{"id":"W4401104224","doi":"10.48550/arxiv.2402.18551","title":"Implicit Optimization Bias of Next-Token Prediction in Linear Models","year":2024,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Intelligent Tutoring Systems and Adaptive Learning","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Natural Sciences and Engineering Research Council of Canada; Canadian Institute for Advanced Research","keywords":"Security token; Computer science; Psychology; Computer security","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":[],"consensus_categories":[],"category_scores_codex":[0.0003640405,0.0002166822,0.0002957596,0.0005016728,0.00004797838,0.00009098977,0.0007535121,0.0002478219,0.00001025155],"category_scores_gemma":[0.00002109233,0.0002438651,0.0001554896,0.0006439253,0.0000282877,0.0004126797,0.001270822,0.0006065895,0.00003243469],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002257718,"about_ca_system_score_gemma":0.0001465968,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0005669391,"about_ca_topic_score_gemma":0.00001335912,"domain_scores_codex":[0.9984518,0.0001251862,0.0003334121,0.0007573522,0.0001064863,0.0002257333],"domain_scores_gemma":[0.998904,0.00005732202,0.0002457645,0.0005700486,0.0001615449,0.00006126703],"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":[0.000006185436,0.00002252515,0.000466922,0.0001117317,0.00002780738,0.00004124068,0.0003717858,0.7615375,0.00002753651,0.2372285,0.00002118626,0.0001371258],"study_design_scores_gemma":[0.000114834,0.00003854978,0.00009696051,0.0004733096,0.00002166756,0.000002290416,0.00008164607,0.9818834,0.0001145516,0.0167985,0.0001907835,0.0001834902],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.09265414,0.0001053522,0.9033073,0.00002527772,0.0009982863,0.00024103,0.00001152767,0.0001781459,0.002478917],"genre_scores_gemma":[0.9910396,0.0001043913,0.003590432,0.000009537181,0.0001035402,0.000001145147,0.0000115884,0.00001822845,0.005121537],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8997169,"threshold_uncertainty_score":0.9944531,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.1304284785535108,"score_gpt":0.2063121845599387,"score_spread":0.07588370600642794,"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."}}