{"id":"W7082259453","doi":"10.48448/jwf5-y993","title":"Combining Domain and Alignment Vectors Provides Better Knowledge-Safety Trade-offs in LLMs","year":2025,"lang":"en","type":"other","venue":"Underline Science Inc.","topic":"Geochemistry and Geologic Mapping","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"Mila - Quebec Artificial Intelligence Institute; Polytechnique Montréal","funders":"","keywords":"SAFER; Domain (mathematical analysis); Interpolation (computer graphics); Code (set theory); Similarity (geometry); Simple (philosophy); Subject-matter expert; Domain model; Software deployment","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":[0.002400317,0.001341649,0.0006366883,0.00114636,0.0004522928,0.001815375,0.001533113,0.001264851,0.004735789],"category_scores_gemma":[0.01296548,0.0005166821,0.0009011878,0.0009265976,0.0007416508,0.003529888,0.002756166,0.001984963,0.001987565],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001275362,"about_ca_system_score_gemma":0.001871809,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.006885914,"about_ca_topic_score_gemma":0.01004049,"domain_scores_codex":[0.9986466,0.0003875722,0.0000917765,0.0003739405,0.0003461019,0.0001540886],"domain_scores_gemma":[0.9956114,0.001980746,0.000300067,0.001259246,0.0006467099,0.000201885],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"not_applicable","study_design_scores_codex":[0.000601977,0.0004606536,0.0182458,0.0003562992,0.0001628689,0.0002233692,0.0004127018,0.7113159,0.01627074,0.009963179,0.00808783,0.2338986],"study_design_scores_gemma":[0.00003797144,0.0001638505,0.001829414,0.00004786647,0.00004809333,0.00006254669,0.0001868197,0.9652746,0.01452768,0.01138744,0.006409522,0.00002412746],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.4346167,0.0008870058,0.5202927,0.001260367,0.000177336,0.0001896446,0.00177774,0.02346746,0.01733101],"genre_scores_gemma":[0.7666844,0.0002590499,0.2222602,0.0003801787,0.00002393239,0.0001243206,0.003705485,0.00204257,0.004519841],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.006885914,"threshold_uncertainty_score":0.0158428,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.01176255926838132,"score_gpt":0.2482184599359145,"score_spread":0.2364559006675332,"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."}}