{"id":"W4416408255","doi":"10.48550/arxiv.2507.19219","title":"How Much Do Large Language Model Cheat on Evaluation? Benchmarking Overestimation under the One-Time-Pad-Based Framework","year":2025,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Topic Modeling","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Instituto de Ciencias del Mar y Limnología, Universidad Nacional Autónoma de México; Innovation and Technology Fund; Institute for Catastrophic Loss Reduction; University of Science and Technology of China; Microsoft Research; National Natural Science Foundation of China; Hong Kong Polytechnic University","keywords":"Benchmarking; Benchmark (surveying); Key (lock); Quality (philosophy); Test (biology); Measure (data warehouse)","routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":true},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0009286606,0.0003410615,0.0002929865,0.000296355,0.0003475019,0.0004429964,0.001877997,0.0004095715,0.00004564536],"category_scores_gemma":[0.0001294812,0.0003361128,0.0002070972,0.000548279,0.00004097711,0.0003259382,0.001167681,0.0009275934,0.00002527028],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0005665376,"about_ca_system_score_gemma":0.0005521316,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000045353,"about_ca_topic_score_gemma":0.00001859349,"domain_scores_codex":[0.9974769,0.0003406374,0.0001945708,0.001208156,0.0003886151,0.0003911276],"domain_scores_gemma":[0.9968409,0.0003843598,0.000246633,0.00219203,0.0002465339,0.00008954854],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00000942045,0.00005924967,0.00007037293,0.00003668673,0.00004022668,0.000006328336,0.0003450889,0.7436295,0.00001196799,0.2550294,0.00007021116,0.0006916642],"study_design_scores_gemma":[0.0003617778,0.00001532639,0.0001362905,0.000341353,0.00009623896,2.173187e-7,0.0001220781,0.8651055,0.00009399912,0.1334337,0.00001194966,0.0002815527],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.09684813,0.00005839344,0.8984024,0.001422545,0.0003715639,0.000586516,0.00002480858,0.0001960686,0.002089608],"genre_scores_gemma":[0.971647,0.0000113505,0.02641644,0.0006995999,0.0001110227,0.000005840181,0.00004111861,0.0000139092,0.001053738],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8747988,"threshold_uncertainty_score":0.9999091,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.08765754670620739,"score_gpt":0.2418694500749871,"score_spread":0.1542119033687797,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}