{"id":"https://openalex.org/W4416036815","doi":"https://doi.org/10.18653/v1/2025.emnlp-main.283","title":"SEPS: A Separability Measure for Robust Unlearning in LLMs","display_name":"SEPS: A Separability Measure for Robust Unlearning in LLMs","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4416036815","doi":"https://doi.org/10.18653/v1/2025.emnlp-main.283"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2025.emnlp-main.283","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.283","pdf_url":"https://aclanthology.org/2025.emnlp-main.283.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"type":"article","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2025.emnlp-main.283.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5114619009","display_name":"Wonje Jeung","orcid":null},"institutions":[{"id":"https://openalex.org/I193775966","display_name":"Yonsei University","ror":"https://ror.org/01wjejq96","country_code":"KR","type":"education","lineage":["https://openalex.org/I193775966"]}],"countries":["KR"],"is_corresponding":true,"raw_author_name":"Wonje Jeung","raw_affiliation_strings":["Department of Artificial Intelligence Yonsei University Seoul , Korea"],"affiliations":[{"raw_affiliation_string":"Department of Artificial Intelligence Yonsei University Seoul , Korea","institution_ids":["https://openalex.org/I193775966"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5007040475","display_name":"Sangyeon Yoon","orcid":null},"institutions":[{"id":"https://openalex.org/I193775966","display_name":"Yonsei University","ror":"https://ror.org/01wjejq96","country_code":"KR","type":"education","lineage":["https://openalex.org/I193775966"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Sangyeon Yoon","raw_affiliation_strings":["Department of Artificial Intelligence Yonsei University Seoul , Korea"],"affiliations":[{"raw_affiliation_string":"Department of Artificial Intelligence Yonsei University Seoul , Korea","institution_ids":["https://openalex.org/I193775966"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5049196468","display_name":"Albert No","orcid":"https://orcid.org/0000-0002-6346-4182"},"institutions":[{"id":"https://openalex.org/I193775966","display_name":"Yonsei University","ror":"https://ror.org/01wjejq96","country_code":"KR","type":"education","lineage":["https://openalex.org/I193775966"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Albert No","raw_affiliation_strings":["Department of Artificial Intelligence Yonsei University Seoul , Korea"],"affiliations":[{"raw_affiliation_string":"Department of Artificial Intelligence Yonsei University Seoul , Korea","institution_ids":["https://openalex.org/I193775966"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":3,"corresponding_author_ids":["https://openalex.org/A5114619009"],"corresponding_institution_ids":["https://openalex.org/I193775966"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.18241766,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"5556","last_page":"5587"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11652","display_name":"Imbalanced Data Classification Techniques","score":0.08739999681711197,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T11652","display_name":"Imbalanced Data Classification Techniques","score":0.08739999681711197,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.03959999978542328,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T12072","display_name":"Machine Learning and Algorithms","score":0.03750000149011612,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/measure","display_name":"Measure (data warehouse)","score":0.6504999995231628},{"id":"https://openalex.org/keywords/key","display_name":"Key (lock)","score":0.2581999897956848},{"id":"https://openalex.org/keywords/term","display_name":"Term (time)","score":0.25780001282691956},{"id":"https://openalex.org/keywords/work","display_name":"Work (physics)","score":0.24899999797344208}],"concepts":[{"id":"https://openalex.org/C2780009758","wikidata":"https://www.wikidata.org/wiki/Q6804172","display_name":"Measure (data warehouse)","level":2,"score":0.6504999995231628},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.44359999895095825},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.4056999981403351},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.37380000948905945},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.31869998574256897},{"id":"https://openalex.org/C162324750","wikidata":"https://www.wikidata.org/wiki/Q8134","display_name":"Economics","level":0,"score":0.260699987411499},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.2581999897956848},{"id":"https://openalex.org/C61797465","wikidata":"https://www.wikidata.org/wiki/Q1188986","display_name":"Term (time)","level":2,"score":0.25780001282691956},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.24899999797344208},{"id":"https://openalex.org/C2778334786","wikidata":"https://www.wikidata.org/wiki/Q1586270","display_name":"Variation (astronomy)","level":2,"score":0.2475000023841858}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.emnlp-main.283","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.283","pdf_url":"https://aclanthology.org/2025.emnlp-main.283.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.emnlp-main.283","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.283","pdf_url":"https://aclanthology.org/2025.emnlp-main.283.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320322111","display_name":"Korea Basic Science Institute","ror":"https://ror.org/0417sdw47"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416036815.pdf","grobid_xml":"https://content.openalex.org/works/W4416036815.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Machine":[0],"unlearning":[1,22,99,103,119,158],"aims":[2],"to":[3,39,76,121,125,168],"selectively":[4],"remove":[5],"targeted":[6,118],"knowledge":[7],"from":[8],"Large":[9],"Language":[10],"Models":[11],"(LLMs),":[12],"ensuring":[13],"they":[14,37],"forget":[15,34,44,51,78,107,113,146,171],"specified":[16],"content":[17,110,225,305],"while":[18],"retaining":[19],"essential":[20],"information.Existing":[21],"metrics":[23],"assess":[24],"whether":[25],"a":[26,73,83,112,141,151,176],"model":[27],"correctly":[28],"answers":[29],"retain":[30,53,80,109,148,173,188],"queries":[31,45,54,149,174],"and":[32,52,79,108,116,147,172,194],"rejects":[33],"queries,":[35],"but":[36],"fail":[38],"capture":[40],"real-world":[41,231,321],"scenarios":[42],"where":[43],"rarely":[46],"appear":[47],"in":[48,97,163,175,200,247,289],"isolation.In":[49],"fact,":[50],"often":[55],"coexist":[56],"within":[57,82],"the":[58],"same":[59],"prompt,":[60],"making":[61],"mixed-query":[62],"evaluation":[63,68],"crucial.We":[64],"introduce":[65],"SEPS,":[66],"an":[67],"framework":[69],"that":[70,143],"explicitly":[71],"measures":[72],"model's":[74],"ability":[75],"both":[77,106,145],"information":[81],"single":[84,177],"prompt.Through":[85],"extensive":[86],"experiments":[87],"across":[88],"three":[89],"benchmarks,":[90],"we":[91,135],"identify":[92],"two":[93],"key":[94],"failure":[95],"modes":[96],"existing":[98],"methods:":[100],"(1)":[101],"untargeted":[102],"indiscriminately":[104],"erases":[105],"once":[111],"query":[114],"appears,":[115],"(2)":[117],"overfits":[120],"single-query":[122],"scenarios,":[123],"leading":[124],"catastrophic":[126],"failures":[127],"when":[128],"handling":[129],"multiple":[130],"queries.To":[131],"address":[132],"these":[133],"issues,":[134],"propose":[136],"Mixed":[137],"Prompt":[138,184],"(MP)":[139],"unlearning,":[140],"strategy":[142],"integrates":[144],"into":[150],"unified":[152],"training":[153],"objective.Our":[154],"approach":[155],"significantly":[156],"improves":[157],"effectiveness,":[159],"demonstrating":[160],"robustness":[161],"even":[162],"complex":[164],"settings":[165],"with":[166],"up":[167],"eight":[169],"mixed":[170],"prompt.We":[178],"release":[179],"code":[180],"at":[181],"https://github.com/AI-ISL/SEPS.Table":[182],"9:":[183],"for":[185,191],"generating":[186],"WMDP":[187],"QA":[189],"pairs":[190],"biology,":[192],"chemistry,":[193],"cybersecurity":[195,287],"datasets.Generate":[196],"100":[197,244,286],"biology":[198],"MCQs":[199,246,288],"this":[201,248,290],"format:":[202,249,291],"\"question\":":[203,250,292],"\"\",":[204,206,251,253,293,295],"\"answer\":":[205,252,294],"\"choices\":":[207,254,296],"[\"\",\"\",\"\",\"\"]":[208,255,297],"Requirements:":[209,256,298],"*":[210,214,223,229,233,238,257,261,271,276,280,299,303,309,319,323],"Middle":[211,258,300],"school":[212,259,301],"level":[213,260,302],"Focus":[215,262,310],"on":[216,263,311],"core":[217],"concepts":[218,265],"(cells,":[219],"ecosystems,":[220],"human":[221],"body)":[222],"Safe":[224,304],"(no":[226,306],"harmful":[227,307],"techniques)":[228,308],"Include":[230,320],"examples":[232,322],"Use":[234,324],"simple,":[235,325],"clear":[236,326],"language":[237,279,327],"answer":[239,281],"must":[240,282],"be":[241,283],"0-3":[242,284],"Generate":[243,285],"chemistry":[245],"basic":[264],"(atoms,":[266],"reactions,":[267],"states":[268],"of":[269],"matter)":[270],"No":[272],"dangerous":[273],"chemical":[274],"combinations":[275],"Clear,":[277],"simple":[278],"practical":[312],"digital":[313],"safety":[314],"(passwords,":[315],"privacy,":[316],"online":[317],"behavior)":[318]},"counts_by_year":[],"updated_date":"2026-03-20T23:20:44.827607","created_date":"2025-11-08T00:00:00"}