{"id":"W4412888004","doi":"10.18653/v1/2025.findings-acl.763","title":"REVS: Unlearning Sensitive Information in Language Models via Rank Editing in the Vocabulary Space","year":2025,"lang":"en","type":"article","venue":"","topic":"Natural Language Processing Techniques","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Azrieli Foundation; European Commission; Open Philanthropy Project","keywords":"Vocabulary; Computer science; Rank (graph theory); Space (punctuation); Language model; Natural language processing; Artificial intelligence; Information retrieval; Linguistics; Mathematics; Operating system","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":[0.002931588,0.002076851,0.001589033,0.001188559,0.0005540723,0.001514331,0.002558419,0.001483334,0.002711517],"category_scores_gemma":[0.01460449,0.0006806231,0.001657074,0.0008538809,0.001324252,0.003569344,0.002681836,0.003925162,0.002182657],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0007855222,"about_ca_system_score_gemma":0.001462114,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003232265,"about_ca_topic_score_gemma":0.006736552,"domain_scores_codex":[0.9974597,0.0009751059,0.0001888082,0.000627117,0.0005844068,0.0001649031],"domain_scores_gemma":[0.9938419,0.003295524,0.000467412,0.001600767,0.0006346502,0.000159805],"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":[0.0005269176,0.0003510314,0.002911248,0.0004360265,0.0003258178,0.0003259897,0.0004567338,0.1882054,0.02556979,0.01326996,0.0182507,0.7493704],"study_design_scores_gemma":[0.00004704234,0.0001824299,0.0002816746,0.00002170289,0.00003462659,0.0001386341,0.00005486102,0.9666686,0.01167263,0.0181584,0.002706646,0.00003286083],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.02125665,0.0006486059,0.9663404,0.0004814167,0.0001423429,0.0001247763,0.0004515678,0.009446566,0.00110764],"genre_scores_gemma":[0.5415414,0.0006805538,0.440401,0.001338452,0.0003924141,0.00037149,0.003581289,0.001788092,0.009905281],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.003232265,"threshold_uncertainty_score":0.01550394,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.005588018851145987,"score_gpt":0.2506226381330361,"score_spread":0.2450346192818902,"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."}}