{"id":"W4404402193","doi":"10.48550/arxiv.2411.06824","title":"Combining Domain and Alignment Vectors to Achieve Better Knowledge-Safety Trade-offs in LLMs","year":2024,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Digital Rights Management and Security","field":"Computer Science","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; Canadian Institute for Advanced Research","keywords":"Domain (mathematical analysis); Computer science; Business; Risk analysis (engineering); Mathematics","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":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0003551846,0.0003647458,0.0003705764,0.0005196111,0.00009238573,0.0003692211,0.001240319,0.0001716649,0.000007961726],"category_scores_gemma":[0.000007870512,0.0003729198,0.0001375406,0.000716688,0.00008135624,0.0003193018,0.005041344,0.0005363023,0.000115147],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002814905,"about_ca_system_score_gemma":0.00006717683,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006555289,"about_ca_topic_score_gemma":0.0002096077,"domain_scores_codex":[0.9978127,0.0001156038,0.0002542445,0.001286646,0.0001136932,0.0004171256],"domain_scores_gemma":[0.9988611,0.00010789,0.00007996938,0.0007172229,0.00001829371,0.0002155668],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.00006085376,0.0002458207,0.001670697,0.0002393196,0.0001457409,0.0009090222,0.004980483,0.004042667,0.00001432633,0.9811859,0.001181843,0.005323327],"study_design_scores_gemma":[0.0009380668,0.0002357277,0.00636042,0.0005626847,0.00007170632,0.000003141477,0.0002057632,0.08014089,0.00006077002,0.8909354,0.0193302,0.001155238],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8972231,0.0001239346,0.04645875,0.001696802,0.001273698,0.0007288523,0.00001815008,0.0003035688,0.0521732],"genre_scores_gemma":[0.9974605,0.00004876986,0.0006308579,0.0002189666,0.00004738944,0.000001939595,0.00000849899,0.00001855968,0.001564539],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1002374,"threshold_uncertainty_score":0.9998723,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.03512000690300871,"score_gpt":0.1879398007447607,"score_spread":0.152819793841752,"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."}}