{"id":"https://openalex.org/W4416036646","doi":"https://doi.org/10.18653/v1/2025.emnlp-main.452","title":"A Fully Probabilistic Perspective on Large Language Model Unlearning: Evaluation and Optimization","display_name":"A Fully Probabilistic Perspective on Large Language Model Unlearning: Evaluation and Optimization","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4416036646","doi":"https://doi.org/10.18653/v1/2025.emnlp-main.452"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2025.emnlp-main.452","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.452","pdf_url":"https://aclanthology.org/2025.emnlp-main.452.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":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2025.emnlp-main.452.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5038610148","display_name":"Anda Cheng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Anda Cheng","raw_affiliation_strings":["Ant Group"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Ant Group","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100352894","display_name":"Wei Huang","orcid":"https://orcid.org/0009-0004-9458-7769"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wei Huang","raw_affiliation_strings":["Ant Group"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Ant Group","institution_ids":[]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5028637336","display_name":"Yinggui Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yinggui Wang","raw_affiliation_strings":["Ant Group"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Ant Group","institution_ids":[]}]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.29712483,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"8943","last_page":"8954"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T12072","display_name":"Machine Learning and Algorithms","score":0.321399986743927,"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/T12072","display_name":"Machine Learning and Algorithms","score":0.321399986743927,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.10520000010728836,"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/T12535","display_name":"Machine Learning and Data Classification","score":0.09709999710321426,"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/perspective","display_name":"Perspective (graphical)","score":0.5952000021934509},{"id":"https://openalex.org/keywords/probabilistic-logic","display_name":"Probabilistic logic","score":0.5587999820709229},{"id":"https://openalex.org/keywords/probabilistic-relevance-model","display_name":"Probabilistic relevance model","score":0.38839998841285706},{"id":"https://openalex.org/keywords/statistical-model","display_name":"Statistical model","score":0.36160001158714294},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.3440999984741211}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6444000005722046},{"id":"https://openalex.org/C12713177","wikidata":"https://www.wikidata.org/wiki/Q1900281","display_name":"Perspective (graphical)","level":2,"score":0.5952000021934509},{"id":"https://openalex.org/C49937458","wikidata":"https://www.wikidata.org/wiki/Q2599292","display_name":"Probabilistic logic","level":2,"score":0.5587999820709229},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.49799999594688416},{"id":"https://openalex.org/C143017306","wikidata":"https://www.wikidata.org/wiki/Q3318133","display_name":"Probabilistic relevance model","level":4,"score":0.38839998841285706},{"id":"https://openalex.org/C114289077","wikidata":"https://www.wikidata.org/wiki/Q3284399","display_name":"Statistical model","level":2,"score":0.36160001158714294},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.35910001397132874},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.3440999984741211},{"id":"https://openalex.org/C527412718","wikidata":"https://www.wikidata.org/wiki/Q855395","display_name":"Interpretation (philosophy)","level":2,"score":0.2865000069141388},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.2793999910354614},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.266400009393692},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.2508000135421753}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.emnlp-main.452","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.452","pdf_url":"https://aclanthology.org/2025.emnlp-main.452.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.452","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.452","pdf_url":"https://aclanthology.org/2025.emnlp-main.452.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":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416036646.pdf","grobid_xml":"https://content.openalex.org/works/W4416036646.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"Language":[1],"Model":[2],"Unlearning":[3],"(LLMU)":[4],"is":[5],"a":[6,41,57,87,96],"promising":[7],"way":[8],"to":[9,82,101,110,129],"remove":[10],"private":[11],"or":[12],"sensitive":[13],"information":[14,91,143],"from":[15,117],"large":[16],"language":[17],"models.However,":[18],"the":[19,103],"comprehensive":[20],"evaluation":[21,28,59,148],"of":[22,34,90,105,114],"LLMU":[23,54],"remains":[24],"underexplored.The":[25],"dominant":[26],"deterministic":[27],"can":[29],"yield":[30],"overly":[31],"optimistic":[32],"assessments":[33],"unlearning":[35,153],"efficacy.To":[36],"mitigate":[37],"this,":[38],"we":[39,94],"propose":[40],"Fully":[42],"Probabilistic":[43],"Evaluation":[44],"(FPE)":[45],"framework":[46],"that":[47,138],"incorporates":[48],"input":[49],"and":[50,70,132,150,159],"output":[51],"distributions":[52],"in":[53,80],"evaluation.FPE":[55],"obtains":[56],"probabilistic":[58],"result":[60],"by":[61,155,161],"querying":[62],"unlearned":[63,115,142],"models":[64],"with":[65,125],"various":[66],"semantically":[67],"similar":[68],"inputs":[69],"multiple":[71],"sampling":[72],"attempts.We":[73],"introduce":[74,95],"an":[75],"Input":[76],"Distribution":[77],"Sampling":[78],"method":[79],"FPE":[81,139],"select":[83],"high-quality":[84],"inputs,":[85],"enabling":[86],"stricter":[88],"measure":[89],"leakage":[92,144],"risks.Furthermore,":[93],"Contrastive":[97],"Embedding":[98],"Loss":[99],"(CEL)":[100],"advance":[102],"performance":[104],"LLMU.CEL":[106],"employs":[107],"contrastive":[108],"learning":[109],"distance":[111],"latent":[112],"representations":[113],"samples":[116,121],"adaptively":[118],"clustered":[119],"contrast":[120],"while":[122,167],"aligning":[123],"them":[124],"random":[126],"vectors,":[127],"leading":[128],"improved":[130],"efficacy":[131],"robustness":[133,160],"for":[134],"LLMU.Our":[135],"experiments":[136],"show":[137],"uncovers":[140],"more":[141],"risks":[145],"than":[146],"prior":[147],"methods,":[149],"CEL":[151],"improves":[152],"effectiveness":[154],"at":[156,162],"least":[157,163],"50.1%":[158],"37.2%":[164],"on":[165],"Llama-2-7B":[166],"retaining":[168],"high":[169],"model":[170],"utility.":[171]},"counts_by_year":[],"updated_date":"2026-08-21T09:56:20.448147","created_date":"2025-11-08T00:00:00"}