{"id":"W4412944786","doi":"10.18653/v1/2025.acl-long.1141","title":"Language Models Resist Alignment: Evidence From Data Compression","year":2025,"lang":"en","type":"article","venue":"","topic":"Natural Language Processing Techniques","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"ca_institutions":"Artificial Intelligence in Medicine (Canada)","funders":"National Natural Science Foundation of China","keywords":"Resist; Computer science; Compression (physics); Data compression; Artificial intelligence; Materials science; Composite material","routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0132027,0.001046828,0.001816105,0.002612076,0.00226149,0.003715344,0.002668069,0.003271726,0.006367],"category_scores_gemma":[0.1550892,0.001082547,0.000777858,0.004141482,0.00459418,0.01022642,0.00366363,0.004356571,0.0042131],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0007373221,"about_ca_system_score_gemma":0.001709836,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001596404,"about_ca_topic_score_gemma":0.001536097,"domain_scores_codex":[0.9887488,0.006019089,0.0006939365,0.001471637,0.002465416,0.0006011972],"domain_scores_gemma":[0.7635953,0.1692414,0.01097848,0.04284872,0.01177317,0.001562913],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.01011159,0.001352308,0.1016557,0.001433784,0.001146041,0.003564297,0.003824538,0.07376916,0.02061055,0.1612615,0.05910567,0.5621648],"study_design_scores_gemma":[0.001075548,0.0007013934,0.01975388,0.0002684484,0.0006770305,0.002957121,0.002071403,0.4415557,0.03123862,0.4828252,0.01665425,0.0002215992],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.6673137,0.007058656,0.2573754,0.01825211,0.001630532,0.0002884083,0.003186441,0.003401013,0.0414938],"genre_scores_gemma":[0.9545805,0.001420084,0.03444781,0.00170174,0.0007618286,0.0001600212,0.003243964,0.0005633123,0.003120761],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.0132027,"threshold_uncertainty_score":0.06982338,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.05182750288185463,"score_gpt":0.3401692022134332,"score_spread":0.2883416993315785,"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."}}