{"id":"W4417298816","doi":"10.48550/arxiv.2505.14685","title":"Language Models use Lookbacks to Track Beliefs","year":2025,"lang":"en","type":"preprint","venue":"ArXiv.org","topic":"Multimodal Machine Learning Applications","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Council for Higher Education; Azrieli Foundation; European Commission; Open Philanthropy Project; National Science Foundation","keywords":"Character (mathematics); Construct (python library); Visibility; Recall; Encoding (memory); State (computer science); Language model; Resolution (logic); Relation (database)","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","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0002693281,0.0003820706,0.0003845741,0.0002642628,0.0001558104,0.0003011201,0.002729115,0.0002902803,0.0000268756],"category_scores_gemma":[0.0001801128,0.0004014983,0.00017878,0.0004668436,0.00003473428,0.0003066711,0.00338072,0.001122831,0.0007952576],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001205687,"about_ca_system_score_gemma":0.0002148125,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00341865,"about_ca_topic_score_gemma":0.0001013703,"domain_scores_codex":[0.9974529,0.0001228301,0.0004118214,0.001256066,0.0003222578,0.0004341867],"domain_scores_gemma":[0.9963865,0.0002310037,0.0001682392,0.002836296,0.0001505644,0.000227422],"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":[0.00002181472,0.0006789401,0.2032141,0.0004349882,0.0002811292,0.00008339909,0.02393587,0.6588346,0.002083914,0.03165662,0.01357821,0.06519635],"study_design_scores_gemma":[0.0004403033,0.00005360903,0.4829059,0.0004138025,0.00005295794,0.000006713545,0.00004142903,0.5018914,0.001217904,0.003028263,0.008710929,0.001236767],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.5815058,0.0001130387,0.4097239,0.003511678,0.0003736631,0.0006400791,0.0000474634,0.0007311745,0.003353143],"genre_scores_gemma":[0.9064794,0.00001722961,0.08496651,0.002095542,0.0001727291,0.0002827144,0.0000474562,0.00002884503,0.005909558],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3249736,"threshold_uncertainty_score":0.9999827,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.05042120623825223,"score_gpt":0.3218039472206062,"score_spread":0.271382740982354,"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."}}