{"cpu_cores": "4", "evaluation_dataset": "HarmBench behaviors_text_all", "evaluation_sample_count": "400", "evaluation_split": "all 400 released text behaviors in canonical CSV order", "frozen_scale": "unchanged", "gold_metric_sha256": "4cf8fe132ae6a9d3043576a25a6369f882b739a983351a70b8db921b52d222d6", "gpu_class": "NVIDIA_H100_80GB_OR_EQUIVALENT", "gpu_count": "1", "memory_gb": "32", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "paper-selected full scale", "paper_title": "The Devil behind the mask: An emergent safety vulnerability of Diffusion LLMs", "project_id": "013-2026.iclr.10007159", "reference_value": "96.3", "repository_commit": "56ba341b3377423e6da212013aec6e825710a64c", "repository_url": "https://github.com/ZichenWen1/DIJA", "scale_class": "full_scale", "screening_domain": "llm_foundation_models", "storage_gb": "90", "target_contract_sha256": "8763a6bfdf5a3d9229cef2956c32beae3d081e7f44f0b9ed2fc94fa0e0adc3a3", "target_metric": "keyword-based attack success rate (ASR-k)", "target_task": "DIJA jailbreak attack generation against LLaDA-8B-Instruct on all HarmBench text behaviors", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "CodeMixBench CM-MMLU zh-en", "evaluation_sample_count": "1,133", "evaluation_split": "complete mmlu_chieng test file, 1,133 examples", "frozen_scale": "unchanged", "gold_metric_sha256": "23b81270dfa51727e2485c312f548c1cf46bde2326e378d270e4b372f51c91e6", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "paper-selected full scale", "paper_title": "CodeMixBench: Evaluating Code-Mixing Capabilities of LLMs Across 18 Languages", "project_id": "016-2025.emnlp.109", "reference_value": "47.4", "repository_commit": "1f3c9bb3ea356e9240389b4a6abbc3b9f8bffdaf", "repository_url": "https://github.com/Jeromeyluck/CodeMixBench", "scale_class": "full_scale", "screening_domain": "llm_foundation_models", "storage_gb": "90", "target_contract_sha256": "79aa2cae76b488fa6622abe140c3f8417068e8fe34922460eb7656ba672bcf27", "target_metric": "accuracy", "target_task": "One-shot Chinese-English code-mixed MMLU multiple-choice answering", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "WinoBias", "evaluation_sample_count": "160", "evaluation_split": "dev", "frozen_scale": "unchanged", "gold_metric_sha256": "2651f828bc6818367f999e53860ee65200de64e5b89d5b0f70671d493a5aca02", "gpu_class": "NVIDIA_H100_80GB_OR_EQUIVALENT", "gpu_count": "1", "memory_gb": "16", "metric_direction": "maximize", "metric_unit": "ratio", "original_scale": "paper-selected full scale", "paper_title": "Multi-component Causal Tracing in Large Language Models", "project_id": "017-2026.acl-long.154", "reference_value": "0.115", "repository_commit": "bc7777b8f87afb155bc489ce582cafd0cb360596", "repository_url": "https://github.com/ZiruiYan/multi-component-causal-tracing", "scale_class": "full_scale", "screening_domain": "llm_foundation_models", "storage_gb": "2", "target_contract_sha256": "e220163da08f4260a1d4e85ba4e8d481ad4ab30518aebda7383fa5cf27bbadee", "target_metric": "average causal-tracing metric l(D,m)", "target_task": "Sparse multi-component causal tracing over GPT-2 attention heads", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "Compositional-ARC", "evaluation_sample_count": "8,546", "evaluation_split": "complete systematicity test split, 8,546 episodes", "frozen_scale": "train episodes=4,096; epochs=40", "gold_metric_sha256": "e730b21b15cd14f4211cd7ab9af898912175a5493fa68fa7e60e87ea33139cec", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "train episodes=82,908; epochs=300", "paper_title": "Compositional-ARC: Assessing Systematic Generalization in Abstract Spatial Reasoning", "project_id": "018-2026.iclr.10008090", "reference_value": "78.26", "repository_commit": "ff0a7d16c678169e5248f46bb45d28d053d0f841", "repository_url": "https://github.com/mainlp/C-ARC", "scale_class": "reduced_pipeline_full_target", "screening_domain": "llm_foundation_models", "storage_gb": "90", "target_contract_sha256": "cbe7fa63388de7e773599da3a00f6b44e4334540b2b665b71638d6543007380f", "target_metric": "exact match accuracy", "target_task": "MLC systematic generalization to held-out visual transformation compositions", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "M3CoT", "evaluation_sample_count": "2,318", "evaluation_split": "test", "frozen_scale": "unchanged", "gold_metric_sha256": "729973939783f196f654ccae3ee0bc1355d47242b951cc2bb5c779f6e85704a0", "gpu_class": "NVIDIA_H100_80GB_OR_EQUIVALENT", "gpu_count": "1", "memory_gb": "16", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "paper-selected full scale", "paper_title": "SinkTrack: Attention Sink based Context Anchoring for Large Language Models", "project_id": "019-2026.iclr.10010470", "reference_value": "48.25", "repository_commit": "b6ba69105904b77bee810eb324b20103cbc33b0f", "repository_url": "https://github.com/67L1/SinkTrack", "scale_class": "full_scale", "screening_domain": "llm_foundation_models", "storage_gb": "30", "target_contract_sha256": "8ef316f813789fdf4c523a8d5676b0079dbcb68f3584b17ce2317ba9355aa243", "target_metric": "accuracy", "target_task": "Training-free multimodal multiple-choice reasoning with SinkTrack context anchoring", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "GLOT diagnostic relational needle stress test", "evaluation_sample_count": "2,000", "evaluation_split": "2,000-example generated test split at 90% distractors", "frozen_scale": "train examples=1,000", "gold_metric_sha256": "810dc58308213463ac3da70e0bcb06989eb59e8d9c997badde5f96e0a6793e3f", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "train examples=10,000", "paper_title": "Towards Improved Sentence Representations using Token Graphs", "project_id": "026-2026.iclr.10007029", "reference_value": "98.8", "repository_commit": "59f897e1c8be735fd99c931d76886ac23dd0ec34", "repository_url": "https://github.com/ipsitmantri/GLOT", "scale_class": "reduced_pipeline_full_target", "screening_domain": "llm_foundation_models", "storage_gb": "90", "target_contract_sha256": "d12d74e85d2487d56b82116b5701dcd6afbd44141c901fb5aa780ce824c91c24", "target_metric": "accuracy", "target_task": "Relational signal classification with 90% distractor tokens", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "AIME 2024", "evaluation_sample_count": "30", "evaluation_split": "train split; all 30 questions", "frozen_scale": "train rows=256; rollouts=129; updates=65", "gold_metric_sha256": "71f76897feca8681b1ca98957376eb9841e2bda71965fc2cccf0b2a83a8a886c", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "64", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "train rows=40,315; rollouts=256; updates=256", "paper_title": "Prosperity before Collapse: How Far Can Off-Policy RL Reach with Stale Data on LLMs?", "project_id": "038-2026.iclr.10010317", "reference_value": "14", "repository_commit": "af54a3e8feffb7a66a8258003aebefa37759cec0", "repository_url": "https://github.com/Infini-AI-Lab/M2PO", "scale_class": "reduced_pipeline_full_target", "screening_domain": "llm_foundation_models", "storage_gb": "90", "target_contract_sha256": "319b0a9c9c03bac69c4d43fa76daa27b66142c509817b08540069c685596ed97", "target_metric": "pass@1 averaged over 16 samples", "target_task": "Mathematical problem solving", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "Unified-Feedback", "evaluation_sample_count": "1,000", "evaluation_split": "paper 1,000-example evaluation split", "frozen_scale": "epochs=1", "gold_metric_sha256": "ac8eb7aedc4da74cbf9a6acf48076d516862e41101cbd55a2be5d8ab134f6bb8", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "epochs=2", "paper_title": "APLOT: Robust Reward Modeling via Adaptive Preference Learning with Optimal Transport", "project_id": "040-2025.emnlp.281", "reference_value": "73.84", "repository_commit": "8b3e79a2266959be4efd17bd95f6e80e10a9274e", "repository_url": "https://github.com/BIRlz/APLOT", "scale_class": "reduced_pipeline_full_target", "screening_domain": "llm_foundation_models", "storage_gb": "90", "target_contract_sha256": "f00a4eada6060a81e6dc5a2dc7d88bd7de9d5aaf730468b68ad5e8e9c07db710", "target_metric": "pairwise accuracy", "target_task": "Pairwise reward-model preference prediction", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "RAGTruth QA (MS MARCO)", "evaluation_sample_count": "1,175", "evaluation_split": "QA/MS MARCO; five 25%-test splits with 100 validation examples", "frozen_scale": "responses=939", "gold_metric_sha256": "80ece72a9dd2272b8061885a9234e0dc51b3fabeb2d6827e99b5852564bf8771", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "64", "metric_direction": "maximize", "metric_unit": "ROC-AUC", "original_scale": "responses=989", "paper_title": "Hallucination Detection in LLMs with Topological Divergence on Attention Graphs", "project_id": "044-2026.acl-long.704", "reference_value": "0.76", "repository_commit": "d15709d005a9db0de22df4d81a2c56f97c27ee89", "repository_url": "https://github.com/sb-ai-lab/TOHA", "scale_class": "reduced_pipeline_full_target", "screening_domain": "llm_foundation_models", "storage_gb": "90", "target_contract_sha256": "058076825b2677313dc8058e5af1dd2dfd39992cdfebe069108e8624b5433b86", "target_metric": "mean test ROC-AUC over five splits", "target_task": "Hallucination detection from attention-graph topology", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "Plan-RewardBench", "evaluation_sample_count": "1,171", "evaluation_split": "all seven released evaluation splits", "frozen_scale": "unchanged", "gold_metric_sha256": "28f203f70f9f68f29f74f39fcfab233f7ed48ea80a945a194c58dc1c8f221bd4", "gpu_class": "NVIDIA_H100_80GB_OR_EQUIVALENT", "gpu_count": "1", "memory_gb": "32", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "paper-selected full scale", "paper_title": "Aligning Agents via Planning: A Benchmark for Trajectory-Level Reward Modeling", "project_id": "047-2026.acl-long.1062", "reference_value": "65.88", "repository_commit": "ffe6fbe923bddf6deea54f92d3ffe2f45f38de54", "repository_url": "https://github.com/wyy-1112/Plan-RewardBench", "scale_class": "full_scale", "screening_domain": "agents_tool_use", "storage_gb": "20", "target_contract_sha256": "3d4f4f1be6341936a5ee2e95449651c466d4a55c756087c7c00e3d372d86c897", "target_metric": "macro pairwise accuracy across seven splits", "target_task": "Pairwise preference evaluation of complete tool-using agent trajectories", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "ALFWorld", "evaluation_sample_count": "274", "evaluation_split": "all 140 valid_seen and 134 valid_unseen tasks", "frozen_scale": "SFT rows=512; RL tasks=4; updates=1", "gold_metric_sha256": "9cc4760066455f2636875d4a3d98034de06f32d8e6f2dfb37af007b213fdb3fa", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "64", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "SFT rows=7,838; RL tasks=500; updates=125", "paper_title": "DPEPO: Diverse Parallel Exploration Policy Optimization for LLM-based Agents", "project_id": "052-2026.acl-long.2151", "reference_value": "94.1", "repository_commit": "4edaa1609869e8f2abde14a5f15a0156d6bca743", "repository_url": "https://github.com/LePanda026/Code-for-DPEPO", "scale_class": "reduced_pipeline_full_target", "screening_domain": "agents_tool_use", "storage_gb": "90", "target_contract_sha256": "bed66cbe7042df645f62c9c0cc063a2774ad1eb21aa889aaf5fcd71dbe29eaef", "target_metric": "seen/OOD macro-average success rate", "target_task": "Embodied household task completion with parallel LLM-agent exploration", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "EgoSchema Subset", "evaluation_sample_count": "500", "evaluation_split": "Subset/test; all 500 public-label questions", "frozen_scale": "unchanged", "gold_metric_sha256": "91c25708799f6cfe7b2efb059e834680ba3963063df6fc8805680c70ec2e2c79", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "64", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "paper-selected full scale", "paper_title": "FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token Merging", "project_id": "069-2026.iclr.10010430", "reference_value": "63.4", "repository_commit": "983cce6e30d7a8012442bfc7557d3afa61b3572d", "repository_url": "https://github.com/Fanziyang-v/FlashVID", "scale_class": "full_scale", "screening_domain": "computer_vision", "storage_gb": "90", "target_contract_sha256": "e15f0782a6b83a3f48d61b9f5b512c753f7782a0fe6794fff4f3588948d60f01", "target_metric": "multiple-choice accuracy", "target_task": "Long-video multiple-choice question answering", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "GeoBenchMeta", "evaluation_sample_count": "5,677", "evaluation_split": "GeoBench 2D edits", "frozen_scale": "unchanged", "gold_metric_sha256": "9c13681213e4bcc07f8cc28c078b7d395952e3ef596dec3f8769d2d4ca71eec5", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "64", "metric_direction": "minimize", "metric_unit": "unitless", "original_scale": "paper-selected full scale", "paper_title": "Training-free Geometric Image Editing on Diffusion Models", "project_id": "078-2025.iccv.Zhu_Training-free_Geometric_Image_Editing_on_Diffusion_Models_ICCV_2025", "reference_value": "0.055", "repository_commit": "4c9fdb971572b32edbeac13464659274c28decbb", "repository_url": "https://github.com/CIawevy/FreeFine", "scale_class": "full_scale", "screening_domain": "computer_vision", "storage_gb": "90", "target_contract_sha256": "cb70b8c791ffdf40e8e16e0e9c595804d9800c1ca08655b75d5680641bccba49", "target_metric": "Warp Error", "target_task": "2D geometric image editing", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "Mind's Eye diagnostic benchmark", "evaluation_sample_count": "100", "evaluation_split": "mental_composition diagnostic partition, all 100 items", "frozen_scale": "task families=1", "gold_metric_sha256": "7049002b953c1ee4f3dfe353a529c210d3d16d168c13040b4e92df6a35728b14", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "task families=8", "paper_title": "Mind's Eye: A Benchmark of Visual Abstraction, Transformation and Composition for Multimodal LLMs", "project_id": "083-2026.acl-long.2124", "reference_value": "27.9", "repository_commit": "c5e1b6fbdb3843681dd33a174d7533951e85c487", "repository_url": "https://github.com/microsoft/Mind-s-Eye", "scale_class": "reduced_pipeline_full_target", "screening_domain": "multimodal_learning", "storage_gb": "90", "target_contract_sha256": "0ada348671953a76c5cd786d592a14e67dd41b0e8589329694b9db8b36b1cf27", "target_metric": "accuracy", "target_task": "Mental composition multiple-choice visual reasoning", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "POPE on COCO", "evaluation_sample_count": "9,000", "evaluation_split": "adversarial + popular + random, 3,000 questions each", "frozen_scale": "unchanged", "gold_metric_sha256": "9086e501ccc12cd75105b0df211b5f69465a61fe7c80dd20490cc47a6caad95c", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "32", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "paper-selected full scale", "paper_title": "Through the Magnifying Glass: Adaptive Perception Magnification for Hallucination-Free VLM Decoding", "project_id": "084-2026.acl-long.2059", "reference_value": "87.68", "repository_commit": "7886fa075e27f723cd89697ea56a6a7ec5a03d52", "repository_url": "https://github.com/ShunqiM/PM", "scale_class": "full_scale", "screening_domain": "multimodal_learning", "storage_gb": "90", "target_contract_sha256": "c4f32a6377a0c6b04fd0284d90aadca4a7c1893b68d8dbd3dd86ef0e0cda7e19", "target_metric": "Macro accuracy", "target_task": "object-hallucination evaluation", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "MMSP", "evaluation_sample_count": "10", "evaluation_split": "official MMSP test windows for the first 10 plants", "frozen_scale": "epochs=10", "gold_metric_sha256": "534ba02e65b2ed294a7c92aa8a141d7828481bac4ca411b240bd66d81ab44863", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "minimize", "metric_unit": "normalized_absolute_error", "original_scale": "epochs=100", "paper_title": "UniCA: Unified Covariate Adaptation for Time Series Foundation Model", "project_id": "085-2026.iclr.10010335", "reference_value": "0.193", "repository_commit": "727d3670e9b7993472b19f47e42ac0d75142b127", "repository_url": "https://github.com/hanlu-nju/UniCA", "scale_class": "reduced_pipeline_full_target", "screening_domain": "multimodal_learning", "storage_gb": "90", "target_contract_sha256": "40df89c22476be913acaca1e440abe34acae2c04046aabc277b0c76af45d3685", "target_metric": "MAE", "target_task": "Solar-power forecasting with image and numerical-weather covariates", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "TextVQA", "evaluation_sample_count": "5,000", "evaluation_split": "validation, all 5,000 questions", "frozen_scale": "unchanged", "gold_metric_sha256": "bbd77cd610f2f3ba797422d4e97f1d7ee1c8f7fa546c72f38b5233f456da3b07", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "32", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "paper-selected full scale", "paper_title": "PPE: Positional Preservation Embedding for Token Compression in Multimodal Large Language Models", "project_id": "090-2026.iclr.10009775", "reference_value": "73.78", "repository_commit": "7b6aa7d211f6e40fca53843d51e087067b3de0a6", "repository_url": "https://github.com/MouxiaoHuang/PPE", "scale_class": "full_scale", "screening_domain": "multimodal_learning", "storage_gb": "90", "target_contract_sha256": "329970990a3bc26bb9aa9b8bbdbbcaec52e238a379f5891ca1b0a30a94cc2466", "target_metric": "Accuracy", "target_task": "visual question answering", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "GeoNames cities1000-derived US city-pair task", "evaluation_sample_count": "175,000", "evaluation_split": "175,000 generated validation pairs", "frozen_scale": "training pairs=70,000; epochs=10", "gold_metric_sha256": "1a527ca49c7456ae057449c0104bf1194c5b905e1b2fa49f15d4f961eaecaae1", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "training pairs=700,000; epochs=40", "paper_title": "From Data Statistics to Feature Geometry: How Correlations Shape Superposition", "project_id": "097-2026.iclr.10011287", "reference_value": "97.94", "repository_commit": "d45878ef91b30e27b38a92f47a99beae40a9992a", "repository_url": "https://github.com/LucasPrietoAl/correlations-feature-geometry", "scale_class": "reduced_pipeline_full_target", "screening_domain": "other_nlp", "storage_gb": "90", "target_contract_sha256": "32187e5abc363289e2e5db59afe6bcd076962b521395c55185fa0da6a436e298", "target_metric": "accuracy", "target_task": "Eight-way relative city-direction classification", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "CIFAR-10", "evaluation_sample_count": "10,000", "evaluation_split": "complete 10,000-image test set after five sequential two-class tasks", "frozen_scale": "epochs/task=5; seeds=1", "gold_metric_sha256": "3da2cbedc99d92511bf393c695982301d23ec108a25266c0445e890f487e23a3", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "epochs/task=50; seeds=5", "paper_title": "IDER: IDempotent Experience Replay for Reliable Continual Learning", "project_id": "098-2026.iclr.10009097", "reference_value": "71.02", "repository_commit": "18abadc8477c82fe384847493948acdc81004dc2", "repository_url": "https://github.com/YutingLi0606/Idempotent-Continual-Learning", "scale_class": "reduced_pipeline_full_target", "screening_domain": "other_nlp", "storage_gb": "90", "target_contract_sha256": "873dc9185a424d7c31625a05a223c8763a5612b6d8088969f8db3489a14c80b3", "target_metric": "Final Average Accuracy", "target_task": "Five-task class-incremental learning on Split CIFAR-10", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "GSM8K", "evaluation_sample_count": "1,319", "evaluation_split": "main/test, all 1,319 examples", "frozen_scale": "unchanged", "gold_metric_sha256": "8231df4089400b7a6ef7cc5478bb926d50ee504a0c90685396415a7809975bc4", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "32", "metric_direction": "maximize", "metric_unit": "tokens/step", "original_scale": "paper-selected full scale", "paper_title": "Overcoming Joint Intractability with Lossless Hierarchical Speculative Decoding", "project_id": "099-2026.iclr.10010023", "reference_value": "6.3", "repository_commit": "b8575e2ad42bcf2b32d0c6f3481eb49099453f33", "repository_url": "https://github.com/ZhouYuxuanYX/Hierarchical-Speculative-Decoding", "scale_class": "full_scale", "screening_domain": "other_nlp", "storage_gb": "90", "target_contract_sha256": "569621ec5568049dab60465aa43a5cd1a8bfe6af893bb67418d1775807d9e5a5", "target_metric": "Block Efficiency", "target_task": "hierarchical speculative decoding", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "Processed Single-cell RNA Time Series Data, Multiome", "evaluation_sample_count": "2", "evaluation_split": "Multiome intermediate days 3 and 4, each held out in cross-validation", "frozen_scale": "seeds=1", "gold_metric_sha256": "2594dd6219b7f21fd10a5583806036fb78ebffa41db02f17cfb4a3ed5a70783a", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "minimize", "metric_unit": "distance", "original_scale": "seeds=3", "paper_title": "Learning Explicit Single-Cell Dynamics Using ODE Representations", "project_id": "102-2026.iclr.10010728", "reference_value": "0.742", "repository_commit": "ce2273e5fe6230035158c6a5bef09c8de5c0d20e", "repository_url": "https://github.com/czi-ai/cell-mnn", "scale_class": "reduced_pipeline_full_target", "screening_domain": "other_nlp", "storage_gb": "90", "target_contract_sha256": "1cccdd3515f0a092ba81ff2b48d9e6fbf6370c66781456a6b3db676b54fe3959", "target_metric": "1-Wasserstein distance (EMD)", "target_task": "Single-cell distribution interpolation at held-out time points", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "TextTopoOOD Cora", "evaluation_sample_count": "3,708", "evaluation_split": "Cora 1,000 ID test nodes and all 2,708 Feature Shift 0.90 OOD nodes", "frozen_scale": "datasets=1; seeds=1", "gold_metric_sha256": "fcae5365677fa9a7d419497bcd73cbe458cf26881f44abb6a676e4e0d44f1829", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "datasets=11; seeds=3", "paper_title": "Text Meets Topology: Rethinking Out-of-distribution Detection in Text-Rich Networks", "project_id": "103-2025.emnlp.280", "reference_value": "94.3", "repository_commit": "97f52ac21c78913dc21093590ac8cffd09529f58", "repository_url": "https://github.com/DannyW618/TNT", "scale_class": "reduced_pipeline_full_target", "screening_domain": "other_nlp", "storage_gb": "90", "target_contract_sha256": "bd0551eb6170b84232489ac06df79f8235f498c7943bbfe8168b3881450e1093", "target_metric": "AUROC", "target_task": "out-of-distribution detection in a text-rich citation network", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "SynLogic-Easy", "evaluation_sample_count": "270", "evaluation_split": "complete easy validation split", "frozen_scale": "train rows=1,024; model=Qwen2.5-1.5B Base; train steps=32", "gold_metric_sha256": "ff47791789021eda9b88b9542f4a55a3beac3e47140a8b5a6a429637c33304d7", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "train rows=15,837; model=Qwen2.5-7B Base; train steps=300", "paper_title": "Mirage or Method? How Model-Task Alignment Induces Divergent RL Conclusions", "project_id": "104-2026.iclr.10011413", "reference_value": "42.6", "repository_commit": "aa9ef253f39a4202594ae7e9bb7176ec5eb9e1f3", "repository_url": "https://github.com/hkust-nlp/model-task-align-rl", "scale_class": "reduced_pipeline_full_target", "screening_domain": "reinforcement_learning", "storage_gb": "90", "target_contract_sha256": "20fd9c02dcc0bf719498c7b40014a0364189976beeb440c3d2c12d9013772689", "target_metric": "avg@8 accuracy", "target_task": "Zero-shot synthetic logical reasoning after ground-truth-reward RL", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "WikiText-103 raw", "evaluation_sample_count": "1,000", "evaluation_split": "repository validation shard, first 1,000 contiguous 2,048-token blocks", "frozen_scale": "train steps=915", "gold_metric_sha256": "74bb398553701a4fed433c0f188776fd9be4bac13063d9446e311b5330f0961b", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "minimize", "metric_unit": "perplexity", "original_scale": "train steps=9,000", "paper_title": "Context-aware Biases for Length Extrapolation", "project_id": "112-2025.emnlp.1545", "reference_value": "21.48", "repository_commit": "05d2c62b3d2602ab5cc780474f485b98f95a7e33", "repository_url": "https://github.com/AlgonetLabs/Cable", "scale_class": "reduced_pipeline_full_target", "screening_domain": "information_retrieval", "storage_gb": "90", "target_contract_sha256": "c705c8cdccbf7db78fabd0b226b5462e8f410b40e4e039420345fed28b5a9375", "target_metric": "perplexity", "target_task": "WikiText-103 next-token prediction at context length 2048", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "BEIR SciFact", "evaluation_sample_count": "300", "evaluation_split": "test, all 300 queries", "frozen_scale": "predictor epochs=50", "gold_metric_sha256": "c8d85d04fc553ee69d5b2a7f2809cb299d7ea18b1566c4f2291d13132171b886", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "fraction", "original_scale": "predictor epochs=200", "paper_title": "Learning to Select: Query-Aware Adaptive Dimension Selection for Dense Retrieval", "project_id": "114-2026.acl-long.849", "reference_value": "0.839", "repository_commit": "d7d7254d0ba9c9bd81745e3e6a7fd1075419d49e", "repository_url": "https://github.com/wzy2001wzy/learning2select", "scale_class": "reduced_pipeline_full_target", "screening_domain": "information_retrieval", "storage_gb": "90", "target_contract_sha256": "1acf84d6175fb58c513daf98d7b73522639d09e9d569bf5b17f0c6d77cca6a55", "target_metric": "NDCG@10", "target_task": "Query-aware dense retrieval dimension selection", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "l3-unc/CausalDiagnosticity", "evaluation_sample_count": "1,000", "evaluation_split": "fact_check/test official first 1,000 rows", "frozen_scale": "unchanged", "gold_metric_sha256": "04325c6977736bcf5a66f43f986daccd5564510a9f31d68e2e7a32ff94a486fb", "gpu_class": "NVIDIA_H100_80GB_OR_EQUIVALENT", "gpu_count": "1", "memory_gb": "32", "metric_direction": "maximize", "metric_unit": "proportion", "original_scale": "paper-selected full scale", "paper_title": "A Causal Lens for Evaluating Faithfulness Metrics", "project_id": "118-2025.emnlp.1496", "reference_value": "0.828", "repository_commit": "3072a5092d6aeed8c1bd697863b920accca1c4c0", "repository_url": "https://github.com/KeremZaman/CausalDiagnosticity", "scale_class": "full_scale", "screening_domain": "knowledge_reasoning", "storage_gb": "20", "target_contract_sha256": "dfba2cc98abeddeadeacde93457674a43319f29e3b46e6558d15a3ea13983a7b", "target_metric": "continuous Filler Tokens diagnosticity", "target_task": "Measure whether continuous Filler Tokens scores rank faithful ICE explanations above paired unfaithful ICE explanations on FactCheck", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "AIME 2024 evaluation data bundled with GIFT/Qwen2.5-Math", "evaluation_sample_count": "30", "evaluation_split": "test, all 30 problems", "frozen_scale": "train rows=500; global batch=64", "gold_metric_sha256": "7eb5ec8dcd7a07ee045a8ef41dd5887842fa7707913fb87318089401345b0919", "gpu_class": "NVIDIA_H100_80GB_OR_EQUIVALENT", "gpu_count": "1", "memory_gb": "64", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "train rows=2,000; global batch=256", "paper_title": "GIFT: Guided Fine-Tuning and Transfer for Enhancing Instruction-Tuned Language Models", "project_id": "119-2026.acl-long.358", "reference_value": "12.7", "repository_commit": "277d059aae6a61e0d08b6287d2e6c1ca042b0f60", "repository_url": "https://github.com/sustech-nlp/gift", "scale_class": "reduced_pipeline_full_target", "screening_domain": "knowledge_reasoning", "storage_gb": "90", "target_contract_sha256": "49114d5d2ab8d3783773ca6087f214b101ca2338b90263ad33ece8687c1f9260", "target_metric": "Average@16 accuracy", "target_task": "AIME 2024 mathematical reasoning with sixteen stochastic solutions per problem", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "HardcoreLogic Hitori 5x5", "evaluation_sample_count": "50", "evaluation_split": "complete HardcoreLogic Hitori 5x5 subtask, 50 puzzles", "frozen_scale": "unchanged", "gold_metric_sha256": "b3b34d9ee9493ad3eb1d00f72943d488570be7be23fd738bed1c095bd2f7d84d", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "paper-selected full scale", "paper_title": "HardcoreLogic: Challenging Large Reasoning Models with Long-tail Logic Puzzle Games", "project_id": "121-2026.iclr.10011195", "reference_value": "21.5", "repository_commit": "f5428ceb08a6dd9cc6b4ef6671e3c7cd68e2274d", "repository_url": "https://github.com/ljcleo/hardcore-logic", "scale_class": "full_scale", "screening_domain": "knowledge_reasoning", "storage_gb": "90", "target_contract_sha256": "bb480da72a94f174a3ca91cafb7a2f2b912bf7fa40392af32a85c602e33b2e1b", "target_metric": "accuracy", "target_task": "Solve HardcoreLogic Hitori 5x5 puzzles with locally validated structured outputs", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "MMLU wrong-option-lengthened (WL)", "evaluation_sample_count": "14,042", "evaluation_split": "complete 57-subject test split, 14,042 examples", "frozen_scale": "unchanged", "gold_metric_sha256": "1e831c35c7eba6a23a95ed5d7c1ade35f3242bbee7328815ba554b387f89c8f4", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "paper-selected full scale", "paper_title": "Large Language Models Badly Generalize across Option Length, Problem Types, and Irrelevant Noun Replacements", "project_id": "124-2025.emnlp.1362", "reference_value": "36.3", "repository_commit": "05ea757c4c1391610099069b63a230307379eb48", "repository_url": "https://github.com/Qihoo360/LLMs-Generalization-Test", "scale_class": "full_scale", "screening_domain": "evaluation_benchmark", "storage_gb": "90", "target_contract_sha256": "9ec1f2dff859ee5c01aa7fd4a06cc28d5073a92d899e99d9ab7ffe9351d80be9", "target_metric": "exact-match accuracy", "target_task": "Five-shot MMLU multiple-choice answering after lengthening one wrong option", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "MATH-500", "evaluation_sample_count": "500", "evaluation_split": "test, all 500 problems", "frozen_scale": "train rows=512; train steps=32", "gold_metric_sha256": "0191f661db6e793cfea209670bd7a8ae19886bbc971ae1a2d0d698425fd256d2", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "train rows=40,315; train steps=2,000", "paper_title": "SPPO: Sequence-Level PPO for Long-Horizon Reasoning Tasks", "project_id": "128-2026.acl-long.356", "reference_value": "83.78", "repository_commit": "55e8a7236dabbd7c421ab987eebeb1a067b4c436", "repository_url": "https://github.com/sustech-nlp/SPPO", "scale_class": "reduced_pipeline_full_target", "screening_domain": "optimization_theory", "storage_gb": "90", "target_contract_sha256": "1f21c84eeed02aac97f2762fe08ccacbd97c83c79bb5e0bedf67977a7889307b", "target_metric": "Average@16 accuracy", "target_task": "Long-horizon mathematical problem solving after sequence-level PPO", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "MMPR-mini", "evaluation_sample_count": "797", "evaluation_split": "released MMPR-mini test split, all 797 rows", "frozen_scale": "train rows=96", "gold_metric_sha256": "62c9e97332152bb7bd73e6fb986e75e4a1647dbeae2b75da5c924571b2bb3353", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "64", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "train rows=7,183", "paper_title": "Counteracting the Matthew Effect in Self-Improvement of LVLMs through Head-Tail Re-balancing", "project_id": "132-2026.acl-long.1010", "reference_value": "45.8", "repository_commit": "1d852d8526436ddaa43241219d76da114cc997f7", "repository_url": "https://github.com/XinGuo2002/Head-Tail-Rebalancing", "scale_class": "reduced_pipeline_full_target", "screening_domain": "optimization_theory", "storage_gb": "90", "target_contract_sha256": "f78bb05ffc8c5228830691d79610c54566db4d5e5b81a1ee700f4596fd69dfda", "target_metric": "final accuracy", "target_task": "Multimodal mathematical reasoning on MMPR-mini", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "GAVEL released misuse and benign dialogue evaluation", "evaluation_sample_count": "2,800", "evaluation_split": "data/eval/data, phishing positive cases and released negative/benign controls", "frozen_scale": "epochs=10", "gold_metric_sha256": "f7b146acbe32a887a544f5f56d08d44adb4990b611206775daa5169a6953c011", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "score", "original_scale": "epochs=25", "paper_title": "GAVEL: Towards Rule-Based Safety through Activation Monitoring", "project_id": "134-2026.iclr.10008373", "reference_value": "0.99", "repository_commit": "cb74332f9f1b5374d0156a16635428ccdeaa15a4", "repository_url": "https://github.com/Offensive-AI-Lab/gavel", "scale_class": "reduced_pipeline_full_target", "screening_domain": "safety_alignment", "storage_gb": "90", "target_contract_sha256": "84b15641e4d25cb35506317eccd0e7f468904e2ae739544fc911480fa2e296b2", "target_metric": "ROC-AUC", "target_task": "Activation-based detection of phishing dialogues", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "CIFAR-10", "evaluation_sample_count": "50,000", "evaluation_split": "complete 50,000-image training split as FID reference; 50,000 generated samples", "frozen_scale": "train steps=100,000; seeds=1", "gold_metric_sha256": "5b617dad2228019a7f50c4b3b4ee1c3af1d159aef9f3b13c66509a30bfd0a2d5", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "minimize", "metric_unit": "FID", "original_scale": "train steps=400,000; seeds=5", "paper_title": "AlignFlow: Improving Flow-based Generative Models with Semi-Discrete Optimal Transport", "project_id": "136-2026.iclr.10007474", "reference_value": "3.71", "repository_commit": "4fd358d1a27968087f26a560f80aa27483b318cc", "repository_url": "https://github.com/konglk1203/AlignFlow", "scale_class": "reduced_pipeline_full_target", "screening_domain": "safety_alignment", "storage_gb": "90", "target_contract_sha256": "0aefb5b4e66113421bbdf1877d54a608fa5b27eb1f6cb10169e85b799b116968", "target_metric": "FID-50K", "target_task": "Unconditional CIFAR-10 image generation with flow matching", "wall_time_hours": "24"}
{"cpu_cores": "12", "evaluation_dataset": "CIFAR-10", "evaluation_sample_count": "50,000", "evaluation_split": "all 50,000 training images as the FID reference distribution", "frozen_scale": "optimization steps=50,000; global batch=128", "gold_metric_sha256": "c70bee0df31cc9867393aaf13f07cb42027febe3d666d702233eade6e599720f", "gpu_class": "NVIDIA_H100_80GB_OR_EQUIVALENT", "gpu_count": "1", "memory_gb": "64", "metric_direction": "minimize", "metric_unit": "FID points", "original_scale": "optimization steps=800,000; global batch=1,024", "paper_title": "SoFlow: Solution Flow Models for One-Step Generative Modeling", "project_id": "138-2026.iclr.10008475", "reference_value": "2.86", "repository_commit": "cc47661a0b0843f8f1b6cff3a17107a052d18b21", "repository_url": "https://github.com/zlab-princeton/SoFlow", "scale_class": "reduced_pipeline_full_target", "screening_domain": "generative_models", "storage_gb": "20", "target_contract_sha256": "ba402feccb64b1b5617f4dc3088dd7f57d7ac3765584f83dcf578f0f55d27856", "target_metric": "FID-50K", "target_task": "One-step unconditional 32x32 CIFAR-10 image generation", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "MCC-E synthetic motif-count classification", "evaluation_sample_count": "2,000", "evaluation_split": "official 2,000-example test split", "frozen_scale": "explainer epochs=10", "gold_metric_sha256": "be3d0c4ef16098ddba36d9a3005b36d249871f3c83d7bca7c01e172655f7fcc1", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "fraction", "original_scale": "explainer epochs=100", "paper_title": "ShapeX: Shapelet-Driven Post Hoc Explanations for Time Series Classification Models", "project_id": "144-2025.neurips.26300d2734ce0c6bdac69b4c3187b5a2", "reference_value": "0.6407", "repository_commit": "fb2a4e076d86e067ea3513f11276d69554f2b2a5", "repository_url": "https://github.com/BosonHwang/ShapeX", "scale_class": "reduced_pipeline_full_target", "screening_domain": "time_series", "storage_gb": "90", "target_contract_sha256": "5452d775ade3a04d615b73bcd373a2f0f62fd36820d12a19133501b125b89eb4", "target_metric": "AUPRC", "target_task": "Post-hoc saliency explanation on synthetic time series", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "Phy-SSM synthetic video pendulum", "evaluation_sample_count": "50", "evaluation_split": "50 generated test trajectories; final 80 of 240 retained frames are extrapolation targets", "frozen_scale": "epochs=80; seeds=1", "gold_metric_sha256": "65bc38f852ef214efb948c4885638a768aeab22fa57f13e882852c79139e8aa2", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "minimize", "metric_unit": "paper table value (x10^-2 MSE)", "original_scale": "epochs=500; seeds=3", "paper_title": "A Generalizable Physics-Enhanced State Space Model for Long-Term Dynamics Forecasting in Complex Environments", "project_id": "145-2025.icml.wang25ex", "reference_value": "1.417", "repository_commit": "d6a224a45a7c8fb290df687cfebcc8d8f9482a04", "repository_url": "https://github.com/511205787/Phy_SSM-ICML2025", "scale_class": "reduced_pipeline_full_target", "screening_domain": "time_series", "storage_gb": "90", "target_contract_sha256": "3b32301fab2a706b97fa1ba14135fc999d5d651693b4717b5344eafac9962298", "target_metric": "extrapolation MSE", "target_task": "long-horizon video pendulum forecasting", "wall_time_hours": "24"}
{"cpu_cores": "4", "evaluation_dataset": "BA-2MOTIFS", "evaluation_sample_count": "100", "evaluation_split": "official 10% test split", "frozen_scale": "base epochs=30; SC epochs=30; runs=1", "gold_metric_sha256": "8c52b08736e706b6a5dce9ce10e9b0012e54f2758e41f9ffb769c302acaa8ded", "gpu_class": "NVIDIA H100 80GB", "gpu_count": "1", "memory_gb": "80", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "base epochs=100; SC epochs=100; runs=10", "paper_title": "Self-Consistency Improves the Trustworthiness of Self-Interpretable GNNs", "project_id": "148-2026.iclr.10008008", "reference_value": "99.14", "repository_commit": "a06a5126320b1d15c63681ac34afdf2390451546", "repository_url": "https://github.com/ICDM-UESTC/SelfConsistencyXGNN", "scale_class": "reduced_pipeline_full_target", "screening_domain": "graph_learning", "storage_gb": "90", "target_contract_sha256": "459b38a2b1da9289563f9e3c13d4fa68dbe133395a3e2fd972b16c4ffa7baa7d", "target_metric": "explanation AUC", "target_task": "Ground-truth graph explanation recovery", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "OpenOOD CIFAR-10 inliers plus DTD Texture outliers", "evaluation_sample_count": "100", "evaluation_split": "100 paper-defined random disjoint calibration/test splits; each test has 950 CIFAR-10 inliers and 50 Texture outliers", "frozen_scale": "unchanged", "gold_metric_sha256": "cca608d8b0e08978ea8c85472dc9b9e5531ade8ad995a06a2511ca5e69edd44b", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "32", "metric_direction": "maximize", "metric_unit": "ratio", "original_scale": "paper-selected full scale", "paper_title": "Robust Conformal Outlier Detection under Contaminated Reference Data", "project_id": "2025.icml.bashari25a", "reference_value": "1.448", "repository_commit": "a322beec53f7042a211fbee5c9bfcd205a864fc3", "repository_url": "https://github.com/Meshiba/robust-conformal-od", "scale_class": "full_scale", "screening_domain": "other_nlp", "storage_gb": "90", "target_contract_sha256": "f2a66b7b2ad5379e8167c9019d88f5ca50a072a361889a323421a28e2d2b1237", "target_metric": "Label-Trim relative empirical power", "target_task": "Robust conformal outlier detection with contaminated calibration data", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "ETDataset ETTh1", "evaluation_sample_count": "2,785", "evaluation_split": "official ETTh1 test split, all 2,785 windows at context 2,880 and horizon 96", "frozen_scale": "unchanged", "gold_metric_sha256": "4c483c26b0902db49ee6a76ca5fc2bfd4a161d4c0b027a9e422db12cdf88b0cd", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "16", "metric_direction": "minimize", "metric_unit": "squared standardized target units", "original_scale": "paper-selected full scale", "paper_title": "VisionTS: Visual Masked Autoencoders Are Free-Lunch Zero-Shot Time Series Forecasters", "project_id": "2025.icml.chen25be", "reference_value": "0.353", "repository_commit": "7f34731cea92f6d670f1f5835f2018267d1b6135", "repository_url": "https://github.com/Keytoyze/VisionTS", "scale_class": "full_scale", "screening_domain": "computer_vision", "storage_gb": "90", "target_contract_sha256": "776bc05c1ed9fb82ea0978c506f7c510b059e489a337ff10d62001a1ca0d7351", "target_metric": "mean squared error", "target_task": "ETTh1 multivariate zero-shot long-term forecasting at prediction horizon 96", "wall_time_hours": "24"}
{"cpu_cores": "16", "evaluation_dataset": "SKEMPIv2 filtered interface-homology held-out test split", "evaluation_sample_count": "1,491", "evaluation_split": "complete released test_pdb.pkl split: 1,491 mutations across 81 interface IDs", "frozen_scale": "unchanged", "gold_metric_sha256": "710a0ae422684064c0bc3792985ec9204f3ef2fd8b27e82358534e8081e66b3f", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "32", "metric_direction": "minimize", "metric_unit": "kcal/mol", "original_scale": "paper-selected full scale", "paper_title": "Predicting mutational effects on protein binding from folding energy", "project_id": "2025.icml.deng25d", "reference_value": "1.72", "repository_commit": "b19a5d72637a6b581b16f38af39383658e5ae1a5", "repository_url": "https://github.com/LDeng0205/StaB-ddG", "scale_class": "full_scale", "screening_domain": "scientific_ml", "storage_gb": "90", "target_contract_sha256": "f4a9a345be03b1aab39afb0b844db014a0cf41ff6bc470bff2825ab6c8e9b4a6", "target_metric": "overall root mean squared error", "target_task": "Predict mutational effects on protein-protein binding energy for the complete released SKEMPIv2 held-out test split", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "TA-BG 2D 40-component Gaussian mixture", "evaluation_sample_count": "1,000", "evaluation_split": "complete paper evaluation of 1,000 model samples at target temperature T=1.0", "frozen_scale": "pretrain steps=12,500; pretrain runs=1; steps/transition=5,000; buffer/transition=500,000; resamples/transition=500,000", "gold_metric_sha256": "4c731540af8cbd8547569467a86806d21ac77691dbf6dff6e020b65cf2b51f71", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "16", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "pretrain steps=50,000; pretrain runs=4; steps/transition=20,000; buffer/transition=2,000,000; resamples/transition=2,000,000", "paper_title": "Temperature-Annealed Boltzmann Generators", "project_id": "2025.icml.schopmans25a", "reference_value": "96.89", "repository_commit": "2f752359fdd548acf252ef3323078921d7167bc5", "repository_url": "https://github.com/aimat-lab/TA-BG", "scale_class": "reduced_pipeline_full_target", "screening_domain": "other_nlp", "storage_gb": "90", "target_contract_sha256": "990af4efecc0217df596a3a4b6365fe6c96bfce8a0be7419732e56dd1d76ca0d", "target_metric": "reverse effective sample size", "target_task": "Variational sampling of the 2D 40-component Gaussian mixture with a temperature-annealed Boltzmann generator", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "WikiText-2 raw", "evaluation_sample_count": "4,358", "evaluation_split": "test; all 4,358 rows in canonical order", "frozen_scale": "unchanged", "gold_metric_sha256": "219d380293be13291d58533bb0e85f6909f9c7bfc7d4091e003567a72644d02b", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "64", "metric_direction": "minimize", "metric_unit": "perplexity", "original_scale": "paper-selected full scale", "paper_title": "FlatQuant: Flatness Matters for LLM Quantization", "project_id": "2025.icml.sun25l", "reference_value": "8.46", "repository_commit": "9d88ffcb7d2c6bda59fb5c44dad36adc101aadb1", "repository_url": "https://github.com/ruikangliu/FlatQuant", "scale_class": "full_scale", "screening_domain": "llm_foundation_models", "storage_gb": "90", "target_contract_sha256": "1b2476bc9e16899f772057223d58885faa0532435e9784b895ea4cc841480297", "target_metric": "WikiText-2 perplexity", "target_task": "WikiText-2 language-model perplexity after FlatQuant W4A4KV4 post-training quantization", "wall_time_hours": "24"}
{"cpu_cores": "16", "evaluation_dataset": "Arrow Pointing Extrapolation", "evaluation_sample_count": "10,240", "evaluation_split": "test/ext; all 10,240 images at 384x384 generated with seed 200043", "frozen_scale": "runs=1", "gold_metric_sha256": "57c1c705d270759bb19d71e08c29ed51be9ca17a8771f90c252b1d01289c5c2a", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "64", "metric_direction": "maximize", "metric_unit": "fraction", "original_scale": "runs=5", "paper_title": "pLSTM: parallelizable Linear Source Transition Mark networks", "project_id": "2025.neurips.0d5c9d43eb13c98db21eeb92c7c986d8", "reference_value": "0.778", "repository_commit": "ac539b61890a793be873472f16d33b7d78f994d4", "repository_url": "https://github.com/ml-jku/plstm_experiments", "scale_class": "reduced_pipeline_full_target", "screening_domain": "computer_vision", "storage_gb": "90", "target_contract_sha256": "edc2ac6ca79c7c44aeb4dbeb5bf94444e9834322282915c4992b89caa5368bb8", "target_metric": "Extrapolation test top-1 accuracy", "target_task": "Arrow Pointing Extrapolation binary image classification", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "ScienceQA", "evaluation_sample_count": "4,241", "evaluation_split": "test, all 4,241 examples", "frozen_scale": "unchanged", "gold_metric_sha256": "39489aded293b41b8bd5a7675932b6c83ddd0c8b6138f878030bb51360b6f274", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "32", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "paper-selected full scale", "paper_title": "RobustMerge: Parameter-Efficient Model Merging for MLLMs with Direction Robustness", "project_id": "2025.neurips.67101f97dc23fcc10346091181fff6cb", "reference_value": "73.43", "repository_commit": "12128022190516cc93a549cc4931112ce4bdcda3", "repository_url": "https://github.com/AuroraZengfh/RobustMerge", "scale_class": "full_scale", "screening_domain": "multimodal_learning", "storage_gb": "90", "target_contract_sha256": "6eba56b9809bb63d71bb3152eb7f137e9e658ad96b20a8b9085eab67377aeb10", "target_metric": "multiple-choice accuracy", "target_task": "ScienceQA multiple-choice visual question answering after eight-task parameter-efficient model merging", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "AI2 ARC-Easy", "evaluation_sample_count": "2,376", "evaluation_split": "ARC-Easy test, all 2,376 examples", "frozen_scale": "unchanged", "gold_metric_sha256": "23e9770c36a8f9bcf697b3e8db7fcdf104fb84b92c7ab3f95871bc456ce90eb8", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "32", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "paper-selected full scale", "paper_title": "A Simple Linear Patch Revives Layer-Pruned Large Language Models", "project_id": "2025.neurips.e323ebf7c91a4ca061612ba4a2f2164d", "reference_value": "54.84", "repository_commit": "d1274e8782676d13f1f006cae4405db5d0f96721", "repository_url": "https://github.com/chenxinrui-tsinghua/LinearPatch", "scale_class": "full_scale", "screening_domain": "other_nlp", "storage_gb": "90", "target_contract_sha256": "0b9f4c74f4762ce3003c6a746284c96c1c93b7f654b0e745de30d0d0c70ed22a", "target_metric": "raw multiple-choice accuracy", "target_task": "ARC-Easy zero-shot multiple-choice question answering after training-free contiguous layer pruning with LinearPatch[L]", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "GSM8K", "evaluation_sample_count": "1,319", "evaluation_split": "main/test; all 1,319 examples", "frozen_scale": "runs=1", "gold_metric_sha256": "ceab97a16ca9210a5ed9c5e133edf562a5f401a2483159107a459c02b5ad55f7", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "32", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "runs=3", "paper_title": "Parallel Test-Time Scaling for Latent Reasoning Models", "project_id": "2026.acl-long.2069", "reference_value": "46.7", "repository_commit": "23e756b6c27bf884aaead146f85b2cf0aecfc34d", "repository_url": "https://github.com/ModalityDance/LatentTTS", "scale_class": "reduced_pipeline_full_target", "screening_domain": "reinforcement_learning", "storage_gb": "90", "target_contract_sha256": "79364f9cc922f430d22c988446560809724bbaad106680aedd674213bac95d6a", "target_metric": "Coverage@4", "target_task": "GSM8K arithmetic reasoning with parallel latent test-time sampling", "wall_time_hours": "24"}
{"cpu_cores": "8", "evaluation_dataset": "ScienceQA", "evaluation_sample_count": "4,241", "evaluation_split": "test; all 4,241 IDs", "frozen_scale": "unchanged", "gold_metric_sha256": "99f48dee51029065bcfb97b2877d057829ab2e84fb95b3ab09b2658a64f8b6d9", "gpu_class": "NVIDIA_H100_80GB", "gpu_count": "1", "memory_gb": "64", "metric_direction": "maximize", "metric_unit": "percent", "original_scale": "paper-selected full scale", "paper_title": "VisionTrim: Unified Vision Token Compression for Training-Free MLLM Acceleration", "project_id": "2026.iclr.10011500", "reference_value": "71", "repository_commit": "308e73a9b8437882c5c1056a32dbc60fd41efd53", "repository_url": "https://github.com/hanxunyu/VisionTrim", "scale_class": "full_scale", "screening_domain": "multimodal_learning", "storage_gb": "90", "target_contract_sha256": "116156d6631de82f3e3c15ec2d192ea18ccec5eee336eb7f8dc0866e69e42d8e", "target_metric": "ScienceQA test accuracy", "target_task": "ScienceQA multiple-choice question answering with vision-token compression", "wall_time_hours": "24"}