{"id":"W7103204082","doi":"","title":"Value Drifts: Tracing Value Alignment During LLM Post-Training","year":2025,"lang":"","type":"article","venue":"ArXiv.org","topic":"Ethics and Social Impacts of AI","field":"Social Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Samsung; Alliance de recherche numérique du Canada; Natural Sciences and Engineering Research Council of Canada; McGill University; Canadian Institute for Advanced Research","keywords":"Preference; Preference elicitation; Selection (genetic algorithm); Value (mathematics); Field (mathematics); Tracing","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":[0.007661186,0.0008008624,0.0007422511,0.0008104367,0.0009301371,0.001855898,0.001544415,0.001073243,0.003109716],"category_scores_gemma":[0.07475094,0.0004977695,0.0004306768,0.0008796728,0.001372841,0.003512169,0.002327538,0.002778749,0.001015711],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001726914,"about_ca_system_score_gemma":0.001496259,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.005135775,"about_ca_topic_score_gemma":0.008058976,"domain_scores_codex":[0.9965294,0.001761038,0.0001549015,0.0006656732,0.0005727044,0.0003162198],"domain_scores_gemma":[0.9717484,0.0180772,0.001704711,0.004955426,0.002682905,0.0008314609],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.001684771,0.001293629,0.1475129,0.0006236524,0.0003044494,0.0004550544,0.004147499,0.3580217,0.01809483,0.02534656,0.02076192,0.421753],"study_design_scores_gemma":[0.00007796228,0.0005703862,0.02012696,0.0001392746,0.00005080069,0.0001305588,0.001109211,0.916939,0.0149126,0.03743232,0.008435728,0.00007516208],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8228918,0.0009537625,0.1595771,0.002302862,0.0002052807,0.0002409168,0.0009956325,0.002953254,0.009879405],"genre_scores_gemma":[0.9702026,0.0000915468,0.02664689,0.0002513815,0.00001736354,0.0001210439,0.001047177,0.0002724635,0.001349602],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.007661186,"threshold_uncertainty_score":0.04051667,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.05222465522300903,"score_gpt":0.3494846867468943,"score_spread":0.2972600315238853,"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."}}