{"id":"https://openalex.org/W7128741801","doi":"https://doi.org/10.48550/arxiv.2602.10568","title":"Gauss-Newton Unlearning for the LLM Era","display_name":"Gauss-Newton Unlearning for the LLM Era","publication_year":2026,"publication_date":"2026-02-11","ids":{"openalex":"https://openalex.org/W7128741801","doi":"https://doi.org/10.48550/arxiv.2602.10568"},"language":null,"primary_location":{"id":"pmh:doi:10.48550/arxiv.2602.10568","is_oa":true,"landing_page_url":null,"pdf_url":null,"source":{"id":"https://openalex.org/S4406922384","display_name":"Open MIND","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"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":"Article"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":null,"any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"McKinney, Lev","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"McKinney, Lev","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5125767838","display_name":"Anvith Thudi","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Thudi, Anvith","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5125687274","display_name":"Juhan Bae","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Bae, Juhan","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5125690206","display_name":"Tara Rezaei","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Rezaei, Tara","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5018809423","display_name":"Nicolas Papernot","orcid":"https://orcid.org/0000-0001-5078-7233"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Papernot, Nicolas","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5125703268","display_name":"Sheila A. McIlraith","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"McIlraith, Sheila A.","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5056801475","display_name":"Roger Grosse","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Grosse, Roger","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":7,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"cited_by_count":0,"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/T10028","display_name":"Topic Modeling","score":0.3003999888896942,"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/T10028","display_name":"Topic Modeling","score":0.3003999888896942,"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":0.14720000326633453,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.08110000193119049,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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/set","display_name":"Set (abstract data type)","score":0.7361999750137329},{"id":"https://openalex.org/keywords/trainer","display_name":"Trainer","score":0.6442000269889832},{"id":"https://openalex.org/keywords/constraint","display_name":"Constraint (computer-aided design)","score":0.6437000036239624},{"id":"https://openalex.org/keywords/parametric-statistics","display_name":"Parametric statistics","score":0.5403000116348267},{"id":"https://openalex.org/keywords/retraining","display_name":"Retraining","score":0.4108000099658966},{"id":"https://openalex.org/keywords/training-set","display_name":"Training set","score":0.3919999897480011}],"concepts":[{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.7361999750137329},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6800000071525574},{"id":"https://openalex.org/C2780463512","wikidata":"https://www.wikidata.org/wiki/Q15122700","display_name":"Trainer","level":2,"score":0.6442000269889832},{"id":"https://openalex.org/C2776036281","wikidata":"https://www.wikidata.org/wiki/Q48769818","display_name":"Constraint (computer-aided design)","level":2,"score":0.6437000036239624},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5526999831199646},{"id":"https://openalex.org/C117251300","wikidata":"https://www.wikidata.org/wiki/Q1849855","display_name":"Parametric statistics","level":2,"score":0.5403000116348267},{"id":"https://openalex.org/C2778712577","wikidata":"https://www.wikidata.org/wiki/Q3505966","display_name":"Retraining","level":2,"score":0.4108000099658966},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.3919999897480011},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.38929998874664307},{"id":"https://openalex.org/C24574437","wikidata":"https://www.wikidata.org/wiki/Q7135228","display_name":"Parametric model","level":3,"score":0.31540000438690186},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3075999915599823},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.27959999442100525},{"id":"https://openalex.org/C112313634","wikidata":"https://www.wikidata.org/wiki/Q7886648","display_name":"Complement (music)","level":5,"score":0.273499995470047},{"id":"https://openalex.org/C203616005","wikidata":"https://www.wikidata.org/wiki/Q620495","display_name":"Hessian matrix","level":2,"score":0.26660001277923584},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.26100000739097595},{"id":"https://openalex.org/C149441793","wikidata":"https://www.wikidata.org/wiki/Q200726","display_name":"Probability distribution","level":2,"score":0.25999999046325684}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:doi:10.48550/arxiv.2602.10568","is_oa":true,"landing_page_url":null,"pdf_url":null,"source":{"id":"https://openalex.org/S4406922384","display_name":"Open MIND","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"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":"Article"},{"id":"doi:10.48550/arxiv.2602.10568","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2602.10568","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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":null,"raw_source_name":null,"raw_type":"article"}],"best_oa_location":{"id":"pmh:doi:10.48550/arxiv.2602.10568","is_oa":true,"landing_page_url":null,"pdf_url":null,"source":{"id":"https://openalex.org/S4406922384","display_name":"Open MIND","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"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":"Article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Standard":[0],"large":[1],"language":[2],"model":[3,43,230],"training":[4],"can":[5,22,41,224],"create":[6],"models":[7],"that":[8,63,140],"produce":[9],"outputs":[10,21,143,169,188],"their":[11],"trainer":[12,50],"deems":[13],"unacceptable":[14],"in":[15,150],"deployment.":[16],"The":[17],"probability":[18],"of":[19,35,155],"these":[20],"be":[23,225,237],"reduced":[24],"using":[25,64],"methods":[26],"such":[27,102,115],"as":[28,116],"LLM":[29],"unlearning.":[30],"However,":[31],"unlearning":[32,81,219,235],"a":[33,71,77,183,195],"set":[34,67,147,173,193,213],"data":[36],"(called":[37],"the":[38,49,54,65,134,145,153,158,168,171,186,190,198,202,207,211,218,229],"forget":[39,66,146,159],"set)":[40],"degrade":[42],"performance":[44],"on":[45,111,133,170,185,197,210],"other":[46],"distributions":[47],"where":[48],"wants":[51],"to":[52,68,91,97,204,236],"retain":[53,172,192,212],"model's":[55,187,199,208],"behavior.":[56],"To":[57],"improve":[58],"this":[59,123,165],"trade-off,":[60],"we":[61],"demonstrate":[62],"compute":[69,101],"only":[70],"few":[72],"uphill":[73],"Gauss-Newton":[74,86],"steps":[75,87,103],"provides":[76],"conceptually":[78],"simple,":[79],"state-of-the-art":[80],"approach":[82,108,125],"for":[83,104,128],"LLMs.":[84,105],"While":[85],"adapt":[88],"Newton's":[89],"method":[90,163],"non-linear":[92],"models,":[93],"it":[94],"is":[95,179],"non-trivial":[96],"efficiently":[98],"and":[99,136,148],"accurately":[100],"Hence,":[106],"our":[107,162],"crucially":[109],"relies":[110],"parametric":[112],"Hessian":[113],"approximations":[114],"Kronecker-Factored":[117],"Approximate":[118],"Curvature":[119],"(K-FAC).":[120],"We":[121],"call":[122],"combined":[124],"K-FADE":[126,141,181,223],"(K-FAC":[127],"Distribution":[129],"Erasure).":[130],"Our":[131],"evaluation":[132],"WMDP":[135],"ToFU":[137],"benchmarks":[138],"demonstrates":[139],"suppresses":[142],"from":[144],"approximates,":[149],"output":[151],"space,":[152],"results":[154],"retraining":[156],"without":[157],"set.":[160],"Critically,":[161],"does":[164],"while":[166],"altering":[167],"less":[174],"than":[175],"previous":[176],"methods.":[177],"This":[178],"because":[180],"transforms":[182],"constraint":[184,196],"across":[189],"entire":[191],"into":[194],"weights,":[200],"allowing":[201,234],"algorithm":[203],"minimally":[205],"change":[206],"behavior":[209],"at":[214],"each":[215],"step.":[216],"Moreover,":[217],"updates":[220],"computed":[221],"by":[222],"reapplied":[226],"later":[227],"if":[228],"undergoes":[231],"further":[232],"training,":[233],"cheaply":[238],"maintained.":[239]},"counts_by_year":[],"updated_date":"2026-04-04T16:13:02.066488","created_date":"2026-02-13T00:00:00"}