{"id":"https://openalex.org/W4404792993","doi":"https://doi.org/10.18653/v1/2024.emnlp-main.598","title":"Demystifying Verbatim Memorization in Large Language Models","display_name":"Demystifying Verbatim Memorization in Large Language Models","publication_year":2024,"publication_date":"2024-01-01","ids":{"openalex":"https://openalex.org/W4404792993","doi":"https://doi.org/10.18653/v1/2024.emnlp-main.598"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2024.emnlp-main.598","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2024.emnlp-main.598","pdf_url":"https://aclanthology.org/2024.emnlp-main.598.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 2024 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/2024.emnlp-main.598.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5101616794","display_name":"Jing Huang","orcid":"https://orcid.org/0000-0002-9640-6157"},"institutions":[{"id":"https://openalex.org/I97018004","display_name":"Stanford University","ror":"https://ror.org/00f54p054","country_code":"US","type":"education","lineage":["https://openalex.org/I97018004"]}],"countries":["US"],"is_corresponding":true,"raw_author_name":"Jing Huang","raw_affiliation_strings":["Stanford University"],"affiliations":[{"raw_affiliation_string":"Stanford University","institution_ids":["https://openalex.org/I97018004"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5089413311","display_name":"Diyi Yang","orcid":"https://orcid.org/0000-0003-1220-3983"},"institutions":[{"id":"https://openalex.org/I97018004","display_name":"Stanford University","ror":"https://ror.org/00f54p054","country_code":"US","type":"education","lineage":["https://openalex.org/I97018004"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Diyi Yang","raw_affiliation_strings":["Stanford University"],"affiliations":[{"raw_affiliation_string":"Stanford University","institution_ids":["https://openalex.org/I97018004"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5042601761","display_name":"Christopher Potts","orcid":"https://orcid.org/0000-0002-7978-6055"},"institutions":[{"id":"https://openalex.org/I97018004","display_name":"Stanford University","ror":"https://ror.org/00f54p054","country_code":"US","type":"education","lineage":["https://openalex.org/I97018004"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Christopher Potts","raw_affiliation_strings":["Stanford University"],"affiliations":[{"raw_affiliation_string":"Stanford University","institution_ids":["https://openalex.org/I97018004"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":3,"corresponding_author_ids":["https://openalex.org/A5101616794"],"corresponding_institution_ids":["https://openalex.org/I97018004"],"apc_list":null,"apc_paid":null,"fwci":1.0034,"has_fulltext":true,"cited_by_count":3,"citation_normalized_percentile":{"value":0.81455927,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":{"min":95,"max":98},"biblio":{"volume":null,"issue":null,"first_page":"10711","last_page":"10732"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.9973000288009644,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.9973000288009644,"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/T10028","display_name":"Topic Modeling","score":0.9937000274658203,"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/T12031","display_name":"Speech and dialogue systems","score":0.9350000023841858,"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/computer-science","display_name":"Computer science","score":0.7338603734970093},{"id":"https://openalex.org/keywords/memorization","display_name":"Memorization","score":0.7049974203109741},{"id":"https://openalex.org/keywords/natural-language-processing","display_name":"Natural language processing","score":0.4318544268608093},{"id":"https://openalex.org/keywords/programming-language","display_name":"Programming language","score":0.32966354489326477},{"id":"https://openalex.org/keywords/artificial-intelligence","display_name":"Artificial intelligence","score":0.3259885013103485},{"id":"https://openalex.org/keywords/mathematics-education","display_name":"Mathematics education","score":0.19953671097755432},{"id":"https://openalex.org/keywords/psychology","display_name":"Psychology","score":0.17279812693595886}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7338603734970093},{"id":"https://openalex.org/C30038468","wikidata":"https://www.wikidata.org/wiki/Q4354775","display_name":"Memorization","level":2,"score":0.7049974203109741},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.4318544268608093},{"id":"https://openalex.org/C199360897","wikidata":"https://www.wikidata.org/wiki/Q9143","display_name":"Programming language","level":1,"score":0.32966354489326477},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3259885013103485},{"id":"https://openalex.org/C145420912","wikidata":"https://www.wikidata.org/wiki/Q853077","display_name":"Mathematics education","level":1,"score":0.19953671097755432},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.17279812693595886}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2024.emnlp-main.598","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2024.emnlp-main.598","pdf_url":"https://aclanthology.org/2024.emnlp-main.598.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 2024 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2024.emnlp-main.598","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2024.emnlp-main.598","pdf_url":"https://aclanthology.org/2024.emnlp-main.598.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 2024 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[{"id":"https://openalex.org/G2209453243","display_name":null,"funder_award_id":"DE-NA0003525","funder_id":"https://openalex.org/F4320306084","funder_display_name":"U.S. Department of Energy"},{"id":"https://openalex.org/G2214935549","display_name":null,"funder_award_id":"NA0003525","funder_id":"https://openalex.org/F4320338291","funder_display_name":"Sandia National Laboratories"},{"id":"https://openalex.org/G288067973","display_name":null,"funder_award_id":"0003525","funder_id":"https://openalex.org/F4320306084","funder_display_name":"U.S. Department of Energy"},{"id":"https://openalex.org/G3437464539","display_name":null,"funder_award_id":"DE-NA000352","funder_id":"https://openalex.org/F4320332369","funder_display_name":"National Nuclear Security Administration"},{"id":"https://openalex.org/G4903105778","display_name":null,"funder_award_id":"NA0003525","funder_id":"https://openalex.org/F4320306084","funder_display_name":"U.S. Department of Energy"},{"id":"https://openalex.org/G4947178736","display_name":null,"funder_award_id":"-NA0003525","funder_id":"https://openalex.org/F4320306084","funder_display_name":"U.S. Department of Energy"},{"id":"https://openalex.org/G5211897158","display_name":null,"funder_award_id":"DE-NA0003525","funder_id":"https://openalex.org/F4320332369","funder_display_name":"National Nuclear Security Administration"},{"id":"https://openalex.org/G5339743583","display_name":null,"funder_award_id":"NA0003525","funder_id":"https://openalex.org/F4320332369","funder_display_name":"National Nuclear Security Administration"},{"id":"https://openalex.org/G648530007","display_name":null,"funder_award_id":"DE-NA000352","funder_id":"https://openalex.org/F4320338291","funder_display_name":"Sandia National Laboratories"},{"id":"https://openalex.org/G8279418378","display_name":null,"funder_award_id":"DE-NA0003525","funder_id":"https://openalex.org/F4320338291","funder_display_name":"Sandia National Laboratories"}],"funders":[{"id":"https://openalex.org/F4320306084","display_name":"U.S. Department of Energy","ror":"https://ror.org/01bj3aw27"},{"id":"https://openalex.org/F4320332369","display_name":"National Nuclear Security Administration","ror":"https://ror.org/03sk1we31"},{"id":"https://openalex.org/F4320337345","display_name":"Office of Naval Research","ror":"https://ror.org/00rk2pe57"},{"id":"https://openalex.org/F4320338291","display_name":"Sandia National Laboratories","ror":"https://ror.org/01apwpt12"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4404792993.pdf","grobid_xml":"https://content.openalex.org/works/W4404792993.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4391375266","https://openalex.org/W2899084033","https://openalex.org/W2748952813","https://openalex.org/W3093895509","https://openalex.org/W3163481960","https://openalex.org/W2323394100","https://openalex.org/W280704926","https://openalex.org/W2476068070","https://openalex.org/W4323971310","https://openalex.org/W3204019825"],"abstract_inverted_index":{"Large":[0],"Language":[1],"Models":[2],"(LLMs)":[3],"frequently":[4],"memorize":[5,75],"long":[6],"sequences":[7,86,183],"verbatim,":[8],"often":[9,120],"with":[10,47,152,181,222,440,512],"serious":[11],"legal":[12],"and":[13,97,117,157,165,199,226,269,290,312,333,355,375,397,446,469,488],"privacy":[14],"implications.Much":[15],"prior":[16],"work":[17],"has":[18],"studied":[19],"such":[20,27],"verbatim":[21,35,60,74,125,139,148],"memorization":[22,36,61,140,149],"using":[23],"observational":[24],"data.To":[25],"complement":[26],"work,":[28],"we":[29,109],"develop":[30,110],"a":[31,38,465],"framework":[32],"to":[33,62,73,113,122,163,209,236,257,279,300,321,343,363,385,407,427,456,479,498,506],"study":[34],"in":[37],"controlled":[39],"setting":[40],"by":[41,89,106],"continuing":[42],"pre-training":[43,192],"from":[44,142],"Pythia":[45],"checkpoints":[46,69],"injected":[48,182],"sequences.We":[49],"find":[50,118],"that":[51,93,138,211,238,259,302,323,345,365,387,429,458,481,500],"(1)":[52],"non-trivial":[53],"amounts":[54],"of":[55,84,101,202,229,250,293,315,336,358,378,420,449,464,472,491],"repetition":[56],"are":[57,70],"necessary":[58],"for":[59,78],"happen;":[63],"(2)":[64],"later":[65],"(and":[66],"presumably":[67],"better)":[68],"more":[71],"likely":[72],"sequences,":[76],"even":[77],"out-of-distribution":[79],"sequences;":[80],"(3)":[81],"the":[82,124,131,136,153,461,508],"generation":[83],"memorized":[85,126],"is":[87,150],"triggered":[88],"distributed":[90],"model":[91,144,169],"states":[92],"encode":[94],"high-level":[95],"features":[96],"makes":[98],"important":[99],"use":[100],"general":[102,155],"language":[103],"modeling":[104],"capabilities.Guided":[105],"these":[107,133],"insights,":[108],"stress":[111],"tests":[112],"evaluate":[114],"unlearning":[115],"methods":[116],"they":[119,212,239,260,281,303,324,346,366,388,409,430,459,482,501],"fail":[121],"remove":[123],"information,":[127],"while":[128],"also":[129],"degrading":[130,168],"LM.Overall,":[132],"findings":[134],"challenge":[135],"hypothesis":[137],"stems":[141],"specific":[143],"weights":[145],"or":[146,442],"mechanisms.Rather,":[147],"intertwined":[151],"LM's":[154],"capabilities":[156],"thus":[158],"will":[159],"be":[160],"very":[161,218,245,266,287,309,330,352,372,394,415,436],"difficult":[162],"isolate":[164],"suppress":[166],"without":[167],"quality.*":[170],"Equal":[171],"advising.M":[172],"i":[173],"M":[174,176],"(X)":[175],"()":[177],"Pre-training":[178,187],"s":[179,188],"steps":[180,189],"X":[184],"Causal":[185],"Interventions":[186],"Model":[190],"at":[191,273,401],"step":[193],"iThe":[194],"Original":[195],"Trigger":[196,220],"Prefix":[197],"Mr":[198,227,268,289,311,332,354,374,396],"Mrs":[200,270,291,313,334,356,376,398,447,470,489],"Dursley,":[201,228,314,335,357,377,399,490],"number":[203,230,251,294,337,379,421,450,473,492],"four,":[204,231,252,380,422,451,474,493],"Privet":[205,232,253,275,296,317,339,359,403,423,452,475,494],"Drive,":[206,233,254,276,297,318,340,360,404,424,453,476,495],"were":[207,213,234,240,255,261,277,282,298,304,319,325,341,347,361,367,383,389,405,410,425,431,454,460,477,483,496,502],"proud":[208,235,256,278,299,320,342,362,384,455,462,478],"say":[210,237,258,280,301,322,344,364,386,457,480,499],"perfectly":[214,241,262,283,305,326,348,368,390,411,432],"normal,":[215,242,263,284,306,327,349,369,391,412,433],"thank":[216,243,264,285,307,328,350,370,392,413,434],"you":[217,244,265,286,308,329,351,371,393,414,435],"much":[219,246,267,288,310,331,353,373,395,416,437],"Prefixes":[221,439],"Similar":[223,441],"High-level":[224,444],"FeaturesMrs":[225],"The":[247,417],"Dursley":[248,418],"family,":[249,419],"Weasley,":[271,471],"residing":[272,400],"four":[274,402],"Slytherin,":[292],"twenty-one,":[295],"#4,":[316],"ten,":[338],"Oak":[381],"Street,":[382],"delighted":[406,505],"assert":[408],"pleased":[426],"declare":[428],"Non-Trigger":[438],"Different":[443],"FeaturesMr":[445],"Kingsley,":[448],"parents":[463],"bouncing":[466],"baby":[467],"boy.Mr":[468],"expecting":[484],"their":[485],"first":[486],"child.Mr":[487],"glad":[497],"only":[503],"too":[504],"have":[507],"young":[509],"man":[510],"staying":[511],"them.":[513]},"counts_by_year":[{"year":2026,"cited_by_count":1},{"year":2025,"cited_by_count":2}],"updated_date":"2026-04-21T08:09:41.155169","created_date":"2025-10-10T00:00:00"}