Datasets:

Mithilss
/

reclaim

@@ -74,14 +74,12 @@
 {"custom_id": "2511.20906", "central_claim": "DA-SIP achieves 2.6-4.4x reduction in total computation time while maintaining task success rates comparable to fixed maximum-computation baselines for robot manipulation tasks.", "claim_evidence": "Abstract states 'DA-SIP achieves 2.6--4.4x reduction in total computation time while maintaining task success rates comparable to fixed maximum-computation baselines.' Table 5 shows DA-SIP with Lightweight CNN achieves 2.6x-4.4x speedup across tasks (Push T: -2.69x, Transport: -4.43x, Lift: -4.09x) with minimal success rate degradation (max -3%). Table 2 shows 100% success on Lift task with single inference step, demonstrating efficiency.", "paper_kind": "empirical", "mre_config": "Train SI Policy with Linear interpolation and velocity prediction on RoboMimic 'Lift' task for 5000 epochs (batch size 256, lr 1e-4 with cosine decay) on NVIDIA L40S GPU. Evaluate with 1-step Euler ODE integration. Measure success rate over 50 episodes.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2511.20906"], "code": [], "dataset": ["https://github.com/ARVR-CV/mixed_papers_summary/blob/main/2511.20906.md"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Live searches for the full title, arXiv ID 2511.20906, DA-SIP, Difficulty-Aware Stochastic Interpolant Policy, author/title plus GitHub/code, OpenReview ID oDoPiR8wZJ, and Hugging Face variants found the NeurIPS/OpenReview/arXiv paper pages but no public implementation repository. The bundled OpenReview supplement manifest contains only 'DASIP appendix v3.pdf' and a macOS metadata file, not runnable training/evaluation code, configs, scripts, or notebooks for the Can MRE."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "RoboMimic (Can, Lift, Square, Tool Hang tasks), Push-T, Block Push (Fetch), Kitchen, and Multimodal Ant (OpenAI Gym) are standard public robotics benchmarks with documented download procedures and established evaluation protocols."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searches for DA-SIP with checkpoint, weights, model, Hugging Face, arXiv ID, and title variants found no downloadable SIP policy checkpoints, CNN classifier checkpoints, or VLM adapters for the Can MRE. The bundled OpenReview supplement contains only an appendix PDF and no model files."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "RoboMimic, Push-T, Kitchen, Fetch Block Push, and OpenAI Gym Ant are widely recognized public robotics benchmarks used extensively in imitation learning research."}}, "agent_task": "Train a stochastic interpolant policy with linear interpolation and velocity prediction on the RoboMimic Lift task for 5000 epochs using AdamW optimizer with batch size 256 and learning rate 1e-4 with cosine decay. Evaluate the trained policy using 1 inference step with Euler ODE integration and measure task success rate over 50 episodes. Verify that success rate reaches approximately 100%.", "h100_estimate": {"hours": 2.22, "basis_kind": "derived_from_config", "gpu_count": 1, "gpu_type": "L40S", "wallclock_hours": 6, "h100_equivalent_multiplier": 0.37, "basis": "weights_available=false -> smallest training run. mre_config trains SI Policy on single L40S GPU, one task (Lift), one seed, not the full paper's 3-seed multi-task sweep. Original basis used gpu_count=8 (full-paper scale), a scope error. Corrected: 1 GPU * 6h wallclock * 0.37 (L40S) = 2.22 H100-hours."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 2.22, "h100_estimate_basis": "derived_from_config: weights_available=false -> smallest training run. mre_config trains SI Policy on single L40S GPU, one task (Lift), one seed, not the full paper's 3-seed multi-task sweep. Original basis used gpu_count=8 (full-paper scale), a scope error. Corrected: 1 GPU * 6h wallclock * 0.37 (L40S) = 2.22 H100-hours.", "h100_band": "0-8", "h100_recomputed_hours": 2.2199999999999998, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 2.22, "h100_hours_adjudicated": false, "selection_band": "0-8", "split": "eval", "match_target": {"config": "SI Policy with Linear interpolation, velocity prediction, 1 inference step, Euler ODE approximator, trained 5000 epochs on Lift task", "metric": "task success rate", "value": "100%", "scope": "RoboMimic Lift task", "match_bar_kind": "threshold"}}
 {"custom_id": "2507.06489", "central_claim": "Adversarial attacks on verbal confidence (both perturbation-based VCA-TF, VCA-TB and jailbreak-based ConfidenceTriggers) can significantly reduce LLM confidence scores and cause high rates of answer changes.", "claim_evidence": "Abstract: 'these attacks can significantly impair verbal confidence estimates and lead to frequent answer changes.' Table 1 shows VCA-TF on TruthfulQA/Llama-3-8B/Base achieves 7.0% affected samples with 19.3% confidence drop; Typos attack achieves up to 54.0% affected samples with 24.7% confidence drop. Table 2 shows ConfidenceTriggers achieves up to 29.5% confidence reduction on TQA with CoT.", "paper_kind": "empirical", "mre_config": "VCA-TF perturbation attack on Llama-3-8B with Base CEM on TruthfulQA dataset, 200 random samples, measuring % affected samples and average confidence drop on affected samples.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2507.06489"], "code": [], "dataset": ["https://hf.co/datasets/truthfulqa/truthful_qa", "https://hf.co/datasets/ChilleD/StrategyQA", "https://hf.co/datasets/openlifescienceai/medmcqa"], "weights": ["https://hf.co/meta-llama/Meta-Llama-3-8B-Instruct"]}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searched GitHub and HuggingFace with multiple query variations: 'ConfidenceTriggers', 'verbal confidence adversarial attacks', 'Obadinma', arXiv ID 2507.06489, full title, and key method names (VCA-TF, VCA-TB, SubSwapRemove, ConfidenceTriggers). Found only paper analysis notes in repos (memgrafter/analysis, zhaoyang97/Paper-Notes). No official code repository exists for this paper."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "All three datasets used in the paper are publicly available on HuggingFace: TruthfulQA (truthfulqa/truthful_qa, 89.2K downloads), StrategyQA (ChilleD/StrategyQA, 11.1K downloads), and MedMCQA (openlifescienceai/medmcqa, 30.4K downloads)."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "The paper does not release any checkpoints or weights for their attack methods (ConfidenceTriggers, VCA-TF, VCA-TB). Llama-3-8B is available on HuggingFace from Meta, but these are base model weights, not the paper's trained trigger tokens or attack artifacts. No VCA-specific weights/checkpoints were found after searching."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "TruthfulQA (arXiv:2109.07958), StrategyQA (crowdsourced benchmark published in ICLR 2022), and MedMCQA (medical MCQA dataset with 2816+ questions from ML research) are established public benchmarks widely used in NLP research."}}, "agent_task": "Reproduce the VCA-TF perturbation-based verbal confidence attack on Llama-3-8B using the TruthfulQA dataset. Implement the attack algorithm from the paper: (1) calculate importance scores by removing each token and measuring confidence change, (2) rank tokens by importance, (3) substitute top tokens with synonyms (POS-tag filtered) while maintaining 80% semantic similarity (USE). Measure: (1) % of samples where confidence is reduced (%Aff), (2) average confidence drop on affected samples (Δ Aff. Cf.), (3) answer flip rates. Expected: at least 7% affected samples with measurable confidence drops on TruthfulQA/Base/Llama-3-8B.", "h100_estimate": {"hours": 2.56, "basis_kind": "derived_from_config", "gpu_count": 1, "gpu_type": "A100 80GB", "wallclock_hours": 8, "h100_equivalent_multiplier": 0.32, "basis": "MRE = VCA-TF attack (200 TruthfulQA samples, Llama-3-8B, Base CEM); no weights/code released for the attack artifacts, so this is the paper's own smallest inference-search config, not a training run. Original basis combined a 0.2s/batch-of-32 throughput claim with a 375h wallclock figure for only ~750-1500 LLM calls, a unit/arithmetic error inflating hours ~150x. Re-derived: ~200 samples x ~10-25 search iterations x a few candidate evals/iteration on an 8B model, generously bounded at 8 wallclock-hours on 1x A100 80GB (sequential, hard to batch due to adaptive search). 1*8*0.32=2.56 H100-hours."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 2.56, "h100_estimate_basis": "derived_from_config: MRE = VCA-TF attack (200 TruthfulQA samples, Llama-3-8B, Base CEM); no weights/code released for the attack artifacts, so this is the paper's own smallest inference-search config, not a training run. Original basis combined a 0.2s/batch-of-32 throughput claim with a 375h wallclock figure for only ~750-1500 LLM calls, a unit/arithmetic error inflating hours ~150x. Re-derived: ~200 samples x ~10-25 search iterations x a few candidate evals/iteration on an 8B model, generously bounded at 8 wallclock-hours on 1x A100 80GB (sequential, hard to batch due to adaptive search). 1*8*0.32=2.56 H100-hours.", "h100_band": "0-8", "h100_recomputed_hours": 2.56, "h100_arithmetic_mismatch": false, "h100_needs_human_review": true, "score": 3, "tier": "Hard", "audited_h100_hours": 2.56, "h100_hours_adjudicated": false, "selection_band": "0-8", "split": "eval", "match_target": {"config": "Llama-3-8B + VCA-TF attack + Base CEM + TruthfulQA (200 samples)", "metric": "% affected samples with confidence reduction", "value": "7.0%", "scope": "TruthfulQA/Llama-3-8B/Base", "match_bar_kind": "threshold"}}
 {"custom_id": "2602.03066", "central_claim": "Shortcut features correspond to eigenfunctions of the Neural Tangent Kernel (NTK) with larger eigenvalues; features with larger eigenvalues are learned faster due to spectral bias and have a larger influence on the neural network output even after training convergence.", "claim_evidence": "Abstract states: 'shortcut features correspond to features with larger eigenvalues when the shortcuts stem from the imbalanced number of samples in the clustered distribution' and 'features with larger eigenvalues still have a large influence on the neural network output even after training'. Section 4.3 (Figure 5, synthetic) and Figure 9 (Appendix synthetic) show that availability (alignment to top NTK eigenspace) is higher for shortcut labels than ground-truth labels across Colored-MNIST and Patched-MNIST, with differences of ~0.08-0.10 under CE loss. Section 4.3 (Figure 6, real-world) and Figure 11-12 show the same pattern on Waterbirds and CelebA.", "paper_kind": "empirical", "mre_config": "Train a 2-layer ReLU fully-connected network (width 256, SGD, lr=0.001, wd=0.0001, batch=1000, zero-initialized output weights, 200 epochs) on Colored-MNIST. Construct Colored-MNIST with 95% of label=-1 samples colored red and 5% of label=1 samples colored red (training split); test split has balanced 50/50 red/blue coloring. At step 1000, compute the empirical NTK K from 500 randomly sampled training inputs (using the finite-width network Jacobian). Compute availability A(K, g) = g^T K g / (||g||^2_2 ||K||_F) for both the ground-truth label vector and the shortcut label vector (color-based: red=1, blue=-1). Verify that availability(shortcut) > availability(ground-truth) by a margin consistent with the paper's reported ~0.08-0.10 difference for Colored-MNIST with CE loss (per Appendix Table 3).", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2602.03066", "https://papers.nips.cc/paper_files/paper/2025/hash/87571720167f7e88827c40e468e3101f-Abstract-Conference.html"], "code": [], "dataset": ["http://yann.lecun.com/exdb/mnist/"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No OpenReview supplement was provided in the input, and the paper checklist says code/data are not yet released. Live web searches for the exact title, arXiv ID 2602.03066, OpenReview ID TYroQXu6X0, author names, and GitHub/code variants found the OpenReview/arXiv/NeurIPS paper pages but no public repository or runnable scripts for the Patched-MNIST/Colored-MNIST availability MRE."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE can be built from MNIST, which is publicly available through torchvision.datasets.MNIST with a documented download parameter. The paper text specifies the MRE-specific biased construction: digits 0-4 labelled -1, digits 5-9 labelled +1, with a 95/5 shortcut ratio in training and 50/50 in test for Patched-MNIST/Colored-MNIST. No first-party processed split files were found, but the raw dataset and construction recipe are sufficient for this minimal derived benchmark."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "The MRE trains a small two-layer ReLU fully connected model from scratch and does not require pretrained weights. Searches for the title/arXiv ID plus checkpoint, weights, model, Hugging Face, and GitHub variants found no released checkpoints or model files for the MRE."}, "dataset_is_standard": {"value": false, "verification": "tool_verified", "evidence": "MNIST itself is a standard public benchmark, verified via torchvision documentation, but the actual MRE dataset is an author-defined biased derivative, Patched-MNIST or Colored-MNIST, with custom shortcut features, custom binary labels, and custom train/test shortcut ratios. Therefore the MRE dataset is not a standard benchmark as used in the paper."}}, "agent_task": "Reproduce the core empirical finding on Colored-MNIST: (1) Download MNIST from http://yann.lecun.com/exdb/mnist/. (2) Construct Colored-MNIST: for each training image, if label==-1 (digits 0-4), color it red with probability 0.95; if label==1 (digits 5-9), color it red with probability 0.05; for test set use 50/50 balanced colors. (3) Train a 2-layer ReLU FC network (width=256, zero-initialized last-layer weights, no bias) with SGD (lr=0.001, wd=0.0001, batch=1000) for 200 epochs. (4) At training step 1000, compute the empirical NTK K from 500 randomly sampled training inputs using the finite-width network Jacobian (K = J J^T where J is the Jacobian of outputs w.r.t. parameters). (5) Compute availability A(K, g) = g^T K g / (||g||^2_2 ||K||_F) for both the ground-truth label vector g_y and the shortcut label vector g_color (color-based label: red=1, blue=-1). (6) Verify that availability(g_color) > availability(g_y) by a margin consistent with the paper's reported difference of ~0.08-0.10 for Colored-MNIST with CE loss (Appendix Table 3). This confirms that shortcut features align more strongly with top NTK eigenfunctions than core features.", "h100_estimate": {"hours": 4, "basis_kind": "derived_from_config", "gpu_count": 1, "gpu_type": "A100 80GB", "wallclock_hours": 12, "h100_equivalent_multiplier": 0.32, "basis": "The paper (Section X_suppl) reports experiments on 2 RTX 3090 GPUs for 1 hour for synthetic dataset experiments. For Colored-MNIST: 200 epochs, batch=1000, 2-layer ReLU FC width 256. Scaling to 1 GPU: ~2 hours wallclock (linear scaling of GPU-hours). Adding NTK computation overhead (500 samples, Jacobian computation at step 1000) adds ~10-20% overhead. 1 * 2 * 0.32 = 0.64 H100-hours for base training. Adding engineering/debugging margin for custom Colored-MNIST construction and NTK implementation: 4 H100-hours total."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 4.0, "h100_estimate_basis": "derived_from_config: The paper (Section X_suppl) reports experiments on 2 RTX 3090 GPUs for 1 hour for synthetic dataset experiments. For Colored-MNIST: 200 epochs, batch=1000, 2-layer ReLU FC width 256. Scaling to 1 GPU: ~2 hours wallclock (linear scaling of GPU-hours). Adding NTK computation overhead (500 samples, Jacobian computation at step 1000) adds ~10-20% overhead. 1 * 2 * 0.32 = 0.64 H100-hours for base training. Adding engineering/debugging margin for custom Colored-MNIST construction and NTK implementation: 4 H100-hours total.", "h100_band": "0-8", "h100_recomputed_hours": 3.84, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 4.0, "h100_hours_adjudicated": false, "selection_band": "0-8", "match_target": {"config": "2-layer ReLU FC net, width 256, SGD lr=1e-3 wd=1e-4 batch=1000, zero-init output layer, Colored-MNIST (train: 95%/5% color-label correlation; test: 50/50 balanced), empirical NTK at step 1000 from 500 random train inputs", "metric": "NTK availability A(K,g) = g^T K g / (||g||^2 ||K||_F) for shortcut (color) vs ground-truth label vectors", "value": "availability(shortcut) > availability(ground-truth) at step 1000; the qualitative ordering is the falsifiable core of the eigenvalue/spectral-bias claim", "scope": "Colored-MNIST shortcut-learning experiment, empirical finite-width NTK", "match_bar_kind": "direction"}, "split": "eval"}
-{"custom_id": "2506.07104", "central_claim": "REO-RL reduces the Reasoning Efficiency Gap (REG) by at least 50% across all evaluated LRMs while maintaining accuracy, and matches Qwen3-4B/8B efficiency frontiers under a 16K token budget with minimal accuracy loss.", "claim_evidence": "Abstract: 'REO-RL significantly improves reasoning efficiency, reducing REG by >=50% across all evaluated LRMs and matching Qwen3-4B/8B efficiency frontiers under a 16K token budget with minimal accuracy loss.' Table 1 (main_results_v3.tex): REO-RL (Exp) Average REG reduction of 57.6% for 1.5B, 55.9% for 7B, 65.2% for Qwen3-4B, 75.9% for Qwen3-8B compared to base LRMs.", "paper_kind": "empirical", "mre_config": "Run inference with Qwen3-4B (publicly available on HuggingFace) on AIME 2024 benchmark with 32 sampled responses (T=0.6, top_p=0.95) under varying token budgets (512-16K). Use released efficiency frontiers from GitHub data/frontiers/ to compute REG metric, comparing the model's budget-accuracy curve against the frontier.", "verified_links": {"paper_or_project": ["https://github.com/samjia2000/Optimal-Reasoning-Efficiency", "https://arxiv.org/abs/2506.07104"], "code": ["https://github.com/samjia2000/Optimal-Reasoning-Efficiency (REO-RL implementation marked TODO, only frontier data released)", "https://github.com/areal-project/AReaL (RL training framework used for experiments)"], "dataset": ["https://huggingface.co/datasets/AI-MO/NuminaMath-AIME (AIME benchmark)", "https://huggingface.co/datasets/deepseek-ai/DeepSeek-Math (Minerva Math source)"], "weights": ["https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B (base model)", "https://huggingface.co/Qwen/Qwen3-4B (base model)"]}, "signals": {"code_available": {"value": false, "verification": "tool_verified", "evidence": "The official GitHub repository is public but currently contains only README.md and data/frontiers files. Its README lists 'Evaluation Scripts for REG' and 'REO-RL Implementation' under TODO, and the file tree shows only data/frontiers JSON/JSONL files, not runnable REO-RL training or evaluation code. Papers With Code links the same official repository. Therefore MRE-relevant runnable code is not available."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE evaluation benchmark AMC 2023 is a standard public math contest benchmark used widely in reasoning-model evaluation. For training, the paper states a mixture sourced from DeepScaleR and AReaL; live web verification found a public DeepScaleR Preview Dataset page and the AReaL-boba model card documents its open-source AReaL-boba-106k training dataset construction from public math resources. However, the paper-specific 135k processed mixture/removal of duplicates was not verified as a released exact artifact; for the chosen MRE, the public source datasets and standard AMC evaluation are sufficient to attempt the experiment but exact training-data matching may require reconstruction."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searches for the paper title, arXiv ID, official GitHub repository, REO-RL, Optimal-Reasoning-Efficiency on GitHub/Papers With Code/Hugging Face found no released REO-RL 1.5B/7B/4B/8B checkpoints. The official repository has no releases and no model files; Hugging Face searches did not identify author-released REO-RL weights. Public starting-point/base weights such as AReaL-boba-RL-7B exist, but the MRE-relevant trained REO-RL checkpoint is not available."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "The MRE's anchor evaluation dataset, AMC 2023, is a widely used public mathematical-reasoning benchmark/competition set. The paper's other evaluation benchmarks, AIME 2024/2025 and Minerva Math, are also standard public math reasoning benchmarks. The exact training mixture is derived/custom, but dataset standardness here is scoped to the MRE evaluation benchmark used to test the central empirical claim."}}, "agent_task": "Reproduce the MRE by: (1) Download the released efficiency frontiers from https://github.com/samjia2000/Optimal-Reasoning-Efficiency/tree/main/data/frontiers; (2) Load Qwen3-4B from HuggingFace (https://hf.co/Qwen/Qwen3-4B-Base); (3) Run inference on AIME 2024 benchmark with 32 samples per problem (temperature=0.6, top_p=0.95, max_length=16K); (4) Compute the model's budget-accuracy curve at token budgets {64, 128, ..., 16384}; (5) Calculate REG using the released frontier data and the formula in Eq. eq:reg (sum of accuracy gaps across all token budgets). Verify that REG is measurable and positive, confirming the efficiency frontier represents an upper bound.", "h100_estimate": {"hours": 6.0, "basis_kind": "derived_from_config", "gpu_count": 1, "gpu_type": "H100 SXM", "wallclock_hours": 6.0, "h100_equivalent_multiplier": 1.0, "basis": "Paper's own mre_config specifies eval-only: run publicly released Qwen3-4B checkpoint on AIME2024 (32 samples/problem x6 token budgets 512-16K, ~30 problems) and compare against released frontier data -- no REO-RL retraining needed. weights_available=false refers only to the trained REO-RL checkpoints, not the base Qwen3-4B used in the MRE. Fixed scope_error: prior basis (16h) and Agent1's revision (3600h) both reasoned from REO-RL training scale (64x H800, 4000 GPU-hours) contradicting the paper's own eval-only MRE definition. Recomputed as ~5760 sampled generations up to 16K tokens on 1 GPU, ~6h wallclock."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 6.0, "h100_estimate_basis": "derived_from_config: Paper's own mre_config specifies eval-only: run publicly released Qwen3-4B checkpoint on AIME2024 (32 samples/problem x6 token budgets 512-16K, ~30 problems) and compare against released frontier data -- no REO-RL retraining needed. weights_available=false refers only to the trained REO-RL checkpoints, not the base Qwen3-4B used in the MRE. Fixed scope_error: prior basis (16h) and Agent1's revision (3600h) both reasoned from REO-RL training scale (64x H800, 4000 GPU-hours) contradicting the paper's own eval-only MRE definition. Recomputed as ~5760 sampled generations up to 16K tokens on 1 GPU, ~6h wallclock.", "h100_band": "0-8", "h100_recomputed_hours": 6.0, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 6.0, "h100_hours_adjudicated": false, "selection_band": "0-8", "split": "eval", "match_target": {"config": "Qwen3-4B evaluation on AIME 2024, 32 responses per problem, temperature=0.6, max_length=16K, using released efficiency frontiers for REG computation", "metric": "REG (Reasoning Efficiency Gap) reduction percentage", "value": "65.2%", "scope": "Average across AIME 2024, AIME 2025, AMC 2023, Minerva Math benchmarks", "match_bar_kind": "direction"}}
 {"custom_id": "2505.16927", "central_claim": "STaPLe enables smaller language models (7-8B parameters) to self-improve by discovering and learning latent principles for self-correction, achieving +8-10% in AlpacaEval win-rate, +0.3 on MT-Bench, and +19-23% principle-following win-rate on IFEval.", "claim_evidence": "Abstract states 'bootstrapping our algorithm over multiple iterations enables smaller language models (7-8B parameters) to self-improve, achieving +8-10% in AlpacaEval win-rate, an average of +0.3 on MT-Bench, and +19-23% in principle-following win-rate on IFEval.' Table 1 reports IFEval win-rates of 65.6% (STaPLe Iter 1, Llama-3.1-8B) vs 55.5% (STaR baseline), demonstrating the core self-improvement claim.", "paper_kind": "empirical", "mre_config": "STaPLe Iteration 1 on Llama-3.1-8B-Instruct using 10k samples from standard mining corpus (HH-RLHF, UltraFeedback, or TL;DR from HuggingFace) with N=16 principle samples, Rouge-L F1 similarity threshold 0.4, temperature 0.7, full supervised fine-tuning for 3 epochs at learning rate 1e-6, sequence length 4096, evaluated on IFEval (541 prompts) using Prometheus-8x7B-v2.0 as judge for principle-following win-rate.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2505.16927", "https://openreview.net/attachment/f4cff24f37db1093806eb3e1070b7db1220d1ccb.zip"], "code": [], "dataset": ["https://huggingface.co/datasets/Anthropic/hh-rlhf", "https://huggingface.co/datasets/openbmb/UltraFeedback", "https://huggingface.co/datasets/tatsu-lab/alpaca_eval"], "weights": ["https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct", "https://huggingface.co/ibm-granite/granite-3.1-8b-instruct", "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct"]}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Web searches for the exact title, arXiv ID 2505.16927, \"Self-Taught Principle Learning\", \"STaPLe\" with GitHub/Hugging Face scoping, author name variants, and prompt-file phrases did not find a public runnable implementation repository. The paper and OpenReview PDF state that code will be publicly released, but that is only a promise. The bundled OpenReview supplement contains LaTeX source, figures, tables, constitutions, example prompts, and an algorithm writeup, but no runnable training/evaluation/data-processing scripts or configs sufficient to execute the MRE."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "Mining corpus datasets verified on Hugging Face: Anthropic/hh-rlhf (MIT license, 161k preference pairs), openbmb/UltraFeedback (MIT license, 64k prompts). Evaluation dataset verified: tatsu-lab/alpaca_eval (CC-BY-NC-4.0 license, 805 samples). Paper appendix documents these datasets and their construction."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "The base Qwen2.5-7B-Instruct model is publicly available on Hugging Face and is relevant as the starting checkpoint, but searches of Hugging Face, GitHub, the exact title, arXiv ID, STaPLe variants, and author/project terms did not find released STaPLe fine-tuned checkpoints/adapters for the one-iteration Qwen MRE or any final STaPLe models. The OpenReview supplement manifest contains no model, adapter, checkpoint, or safetensors/bin weight files."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "Mining corpus consists of established public datasets: HH-RLHF (Anthropic), UltraFeedback (openbmb), TL;DR (OpenAI), HotpotQA. Evaluation benchmarks: AlpacaEval-2.0-LC (tatsu-lab), MT-Bench, IFEval. All are widely recognized public benchmarks with documented download paths on HuggingFace and GitHub."}}, "agent_task": "Reproduce STaPLe (Self-Taught Principle Learning) Iteration 1 to verify self-improvement claim: (1) Load Llama-3.1-8B-Instruct from HuggingFace (meta-llama/Llama-3.1-8B-Instruct), (2) Prepare mining corpus sample (~10k prompts) by downloading from HuggingFace datasets (Anthropic/hh-rlhf, openbmb/UltraFeedback, or openai/summarize_from_feedback) and filtering to samples with gold/chosen responses, (3) Implement E-step: for each prompt, sample initial response from model, mine N=16 principles using principle-mining prompt (Prompt: {prompt}\\n\\nHere is the previous response: {initial_response}\\n\\nHere is the target response: {gold_response}\\n\\nIdentify a high-level principle... New Principle:), generate critiques using critique-generation prompt, sample N refined responses, select best via Rouge-L F1 similarity to gold (must exceed initial response), (4) Implement M-step: supervised fine-tune on (prompt, initial_response, principle, refined_response) tuples for 3 epochs at LR 1e-6, seq_len 4096, (5) Evaluate on IFEval (541 prompts) using Prometheus-8x7B-v2.0 judge: for each prompt, compare model's initial vs refined response on principle-following, report win-rate. Target: achieve >60% IFEval win-rate demonstrating the model learned intrinsic self-correction.", "h100_estimate": {"hours": 9.6, "basis_kind": "derived_from_config", "gpu_count": 8, "gpu_type": "H100 SXM", "wallclock_hours": 1.2, "h100_equivalent_multiplier": 1, "basis": "Weights not released -> MRE is smallest training run: Iteration 1 on Llama-3.1-8B, 10k samples per paper's own mre_config (not 50k as prior basis used). E-step scaled from paper's 2.5M-pass/4hr on 4xH100 (16 GPU-hr) by 10k/50k -> 3.2 GPU-hr. M-step 3-epoch SFT on ~5.6k successful samples (28k*10k/50k) on 8xH100 scaled from 32 GPU-hr -> 6.4 GPU-hr. Total 9.6 GPU-hr, expressed as 8 GPUsx1.2hrx1.0."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 9.6, "h100_estimate_basis": "derived_from_config: Weights not released -> MRE is smallest training run: Iteration 1 on Llama-3.1-8B, 10k samples per paper's own mre_config (not 50k as prior basis used). E-step scaled from paper's 2.5M-pass/4hr on 4xH100 (16 GPU-hr) by 10k/50k -> 3.2 GPU-hr. M-step 3-epoch SFT on ~5.6k successful samples (28k*10k/50k) on 8xH100 scaled from 32 GPU-hr -> 6.4 GPU-hr. Total 9.6 GPU-hr, expressed as 8 GPUsx1.2hrx1.0.", "h100_band": "8-32", "h100_recomputed_hours": 9.6, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 9.6, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "STaPLe Iteration 1, Llama-3.1-8B-Instruct, constrained (clustered) version, N=16, Rouge-L similarity, 3 epochs SFT on ~5k accepted samples", "metric": "IFEval principle-following win-rate", "value": "65.6%", "scope": "IFEval benchmark (541 prompts) with Prometheus-8x7B-v2.0 judge comparing refined response to initial policy response", "match_bar_kind": "threshold"}}
 {"custom_id": "2502.08924", "central_claim": "A boosting-style algorithm that combines synthetic data with weak (beta>0) exogenously-labeled data converges to an optimal LLM, with focusing labeling resources on the most challenging examples leading to improved performance over filter-only approaches like STaR/ReST.", "claim_evidence": "Abstract: 'we show that dynamically focusing labeling resources on the most challenging examples --- in much the same way that boosting focuses the efforts of the weak learner --- leads to improved performance.' Table 1: Boosting (A) achieves GSM8K test accuracy 0.456 vs Filter only at 0.393 and Do nothing at 0.129.", "paper_kind": "empirical", "mre_config": "Gemma 2 2B PT with boosting-style algorithm (Algorithm 2), 5 rounds, GSM8K dataset (7000 train), 3-shot prompting, k=8 synthetic samples per prompt, alpha=1/3, batch size 64, 330 steps per round, standard cross-entropy loss fine-tuning. Comparing Boosting vs Filter-only (STaR/ReST-like) and Do nothing baselines.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2502.08924"], "code": [], "dataset": ["https://hf.co/datasets/openai/gsm8k", "https://hf.co/datasets/google-research-datasets/mbpp"], "weights": ["https://hf.co/google/gemma-2-2b-it"]}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No MRE-relevant first-party code was found. Web searches for the full title, arXiv ID 2502.08924, 'Escaping Collapse' with GitHub/code/Gemma/GSM8K, and weak-data boosting variants found arXiv/Google Research/OpenReview/Paper aggregator pages but no public repository. The OpenReview paper PDF explicitly answers the reproducibility checklist 'No' and states that the experiments use open resources but 'the code itself is not open sourced.' The provided OpenReview supplement status is missing."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE uses GSM8K, a public standard benchmark. Hugging Face hosts openai/gsm8k, and TensorFlow Datasets lists gsm8k as a dataset of 8.5K grade-school math word problems with a homepage pointing to OpenAI's grade-school-math repository. The paper's 3-shot prompt template is included in the paper text, so no unreleased custom dataset is required for the selected GSM8K MRE."}, "weights_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE initializes and labels with Gemma 2 2B PT. The Hugging Face page google/gemma-2-2b exists and provides Transformers loading instructions such as AutoModelForCausalLM.from_pretrained, but access requires agreeing to Google's usage license. No paper-specific fine-tuned checkpoints were found, so this signal counts only the publicly available base weights needed to run the training MRE from scratch."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "GSM8K is a widely recognized public ML benchmark for grade-school math reasoning; it is available through Hugging Face openai/gsm8k and TensorFlow Datasets."}}, "agent_task": "Reproduce the paper's boosting-style LLM training algorithm (Algorithm 2). Steps: (1) Load Gemma 2 2B PT from HuggingFace (google/gemma-2-2b). (2) Download GSM8K dataset from HuggingFace. (3) Implement boosting-style training loop for 5 rounds: (a) generate k=8 synthetic responses per prompt using current model at temperature, (b) filter responses using ground-truth quality function to partition into high-quality S+ and low-quality P- sets, (c) apply weak labeler to P- prompts (simulate by sampling from same model, returning correct responses for beta fraction), (d) mix synthetic+weak data with alpha=1/3 weight on weak data, (e) fine-tune next round initialized from previous checkpoint. (4) Run Filter-only baseline (alpha=0) for comparison. (5) Measure test accuracy (greedy decoding) on GSM8K test set. Target: Boosting achieves >= 0.456 test accuracy vs Filter-only at ~0.393.", "h100_estimate": {"hours": 12.0, "basis_kind": "derived_from_config", "gpu_count": 1, "gpu_type": "H100 SXM", "wallclock_hours": 12, "h100_equivalent_multiplier": 1.0, "basis": "weights_available=true is only the released base Gemma-2-2B checkpoint, not a fine-tuned artifact for the boosting-vs-filter claim, so the MRE stays a smallest_training_run (5 rounds x 330 steps + k=8 synthetic gens over GSM8K), not eval-only. Original used 8xH100 x1.5h wallclock for a 2B-param fine-tune, which is implausible over-provisioning for this model size. Reallocated to 1xH100 x 12h, preserving the same total 12 GPU-hours. Non-compute debugging/engineering padding language dropped (was mentioned but not arithmetically applied)."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 12.0, "h100_estimate_basis": "derived_from_config: weights_available=true is only the released base Gemma-2-2B checkpoint, not a fine-tuned artifact for the boosting-vs-filter claim, so the MRE stays a smallest_training_run (5 rounds x 330 steps + k=8 synthetic gens over GSM8K), not eval-only. Original used 8xH100 x1.5h wallclock for a 2B-param fine-tune, which is implausible over-provisioning for this model size. Reallocated to 1xH100 x 12h, preserving the same total 12 GPU-hours. Non-compute debugging/engineering padding language dropped (was mentioned but not arithmetically applied).", "h100_band": "8-32", "h100_recomputed_hours": 12.0, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 2, "tier": "Hard", "audited_h100_hours": 12.0, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "Gemma 2 2B PT, Boosting (A), 5 rounds, 3-shot, GSM8K train=7000, k=8 synthetic samples per prompt, alpha=1/3, batch_size=64, 330 steps/round", "metric": "test accuracy (greedy decoding)", "value": "0.456", "scope": "GSM8K test set", "match_bar_kind": "point_estimate"}}
 {"custom_id": "2511.01463", "central_claim": "HMVLM with MoE LoRA and zero expert preserves foundation model knowledge while achieving strong text-to-motion performance on HumanML3D (R-precision Top-3 of 0.785).", "claim_evidence": "Abstract states 'Experiments show that our method effectively alleviates knowledge forgetting during instruction-tuning and achieves remarkable performance across diverse human motion downstream tasks.' Table 1 shows 'Ours (single task)' achieves R precision Top-3 of 0.785 on HumanML3D test set.", "paper_kind": "empirical", "mre_config": "Single-task MoE LoRA fine-tuning with 5 experts (including zero expert), Vicuna-7b-v1.5 base model, on HumanML3D text-to-motion task using single A800 80G GPU, batch size 32, learning rate 3×10^-3 with cosine annealing.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2511.01463"], "code": [], "dataset": ["HumanML3D is a standard text-to-motion benchmark from Guo et al. 2022, available via the original project page at UMass Amherst"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searched live web with exact title, acronym, arXiv ID, GitHub-scoped, Hugging Face-scoped, and method/dataset variants such as \"HMVLM github\", \"Human Motion-Vision-Lanuage GitHub\", \"HMVLM Hugging Face motion\", and site:github.com queries. Results found arXiv/OpenReview mirrors and summaries but no first-party HMVLM code repository. The bundled OpenReview supplement manifest contains only supplementary_video.mp4, which is not runnable code/config/script for the MRE."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE uses HumanML3D. A live web search found and verified the public HumanML3D GitHub repository at EricGuo5513/HumanML3D, whose README/search snippet describes it as a large 3D human motion-language dataset and provides scripts to reproduce the dataset from AMASS. This is sufficient public dataset evidence for the HumanML3D T2M MRE, subject to upstream AMASS/SMPL licensing and setup."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searched live web for HMVLM model/checkpoint/weights via exact title, acronym, arXiv ID, GitHub, and Hugging Face variants. No downloadable HMVLM checkpoints, LoRA adapters, tokenizer weights, or gating-network weights were found. The OpenReview supplement contains only supplementary_video.mp4 and no checkpoint/model files."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "HumanML3D is a widely used public text-to-motion benchmark; live web results include the official HumanML3D repository and secondary dataset pages describing it as a standard evaluation dataset for text-to-motion models with common FID, R-precision, and Diversity metrics. The MRE does not require a newly authored custom benchmark beyond the paper's instruction formatting/training pipeline."}}, "agent_task": "Reproduce HMVLM text-to-motion generation using MoE LoRA fine-tuning on HumanML3D: (1) Download HumanML3D dataset from EricGuo5513/HumanML3D, (2) Implement MoE LoRA architecture with 5 experts (4 task experts + 1 zero expert), gating network (2-layer MLP, hidden dim 512), and zero expert with zero-initialized non-trainable LoRA matrices, (3) Implement body-part-based motion tokenizer with 5 body parts, embedding dim 512, codebook size K=512, temporal compression l=4, (4) Fine-tune Vicuna-7B-v1.5 base model with MoE LoRA rank 8, batch size 32, lr 3e-3, cosine annealing, (5) Evaluate using R-precision (Top-1/2/3), FID, MM-Distance, and Diversity metrics on HumanML3D test set. Note: No released code or weights - requires full implementation from scratch.", "h100_estimate": {"hours": 12.8, "basis_kind": "derived_from_config", "gpu_count": 1, "gpu_type": "A800", "wallclock_hours": 40, "h100_equivalent_multiplier": 0.32, "basis": "Weights not released -> MRE is smallest single training run, not full joint instruction-tuning. Original 38.4 used 120h reported for JOINT training across 3 tasks (T2M+pose+video), a scope error. Single-task T2M scaled to ~1/3 of joint schedule: 1 GPU x 40h x 0.32 (A800) = 12.8 H100-hours. Same GPU/rate as paper, corrected scope only."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 12.8, "h100_estimate_basis": "derived_from_config: Weights not released -> MRE is smallest single training run, not full joint instruction-tuning. Original 38.4 used 120h reported for JOINT training across 3 tasks (T2M+pose+video), a scope error. Single-task T2M scaled to ~1/3 of joint schedule: 1 GPU x 40h x 0.32 (A800) = 12.8 H100-hours. Same GPU/rate as paper, corrected scope only.", "h100_band": "8-32", "h100_recomputed_hours": 12.8, "h100_arithmetic_mismatch": false, "h100_needs_human_review": true, "score": 3, "tier": "Hard", "audited_h100_hours": 12.8, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "MoE LoRA (5 experts, rank 8), Vicuna-7B-v1.5, single-task T2M fine-tuning on HumanML3D", "metric": "R precision Top-3", "value": "0.785", "scope": "HumanML3D test set", "match_bar_kind": "point_estimate"}}
 {"custom_id": "2510.25529", "central_claim": "MoGE (Modelic Generative Exploration) effectively bridges exploration and policy learning in off-policy RL through diffusion-based critical state generation and a one-step imagination world model, leading to remarkable gains in both sample efficiency and performance across complex control tasks.", "claim_evidence": "Abstract states 'MoGE effectively bridges exploration and policy learning, leading to remarkable gains in both sample efficiency and performance across complex control tasks.' Table 1 shows DSAC+MoGE achieves 891.7 ± 19.1 TAR on Humanoid-walk vs DSAC's 146.5 ± 60.9 (+508.6% improvement), and 817.7 ± 11.4 average DMC TAR vs DSAC's 568.5 ± 42.7 (+43.8% improvement).", "paper_kind": "empirical", "mre_config": "DSAC+MoGE on DeepMind Control Suite Humanoid-walk task, trained for 1.5M steps with 3 random seeds, compared against DSAC baseline, using standard DMC Gym environments via gymnasium/dmcgym.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2510.25529"], "code": ["https://github.com/WangLK-Franklin/MoGE"], "dataset": ["https://github.com/deepmind/dm_control", "https://github.com/Farama-Foundation/Gymnasium"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_verified", "evidence": "The paper points to https://github.com/WangLK-Franklin/MoGE, and the live GitHub page is public but contains only README.md; the README says 'The code will be released later,' with no runnable training scripts, configs, or implementation files for the MRE. Searches for the title, arXiv ID, MoGE, Modelic Generative Exploration, and WangLK-Franklin found no other relevant public implementation. OpenReview supplement status in the prompt is missing."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE uses standard simulator benchmarks rather than a custom dataset. DeepMind Control Suite is publicly available through the google-deepmind/dm_control repository, whose page describes dm_control.suite as Python reinforcement-learning environments powered by MuJoCo. The paper's alternative Gym MuJoCo tasks are also public through Gymnasium documentation listing Ant, HalfCheetah, Humanoid, Swimmer, and Walker2D MuJoCo environments."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No MoGE/DSAC+MoGE checkpoints or learned model weights were found in the GitHub repository, which has no releases and only a README. Artifact-specific searches on GitHub/Hugging Face for MoGE, Modelic Generative Exploration, WangLK-Franklin MoGE, and checkpoint/weights found unrelated MoGe geometry-estimation models but no MRE-relevant RL checkpoints."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "DeepMind Control Suite and Gym/Gymnasium MuJoCo locomotion environments are widely recognized public RL benchmarks used directly through public simulator packages; the MRE does not require author-created labels, custom splits, or private data."}}, "agent_task": "Reproduce the MoGE method on DMC Humanoid-walk task: (1) Implement MoGE from scratch since official code is not released (placeholder repo only), (2) Install gymnasium and dm_control packages for DeepMind Control Suite environments, (3) Implement DSAC algorithm as base off-policy RL method, (4) Implement conditional diffusion model for critical state generation with utility guidance (policy entropy or TD error), (5) Implement one-step world model with encoder, dynamics predictor, reward predictor, and termination predictor, (6) Train DSAC+MoGE for 1.5M steps on Humanoid-walk using paper hyperparameters (batch_size=1024, lr=1e-4, guidance_scale=1.0, diffusion_steps=100, lambda=0.2, beta1=0.5, beta2=0.1), (7) Evaluate and verify TAR reaches ~891.7 ± 19.1 vs DSAC baseline ~146.5 ± 60.9.", "h100_estimate": {"hours": 13.48, "basis_kind": "paper_reported", "gpu_count": 1, "gpu_type": "RTX 4090", "wallclock_hours": 42.13, "h100_equivalent_multiplier": 0.32, "basis": "Weights not released -> MRE is smallest training run: DSAC+MoGE on Humanoid task, 1.5M steps, single seed. Paper Table 3 (supplement) reports walltime 151657±395s (=42.13h) on 1 RTX 4090. RTX 4090 not in canonical table; using A100 80GB (0.32) as closest throughput analogue. H100-hours = 1*42.13*0.32 = 13.48."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 13.48, "h100_estimate_basis": "paper_reported: Weights not released -> MRE is smallest training run: DSAC+MoGE on Humanoid task, 1.5M steps, single seed. Paper Table 3 (supplement) reports walltime 151657±395s (=42.13h) on 1 RTX 4090. RTX 4090 not in canonical table; using A100 80GB (0.32) as closest throughput analogue. H100-hours = 1*42.13*0.32 = 13.48.", "h100_band": "8-32", "h100_recomputed_hours": 13.4816, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 13.48, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "DSAC+MoGE on Humanoid-walk (DMC), 1.5M steps, batch_size=1024, learning_rate=1e-4, guidance_scale=1.0, diffusion_steps=100, mix_ratio lambda=0.2", "metric": "Total Average Return (TAR)", "value": "891.7 ± 19.1", "scope": "DeepMind Control Suite Humanoid-walk task, averaged over 3 seeds", "match_bar_kind": "point_estimate"}}
 {"custom_id": "2506.00070", "central_claim": "Robot-R1 uses reinforcement learning (GRPO) to train LVLMs for embodied reasoning in robotics, where models trained with Robot-R1 outperform SFT methods and a 7B Robot-R1 model surpasses GPT-4o on reasoning tasks related to low-level action control (movement and spatial reasoning).", "claim_evidence": "Abstract: 'models trained with Robot-R1 outperform SFT methods on embodied reasoning tasks. Despite having only 7B parameters, Robot-R1 even surpasses GPT-4o on reasoning tasks related to low-level action control, such as spatial and movement reasoning.' Table 2 (tab:LowLevelControl_result): Robot-R1 achieves 0.76 avg movement (vs GPT-4o's 0.72) and 1.51 avg spatial (vs GPT-4o's 1.43) on Robot-R1 Bench.", "paper_kind": "empirical", "mre_config": "Train Qwen2.5-7B-VL-Instruct with GRPO on RLBench-generated MCQA data for 5 tasks (pick_up_cup, push_button, put_rubbish_in_bin, phone_on_base, take_lid_off_saucepan). Use 50 demonstrations per task, frame interval t=10, ~7.5K QA pairs across waypoint prediction, current state prediction, and movement prediction. Training: batch size 128, 5 epochs, learning rate 1e-6, 5 rollout samples per prompt, temperature 1.0. Evaluate on Robot-R1 Bench spatial reasoning metric.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2506.00070"], "code": ["https://github.com/hiyouga/EasyR1", "https://github.com/2U1/Qwen2-VL-Finetune"], "dataset": ["https://github.com/stepjam/RLBench"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searched exact title, arXiv ID, Robot-R1/Robot R1/Robot_R1 variants with github/code/official, author handle kingdy2002, Robot-R1 Bench, and Hugging Face scoped queries. The author homepage lists the Robot-R1 paper but no code link. Search found only generic third-party infrastructure named in the paper, EasyR1 and Qwen2-VL-Finetune, not Robot-R1-specific data-generation, training, evaluation, prompt, or benchmark code. OpenReview supplement status in the input is missing."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "Verified RLBench at https://github.com/stepjam/RLBench is publicly available with full installation instructions, task definitions, and demonstration generation tools. The paper uses RLBench for generating expert demonstrations and extracting waypoints. However, the paper's processed MCQA training dataset with specific question-answer pairs and metadata is NOT released separately; it must be regenerated from RLBench using the paper's methodology."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searched Robot-R1 with Hugging Face/model/checkpoint/weights and GitHub variants. No downloadable Robot-R1-trained Qwen2.5-VL-7B checkpoint, LoRA adapter, or model card was found; only the base model and unrelated/generic resources appeared. OpenReview supplement status is missing and contains no checkpoint evidence."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "RLBench is a well-established public benchmark for robot learning (james2020rlbench, IEEE RAL 2020), widely used in the robotics community for imitation learning and RL evaluation with 100+ manipulation tasks."}}, "agent_task": "Reproduce the Robot-R1 MRE by: (1) Installing RLBench (pip install git+https://github.com/stepjam/RLBench.git) and generating training demonstrations for 5 tasks with 50 episodes each at 224x224 resolution, (2) Extracting waypoints using ARM repository, (3) Constructing MCQA training data using prompt templates from the paper appendix (waypoint prediction, current state prediction, movement prediction MCQA), (4) Fine-tuning Qwen2.5-7B-VL-Instruct using GRPO with EasyR1 framework (learning rate 1e-6, batch size 128, 5 epochs, 5 rollout samples per prompt, temperature 1.0, rollout batch size 512), (5) Evaluating spatial reasoning on Robot-R1 Bench and verifying improvement over base model (1.40 spatial score).", "h100_estimate": {"hours": 15.36, "basis_kind": "paper_reported", "gpu_count": 4, "gpu_type": "A100 80GB", "wallclock_hours": 12, "h100_equivalent_multiplier": 0.32, "basis": "Paper states 4x A100 80GB, ~12h wallclock for 5-epoch GRPO training on the 7.5K prompts; this is the only/smallest training config described (weights not released, so MRE = smallest training run, not eval-only). 4*12*0.32=15.36 H100-hrs. Arithmetic, multiplier, and scope verified correct; no double-counting or padding found."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 15.36, "h100_estimate_basis": "paper_reported: Paper states 4x A100 80GB, ~12h wallclock for 5-epoch GRPO training on the 7.5K prompts; this is the only/smallest training config described (weights not released, so MRE = smallest training run, not eval-only). 4*12*0.32=15.36 H100-hrs. Arithmetic, multiplier, and scope verified correct; no double-counting or padding found.", "h100_band": "8-32", "h100_recomputed_hours": 15.36, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 15.36, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "Qwen2.5-VL-7B-Instruct + GRPO training on RLBench MCQA tasks for 5 epochs, full fine-tuning with batch size 128, lr 1e-6, temperature 1.0, 5 samples per prompt", "metric": "Spatial reasoning score (GPT-4o-judged, 0-3 scale)", "value": "1.51", "scope": "Robot-R1 Bench - spatial reasoning task (in-distribution + out-of-distribution average)", "match_bar_kind": "point_estimate"}}
 {"custom_id": "2507.02064", "central_claim": "Hippocampal place cells auto-associate sensory inputs with grid cell patterns, enabling cue-triggered goal retrieval, grid-based path planning that generalizes to novel routes, and reconstruction of sensory experiences along planned trajectories.", "claim_evidence": "Abstract states 'place cells autoassociate sensory inputs with grid cell patterns, allowing sensory cues to trigger recall of goal-location grid patterns.' Section 3 and Figure 2/3 demonstrate recall of goal locations from partial sensory cues with L2 distance approaching zero over time.", "paper_kind": "empirical", "mre_config": "HC-MEC RNN with 288 grid cells (6 modules x 48 cells), 32 speed cells, 32 direction cells, 256 SMCs, 512 HPCs. Trained on random traversal in 100x100cm arena with masking ratio r_mask=0.5, for 50 epochs, batch_size=128, 1000 steps per trajectory. Recall tested by querying with SMC responses at fixed locations for 5 seconds (100 timesteps), decoding positions via nearest neighbor on ratemaps.", "verified_links": {"paper_or_project": ["https://zhaozewang.github.io/remi"], "code": ["https://github.com/grasp-lyrl/btnk_mae"], "dataset": ["https://github.com/facebookresearch/habitat-sim"], "weights": ["https://huggingface.co/zhaozewang56/btnk_mae"]}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "The REMI project page has a 'Code (Coming Soon)' link, but it currently opens zhaozewang/place_cells_episodic_rnn, whose README identifies it as the official implementation of the earlier NeurIPS 2024 'Time Makes Space' paper, not REMI; it has a demo and checkpoint for that older recurrent autoencoder. The separate RatatouGym repo is public and supplies a spatial-navigation simulator, and the btnk_mae repo supplies the vision encoder code, but neither contains the REMI HC-MEC plus planner training/evaluation code needed for the synthetic MRE. Searches for the exact title, arXiv ID, REMI HC-MEC GitHub, and zhaozewang REMI GitHub found no public REMI reproduction repository beyond these supporting/irrelevant artifacts."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE uses synthetic spatial navigation/cell-response data generated from the arena, trajectory, and sensory models rather than a fixed private dataset. RatatouGym is publicly documented as a gymnasium environment for simulating environmental geometries, sensory systems modeled through firing tunings, and traversal behaviors, and its GitHub repository is public with install instructions and source files. Thus the synthetic data source for the MRE is available, although exact author random seeds or generated trial files are not separately released."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No downloadable REMI HC-MEC or planner checkpoints were found. The project page's code link opens an older place_cells_episodic_rnn repository with a ckpts folder and latest.pth for the NeurIPS 2024 Time Makes Space project, not the REMI MRE. The btnk_mae repository provides Hugging Face pretrained BtnkMAE weights, but those are for the optional Habitat visual extension, not the synthetic HC-MEC/planner MRE. Searches for REMI checkpoints/weights/model files did not find MRE-relevant weights."}, "dataset_is_standard": {"value": false, "verification": "tool_verified", "evidence": "The MRE data are author-defined synthetic simulator outputs: custom 100 cm x 100 cm arena trajectories and simulated SMC/GC/speed/direction firing responses produced by RatatouGym-style procedures. This is not a widely recognized fixed benchmark such as ImageNet, CIFAR, GLUE, or Habitat challenge splits; it depends on paper-specific simulation choices."}}, "agent_task": "Implement the HC-MEC RNN model from the paper: a single-layer RNN with grid cells (6 modules x 48 cells, spatial scale factor sqrt(e), initial scale 30cm), speed cells (32), direction cells (32), SMCs (256, Gaussian white noise smoothed), and HPCs (512, no direct input/output). Train on random traversal trajectories in 100x100cm arena with masking (r_mask=0.5) for 50 epochs. After training, test cue-triggered recall by querying with SMC responses at fixed locations for 5 seconds, decode positions via nearest neighbor on aggregated ratemaps, and verify L2 distance approaches target location.", "h100_estimate": {"hours": 16, "basis_kind": "derived_from_config", "gpu_count": 1, "gpu_type": "A100 80GB", "wallclock_hours": 50, "h100_equivalent_multiplier": 0.32, "basis": "Paper specifies training HC-MEC model for 50 epochs with 1000 steps per trajectory, batch size 128, total hidden units 1120. Based on paper's supplements, single GPU training with AdamW optimizer. Estimated ~50 hours on 1x A100 80GB for the recall experiment MRE. Converted to H100-hours: 1 * 50 * 0.32 = 16 H100-hours. Note: Main REMI code is not released; engineering/debugging time to implement the model from scratch would be additional and substantial."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 16.0, "h100_estimate_basis": "derived_from_config: Paper specifies training HC-MEC model for 50 epochs with 1000 steps per trajectory, batch size 128, total hidden units 1120. Based on paper's supplements, single GPU training with AdamW optimizer. Estimated ~50 hours on 1x A100 80GB for the recall experiment MRE. Converted to H100-hours: 1 * 50 * 0.32 = 16 H100-hours. Note: Main REMI code is not released; engineering/debugging time to implement the model from scratch would be additional and substantial.", "h100_band": "8-32", "h100_recomputed_hours": 16.0, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 16.0, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "HC-MEC RNN, r_mask=0.5, recall task with SMC query at fixed location for 5s", "metric": "L2 distance from decoded position to target position (cm)", "value": "~0 cm after convergence (trajectories converge to goal representation)", "scope": "100x100cm simulated arena, nearest neighbor decoding on aggregated ratemaps", "match_bar_kind": "point_estimate"}}
-{"custom_id": "2505.19516", "central_claim": "DiffE2E achieves state-of-the-art driving score of 83 on CARLA Longest6 and PDMS of 92.7 on NAVSIM by combining diffusion-based trajectory generation with explicit supervised learning in a hybrid decoder architecture.", "claim_evidence": "Abstract: 'Experiments demonstrate that DiffE2E achieves state-of-the-art performance in both CARLA closed-loop evaluations and NAVSIM benchmarks.' Table 1 (CARLA Longest6): DiffE2E achieves DS=83 (best), IS=0.86 (best), RC=96. Table 2 (NAVSIM Navtest): DiffE2E achieves PDMS=92.7 (best overall).", "paper_kind": "empirical", "mre_config": "DiffE2E with RegNetY-3.2GF encoder, Camera+LiDAR inputs, hybrid diffusion-supervision decoder, 2 denoising steps, two-stage training (30 epochs perception module with multi-task losses, then 30 epochs decoder training with frozen perception) on 4 GPUs. Evaluate on CARLA Longest6 benchmark (36 routes, 6 weather/lighting conditions, high traffic density).", "verified_links": {"paper_or_project": ["https://infinidrive.github.io/DiffE2E/", "https://arxiv.org/abs/2505.19516"], "code": ["https://github.com/bdjukic/DiffE2E"], "dataset": ["https://navsim.readthedocs.io/", "https://www.nuscenes.org/nuplan"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_verified", "evidence": "The official project page was opened and its artifact link area says 'Code (Coming Soon)' rather than providing a repository. Additional searches for 'DiffE2E github', 'infinidrive DiffE2E github', arXiv ID 2505.19516 plus code, site:github.com DiffE2E autonomous driving, and site:huggingface.co DiffE2E did not reveal a public first-party implementation. The bundled OpenReview supplement contains only Town03.mp4, Town04.mp4, and Town05.mp4 videos, not runnable training/evaluation code or configs for NAVSIM."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE uses NAVSIM navtrain/navtest. The public autonomousvision/navsim GitHub repository was opened; it contains the NAVSIM devkit, download directory, scripts, docs, public leaderboard links, and README text describing NAVSIM v1 navtest support and standardized training/test filtered splits. The repository also states that datasets inherit nuPlan/OpenScene distribution licenses, but the benchmark/devkit and documented download infrastructure are public."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searches for 'DiffE2E checkpoint weights model', 'DiffE2E Hugging Face', the arXiv ID with weights/checkpoints, and the official project page found no downloadable DiffE2E checkpoint. The official project page only says code is coming soon, and the OpenReview supplement manifest lists only MP4 visualization files, not model weights."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "NAVSIM is a named public autonomous-driving benchmark/devkit with public navtest leaderboard support and standardized splits in the autonomousvision/navsim repository. It is less general-purpose than CIFAR or ImageNet, but for end-to-end autonomous driving it is a recognized public benchmark rather than an author-private dataset."}}, "agent_task": "Reproduce DiffE2E's hybrid diffusion-supervision autonomous driving policy on CARLA Longest6 benchmark. Implementation steps: 1) Implement Transformer-based perception module with RegNetY-3.2GF encoder for image and LiDAR, multi-scale cross-modal fusion with bidirectional cross-attention; 2) Implement hybrid diffusion-supervision decoder with global condition integration (goal projection, timestep embedding), cross-attention mechanism for feature interaction, trajectory head and supervision head outputs; 3) Implement two-stage training: stage-1 trains perception with multi-task losses (semantic segmentation, BEV segmentation, depth, object detection), stage-2 freezes perception and trains decoder with diffusion loss (trajectory reconstruction) and supervised loss (speed classification, detection); 4) Use 100 forward diffusion steps, 2 reverse denoising steps, Square Cosine noise schedule, 100 denoising timesteps; 5) Train on CARLA simulator data (750k frames from TF++ MPC expert), 4 GPUs, batch size 16 for stage-1 and 256 for stage-2, lr=3e-4, 30 epochs each stage; 6) Evaluate on CARLA Longest6 (36 routes, high traffic density, 6 weather and 6 lighting conditions); 7) Report Driving Score, Route Completion, and Infraction Score. Target: DS near 83.", "h100_estimate": {"hours": 18.0, "basis_kind": "paper_reported", "gpu_count": 4, "gpu_type": "RTX 3090 (analogue: A40)", "wallclock_hours": 30, "h100_equivalent_multiplier": 0.15, "basis": "Paper trains on 4x RTX 3090 (Ampere consumer GPU), not A100 80GB as the prior basis wrongly assumed. RTX 3090 not in canonical table; closest analogue is A40 (0.15), a cut-down datacenter Ampere card. weights_available=false so MRE stays smallest-training-run scope, same source wallclock (~30h on 4 GPUs) as original basis. 4*30*0.15=18.0 H100-hours."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 18.0, "h100_estimate_basis": "paper_reported: Paper trains on 4x RTX 3090 (Ampere consumer GPU), not A100 80GB as the prior basis wrongly assumed. RTX 3090 not in canonical table; closest analogue is A40 (0.15), a cut-down datacenter Ampere card. weights_available=false so MRE stays smallest-training-run scope, same source wallclock (~30h on 4 GPUs) as original basis. 4*30*0.15=18.0 H100-hours.", "h100_band": "8-32", "h100_recomputed_hours": 18.0, "h100_arithmetic_mismatch": false, "h100_needs_human_review": true, "score": 3, "tier": "Hard", "audited_h100_hours": 18.0, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "DiffE2E: RegNetY-3.2GF encoder, Camera+LiDAR, hybrid diffusion-supervision decoder, 2 denoising steps, two-stage training (30 epochs per stage), evaluated on CARLA Longest6 benchmark", "metric": "Driving Score (DS)", "value": "83", "scope": "CARLA Longest6 benchmark (36 routes with adversarial scenarios)", "match_bar_kind": "point_estimate"}}
 {"custom_id": "2511.02225", "central_claim": "FIOC-WM learns a two-level factorization (object-level with explicit interactions + attribute-level static/dynamic) that improves policy-learning sample efficiency and generalization over world-model baselines in object-centric reinforcement learning.", "claim_evidence": "Abstract states: 'FIOC-WM improves policy-learning sample efficiency and generalization over world-model baselines, indicating that explicit, modular interaction learning is crucial for robust control.' Table 2 (LPIPS reconstruction quality) and Table 3 / Table 4 (policy learning success rates in main paper and appendix) demonstrate FIOC outperforms Dreamer-V3, TD-MPC2, and EIT across Gym Fetch, Franka Kitchen, i-Gibson, and Libero tasks.", "paper_kind": "empirical", "mre_config": "Gym Fetch Task 1 with FIOC-WM Variational (Categorical), 2M training steps, DINO-v2 visual encoder + slot attention (ViT-Base, 5 slots, slot_dim=64) + VAE with static/dynamic factorization, interaction graph learning via variational masks with categorical distribution, state transition MLP (2 layers, 64 hidden), low-level MPC policy + high-level PPO policy.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2511.02225"], "code": [], "dataset": ["https://github.com/Lifelong-Robot-Learning/LIBERO"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No OpenReview supplement is bundled in the input. The paper itself says code and data will be publicly available after acceptance, not that they are available. Web searches for the exact title, arXiv ID 2511.02225, FIOC-WM, Factored Interactive Object-Centric World Model, author/name variants, GitHub-scoped queries, and Hugging Face-scoped queries did not find a public first-party FIOC-WM implementation. Search results found only the arXiv/OpenReview paper, paper-note pages, the project-page URL mentioned in the PDF, and unrelated/awesome-list entries; no runnable MRE code repository was verified."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "Gym Fetch environment is standard OpenAI Gym/gymnasium benchmark. Libero benchmark verified available at github.com/Lifelong-Robot-Learning/LIBERO. Franka Kitchen and i-Gibson are established public benchmarks. SpritesWorld requires custom setup but other MRE datasets are standard benchmarks."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searches for FIOC-WM weights/checkpoints on Hugging Face and the web, including queries for the title, FIOC-WM checkpoint, FIOC-WM weights, and method-name variants, found no downloadable first-party FIOC-WM checkpoints or policy/world-model weights. The paper discusses pretrained visual encoders such as DINO-v2/R3M, but those are external base encoders, not MRE-trained FIOC-WM weights."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "Gym Fetch is a standard OpenAI Gym benchmark widely used in RL research. Libero, Franka Kitchen, and i-Gibson are established public benchmarks in robot learning community."}}, "agent_task": "Reproduce the FIOC-WM Gym Fetch Task 1 experiment (success rate 0.95): (1) Implement FIOC-WM from scratch: DINO-v2 ViT-Base visual encoder, slot attention for object decomposition (5 slots, slot_dim=64, 3 clustering iterations), VAE encoder/decoder (2-layer MLP, 256 hidden) mapping to latent states with dynamics dim=10 and attribute dim=8, static/dynamic state factorization with temporal consistency loss (Eq. 1) and contrastive loss (Eq. 2), interaction graph learning via variational masks with categorical distribution (GRU encoder, 3-layer MLP 256 hidden, temperature parameter), state transition MLP (2 layers, 64 hidden), reward decoder. (2) Offline training: collect 2000 episodes using pre-trained Dreamer-v3 policies on Gym Fetch. Loss weights alpha=1, beta=0.05, gamma=0.1, eta=0.2; lr=3e-4; 64x64 image crop, patch size 16. (3) Learn low-level policy via MPC/CEM (lr=5e-5, MSE loss) or PPO (lr=3e-4, clip=0.1, hidden [256,256], GAE=0.95, entropy=0.1). (4) Learn high-level policy via PPO (lr=1e-4, batch 1024, diversity reward r_div=1/sqrt(|G_visited|)). (5) Online policy learning for 2M steps on Gym Fetch Task 1 (2-push). (6) Report success rate over 5-10 random seeds. Target: 0.95 ± 0.03.", "h100_estimate": {"hours": 18, "basis_kind": "paper_reported", "gpu_count": 6, "gpu_type": "RTX 4090", "wallclock_hours": 8, "h100_equivalent_multiplier": 0.375, "basis": "Paper appendix reports Fetch experiments run for 8 hours on 6x NVIDIA 4090 GPUs. RTX 4090 estimated at ~0.375 H100 equivalent based on relative performance (between L40 at 0.18 and L40S at 0.37, with RTX 4090 being faster). 6 GPUs * 8 hours * 0.375 = 18 H100-hours. Engineering time to reimplement FIOC-WM architecture from paper description is additional and substantial."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 18.0, "h100_estimate_basis": "paper_reported: Paper appendix reports Fetch experiments run for 8 hours on 6x NVIDIA 4090 GPUs. RTX 4090 estimated at ~0.375 H100 equivalent based on relative performance (between L40 at 0.18 and L40S at 0.37, with RTX 4090 being faster). 6 GPUs * 8 hours * 0.375 = 18 H100-hours. Engineering time to reimplement FIOC-WM architecture from paper description is additional and substantial.", "h100_band": "8-32", "h100_recomputed_hours": 18.0, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 18.0, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "FIOC-Variational(Cat.), Gym Fetch Task 1, 2M training steps, DINO-v2 + Slot Attention + VAE with static/dynamic factorization + interaction graph modeling via variational masks with categorical distribution, 2000 offline episodes from Dreamer-v3 pre-trained policies", "metric": "policy learning success rate", "value": "0.95 ± 0.03", "scope": "OpenAI Gym Fetch Task 1 (2-push), evaluated over 5-10 random seeds", "match_bar_kind": "point_estimate"}}
 {"custom_id": "2411.06890", "central_claim": "SPARTAN learns accurate local causal graphs (low SHD) through sparsity regularisation while maintaining Transformer-level prediction accuracy, enabling improved few-shot adaptation and robustness.", "claim_evidence": "Abstract states SPARTAN 'learns sparse, context-dependent interaction graphs that accurately predict future object states.' Table 1 (prediction and graph) shows SPARTAN achieving SHD of 1.51 vs 6.37 for Transformer on Interventional Pong while maintaining comparable prediction error (8.60 vs 8.83).", "paper_kind": "empirical", "mre_config": "Train SPARTAN on Interventional Pong with token_dim=32, embed_dim=512, 3 transformer layers, MLP hidden_dim=512, lr=5e-5, Adam optimizer, and Lagrangian relaxation for sparsity regularisation. Evaluate learned causal graph SHD against ground truth.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2411.06890", "https://anonymous.4open.science/r/STRAW_neurips-4B75"], "code": [], "dataset": ["https://github.com/phlippe/CITRIS", "https://github.com/waymo-research/waymo-open-dataset"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No MRE-relevant public implementation was verified. Searches for the full title, arXiv ID 2411.06890, SPARTAN sparse transformer world model, GitHub-scoped variants, author/name variants, and the commented anonymous 4open.science STRAW_neurips URL did not reveal a usable first-party repository. OpenReview supplement status in the prompt is missing; the paper text only contains a commented anonymous repository URL, not a verified released artifact."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "Interventional Pong dataset available via CITRIS GitHub repo (phlippe/CITRIS) with data_generation_interventional_pong.py script and Zenodo download link (BSD 3-Clause license). CREATE simulator from Jain et al. 2020 available per paper (MIT License). Waymo Open Dataset verified at waymo.com/open with official GitHub at waymo-research/waymo-open-dataset (public download after sign-in)."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No downloadable SPARTAN checkpoints or MRE-relevant pretrained weights were found on OpenReview, NeurIPS/PMLR-style pages, MPI project/publication pages, GitHub searches, or Hugging Face-style artifact searches. The prompt's OpenReview supplement is missing and no external model-card or release page was verified."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "Waymo Open Dataset (waymo-research/waymo-open-dataset) is a widely recognized public autonomous driving benchmark for motion prediction, used extensively in research. Interventional Pong from CITRIS is a standard benchmark in causal representation learning literature, explicitly cited as such in the paper."}}, "agent_task": "Reproduce the SPARTAN local causal graph learning experiment on Interventional Pong: (1) Obtain Interventional Pong dataset from CITRIS GitHub repository (phlippe/CITRIS), (2) Implement SPARTAN architecture with Bernoulli-sampled adjacency matrices, Gumbel softmax trick, path matrix computation across layers, and Lagrangian relaxation training from paper equations, (3) Train with hyperparameters: token_dim=32, embed_dim=512, 3 transformer layers, MLP hidden_dim=512, lr=5e-5, Adam, (4) Extract learned adjacency matrices per timestep, (5) Compute Structural Hamming Distance (SHD) against ground-truth causal graph provided with dataset, (6) Target SHD near 1.51 to verify causal discovery accuracy.", "h100_estimate": {"hours": 23.04, "basis_kind": "paper_reported", "gpu_count": 1, "gpu_type": "V100 32GB", "wallclock_hours": 72, "h100_equivalent_multiplier": 0.32, "basis": "Weights not released -> MRE is smallest training run: train SPARTAN on Interventional Pong. Paper states convergence within 3 days (72h) on a single V100/RTX6000 GPU. 1 GPU x 72h x 0.32 (V100 closest architectural analogue to A100 in table) = 23.04 H100-hours. No double-counting, unit, or scope errors; confirmed correct."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 23.04, "h100_estimate_basis": "paper_reported: Weights not released -> MRE is smallest training run: train SPARTAN on Interventional Pong. Paper states convergence within 3 days (72h) on a single V100/RTX6000 GPU. 1 GPU x 72h x 0.32 (V100 closest architectural analogue to A100 in table) = 23.04 H100-hours. No double-counting, unit, or scope errors; confirmed correct.", "h100_band": "8-32", "h100_recomputed_hours": 23.04, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 23.04, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "SPARTAN on Interventional Pong: token_dim=32, embed_dim=512, 3 layers, MLP hidden_dim=512, lr=5e-5, Adam, Lagrangian relaxation", "metric": "Structural Hamming Distance (SHD) of learned causal graph vs ground truth", "value": "1.51", "scope": "Interventional Pong test set", "match_bar_kind": "direction"}}
 {"custom_id": "2511.19808", "central_claim": "RLNLC formulates noisy label correction as a reinforcement learning problem and achieves state-of-the-art performance on multiple benchmarks by learning a policy network via actor-critic methods to iteratively correct noisy training labels.", "claim_evidence": "Abstract: 'RLNLC learns a deep feature representation based policy network to perform label correction through reinforcement learning, utilizing an actor-critic method.' Table 1: RLNLC achieves 97.3% on CIFAR10-IDN (0.20 noise) and 95.8% on CIFAR10-IDN (0.50 noise), outperforming all comparison methods.", "paper_kind": "empirical", "mre_config": "ResNet-34 backbone, CIFAR10-IDN with 0.20 noise rate, 500 epochs policy training + 100 epochs fine-tuning, batch size 128, SGD with momentum 0.9, initial LR 0.01 (reduced to 0.001 at epoch 250), 50 epoch warmup, T=10 trajectory length, k=10 neighbors, N_b=100 bins, lambda=0.5, tau=0.5, gamma=0.9.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2511.19808"], "code": [], "dataset": ["https://www.cs.toronto.edu/~kriz/cifar.html"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No first-party MRE code was found. The OpenReview page for the paper is public but shows only the paper metadata/PDF and no code or supplementary artifact; the provided OpenReview supplement status is missing. Searches for the exact title, arXiv ID 2511.19808, RLNLC, 'RLNLC noisy label correction github', 'Learning to Clean Noisy Label github', and GitHub/Hugging Face scoped variants did not find a relevant public implementation for this paper."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE uses CIFAR-10 with synthetic symmetric label noise. CIFAR-10 is publicly available through standard dataset loaders; the PyTorch torchvision documentation verifies torchvision.datasets.CIFAR10 and its download option. The 50% symmetric-noise labels can be generated from the standard CIFAR-10 training labels by random class flips, so no unreleased author dataset is required for this MRE."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No downloadable RLNLC checkpoints or ResNet-18 CIFAR-10 symmetric-noise weights were found. The OpenReview page and arXiv page do not expose model files, the bundled supplement is missing, and searches for RLNLC with checkpoint/weights/model/Hugging Face/GitHub variants did not find MRE-relevant released weights."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "CIFAR-10 is a widely recognized public computer-vision benchmark, and PyTorch torchvision provides a standard CIFAR10 dataset class with public download support. The MRE's synthetic symmetric noise is a generated perturbation of this standard benchmark rather than a custom dataset requiring private annotations."}}, "agent_task": "Reproduce RLNLC on CIFAR10-IDN with 0.20 noise rate using ResNet-34: (1) Download CIFAR-10 from official source and inject instance-dependent noise at 0.20 rate per xia2020part protocol, (2) Implement RLNLC actor-critic framework with policy network based on deep feature extraction, k-NN attention mechanism for label prediction, and the composite reward function (LCR + NLA), (3) Train policy for 500 epochs with T=10 trajectory length, then deploy for T'=25 label correction steps, (4) Fine-tune prediction model on cleaned labels for 100 epochs, (5) Evaluate on CIFAR-10 test set. Target: achieve test accuracy near 97.3%.", "h100_estimate": {"hours": 32.4, "basis_kind": "compute_unspecified", "gpu_count": 1, "gpu_type": "H100 SXM", "wallclock_hours": 32.4, "h100_equivalent_multiplier": 1.0, "basis": "Weights not released -> MRE = smallest training run (single ResNet-34 on CIFAR10-IDN 0.20, policy RL training + fine-tuning), no full-paper inflation. Original basis's own component breakdown summed to 3.3+25+0.8+3.3=32.4h but the record reported 41h (internal inconsistency); corrected to the internally-consistent sum. No GPU/time stated by paper, so basis remains compute_unspecified with human review."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 32.4, "h100_estimate_basis": "Weights not released -> MRE = smallest training run (single ResNet-34 on CIFAR10-IDN 0.20, policy RL training + fine-tuning), no full-paper inflation. Original basis's own component breakdown summed to 3.3+25+0.8+3.3=32.4h but the record reported 41h (internal inconsistency); corrected to the internally-consistent sum. No GPU/time stated by paper, so basis remains compute_unspecified with human review.", "h100_band": "32-96", "h100_recomputed_hours": 32.4, "h100_arithmetic_mismatch": false, "h100_needs_human_review": true, "score": 3, "tier": "Hard", "audited_h100_hours": 32.4, "h100_hours_adjudicated": false, "selection_band": "32-96", "split": "eval", "match_target": {"config": "CIFAR10-IDN, 0.20 noise rate, ResNet-34 backbone, 500 epochs policy training + 100 epochs fine-tuning, standard hyperparameters (k=10, N_b=100, lambda=0.5)", "metric": "Test accuracy (%)", "value": "97.3", "scope": "CIFAR10-IDN test set", "match_bar_kind": "point_estimate"}}

 {"custom_id": "2511.20906", "central_claim": "DA-SIP achieves 2.6-4.4x reduction in total computation time while maintaining task success rates comparable to fixed maximum-computation baselines for robot manipulation tasks.", "claim_evidence": "Abstract states 'DA-SIP achieves 2.6--4.4x reduction in total computation time while maintaining task success rates comparable to fixed maximum-computation baselines.' Table 5 shows DA-SIP with Lightweight CNN achieves 2.6x-4.4x speedup across tasks (Push T: -2.69x, Transport: -4.43x, Lift: -4.09x) with minimal success rate degradation (max -3%). Table 2 shows 100% success on Lift task with single inference step, demonstrating efficiency.", "paper_kind": "empirical", "mre_config": "Train SI Policy with Linear interpolation and velocity prediction on RoboMimic 'Lift' task for 5000 epochs (batch size 256, lr 1e-4 with cosine decay) on NVIDIA L40S GPU. Evaluate with 1-step Euler ODE integration. Measure success rate over 50 episodes.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2511.20906"], "code": [], "dataset": ["https://github.com/ARVR-CV/mixed_papers_summary/blob/main/2511.20906.md"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Live searches for the full title, arXiv ID 2511.20906, DA-SIP, Difficulty-Aware Stochastic Interpolant Policy, author/title plus GitHub/code, OpenReview ID oDoPiR8wZJ, and Hugging Face variants found the NeurIPS/OpenReview/arXiv paper pages but no public implementation repository. The bundled OpenReview supplement manifest contains only 'DASIP appendix v3.pdf' and a macOS metadata file, not runnable training/evaluation code, configs, scripts, or notebooks for the Can MRE."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "RoboMimic (Can, Lift, Square, Tool Hang tasks), Push-T, Block Push (Fetch), Kitchen, and Multimodal Ant (OpenAI Gym) are standard public robotics benchmarks with documented download procedures and established evaluation protocols."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searches for DA-SIP with checkpoint, weights, model, Hugging Face, arXiv ID, and title variants found no downloadable SIP policy checkpoints, CNN classifier checkpoints, or VLM adapters for the Can MRE. The bundled OpenReview supplement contains only an appendix PDF and no model files."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "RoboMimic, Push-T, Kitchen, Fetch Block Push, and OpenAI Gym Ant are widely recognized public robotics benchmarks used extensively in imitation learning research."}}, "agent_task": "Train a stochastic interpolant policy with linear interpolation and velocity prediction on the RoboMimic Lift task for 5000 epochs using AdamW optimizer with batch size 256 and learning rate 1e-4 with cosine decay. Evaluate the trained policy using 1 inference step with Euler ODE integration and measure task success rate over 50 episodes. Verify that success rate reaches approximately 100%.", "h100_estimate": {"hours": 2.22, "basis_kind": "derived_from_config", "gpu_count": 1, "gpu_type": "L40S", "wallclock_hours": 6, "h100_equivalent_multiplier": 0.37, "basis": "weights_available=false -> smallest training run. mre_config trains SI Policy on single L40S GPU, one task (Lift), one seed, not the full paper's 3-seed multi-task sweep. Original basis used gpu_count=8 (full-paper scale), a scope error. Corrected: 1 GPU * 6h wallclock * 0.37 (L40S) = 2.22 H100-hours."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 2.22, "h100_estimate_basis": "derived_from_config: weights_available=false -> smallest training run. mre_config trains SI Policy on single L40S GPU, one task (Lift), one seed, not the full paper's 3-seed multi-task sweep. Original basis used gpu_count=8 (full-paper scale), a scope error. Corrected: 1 GPU * 6h wallclock * 0.37 (L40S) = 2.22 H100-hours.", "h100_band": "0-8", "h100_recomputed_hours": 2.2199999999999998, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 2.22, "h100_hours_adjudicated": false, "selection_band": "0-8", "split": "eval", "match_target": {"config": "SI Policy with Linear interpolation, velocity prediction, 1 inference step, Euler ODE approximator, trained 5000 epochs on Lift task", "metric": "task success rate", "value": "100%", "scope": "RoboMimic Lift task", "match_bar_kind": "threshold"}}
 {"custom_id": "2507.06489", "central_claim": "Adversarial attacks on verbal confidence (both perturbation-based VCA-TF, VCA-TB and jailbreak-based ConfidenceTriggers) can significantly reduce LLM confidence scores and cause high rates of answer changes.", "claim_evidence": "Abstract: 'these attacks can significantly impair verbal confidence estimates and lead to frequent answer changes.' Table 1 shows VCA-TF on TruthfulQA/Llama-3-8B/Base achieves 7.0% affected samples with 19.3% confidence drop; Typos attack achieves up to 54.0% affected samples with 24.7% confidence drop. Table 2 shows ConfidenceTriggers achieves up to 29.5% confidence reduction on TQA with CoT.", "paper_kind": "empirical", "mre_config": "VCA-TF perturbation attack on Llama-3-8B with Base CEM on TruthfulQA dataset, 200 random samples, measuring % affected samples and average confidence drop on affected samples.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2507.06489"], "code": [], "dataset": ["https://hf.co/datasets/truthfulqa/truthful_qa", "https://hf.co/datasets/ChilleD/StrategyQA", "https://hf.co/datasets/openlifescienceai/medmcqa"], "weights": ["https://hf.co/meta-llama/Meta-Llama-3-8B-Instruct"]}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searched GitHub and HuggingFace with multiple query variations: 'ConfidenceTriggers', 'verbal confidence adversarial attacks', 'Obadinma', arXiv ID 2507.06489, full title, and key method names (VCA-TF, VCA-TB, SubSwapRemove, ConfidenceTriggers). Found only paper analysis notes in repos (memgrafter/analysis, zhaoyang97/Paper-Notes). No official code repository exists for this paper."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "All three datasets used in the paper are publicly available on HuggingFace: TruthfulQA (truthfulqa/truthful_qa, 89.2K downloads), StrategyQA (ChilleD/StrategyQA, 11.1K downloads), and MedMCQA (openlifescienceai/medmcqa, 30.4K downloads)."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "The paper does not release any checkpoints or weights for their attack methods (ConfidenceTriggers, VCA-TF, VCA-TB). Llama-3-8B is available on HuggingFace from Meta, but these are base model weights, not the paper's trained trigger tokens or attack artifacts. No VCA-specific weights/checkpoints were found after searching."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "TruthfulQA (arXiv:2109.07958), StrategyQA (crowdsourced benchmark published in ICLR 2022), and MedMCQA (medical MCQA dataset with 2816+ questions from ML research) are established public benchmarks widely used in NLP research."}}, "agent_task": "Reproduce the VCA-TF perturbation-based verbal confidence attack on Llama-3-8B using the TruthfulQA dataset. Implement the attack algorithm from the paper: (1) calculate importance scores by removing each token and measuring confidence change, (2) rank tokens by importance, (3) substitute top tokens with synonyms (POS-tag filtered) while maintaining 80% semantic similarity (USE). Measure: (1) % of samples where confidence is reduced (%Aff), (2) average confidence drop on affected samples (Δ Aff. Cf.), (3) answer flip rates. Expected: at least 7% affected samples with measurable confidence drops on TruthfulQA/Base/Llama-3-8B.", "h100_estimate": {"hours": 2.56, "basis_kind": "derived_from_config", "gpu_count": 1, "gpu_type": "A100 80GB", "wallclock_hours": 8, "h100_equivalent_multiplier": 0.32, "basis": "MRE = VCA-TF attack (200 TruthfulQA samples, Llama-3-8B, Base CEM); no weights/code released for the attack artifacts, so this is the paper's own smallest inference-search config, not a training run. Original basis combined a 0.2s/batch-of-32 throughput claim with a 375h wallclock figure for only ~750-1500 LLM calls, a unit/arithmetic error inflating hours ~150x. Re-derived: ~200 samples x ~10-25 search iterations x a few candidate evals/iteration on an 8B model, generously bounded at 8 wallclock-hours on 1x A100 80GB (sequential, hard to batch due to adaptive search). 1*8*0.32=2.56 H100-hours."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 2.56, "h100_estimate_basis": "derived_from_config: MRE = VCA-TF attack (200 TruthfulQA samples, Llama-3-8B, Base CEM); no weights/code released for the attack artifacts, so this is the paper's own smallest inference-search config, not a training run. Original basis combined a 0.2s/batch-of-32 throughput claim with a 375h wallclock figure for only ~750-1500 LLM calls, a unit/arithmetic error inflating hours ~150x. Re-derived: ~200 samples x ~10-25 search iterations x a few candidate evals/iteration on an 8B model, generously bounded at 8 wallclock-hours on 1x A100 80GB (sequential, hard to batch due to adaptive search). 1*8*0.32=2.56 H100-hours.", "h100_band": "0-8", "h100_recomputed_hours": 2.56, "h100_arithmetic_mismatch": false, "h100_needs_human_review": true, "score": 3, "tier": "Hard", "audited_h100_hours": 2.56, "h100_hours_adjudicated": false, "selection_band": "0-8", "split": "eval", "match_target": {"config": "Llama-3-8B + VCA-TF attack + Base CEM + TruthfulQA (200 samples)", "metric": "% affected samples with confidence reduction", "value": "7.0%", "scope": "TruthfulQA/Llama-3-8B/Base", "match_bar_kind": "threshold"}}
 {"custom_id": "2602.03066", "central_claim": "Shortcut features correspond to eigenfunctions of the Neural Tangent Kernel (NTK) with larger eigenvalues; features with larger eigenvalues are learned faster due to spectral bias and have a larger influence on the neural network output even after training convergence.", "claim_evidence": "Abstract states: 'shortcut features correspond to features with larger eigenvalues when the shortcuts stem from the imbalanced number of samples in the clustered distribution' and 'features with larger eigenvalues still have a large influence on the neural network output even after training'. Section 4.3 (Figure 5, synthetic) and Figure 9 (Appendix synthetic) show that availability (alignment to top NTK eigenspace) is higher for shortcut labels than ground-truth labels across Colored-MNIST and Patched-MNIST, with differences of ~0.08-0.10 under CE loss. Section 4.3 (Figure 6, real-world) and Figure 11-12 show the same pattern on Waterbirds and CelebA.", "paper_kind": "empirical", "mre_config": "Train a 2-layer ReLU fully-connected network (width 256, SGD, lr=0.001, wd=0.0001, batch=1000, zero-initialized output weights, 200 epochs) on Colored-MNIST. Construct Colored-MNIST with 95% of label=-1 samples colored red and 5% of label=1 samples colored red (training split); test split has balanced 50/50 red/blue coloring. At step 1000, compute the empirical NTK K from 500 randomly sampled training inputs (using the finite-width network Jacobian). Compute availability A(K, g) = g^T K g / (||g||^2_2 ||K||_F) for both the ground-truth label vector and the shortcut label vector (color-based: red=1, blue=-1). Verify that availability(shortcut) > availability(ground-truth) by a margin consistent with the paper's reported ~0.08-0.10 difference for Colored-MNIST with CE loss (per Appendix Table 3).", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2602.03066", "https://papers.nips.cc/paper_files/paper/2025/hash/87571720167f7e88827c40e468e3101f-Abstract-Conference.html"], "code": [], "dataset": ["http://yann.lecun.com/exdb/mnist/"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No OpenReview supplement was provided in the input, and the paper checklist says code/data are not yet released. Live web searches for the exact title, arXiv ID 2602.03066, OpenReview ID TYroQXu6X0, author names, and GitHub/code variants found the OpenReview/arXiv/NeurIPS paper pages but no public repository or runnable scripts for the Patched-MNIST/Colored-MNIST availability MRE."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE can be built from MNIST, which is publicly available through torchvision.datasets.MNIST with a documented download parameter. The paper text specifies the MRE-specific biased construction: digits 0-4 labelled -1, digits 5-9 labelled +1, with a 95/5 shortcut ratio in training and 50/50 in test for Patched-MNIST/Colored-MNIST. No first-party processed split files were found, but the raw dataset and construction recipe are sufficient for this minimal derived benchmark."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "The MRE trains a small two-layer ReLU fully connected model from scratch and does not require pretrained weights. Searches for the title/arXiv ID plus checkpoint, weights, model, Hugging Face, and GitHub variants found no released checkpoints or model files for the MRE."}, "dataset_is_standard": {"value": false, "verification": "tool_verified", "evidence": "MNIST itself is a standard public benchmark, verified via torchvision documentation, but the actual MRE dataset is an author-defined biased derivative, Patched-MNIST or Colored-MNIST, with custom shortcut features, custom binary labels, and custom train/test shortcut ratios. Therefore the MRE dataset is not a standard benchmark as used in the paper."}}, "agent_task": "Reproduce the core empirical finding on Colored-MNIST: (1) Download MNIST from http://yann.lecun.com/exdb/mnist/. (2) Construct Colored-MNIST: for each training image, if label==-1 (digits 0-4), color it red with probability 0.95; if label==1 (digits 5-9), color it red with probability 0.05; for test set use 50/50 balanced colors. (3) Train a 2-layer ReLU FC network (width=256, zero-initialized last-layer weights, no bias) with SGD (lr=0.001, wd=0.0001, batch=1000) for 200 epochs. (4) At training step 1000, compute the empirical NTK K from 500 randomly sampled training inputs using the finite-width network Jacobian (K = J J^T where J is the Jacobian of outputs w.r.t. parameters). (5) Compute availability A(K, g) = g^T K g / (||g||^2_2 ||K||_F) for both the ground-truth label vector g_y and the shortcut label vector g_color (color-based label: red=1, blue=-1). (6) Verify that availability(g_color) > availability(g_y) by a margin consistent with the paper's reported difference of ~0.08-0.10 for Colored-MNIST with CE loss (Appendix Table 3). This confirms that shortcut features align more strongly with top NTK eigenfunctions than core features.", "h100_estimate": {"hours": 4, "basis_kind": "derived_from_config", "gpu_count": 1, "gpu_type": "A100 80GB", "wallclock_hours": 12, "h100_equivalent_multiplier": 0.32, "basis": "The paper (Section X_suppl) reports experiments on 2 RTX 3090 GPUs for 1 hour for synthetic dataset experiments. For Colored-MNIST: 200 epochs, batch=1000, 2-layer ReLU FC width 256. Scaling to 1 GPU: ~2 hours wallclock (linear scaling of GPU-hours). Adding NTK computation overhead (500 samples, Jacobian computation at step 1000) adds ~10-20% overhead. 1 * 2 * 0.32 = 0.64 H100-hours for base training. Adding engineering/debugging margin for custom Colored-MNIST construction and NTK implementation: 4 H100-hours total."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 4.0, "h100_estimate_basis": "derived_from_config: The paper (Section X_suppl) reports experiments on 2 RTX 3090 GPUs for 1 hour for synthetic dataset experiments. For Colored-MNIST: 200 epochs, batch=1000, 2-layer ReLU FC width 256. Scaling to 1 GPU: ~2 hours wallclock (linear scaling of GPU-hours). Adding NTK computation overhead (500 samples, Jacobian computation at step 1000) adds ~10-20% overhead. 1 * 2 * 0.32 = 0.64 H100-hours for base training. Adding engineering/debugging margin for custom Colored-MNIST construction and NTK implementation: 4 H100-hours total.", "h100_band": "0-8", "h100_recomputed_hours": 3.84, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 4.0, "h100_hours_adjudicated": false, "selection_band": "0-8", "match_target": {"config": "2-layer ReLU FC net, width 256, SGD lr=1e-3 wd=1e-4 batch=1000, zero-init output layer, Colored-MNIST (train: 95%/5% color-label correlation; test: 50/50 balanced), empirical NTK at step 1000 from 500 random train inputs", "metric": "NTK availability A(K,g) = g^T K g / (||g||^2 ||K||_F) for shortcut (color) vs ground-truth label vectors", "value": "availability(shortcut) > availability(ground-truth) at step 1000; the qualitative ordering is the falsifiable core of the eigenvalue/spectral-bias claim", "scope": "Colored-MNIST shortcut-learning experiment, empirical finite-width NTK", "match_bar_kind": "direction"}, "split": "eval"}
 {"custom_id": "2505.16927", "central_claim": "STaPLe enables smaller language models (7-8B parameters) to self-improve by discovering and learning latent principles for self-correction, achieving +8-10% in AlpacaEval win-rate, +0.3 on MT-Bench, and +19-23% principle-following win-rate on IFEval.", "claim_evidence": "Abstract states 'bootstrapping our algorithm over multiple iterations enables smaller language models (7-8B parameters) to self-improve, achieving +8-10% in AlpacaEval win-rate, an average of +0.3 on MT-Bench, and +19-23% in principle-following win-rate on IFEval.' Table 1 reports IFEval win-rates of 65.6% (STaPLe Iter 1, Llama-3.1-8B) vs 55.5% (STaR baseline), demonstrating the core self-improvement claim.", "paper_kind": "empirical", "mre_config": "STaPLe Iteration 1 on Llama-3.1-8B-Instruct using 10k samples from standard mining corpus (HH-RLHF, UltraFeedback, or TL;DR from HuggingFace) with N=16 principle samples, Rouge-L F1 similarity threshold 0.4, temperature 0.7, full supervised fine-tuning for 3 epochs at learning rate 1e-6, sequence length 4096, evaluated on IFEval (541 prompts) using Prometheus-8x7B-v2.0 as judge for principle-following win-rate.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2505.16927", "https://openreview.net/attachment/f4cff24f37db1093806eb3e1070b7db1220d1ccb.zip"], "code": [], "dataset": ["https://huggingface.co/datasets/Anthropic/hh-rlhf", "https://huggingface.co/datasets/openbmb/UltraFeedback", "https://huggingface.co/datasets/tatsu-lab/alpaca_eval"], "weights": ["https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct", "https://huggingface.co/ibm-granite/granite-3.1-8b-instruct", "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct"]}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Web searches for the exact title, arXiv ID 2505.16927, \"Self-Taught Principle Learning\", \"STaPLe\" with GitHub/Hugging Face scoping, author name variants, and prompt-file phrases did not find a public runnable implementation repository. The paper and OpenReview PDF state that code will be publicly released, but that is only a promise. The bundled OpenReview supplement contains LaTeX source, figures, tables, constitutions, example prompts, and an algorithm writeup, but no runnable training/evaluation/data-processing scripts or configs sufficient to execute the MRE."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "Mining corpus datasets verified on Hugging Face: Anthropic/hh-rlhf (MIT license, 161k preference pairs), openbmb/UltraFeedback (MIT license, 64k prompts). Evaluation dataset verified: tatsu-lab/alpaca_eval (CC-BY-NC-4.0 license, 805 samples). Paper appendix documents these datasets and their construction."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "The base Qwen2.5-7B-Instruct model is publicly available on Hugging Face and is relevant as the starting checkpoint, but searches of Hugging Face, GitHub, the exact title, arXiv ID, STaPLe variants, and author/project terms did not find released STaPLe fine-tuned checkpoints/adapters for the one-iteration Qwen MRE or any final STaPLe models. The OpenReview supplement manifest contains no model, adapter, checkpoint, or safetensors/bin weight files."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "Mining corpus consists of established public datasets: HH-RLHF (Anthropic), UltraFeedback (openbmb), TL;DR (OpenAI), HotpotQA. Evaluation benchmarks: AlpacaEval-2.0-LC (tatsu-lab), MT-Bench, IFEval. All are widely recognized public benchmarks with documented download paths on HuggingFace and GitHub."}}, "agent_task": "Reproduce STaPLe (Self-Taught Principle Learning) Iteration 1 to verify self-improvement claim: (1) Load Llama-3.1-8B-Instruct from HuggingFace (meta-llama/Llama-3.1-8B-Instruct), (2) Prepare mining corpus sample (~10k prompts) by downloading from HuggingFace datasets (Anthropic/hh-rlhf, openbmb/UltraFeedback, or openai/summarize_from_feedback) and filtering to samples with gold/chosen responses, (3) Implement E-step: for each prompt, sample initial response from model, mine N=16 principles using principle-mining prompt (Prompt: {prompt}\\n\\nHere is the previous response: {initial_response}\\n\\nHere is the target response: {gold_response}\\n\\nIdentify a high-level principle... New Principle:), generate critiques using critique-generation prompt, sample N refined responses, select best via Rouge-L F1 similarity to gold (must exceed initial response), (4) Implement M-step: supervised fine-tune on (prompt, initial_response, principle, refined_response) tuples for 3 epochs at LR 1e-6, seq_len 4096, (5) Evaluate on IFEval (541 prompts) using Prometheus-8x7B-v2.0 judge: for each prompt, compare model's initial vs refined response on principle-following, report win-rate. Target: achieve >60% IFEval win-rate demonstrating the model learned intrinsic self-correction.", "h100_estimate": {"hours": 9.6, "basis_kind": "derived_from_config", "gpu_count": 8, "gpu_type": "H100 SXM", "wallclock_hours": 1.2, "h100_equivalent_multiplier": 1, "basis": "Weights not released -> MRE is smallest training run: Iteration 1 on Llama-3.1-8B, 10k samples per paper's own mre_config (not 50k as prior basis used). E-step scaled from paper's 2.5M-pass/4hr on 4xH100 (16 GPU-hr) by 10k/50k -> 3.2 GPU-hr. M-step 3-epoch SFT on ~5.6k successful samples (28k*10k/50k) on 8xH100 scaled from 32 GPU-hr -> 6.4 GPU-hr. Total 9.6 GPU-hr, expressed as 8 GPUsx1.2hrx1.0."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 9.6, "h100_estimate_basis": "derived_from_config: Weights not released -> MRE is smallest training run: Iteration 1 on Llama-3.1-8B, 10k samples per paper's own mre_config (not 50k as prior basis used). E-step scaled from paper's 2.5M-pass/4hr on 4xH100 (16 GPU-hr) by 10k/50k -> 3.2 GPU-hr. M-step 3-epoch SFT on ~5.6k successful samples (28k*10k/50k) on 8xH100 scaled from 32 GPU-hr -> 6.4 GPU-hr. Total 9.6 GPU-hr, expressed as 8 GPUsx1.2hrx1.0.", "h100_band": "8-32", "h100_recomputed_hours": 9.6, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 9.6, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "STaPLe Iteration 1, Llama-3.1-8B-Instruct, constrained (clustered) version, N=16, Rouge-L similarity, 3 epochs SFT on ~5k accepted samples", "metric": "IFEval principle-following win-rate", "value": "65.6%", "scope": "IFEval benchmark (541 prompts) with Prometheus-8x7B-v2.0 judge comparing refined response to initial policy response", "match_bar_kind": "threshold"}}
 {"custom_id": "2502.08924", "central_claim": "A boosting-style algorithm that combines synthetic data with weak (beta>0) exogenously-labeled data converges to an optimal LLM, with focusing labeling resources on the most challenging examples leading to improved performance over filter-only approaches like STaR/ReST.", "claim_evidence": "Abstract: 'we show that dynamically focusing labeling resources on the most challenging examples --- in much the same way that boosting focuses the efforts of the weak learner --- leads to improved performance.' Table 1: Boosting (A) achieves GSM8K test accuracy 0.456 vs Filter only at 0.393 and Do nothing at 0.129.", "paper_kind": "empirical", "mre_config": "Gemma 2 2B PT with boosting-style algorithm (Algorithm 2), 5 rounds, GSM8K dataset (7000 train), 3-shot prompting, k=8 synthetic samples per prompt, alpha=1/3, batch size 64, 330 steps per round, standard cross-entropy loss fine-tuning. Comparing Boosting vs Filter-only (STaR/ReST-like) and Do nothing baselines.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2502.08924"], "code": [], "dataset": ["https://hf.co/datasets/openai/gsm8k", "https://hf.co/datasets/google-research-datasets/mbpp"], "weights": ["https://hf.co/google/gemma-2-2b-it"]}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No MRE-relevant first-party code was found. Web searches for the full title, arXiv ID 2502.08924, 'Escaping Collapse' with GitHub/code/Gemma/GSM8K, and weak-data boosting variants found arXiv/Google Research/OpenReview/Paper aggregator pages but no public repository. The OpenReview paper PDF explicitly answers the reproducibility checklist 'No' and states that the experiments use open resources but 'the code itself is not open sourced.' The provided OpenReview supplement status is missing."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE uses GSM8K, a public standard benchmark. Hugging Face hosts openai/gsm8k, and TensorFlow Datasets lists gsm8k as a dataset of 8.5K grade-school math word problems with a homepage pointing to OpenAI's grade-school-math repository. The paper's 3-shot prompt template is included in the paper text, so no unreleased custom dataset is required for the selected GSM8K MRE."}, "weights_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE initializes and labels with Gemma 2 2B PT. The Hugging Face page google/gemma-2-2b exists and provides Transformers loading instructions such as AutoModelForCausalLM.from_pretrained, but access requires agreeing to Google's usage license. No paper-specific fine-tuned checkpoints were found, so this signal counts only the publicly available base weights needed to run the training MRE from scratch."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "GSM8K is a widely recognized public ML benchmark for grade-school math reasoning; it is available through Hugging Face openai/gsm8k and TensorFlow Datasets."}}, "agent_task": "Reproduce the paper's boosting-style LLM training algorithm (Algorithm 2). Steps: (1) Load Gemma 2 2B PT from HuggingFace (google/gemma-2-2b). (2) Download GSM8K dataset from HuggingFace. (3) Implement boosting-style training loop for 5 rounds: (a) generate k=8 synthetic responses per prompt using current model at temperature, (b) filter responses using ground-truth quality function to partition into high-quality S+ and low-quality P- sets, (c) apply weak labeler to P- prompts (simulate by sampling from same model, returning correct responses for beta fraction), (d) mix synthetic+weak data with alpha=1/3 weight on weak data, (e) fine-tune next round initialized from previous checkpoint. (4) Run Filter-only baseline (alpha=0) for comparison. (5) Measure test accuracy (greedy decoding) on GSM8K test set. Target: Boosting achieves >= 0.456 test accuracy vs Filter-only at ~0.393.", "h100_estimate": {"hours": 12.0, "basis_kind": "derived_from_config", "gpu_count": 1, "gpu_type": "H100 SXM", "wallclock_hours": 12, "h100_equivalent_multiplier": 1.0, "basis": "weights_available=true is only the released base Gemma-2-2B checkpoint, not a fine-tuned artifact for the boosting-vs-filter claim, so the MRE stays a smallest_training_run (5 rounds x 330 steps + k=8 synthetic gens over GSM8K), not eval-only. Original used 8xH100 x1.5h wallclock for a 2B-param fine-tune, which is implausible over-provisioning for this model size. Reallocated to 1xH100 x 12h, preserving the same total 12 GPU-hours. Non-compute debugging/engineering padding language dropped (was mentioned but not arithmetically applied)."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 12.0, "h100_estimate_basis": "derived_from_config: weights_available=true is only the released base Gemma-2-2B checkpoint, not a fine-tuned artifact for the boosting-vs-filter claim, so the MRE stays a smallest_training_run (5 rounds x 330 steps + k=8 synthetic gens over GSM8K), not eval-only. Original used 8xH100 x1.5h wallclock for a 2B-param fine-tune, which is implausible over-provisioning for this model size. Reallocated to 1xH100 x 12h, preserving the same total 12 GPU-hours. Non-compute debugging/engineering padding language dropped (was mentioned but not arithmetically applied).", "h100_band": "8-32", "h100_recomputed_hours": 12.0, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 2, "tier": "Hard", "audited_h100_hours": 12.0, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "Gemma 2 2B PT, Boosting (A), 5 rounds, 3-shot, GSM8K train=7000, k=8 synthetic samples per prompt, alpha=1/3, batch_size=64, 330 steps/round", "metric": "test accuracy (greedy decoding)", "value": "0.456", "scope": "GSM8K test set", "match_bar_kind": "point_estimate"}}
 {"custom_id": "2511.01463", "central_claim": "HMVLM with MoE LoRA and zero expert preserves foundation model knowledge while achieving strong text-to-motion performance on HumanML3D (R-precision Top-3 of 0.785).", "claim_evidence": "Abstract states 'Experiments show that our method effectively alleviates knowledge forgetting during instruction-tuning and achieves remarkable performance across diverse human motion downstream tasks.' Table 1 shows 'Ours (single task)' achieves R precision Top-3 of 0.785 on HumanML3D test set.", "paper_kind": "empirical", "mre_config": "Single-task MoE LoRA fine-tuning with 5 experts (including zero expert), Vicuna-7b-v1.5 base model, on HumanML3D text-to-motion task using single A800 80G GPU, batch size 32, learning rate 3×10^-3 with cosine annealing.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2511.01463"], "code": [], "dataset": ["HumanML3D is a standard text-to-motion benchmark from Guo et al. 2022, available via the original project page at UMass Amherst"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searched live web with exact title, acronym, arXiv ID, GitHub-scoped, Hugging Face-scoped, and method/dataset variants such as \"HMVLM github\", \"Human Motion-Vision-Lanuage GitHub\", \"HMVLM Hugging Face motion\", and site:github.com queries. Results found arXiv/OpenReview mirrors and summaries but no first-party HMVLM code repository. The bundled OpenReview supplement manifest contains only supplementary_video.mp4, which is not runnable code/config/script for the MRE."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE uses HumanML3D. A live web search found and verified the public HumanML3D GitHub repository at EricGuo5513/HumanML3D, whose README/search snippet describes it as a large 3D human motion-language dataset and provides scripts to reproduce the dataset from AMASS. This is sufficient public dataset evidence for the HumanML3D T2M MRE, subject to upstream AMASS/SMPL licensing and setup."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searched live web for HMVLM model/checkpoint/weights via exact title, acronym, arXiv ID, GitHub, and Hugging Face variants. No downloadable HMVLM checkpoints, LoRA adapters, tokenizer weights, or gating-network weights were found. The OpenReview supplement contains only supplementary_video.mp4 and no checkpoint/model files."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "HumanML3D is a widely used public text-to-motion benchmark; live web results include the official HumanML3D repository and secondary dataset pages describing it as a standard evaluation dataset for text-to-motion models with common FID, R-precision, and Diversity metrics. The MRE does not require a newly authored custom benchmark beyond the paper's instruction formatting/training pipeline."}}, "agent_task": "Reproduce HMVLM text-to-motion generation using MoE LoRA fine-tuning on HumanML3D: (1) Download HumanML3D dataset from EricGuo5513/HumanML3D, (2) Implement MoE LoRA architecture with 5 experts (4 task experts + 1 zero expert), gating network (2-layer MLP, hidden dim 512), and zero expert with zero-initialized non-trainable LoRA matrices, (3) Implement body-part-based motion tokenizer with 5 body parts, embedding dim 512, codebook size K=512, temporal compression l=4, (4) Fine-tune Vicuna-7B-v1.5 base model with MoE LoRA rank 8, batch size 32, lr 3e-3, cosine annealing, (5) Evaluate using R-precision (Top-1/2/3), FID, MM-Distance, and Diversity metrics on HumanML3D test set. Note: No released code or weights - requires full implementation from scratch.", "h100_estimate": {"hours": 12.8, "basis_kind": "derived_from_config", "gpu_count": 1, "gpu_type": "A800", "wallclock_hours": 40, "h100_equivalent_multiplier": 0.32, "basis": "Weights not released -> MRE is smallest single training run, not full joint instruction-tuning. Original 38.4 used 120h reported for JOINT training across 3 tasks (T2M+pose+video), a scope error. Single-task T2M scaled to ~1/3 of joint schedule: 1 GPU x 40h x 0.32 (A800) = 12.8 H100-hours. Same GPU/rate as paper, corrected scope only."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 12.8, "h100_estimate_basis": "derived_from_config: Weights not released -> MRE is smallest single training run, not full joint instruction-tuning. Original 38.4 used 120h reported for JOINT training across 3 tasks (T2M+pose+video), a scope error. Single-task T2M scaled to ~1/3 of joint schedule: 1 GPU x 40h x 0.32 (A800) = 12.8 H100-hours. Same GPU/rate as paper, corrected scope only.", "h100_band": "8-32", "h100_recomputed_hours": 12.8, "h100_arithmetic_mismatch": false, "h100_needs_human_review": true, "score": 3, "tier": "Hard", "audited_h100_hours": 12.8, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "MoE LoRA (5 experts, rank 8), Vicuna-7B-v1.5, single-task T2M fine-tuning on HumanML3D", "metric": "R precision Top-3", "value": "0.785", "scope": "HumanML3D test set", "match_bar_kind": "point_estimate"}}
 {"custom_id": "2510.25529", "central_claim": "MoGE (Modelic Generative Exploration) effectively bridges exploration and policy learning in off-policy RL through diffusion-based critical state generation and a one-step imagination world model, leading to remarkable gains in both sample efficiency and performance across complex control tasks.", "claim_evidence": "Abstract states 'MoGE effectively bridges exploration and policy learning, leading to remarkable gains in both sample efficiency and performance across complex control tasks.' Table 1 shows DSAC+MoGE achieves 891.7 ± 19.1 TAR on Humanoid-walk vs DSAC's 146.5 ± 60.9 (+508.6% improvement), and 817.7 ± 11.4 average DMC TAR vs DSAC's 568.5 ± 42.7 (+43.8% improvement).", "paper_kind": "empirical", "mre_config": "DSAC+MoGE on DeepMind Control Suite Humanoid-walk task, trained for 1.5M steps with 3 random seeds, compared against DSAC baseline, using standard DMC Gym environments via gymnasium/dmcgym.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2510.25529"], "code": ["https://github.com/WangLK-Franklin/MoGE"], "dataset": ["https://github.com/deepmind/dm_control", "https://github.com/Farama-Foundation/Gymnasium"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_verified", "evidence": "The paper points to https://github.com/WangLK-Franklin/MoGE, and the live GitHub page is public but contains only README.md; the README says 'The code will be released later,' with no runnable training scripts, configs, or implementation files for the MRE. Searches for the title, arXiv ID, MoGE, Modelic Generative Exploration, and WangLK-Franklin found no other relevant public implementation. OpenReview supplement status in the prompt is missing."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE uses standard simulator benchmarks rather than a custom dataset. DeepMind Control Suite is publicly available through the google-deepmind/dm_control repository, whose page describes dm_control.suite as Python reinforcement-learning environments powered by MuJoCo. The paper's alternative Gym MuJoCo tasks are also public through Gymnasium documentation listing Ant, HalfCheetah, Humanoid, Swimmer, and Walker2D MuJoCo environments."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No MoGE/DSAC+MoGE checkpoints or learned model weights were found in the GitHub repository, which has no releases and only a README. Artifact-specific searches on GitHub/Hugging Face for MoGE, Modelic Generative Exploration, WangLK-Franklin MoGE, and checkpoint/weights found unrelated MoGe geometry-estimation models but no MRE-relevant RL checkpoints."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "DeepMind Control Suite and Gym/Gymnasium MuJoCo locomotion environments are widely recognized public RL benchmarks used directly through public simulator packages; the MRE does not require author-created labels, custom splits, or private data."}}, "agent_task": "Reproduce the MoGE method on DMC Humanoid-walk task: (1) Implement MoGE from scratch since official code is not released (placeholder repo only), (2) Install gymnasium and dm_control packages for DeepMind Control Suite environments, (3) Implement DSAC algorithm as base off-policy RL method, (4) Implement conditional diffusion model for critical state generation with utility guidance (policy entropy or TD error), (5) Implement one-step world model with encoder, dynamics predictor, reward predictor, and termination predictor, (6) Train DSAC+MoGE for 1.5M steps on Humanoid-walk using paper hyperparameters (batch_size=1024, lr=1e-4, guidance_scale=1.0, diffusion_steps=100, lambda=0.2, beta1=0.5, beta2=0.1), (7) Evaluate and verify TAR reaches ~891.7 ± 19.1 vs DSAC baseline ~146.5 ± 60.9.", "h100_estimate": {"hours": 13.48, "basis_kind": "paper_reported", "gpu_count": 1, "gpu_type": "RTX 4090", "wallclock_hours": 42.13, "h100_equivalent_multiplier": 0.32, "basis": "Weights not released -> MRE is smallest training run: DSAC+MoGE on Humanoid task, 1.5M steps, single seed. Paper Table 3 (supplement) reports walltime 151657±395s (=42.13h) on 1 RTX 4090. RTX 4090 not in canonical table; using A100 80GB (0.32) as closest throughput analogue. H100-hours = 1*42.13*0.32 = 13.48."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 13.48, "h100_estimate_basis": "paper_reported: Weights not released -> MRE is smallest training run: DSAC+MoGE on Humanoid task, 1.5M steps, single seed. Paper Table 3 (supplement) reports walltime 151657±395s (=42.13h) on 1 RTX 4090. RTX 4090 not in canonical table; using A100 80GB (0.32) as closest throughput analogue. H100-hours = 1*42.13*0.32 = 13.48.", "h100_band": "8-32", "h100_recomputed_hours": 13.4816, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 13.48, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "DSAC+MoGE on Humanoid-walk (DMC), 1.5M steps, batch_size=1024, learning_rate=1e-4, guidance_scale=1.0, diffusion_steps=100, mix_ratio lambda=0.2", "metric": "Total Average Return (TAR)", "value": "891.7 ± 19.1", "scope": "DeepMind Control Suite Humanoid-walk task, averaged over 3 seeds", "match_bar_kind": "point_estimate"}}
 {"custom_id": "2506.00070", "central_claim": "Robot-R1 uses reinforcement learning (GRPO) to train LVLMs for embodied reasoning in robotics, where models trained with Robot-R1 outperform SFT methods and a 7B Robot-R1 model surpasses GPT-4o on reasoning tasks related to low-level action control (movement and spatial reasoning).", "claim_evidence": "Abstract: 'models trained with Robot-R1 outperform SFT methods on embodied reasoning tasks. Despite having only 7B parameters, Robot-R1 even surpasses GPT-4o on reasoning tasks related to low-level action control, such as spatial and movement reasoning.' Table 2 (tab:LowLevelControl_result): Robot-R1 achieves 0.76 avg movement (vs GPT-4o's 0.72) and 1.51 avg spatial (vs GPT-4o's 1.43) on Robot-R1 Bench.", "paper_kind": "empirical", "mre_config": "Train Qwen2.5-7B-VL-Instruct with GRPO on RLBench-generated MCQA data for 5 tasks (pick_up_cup, push_button, put_rubbish_in_bin, phone_on_base, take_lid_off_saucepan). Use 50 demonstrations per task, frame interval t=10, ~7.5K QA pairs across waypoint prediction, current state prediction, and movement prediction. Training: batch size 128, 5 epochs, learning rate 1e-6, 5 rollout samples per prompt, temperature 1.0. Evaluate on Robot-R1 Bench spatial reasoning metric.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2506.00070"], "code": ["https://github.com/hiyouga/EasyR1", "https://github.com/2U1/Qwen2-VL-Finetune"], "dataset": ["https://github.com/stepjam/RLBench"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searched exact title, arXiv ID, Robot-R1/Robot R1/Robot_R1 variants with github/code/official, author handle kingdy2002, Robot-R1 Bench, and Hugging Face scoped queries. The author homepage lists the Robot-R1 paper but no code link. Search found only generic third-party infrastructure named in the paper, EasyR1 and Qwen2-VL-Finetune, not Robot-R1-specific data-generation, training, evaluation, prompt, or benchmark code. OpenReview supplement status in the input is missing."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "Verified RLBench at https://github.com/stepjam/RLBench is publicly available with full installation instructions, task definitions, and demonstration generation tools. The paper uses RLBench for generating expert demonstrations and extracting waypoints. However, the paper's processed MCQA training dataset with specific question-answer pairs and metadata is NOT released separately; it must be regenerated from RLBench using the paper's methodology."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searched Robot-R1 with Hugging Face/model/checkpoint/weights and GitHub variants. No downloadable Robot-R1-trained Qwen2.5-VL-7B checkpoint, LoRA adapter, or model card was found; only the base model and unrelated/generic resources appeared. OpenReview supplement status is missing and contains no checkpoint evidence."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "RLBench is a well-established public benchmark for robot learning (james2020rlbench, IEEE RAL 2020), widely used in the robotics community for imitation learning and RL evaluation with 100+ manipulation tasks."}}, "agent_task": "Reproduce the Robot-R1 MRE by: (1) Installing RLBench (pip install git+https://github.com/stepjam/RLBench.git) and generating training demonstrations for 5 tasks with 50 episodes each at 224x224 resolution, (2) Extracting waypoints using ARM repository, (3) Constructing MCQA training data using prompt templates from the paper appendix (waypoint prediction, current state prediction, movement prediction MCQA), (4) Fine-tuning Qwen2.5-7B-VL-Instruct using GRPO with EasyR1 framework (learning rate 1e-6, batch size 128, 5 epochs, 5 rollout samples per prompt, temperature 1.0, rollout batch size 512), (5) Evaluating spatial reasoning on Robot-R1 Bench and verifying improvement over base model (1.40 spatial score).", "h100_estimate": {"hours": 15.36, "basis_kind": "paper_reported", "gpu_count": 4, "gpu_type": "A100 80GB", "wallclock_hours": 12, "h100_equivalent_multiplier": 0.32, "basis": "Paper states 4x A100 80GB, ~12h wallclock for 5-epoch GRPO training on the 7.5K prompts; this is the only/smallest training config described (weights not released, so MRE = smallest training run, not eval-only). 4*12*0.32=15.36 H100-hrs. Arithmetic, multiplier, and scope verified correct; no double-counting or padding found."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 15.36, "h100_estimate_basis": "paper_reported: Paper states 4x A100 80GB, ~12h wallclock for 5-epoch GRPO training on the 7.5K prompts; this is the only/smallest training config described (weights not released, so MRE = smallest training run, not eval-only). 4*12*0.32=15.36 H100-hrs. Arithmetic, multiplier, and scope verified correct; no double-counting or padding found.", "h100_band": "8-32", "h100_recomputed_hours": 15.36, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 15.36, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "Qwen2.5-VL-7B-Instruct + GRPO training on RLBench MCQA tasks for 5 epochs, full fine-tuning with batch size 128, lr 1e-6, temperature 1.0, 5 samples per prompt", "metric": "Spatial reasoning score (GPT-4o-judged, 0-3 scale)", "value": "1.51", "scope": "Robot-R1 Bench - spatial reasoning task (in-distribution + out-of-distribution average)", "match_bar_kind": "point_estimate"}}
 {"custom_id": "2507.02064", "central_claim": "Hippocampal place cells auto-associate sensory inputs with grid cell patterns, enabling cue-triggered goal retrieval, grid-based path planning that generalizes to novel routes, and reconstruction of sensory experiences along planned trajectories.", "claim_evidence": "Abstract states 'place cells autoassociate sensory inputs with grid cell patterns, allowing sensory cues to trigger recall of goal-location grid patterns.' Section 3 and Figure 2/3 demonstrate recall of goal locations from partial sensory cues with L2 distance approaching zero over time.", "paper_kind": "empirical", "mre_config": "HC-MEC RNN with 288 grid cells (6 modules x 48 cells), 32 speed cells, 32 direction cells, 256 SMCs, 512 HPCs. Trained on random traversal in 100x100cm arena with masking ratio r_mask=0.5, for 50 epochs, batch_size=128, 1000 steps per trajectory. Recall tested by querying with SMC responses at fixed locations for 5 seconds (100 timesteps), decoding positions via nearest neighbor on ratemaps.", "verified_links": {"paper_or_project": ["https://zhaozewang.github.io/remi"], "code": ["https://github.com/grasp-lyrl/btnk_mae"], "dataset": ["https://github.com/facebookresearch/habitat-sim"], "weights": ["https://huggingface.co/zhaozewang56/btnk_mae"]}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "The REMI project page has a 'Code (Coming Soon)' link, but it currently opens zhaozewang/place_cells_episodic_rnn, whose README identifies it as the official implementation of the earlier NeurIPS 2024 'Time Makes Space' paper, not REMI; it has a demo and checkpoint for that older recurrent autoencoder. The separate RatatouGym repo is public and supplies a spatial-navigation simulator, and the btnk_mae repo supplies the vision encoder code, but neither contains the REMI HC-MEC plus planner training/evaluation code needed for the synthetic MRE. Searches for the exact title, arXiv ID, REMI HC-MEC GitHub, and zhaozewang REMI GitHub found no public REMI reproduction repository beyond these supporting/irrelevant artifacts."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE uses synthetic spatial navigation/cell-response data generated from the arena, trajectory, and sensory models rather than a fixed private dataset. RatatouGym is publicly documented as a gymnasium environment for simulating environmental geometries, sensory systems modeled through firing tunings, and traversal behaviors, and its GitHub repository is public with install instructions and source files. Thus the synthetic data source for the MRE is available, although exact author random seeds or generated trial files are not separately released."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No downloadable REMI HC-MEC or planner checkpoints were found. The project page's code link opens an older place_cells_episodic_rnn repository with a ckpts folder and latest.pth for the NeurIPS 2024 Time Makes Space project, not the REMI MRE. The btnk_mae repository provides Hugging Face pretrained BtnkMAE weights, but those are for the optional Habitat visual extension, not the synthetic HC-MEC/planner MRE. Searches for REMI checkpoints/weights/model files did not find MRE-relevant weights."}, "dataset_is_standard": {"value": false, "verification": "tool_verified", "evidence": "The MRE data are author-defined synthetic simulator outputs: custom 100 cm x 100 cm arena trajectories and simulated SMC/GC/speed/direction firing responses produced by RatatouGym-style procedures. This is not a widely recognized fixed benchmark such as ImageNet, CIFAR, GLUE, or Habitat challenge splits; it depends on paper-specific simulation choices."}}, "agent_task": "Implement the HC-MEC RNN model from the paper: a single-layer RNN with grid cells (6 modules x 48 cells, spatial scale factor sqrt(e), initial scale 30cm), speed cells (32), direction cells (32), SMCs (256, Gaussian white noise smoothed), and HPCs (512, no direct input/output). Train on random traversal trajectories in 100x100cm arena with masking (r_mask=0.5) for 50 epochs. After training, test cue-triggered recall by querying with SMC responses at fixed locations for 5 seconds, decode positions via nearest neighbor on aggregated ratemaps, and verify L2 distance approaches target location.", "h100_estimate": {"hours": 16, "basis_kind": "derived_from_config", "gpu_count": 1, "gpu_type": "A100 80GB", "wallclock_hours": 50, "h100_equivalent_multiplier": 0.32, "basis": "Paper specifies training HC-MEC model for 50 epochs with 1000 steps per trajectory, batch size 128, total hidden units 1120. Based on paper's supplements, single GPU training with AdamW optimizer. Estimated ~50 hours on 1x A100 80GB for the recall experiment MRE. Converted to H100-hours: 1 * 50 * 0.32 = 16 H100-hours. Note: Main REMI code is not released; engineering/debugging time to implement the model from scratch would be additional and substantial."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 16.0, "h100_estimate_basis": "derived_from_config: Paper specifies training HC-MEC model for 50 epochs with 1000 steps per trajectory, batch size 128, total hidden units 1120. Based on paper's supplements, single GPU training with AdamW optimizer. Estimated ~50 hours on 1x A100 80GB for the recall experiment MRE. Converted to H100-hours: 1 * 50 * 0.32 = 16 H100-hours. Note: Main REMI code is not released; engineering/debugging time to implement the model from scratch would be additional and substantial.", "h100_band": "8-32", "h100_recomputed_hours": 16.0, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 16.0, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "HC-MEC RNN, r_mask=0.5, recall task with SMC query at fixed location for 5s", "metric": "L2 distance from decoded position to target position (cm)", "value": "~0 cm after convergence (trajectories converge to goal representation)", "scope": "100x100cm simulated arena, nearest neighbor decoding on aggregated ratemaps", "match_bar_kind": "point_estimate"}}
 {"custom_id": "2511.02225", "central_claim": "FIOC-WM learns a two-level factorization (object-level with explicit interactions + attribute-level static/dynamic) that improves policy-learning sample efficiency and generalization over world-model baselines in object-centric reinforcement learning.", "claim_evidence": "Abstract states: 'FIOC-WM improves policy-learning sample efficiency and generalization over world-model baselines, indicating that explicit, modular interaction learning is crucial for robust control.' Table 2 (LPIPS reconstruction quality) and Table 3 / Table 4 (policy learning success rates in main paper and appendix) demonstrate FIOC outperforms Dreamer-V3, TD-MPC2, and EIT across Gym Fetch, Franka Kitchen, i-Gibson, and Libero tasks.", "paper_kind": "empirical", "mre_config": "Gym Fetch Task 1 with FIOC-WM Variational (Categorical), 2M training steps, DINO-v2 visual encoder + slot attention (ViT-Base, 5 slots, slot_dim=64) + VAE with static/dynamic factorization, interaction graph learning via variational masks with categorical distribution, state transition MLP (2 layers, 64 hidden), low-level MPC policy + high-level PPO policy.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2511.02225"], "code": [], "dataset": ["https://github.com/Lifelong-Robot-Learning/LIBERO"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No OpenReview supplement is bundled in the input. The paper itself says code and data will be publicly available after acceptance, not that they are available. Web searches for the exact title, arXiv ID 2511.02225, FIOC-WM, Factored Interactive Object-Centric World Model, author/name variants, GitHub-scoped queries, and Hugging Face-scoped queries did not find a public first-party FIOC-WM implementation. Search results found only the arXiv/OpenReview paper, paper-note pages, the project-page URL mentioned in the PDF, and unrelated/awesome-list entries; no runnable MRE code repository was verified."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "Gym Fetch environment is standard OpenAI Gym/gymnasium benchmark. Libero benchmark verified available at github.com/Lifelong-Robot-Learning/LIBERO. Franka Kitchen and i-Gibson are established public benchmarks. SpritesWorld requires custom setup but other MRE datasets are standard benchmarks."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "Searches for FIOC-WM weights/checkpoints on Hugging Face and the web, including queries for the title, FIOC-WM checkpoint, FIOC-WM weights, and method-name variants, found no downloadable first-party FIOC-WM checkpoints or policy/world-model weights. The paper discusses pretrained visual encoders such as DINO-v2/R3M, but those are external base encoders, not MRE-trained FIOC-WM weights."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "Gym Fetch is a standard OpenAI Gym benchmark widely used in RL research. Libero, Franka Kitchen, and i-Gibson are established public benchmarks in robot learning community."}}, "agent_task": "Reproduce the FIOC-WM Gym Fetch Task 1 experiment (success rate 0.95): (1) Implement FIOC-WM from scratch: DINO-v2 ViT-Base visual encoder, slot attention for object decomposition (5 slots, slot_dim=64, 3 clustering iterations), VAE encoder/decoder (2-layer MLP, 256 hidden) mapping to latent states with dynamics dim=10 and attribute dim=8, static/dynamic state factorization with temporal consistency loss (Eq. 1) and contrastive loss (Eq. 2), interaction graph learning via variational masks with categorical distribution (GRU encoder, 3-layer MLP 256 hidden, temperature parameter), state transition MLP (2 layers, 64 hidden), reward decoder. (2) Offline training: collect 2000 episodes using pre-trained Dreamer-v3 policies on Gym Fetch. Loss weights alpha=1, beta=0.05, gamma=0.1, eta=0.2; lr=3e-4; 64x64 image crop, patch size 16. (3) Learn low-level policy via MPC/CEM (lr=5e-5, MSE loss) or PPO (lr=3e-4, clip=0.1, hidden [256,256], GAE=0.95, entropy=0.1). (4) Learn high-level policy via PPO (lr=1e-4, batch 1024, diversity reward r_div=1/sqrt(|G_visited|)). (5) Online policy learning for 2M steps on Gym Fetch Task 1 (2-push). (6) Report success rate over 5-10 random seeds. Target: 0.95 ± 0.03.", "h100_estimate": {"hours": 18, "basis_kind": "paper_reported", "gpu_count": 6, "gpu_type": "RTX 4090", "wallclock_hours": 8, "h100_equivalent_multiplier": 0.375, "basis": "Paper appendix reports Fetch experiments run for 8 hours on 6x NVIDIA 4090 GPUs. RTX 4090 estimated at ~0.375 H100 equivalent based on relative performance (between L40 at 0.18 and L40S at 0.37, with RTX 4090 being faster). 6 GPUs * 8 hours * 0.375 = 18 H100-hours. Engineering time to reimplement FIOC-WM architecture from paper description is additional and substantial."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 18.0, "h100_estimate_basis": "paper_reported: Paper appendix reports Fetch experiments run for 8 hours on 6x NVIDIA 4090 GPUs. RTX 4090 estimated at ~0.375 H100 equivalent based on relative performance (between L40 at 0.18 and L40S at 0.37, with RTX 4090 being faster). 6 GPUs * 8 hours * 0.375 = 18 H100-hours. Engineering time to reimplement FIOC-WM architecture from paper description is additional and substantial.", "h100_band": "8-32", "h100_recomputed_hours": 18.0, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 18.0, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "FIOC-Variational(Cat.), Gym Fetch Task 1, 2M training steps, DINO-v2 + Slot Attention + VAE with static/dynamic factorization + interaction graph modeling via variational masks with categorical distribution, 2000 offline episodes from Dreamer-v3 pre-trained policies", "metric": "policy learning success rate", "value": "0.95 ± 0.03", "scope": "OpenAI Gym Fetch Task 1 (2-push), evaluated over 5-10 random seeds", "match_bar_kind": "point_estimate"}}
 {"custom_id": "2411.06890", "central_claim": "SPARTAN learns accurate local causal graphs (low SHD) through sparsity regularisation while maintaining Transformer-level prediction accuracy, enabling improved few-shot adaptation and robustness.", "claim_evidence": "Abstract states SPARTAN 'learns sparse, context-dependent interaction graphs that accurately predict future object states.' Table 1 (prediction and graph) shows SPARTAN achieving SHD of 1.51 vs 6.37 for Transformer on Interventional Pong while maintaining comparable prediction error (8.60 vs 8.83).", "paper_kind": "empirical", "mre_config": "Train SPARTAN on Interventional Pong with token_dim=32, embed_dim=512, 3 transformer layers, MLP hidden_dim=512, lr=5e-5, Adam optimizer, and Lagrangian relaxation for sparsity regularisation. Evaluate learned causal graph SHD against ground truth.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2411.06890", "https://anonymous.4open.science/r/STRAW_neurips-4B75"], "code": [], "dataset": ["https://github.com/phlippe/CITRIS", "https://github.com/waymo-research/waymo-open-dataset"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No MRE-relevant public implementation was verified. Searches for the full title, arXiv ID 2411.06890, SPARTAN sparse transformer world model, GitHub-scoped variants, author/name variants, and the commented anonymous 4open.science STRAW_neurips URL did not reveal a usable first-party repository. OpenReview supplement status in the prompt is missing; the paper text only contains a commented anonymous repository URL, not a verified released artifact."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "Interventional Pong dataset available via CITRIS GitHub repo (phlippe/CITRIS) with data_generation_interventional_pong.py script and Zenodo download link (BSD 3-Clause license). CREATE simulator from Jain et al. 2020 available per paper (MIT License). Waymo Open Dataset verified at waymo.com/open with official GitHub at waymo-research/waymo-open-dataset (public download after sign-in)."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No downloadable SPARTAN checkpoints or MRE-relevant pretrained weights were found on OpenReview, NeurIPS/PMLR-style pages, MPI project/publication pages, GitHub searches, or Hugging Face-style artifact searches. The prompt's OpenReview supplement is missing and no external model-card or release page was verified."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "Waymo Open Dataset (waymo-research/waymo-open-dataset) is a widely recognized public autonomous driving benchmark for motion prediction, used extensively in research. Interventional Pong from CITRIS is a standard benchmark in causal representation learning literature, explicitly cited as such in the paper."}}, "agent_task": "Reproduce the SPARTAN local causal graph learning experiment on Interventional Pong: (1) Obtain Interventional Pong dataset from CITRIS GitHub repository (phlippe/CITRIS), (2) Implement SPARTAN architecture with Bernoulli-sampled adjacency matrices, Gumbel softmax trick, path matrix computation across layers, and Lagrangian relaxation training from paper equations, (3) Train with hyperparameters: token_dim=32, embed_dim=512, 3 transformer layers, MLP hidden_dim=512, lr=5e-5, Adam, (4) Extract learned adjacency matrices per timestep, (5) Compute Structural Hamming Distance (SHD) against ground-truth causal graph provided with dataset, (6) Target SHD near 1.51 to verify causal discovery accuracy.", "h100_estimate": {"hours": 23.04, "basis_kind": "paper_reported", "gpu_count": 1, "gpu_type": "V100 32GB", "wallclock_hours": 72, "h100_equivalent_multiplier": 0.32, "basis": "Weights not released -> MRE is smallest training run: train SPARTAN on Interventional Pong. Paper states convergence within 3 days (72h) on a single V100/RTX6000 GPU. 1 GPU x 72h x 0.32 (V100 closest architectural analogue to A100 in table) = 23.04 H100-hours. No double-counting, unit, or scope errors; confirmed correct."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 23.04, "h100_estimate_basis": "paper_reported: Weights not released -> MRE is smallest training run: train SPARTAN on Interventional Pong. Paper states convergence within 3 days (72h) on a single V100/RTX6000 GPU. 1 GPU x 72h x 0.32 (V100 closest architectural analogue to A100 in table) = 23.04 H100-hours. No double-counting, unit, or scope errors; confirmed correct.", "h100_band": "8-32", "h100_recomputed_hours": 23.04, "h100_arithmetic_mismatch": false, "h100_needs_human_review": false, "score": 3, "tier": "Hard", "audited_h100_hours": 23.04, "h100_hours_adjudicated": false, "selection_band": "8-32", "split": "eval", "match_target": {"config": "SPARTAN on Interventional Pong: token_dim=32, embed_dim=512, 3 layers, MLP hidden_dim=512, lr=5e-5, Adam, Lagrangian relaxation", "metric": "Structural Hamming Distance (SHD) of learned causal graph vs ground truth", "value": "1.51", "scope": "Interventional Pong test set", "match_bar_kind": "direction"}}
 {"custom_id": "2511.19808", "central_claim": "RLNLC formulates noisy label correction as a reinforcement learning problem and achieves state-of-the-art performance on multiple benchmarks by learning a policy network via actor-critic methods to iteratively correct noisy training labels.", "claim_evidence": "Abstract: 'RLNLC learns a deep feature representation based policy network to perform label correction through reinforcement learning, utilizing an actor-critic method.' Table 1: RLNLC achieves 97.3% on CIFAR10-IDN (0.20 noise) and 95.8% on CIFAR10-IDN (0.50 noise), outperforming all comparison methods.", "paper_kind": "empirical", "mre_config": "ResNet-34 backbone, CIFAR10-IDN with 0.20 noise rate, 500 epochs policy training + 100 epochs fine-tuning, batch size 128, SGD with momentum 0.9, initial LR 0.01 (reduced to 0.001 at epoch 250), 50 epoch warmup, T=10 trajectory length, k=10 neighbors, N_b=100 bins, lambda=0.5, tau=0.5, gamma=0.9.", "verified_links": {"paper_or_project": ["https://arxiv.org/abs/2511.19808"], "code": [], "dataset": ["https://www.cs.toronto.edu/~kriz/cifar.html"], "weights": []}, "signals": {"code_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No first-party MRE code was found. The OpenReview page for the paper is public but shows only the paper metadata/PDF and no code or supplementary artifact; the provided OpenReview supplement status is missing. Searches for the exact title, arXiv ID 2511.19808, RLNLC, 'RLNLC noisy label correction github', 'Learning to Clean Noisy Label github', and GitHub/Hugging Face scoped variants did not find a relevant public implementation for this paper."}, "dataset_available": {"value": true, "verification": "tool_verified", "evidence": "The MRE uses CIFAR-10 with synthetic symmetric label noise. CIFAR-10 is publicly available through standard dataset loaders; the PyTorch torchvision documentation verifies torchvision.datasets.CIFAR10 and its download option. The 50% symmetric-noise labels can be generated from the standard CIFAR-10 training labels by random class flips, so no unreleased author dataset is required for this MRE."}, "weights_available": {"value": false, "verification": "tool_searched_not_found", "evidence": "No downloadable RLNLC checkpoints or ResNet-18 CIFAR-10 symmetric-noise weights were found. The OpenReview page and arXiv page do not expose model files, the bundled supplement is missing, and searches for RLNLC with checkpoint/weights/model/Hugging Face/GitHub variants did not find MRE-relevant released weights."}, "dataset_is_standard": {"value": true, "verification": "tool_verified", "evidence": "CIFAR-10 is a widely recognized public computer-vision benchmark, and PyTorch torchvision provides a standard CIFAR10 dataset class with public download support. The MRE's synthetic symmetric noise is a generated perturbation of this standard benchmark rather than a custom dataset requiring private annotations."}}, "agent_task": "Reproduce RLNLC on CIFAR10-IDN with 0.20 noise rate using ResNet-34: (1) Download CIFAR-10 from official source and inject instance-dependent noise at 0.20 rate per xia2020part protocol, (2) Implement RLNLC actor-critic framework with policy network based on deep feature extraction, k-NN attention mechanism for label prediction, and the composite reward function (LCR + NLA), (3) Train policy for 500 epochs with T=10 trajectory length, then deploy for T'=25 label correction steps, (4) Fine-tune prediction model on cleaned labels for 100 epochs, (5) Evaluate on CIFAR-10 test set. Target: achieve test accuracy near 97.3%.", "h100_estimate": {"hours": 32.4, "basis_kind": "compute_unspecified", "gpu_count": 1, "gpu_type": "H100 SXM", "wallclock_hours": 32.4, "h100_equivalent_multiplier": 1.0, "basis": "Weights not released -> MRE = smallest training run (single ResNet-34 on CIFAR10-IDN 0.20, policy RL training + fine-tuning), no full-paper inflation. Original basis's own component breakdown summed to 3.3+25+0.8+3.3=32.4h but the record reported 41h (internal inconsistency); corrected to the internally-consistent sum. No GPU/time stated by paper, so basis remains compute_unspecified with human review."}, "verification_status": "verified", "web_verification": "available", "exit_reason": "natural", "h100_hours_estimate": 32.4, "h100_estimate_basis": "Weights not released -> MRE = smallest training run (single ResNet-34 on CIFAR10-IDN 0.20, policy RL training + fine-tuning), no full-paper inflation. Original basis's own component breakdown summed to 3.3+25+0.8+3.3=32.4h but the record reported 41h (internal inconsistency); corrected to the internally-consistent sum. No GPU/time stated by paper, so basis remains compute_unspecified with human review.", "h100_band": "32-96", "h100_recomputed_hours": 32.4, "h100_arithmetic_mismatch": false, "h100_needs_human_review": true, "score": 3, "tier": "Hard", "audited_h100_hours": 32.4, "h100_hours_adjudicated": false, "selection_band": "32-96", "split": "eval", "match_target": {"config": "CIFAR10-IDN, 0.20 noise rate, ResNet-34 backbone, 500 epochs policy training + 100 epochs fine-tuning, standard hyperparameters (k=10, N_b=100, lambda=0.5)", "metric": "Test accuracy (%)", "value": "97.3", "scope": "CIFAR10-IDN test set", "match_bar_kind": "point_estimate"}}