{"id":"https://openalex.org/W4414891753","doi":"https://doi.org/10.48550/arxiv.2506.00688","title":"Existing Large Language Model Unlearning Evaluations Are Inconclusive","display_name":"Existing Large Language Model Unlearning Evaluations Are Inconclusive","publication_year":2025,"publication_date":"2025-05-31","ids":{"openalex":"https://openalex.org/W4414891753","doi":"https://doi.org/10.48550/arxiv.2506.00688"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2506.00688","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2506.00688","pdf_url":"https://arxiv.org/pdf/2506.00688","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2506.00688","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5100307168","display_name":"Zhili Feng","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Feng, Zhili","raw_affiliation_strings":["Carnegie Mellon University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University","institution_ids":["https://openalex.org/I74973139"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101618176","display_name":"Ye Xu","orcid":"https://orcid.org/0000-0002-3361-0177","observed_orcids":["https://orcid.org/0000-0002-3361-0177"]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Xu, Yixuan Even","raw_affiliation_strings":["Carnegie Mellon University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University","institution_ids":["https://openalex.org/I74973139"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5028501158","display_name":"Alexander Robey","orcid":"https://orcid.org/0009-0003-5693-2819","observed_orcids":["https://orcid.org/0009-0003-5693-2819"]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Robey, Alexander","raw_affiliation_strings":["Carnegie Mellon University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University","institution_ids":["https://openalex.org/I74973139"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Kirk, Robert","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kirk, Robert","raw_affiliation_strings":["UK AI Security Institute"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"UK AI Security Institute","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5145865735","display_name":"Xander Davies","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I40120149","display_name":"University of Oxford","ror":"https://ror.org/052gg0110","country_code":"GB","type":"education","lineage":["https://openalex.org/I40120149"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Davies, Xander","raw_affiliation_strings":["UK AI Security Institute OATML , University of Oxford"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"UK AI Security Institute OATML , University of Oxford","institution_ids":["https://openalex.org/I40120149"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5029186201","display_name":"Yarin Gal","orcid":"https://orcid.org/0000-0002-2733-2078","observed_orcids":["https://orcid.org/0000-0002-2733-2078"]},"institutions":[{"id":"https://openalex.org/I40120149","display_name":"University of Oxford","ror":"https://ror.org/052gg0110","country_code":"GB","type":"education","lineage":["https://openalex.org/I40120149"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Gal, Yarin","raw_affiliation_strings":["UK AI Security Institute OATML , University of Oxford"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"UK AI Security Institute OATML , University of Oxford","institution_ids":["https://openalex.org/I40120149"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5026712341","display_name":"Avi Schwarzschild","orcid":"https://orcid.org/0000-0003-0997-4867","observed_orcids":["https://orcid.org/0000-0003-0997-4867"]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Schwarzschild, Avi","raw_affiliation_strings":["Carnegie Mellon University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University","institution_ids":["https://openalex.org/I74973139"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5075035644","display_name":"J. Zico Kolter","orcid":"https://orcid.org/0000-0002-8106-5759","observed_orcids":["https://orcid.org/0000-0002-8106-5759"]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Kolter, J. Zico","raw_affiliation_strings":["Carnegie Mellon University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University","institution_ids":["https://openalex.org/I74973139"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":10,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10764","display_name":"Privacy-Preserving Technologies in Data","score":0.9980000257492065,"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/T10764","display_name":"Privacy-Preserving Technologies in Data","score":0.9980000257492065,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.00039999998989515007,"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/T10883","display_name":"Ethics and Social Impacts of AI","score":0.00019999999494757503,"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/machine-unlearning","display_name":"machine unlearning","score":0.9810000061988831},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.984000027179718},{"id":"https://openalex.org/keywords/spurious-correlations","display_name":"spurious correlations","score":0.7170000076293945},{"id":"https://openalex.org/keywords/data-deletion","display_name":"data deletion","score":0.3919999897480011}],"concepts":[{"id":"https://openalex.org/C27158222","wikidata":"https://www.wikidata.org/wiki/Q5532422","display_name":"Generalizability theory","level":2,"score":0.7282000184059143},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6312000155448914},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.5896000266075134},{"id":"https://openalex.org/C97256817","wikidata":"https://www.wikidata.org/wiki/Q1462316","display_name":"Spurious relationship","level":2,"score":0.5871000289916992},{"id":"https://openalex.org/C2777402240","wikidata":"https://www.wikidata.org/wiki/Q6783436","display_name":"Masking (illustration)","level":2,"score":0.5266000032424927},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.4952999949455261},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3765999972820282},{"id":"https://openalex.org/C56739046","wikidata":"https://www.wikidata.org/wiki/Q192060","display_name":"Knowledge management","level":1,"score":0.32019999623298645},{"id":"https://openalex.org/C165064840","wikidata":"https://www.wikidata.org/wiki/Q1321061","display_name":"Matching (statistics)","level":2,"score":0.30660000443458557},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.3001999855041504},{"id":"https://openalex.org/C87868495","wikidata":"https://www.wikidata.org/wiki/Q750843","display_name":"Information processing","level":2,"score":0.28189998865127563},{"id":"https://openalex.org/C2522767166","wikidata":"https://www.wikidata.org/wiki/Q2374463","display_name":"Data science","level":1,"score":0.27709999680519104},{"id":"https://openalex.org/C71745522","wikidata":"https://www.wikidata.org/wiki/Q2476929","display_name":"Confidentiality","level":2,"score":0.26019999384880066},{"id":"https://openalex.org/C175154964","wikidata":"https://www.wikidata.org/wiki/Q380077","display_name":"Task analysis","level":3,"score":0.25110000371932983}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2506.00688","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2506.00688","pdf_url":"https://arxiv.org/pdf/2506.00688","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2506.00688","doi":"https://doi.org/10.48550/arxiv.2506.00688","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2506.00688","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2506.00688","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2506.00688","pdf_url":"https://arxiv.org/pdf/2506.00688","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G6514084706","display_name":"RI: Small: Robustness to Undesirable Behavior in Peer Review","funder_award_id":"2200410","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"}],"funders":[{"id":"https://openalex.org/F4320306076","display_name":"National Science Foundation","ror":"https://ror.org/021nxhr62"},{"id":"https://openalex.org/F4320337345","display_name":"Office of Naval Research","ror":"https://ror.org/00rk2pe57"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4414891753.pdf","grobid_xml":"https://content.openalex.org/works/W4414891753.grobid-xml"},"referenced_works_count":32,"referenced_works":["https://openalex.org/W2985671575","https://openalex.org/W2995273672","https://openalex.org/W3015449890","https://openalex.org/W4385374425","https://openalex.org/W4386555476","https://openalex.org/W4387389584","https://openalex.org/W4387799916","https://openalex.org/W4389216598","https://openalex.org/W4390833824","https://openalex.org/W4392223220","https://openalex.org/W4392539041","https://openalex.org/W4394708837","https://openalex.org/W4399912709","https://openalex.org/W4400572793","https://openalex.org/W4401306886","https://openalex.org/W4401307749","https://openalex.org/W4402588152","https://openalex.org/W4403160620","https://openalex.org/W4403345854","https://openalex.org/W4403443689","https://openalex.org/W4403884617","https://openalex.org/W4403928578","https://openalex.org/W4404569914","https://openalex.org/W4406272124","https://openalex.org/W4410609100","https://openalex.org/W4410609217","https://openalex.org/W4412945388","https://openalex.org/W4413925879","https://openalex.org/W4414903222","https://openalex.org/W4415796188","https://openalex.org/W4415797562","https://openalex.org/W4415797606"],"related_works":["https://openalex.org/W4400572793","https://openalex.org/W7125269111","https://openalex.org/W7123805892","https://openalex.org/W4392223220","https://openalex.org/W4404649372","https://openalex.org/W4415663990","https://openalex.org/W4393905542","https://openalex.org/W4415339025","https://openalex.org/W7169635611","https://openalex.org/W7213140644"],"abstract_inverted_index":{"Machine":[0],"unlearning":[1,18,37,67,124,135],"aims":[2],"to":[3,106,161],"remove":[4],"sensitive":[5],"or":[6],"undesired":[7],"data":[8],"from":[9],"large":[10],"language":[11],"models.":[12],"However,":[13],"recent":[14],"studies":[15],"suggest":[16,114],"that":[17,23,44,54,78,95,115],"is":[19],"often":[20],"shallow,":[21],"claiming":[22],"removed":[24],"knowledge":[25],"can":[26,159],"easily":[27],"be":[28],"recovered.":[29],"In":[30],"this":[31],"work,":[32],"we":[33,52,76,93,129],"critically":[34],"examine":[35],"standard":[36],"evaluation":[38,79,90,117],"practices":[39],"and":[40,108,122,140],"uncover":[41],"key":[42],"limitations":[43],"shake":[45],"our":[46],"trust":[47,107],"in":[48],"those":[49],"findings.":[50],"First,":[51],"show":[53],"some":[55],"evaluations":[56,97],"introduce":[57],"substantial":[58],"new":[59],"information":[60,138],"into":[61],"the":[62,71,86],"model,":[63],"potentially":[64],"masking":[65],"true":[66],"performance":[68],"by":[69],"re-teaching":[70],"model":[72],"during":[73],"testing.":[74],"Second,":[75],"demonstrate":[77],"outcomes":[80],"vary":[81],"significantly":[82],"across":[83],"tasks,":[84],"undermining":[85],"generalizability":[87],"of":[88,151,157],"current":[89,116],"routines.":[91],"Finally,":[92],"find":[94],"many":[96],"rely":[98],"on":[99],"spurious":[100],"correlations,":[101],"making":[102],"their":[103],"results":[104],"difficult":[105],"interpret.":[109],"Taken":[110],"together,":[111],"these":[112,146],"issues":[113],"protocols":[118],"may":[119],"both":[120],"overstate":[121],"understate":[123],"success.":[125],"To":[126],"address":[127],"this,":[128],"propose":[130],"two":[131],"principles":[132,147],"for":[133],"future":[134],"evaluations:":[136],"minimal":[137],"injection":[139],"downstream":[141],"task":[142],"awareness.":[143],"We":[144],"validate":[145],"through":[148],"a":[149],"series":[150],"targeted":[152],"experiments,":[153],"showing":[154],"how":[155],"violations":[156],"each":[158],"lead":[160],"misleading":[162],"conclusions.":[163]},"counts_by_year":[{"year":2026,"cited_by_count":4},{"year":2025,"cited_by_count":6}],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2025-10-10T00:00:00"}