{"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":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00376266,0.001211084,0.0008810645,0.002336586,0.0009747252,0.006191232,0.002923046,0.001992395,0.0047006],"category_scores_gemma":[0.04359813,0.001429438,0.002016026,0.001671258,0.002556987,0.01422361,0.003459708,0.003846432,0.001211913],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002456586,"about_ca_system_score_gemma":0.001538551,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01373768,"about_ca_topic_score_gemma":0.0134495,"domain_scores_codex":[0.9964059,0.001446302,0.0002413217,0.00105096,0.0005877995,0.0002675593],"domain_scores_gemma":[0.9796053,0.01264555,0.002271669,0.003545093,0.001458822,0.0004735405],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","study_design_scores_codex":[0.001236179,0.0003246788,0.07018127,0.001149456,0.0009029334,0.00109969,0.01417462,0.2750159,0.01902373,0.306052,0.009050015,0.3017896],"study_design_scores_gemma":[0.00005582193,0.000152662,0.004304831,0.0001211951,0.000193794,0.0002195888,0.001198082,0.6594239,0.006881164,0.3194401,0.007881615,0.0001272125],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1260739,0.0005511878,0.859211,0.001741438,0.00006222568,0.000126638,0.001892744,0.005187767,0.0051531],"genre_scores_gemma":[0.8247536,0.0002674557,0.1698549,0.0003665775,0.00003407069,0.0001321847,0.002325129,0.0005391232,0.00172697],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.01373768,"threshold_uncertainty_score":0.02731544,"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."}}