{"id":"W4412889746","doi":"10.18653/v1/2025.acl-long.1458","title":"Stochastic Chameleons: Irrelevant Context Hallucinations Reveal Class-Based (Mis)Generalization in LLMs","year":2025,"lang":"en","type":"article","venue":"","topic":"Anomaly Detection Techniques and Applications","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Canadian Institute for Advanced Research; Nvidia","keywords":"Generalization; Context (archaeology); Class (philosophy); Computer science; Cognitive psychology; Artificial intelligence; Psychology; Social psychology; Econometrics; Mathematics; Geography","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":[0.002555168,0.0006032734,0.000643383,0.0004652625,0.000419304,0.001661113,0.001297289,0.0009925725,0.002508077],"category_scores_gemma":[0.02748579,0.0004979527,0.0007782102,0.0003308916,0.00234621,0.004626805,0.002698149,0.002935488,0.0004100105],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000825503,"about_ca_system_score_gemma":0.0005423061,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002263241,"about_ca_topic_score_gemma":0.001897411,"domain_scores_codex":[0.9983889,0.0006107264,0.0001074727,0.0004349257,0.0003214869,0.0001365857],"domain_scores_gemma":[0.9832121,0.009329611,0.001323164,0.005087112,0.0006892186,0.0003588109],"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":[0.001533748,0.0003642737,0.04890298,0.0006194945,0.000347395,0.001596913,0.009701292,0.386594,0.1382746,0.1194768,0.007029701,0.2855589],"study_design_scores_gemma":[0.00003632205,0.0002239138,0.006232761,0.00002168152,0.00004346641,0.0003218403,0.0005303361,0.8188284,0.02715228,0.1447895,0.001765933,0.00005358542],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.5878755,0.0001714527,0.4010606,0.001309046,0.00007026432,0.00006823801,0.0002809372,0.00342312,0.005740745],"genre_scores_gemma":[0.9797353,0.00003464462,0.01894681,0.0001757796,0.00001276911,0.00002575513,0.000130223,0.0001780072,0.0007608389],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.002555168,"threshold_uncertainty_score":0.01351315,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.009901039363788106,"score_gpt":0.2679487081815004,"score_spread":0.2580476688177122,"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."}}