach supports open-ended long text generation with faster decoding compared to autoregressive methods, this study explores its potential as an instruction follower. Given that real-world applications often require collaboration between generative large language models and expert models, and considering that the BERT family has previously driven numerous tasks ahead of current generative models while still holding records in many downstream tasks, it is a promising candidate for further exploration.\nThe contributions of this work are fourfold. First, it introduces a basic method to assess the potential and limitations of using the BERT family as instruction followers. Second, it identifies key factors that could make these models competitive with both encoder-decoder and decoder-only architectures of similar scale and pretraining data. Third, it uncovers various risks and limitations. Finally, it discusses possible reasons behind these challenges and suggests avenues for improvement.\nPreliminary observations indicate that the BERT family may generalize well across tasks and languages when compared to other popular language models with similar sizes and data. Moreover, the proposed method is scalable, allowing for significant performance improvements with additional computational resources and data. However, some issues might arise from the intrinsic constraints of the BERT models, which have not yet reached the power of their decoder-only and encoder-decoder counterparts, and the inherent limitations of non-autoregressive decoding may lead to certain unreliability issues when applied to diverse tasks.\nThe remainder of the paper is structured as follows: Section 2 reviews the background of text generation with the BERT family and instruction tuning. Section 3 details the approach used for post-training BERT models as language generators, and Section 4 describes the experimental settings for instruction fine-tuning. Finally, Sections 5 and 6 delve into the potential benefits and limitations of this approach, analyzing the underlying causes and exploring possible solutions for future enhancements."},"source":{"paper_id":"x8VNtpCu1I","cohort_label":"creative","pairing_group":"11","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[2.0,3.0,4.0],"soundness":[3.0,3.0,4.0],"presentation":[3.0,4.0,4.0],"overall_score":[6.0,8.0,6.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":3.0,"soundness_mean":3.3333333333333335,"presentation_mean":3.6666666666666665,"overall_score_mean":6.666666666666667},"split_seed":42,"split":"validation"} | |
| {"dataset":"CreativeEval","example_id":"zmJDzPh1Dm","group_id":"zmJDzPh1Dm","input":{"task":"Predict multidimensional human evaluations of a research idea described by a paper introduction.","paper_introduction":"In recent years, large-scale pretrained vision-language models like CLIP, Flamingo, and BLIP have gained popularity, and soft-prompt-based methods—often referred to as prompt-tuning—have emerged as a dominant technique for adapting these models to various downstream tasks. For example, one approach learns soft prompts in a continuous space for image classification, while other studies have extended prompt-tuning to address dense prediction and open-vocabulary object detection. Much of the recent research has focused on aligning visual and textual features to enhance model performance, whether by estimating the weight distribution of output embeddings, jointly optimizing prompts across multiple modalities, or applying optimal transport techniques. Some investigations have even interpreted learned soft prompts by mapping them to the nearest words in the embedding space and examined the role of attention mechanisms in these tuning methods.\nDespite these advancements, little attention has been directed toward the intrinsic properties of soft prompts, particularly the norms of the learnable soft-prompt vectors. We believe that these norms are a crucial yet overlooked factor that significantly influences the performance of vision-language models. This paper addresses that gap by asking, “Do we need to normalize the soft prompts in vision-language models?”—a question that, to our knowledge, has not been previously explored.\nTo investigate the role of soft-prompt norms, we introduce two corruption operations, REPLACE and RESCALE, which alter the norms of vectors learned by existing methods. This corruption reveals an intriguing phenomenon: reducing the norms at certain positions within the prompts can enhance performance, whereas increasing the norms typically leads to a decline in performance. We refer to this phenomenon as the Low-Norm Effect. Our analysis across 11 widely used datasets shows that the Low-Norm Effect occurs more frequently in some datasets—such as ImageNet, OxfordPets, and Food101—compared to others like Flowers102. Moreover, there is a negative correlation between the frequency of the Low-Norm Effect and the number of training examples, suggesting that limited data is more likely to trigger this effect. These discrepancies may be related to the degradation of soft-prompt methods when training data is scarce, adversely affecting model performance. However, addressing the Low-Norm Effect is challenging because it does not occur uniformly across all datasets.\nTo leverage the Low-Norm Effect for improved performance, we propose a method called Nemesis, which normalizes the soft-prompt vectors in vision-language models. Initially, we introduce a Position-Uniform Normalization loss that regulates the norms of all prompt vectors—a simple addition to current soft-prompt methods that incurs minimal computational cost. However, this uniform approach can inadvertently normalize vectors that are not affected by the Low-Norm Effect, potentially degrading performance. To mitigate this issue, we propose a refined Position-Aware Normalization loss. This method includes a pre-inference step before each training batch, during which corruption operations generate multiple sets of corrupted prompts at various positions. By comparing these with the original prompts, we can identify the positions that induce the Low-Norm Effect and then selectively normalize only those positions. Extensive experiments demonstrate that the Nemesis method can effectively boost the performance of soft-prompt-based vision-language models across various tasks."},"source":{"paper_id":"zmJDzPh1Dm","cohort_label":"creative","pairing_group":"10","paper":"https://aclanthology.org/2026.eacl-long.297/","code":"https://github.com/lichun-19/creative_eval","source_dataset":"CreativeEval research ideation subset"},"reviewer_scores":{"contribution":[3.0,2.0,3.0],"soundness":[3.0,2.0,4.0],"presentation":[3.0,3.0,3.0],"overall_score":[6.0,6.0,8.0]},"target_descriptors":{"contribution":"Significance, originality, and potential impact of the proposed research idea.","soundness":"Technical validity of the idea and whether its proposed methodology supports its claims.","presentation":"Clarity, organization, and effectiveness with which the research idea is communicated.","overall_score":"Holistic assessment of the research idea's quality, promise, and suitability for further development."},"targets":{"contribution_mean":2.6666666666666665,"soundness_mean":3.0,"presentation_mean":3.0,"overall_score_mean":6.666666666666667},"split_seed":42,"split":"validation"} | |