{"id":"https://openalex.org/W4412886663","doi":"https://doi.org/10.18653/v1/2025.acl-long.245","title":"Improve Safety Training of Large Language Models with Safety-Critical Singular Vectors Localization","display_name":"Improve Safety Training of Large Language Models with Safety-Critical Singular Vectors Localization","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4412886663","doi":"https://doi.org/10.18653/v1/2025.acl-long.245"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2025.acl-long.245","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.245","pdf_url":"https://aclanthology.org/2025.acl-long.245.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 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2025.acl-long.245.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5025695220","display_name":"Peijian Gu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Peijian Gu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5032868162","display_name":"Quan Wang","orcid":"https://orcid.org/0000-0002-5951-6752"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Quan Wang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5023341829","display_name":"Zhendong Mao","orcid":"https://orcid.org/0000-0001-5739-8126"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhendong Mao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"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":0.14085833,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"4941","last_page":"4954"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.9592999815940857,"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":0.9592999815940857,"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/training","display_name":"Training (meteorology)","score":0.6881077289581299},{"id":"https://openalex.org/keywords/computer-science","display_name":"Computer science","score":0.6701765060424805},{"id":"https://openalex.org/keywords/artificial-intelligence","display_name":"Artificial intelligence","score":0.3944269120693207},{"id":"https://openalex.org/keywords/risk-analysis","display_name":"Risk analysis (engineering)","score":0.3276737928390503},{"id":"https://openalex.org/keywords/business","display_name":"Business","score":0.13476067781448364}],"concepts":[{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.6881077289581299},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6701765060424805},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3944269120693207},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.3276737928390503},{"id":"https://openalex.org/C144133560","wikidata":"https://www.wikidata.org/wiki/Q4830453","display_name":"Business","level":0,"score":0.13476067781448364},{"id":"https://openalex.org/C121332964","wikidata":"https://www.wikidata.org/wiki/Q413","display_name":"Physics","level":0,"score":0.0},{"id":"https://openalex.org/C153294291","wikidata":"https://www.wikidata.org/wiki/Q25261","display_name":"Meteorology","level":1,"score":0.0}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.acl-long.245","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.245","pdf_url":"https://aclanthology.org/2025.acl-long.245.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 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.acl-long.245","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.245","pdf_url":"https://aclanthology.org/2025.acl-long.245.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 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[{"id":"https://openalex.org/G4729153174","display_name":null,"funder_award_id":"2023ZD0121200","funder_id":"https://openalex.org/F4320329860","funder_display_name":"National Science and Technology Major Project"},{"id":"https://openalex.org/G5737696022","display_name":null,"funder_award_id":"62376033","funder_id":"https://openalex.org/F4320321001","funder_display_name":"National Natural Science Foundation of China"},{"id":"https://openalex.org/G7854910236","display_name":null,"funder_award_id":"62222212","funder_id":"https://openalex.org/F4320321001","funder_display_name":"National Natural Science Foundation of China"}],"funders":[{"id":"https://openalex.org/F4320321001","display_name":"National Natural Science Foundation of China","ror":"https://ror.org/01h0zpd94"},{"id":"https://openalex.org/F4320329860","display_name":"National Science and Technology Major Project","ror":null}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4412886663.pdf","grobid_xml":"https://content.openalex.org/works/W4412886663.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/W230091440","https://openalex.org/W2390279801","https://openalex.org/W2233261550","https://openalex.org/W4391913857","https://openalex.org/W2358668433","https://openalex.org/W4396701345","https://openalex.org/W2810751659"],"abstract_inverted_index":{"The":[0],"rapid":[1],"advancement":[2],"of":[3,65,92,99,167,188,198],"large":[4],"language":[5],"models":[6],"(LLMs)":[7],"has":[8],"brought":[9],"about":[10],"increased":[11],"concerns":[12],"regarding":[13],"their":[14,100],"safety,":[15,83],"especially":[16],"as":[17,35,96,118],"adversaries":[18],"develop":[19],"jailbreak":[20,162],"techniques":[21],"to":[22,39,82,123,153],"bypass":[23],"LLMs'":[24],"safety":[25,30,66,103,130,171,189,201],"mechanism.Although":[26],"recent":[27],"work":[28],"on":[29,68,158,175,191],"training":[31,67,131,140,190,202],"with":[32,160],"modules":[33,127],"such":[34],"low-rank":[36,110],"adaptation":[37],"(LoRA)":[38],"resist":[40],"jailbreaks":[41],"shows":[42],"promise,":[43],"these":[44],"approaches":[45],"can":[46],"inadvertently":[47],"degrade":[48],"a":[49,57,108,134,147],"model's":[50,86],"general":[51],"utility.In":[52],"this":[53],"paper,":[54],"we":[55,145],"propose":[56,146],"novel":[58],"plugand-play":[59],"method":[60,182],"that":[61,180],"mitigates":[62,184],"the":[63,85,90,97,125,139,165,185,195,199],"impact":[64,187],"model":[69,192],"utility":[70,105,177,196],"by":[71,107],"explicitly":[72],"locating":[73],"and":[74,104,120],"leveraging":[75],"safety-critical":[76,119,143],"singular":[77,94,114],"vectors,":[78],"which":[79],"only":[80],"contribute":[81],"within":[84,128,142],"parameter":[87,156],"space.We":[88],"quantify":[89],"safety-criticality":[91],"each":[93],"vector":[95],"difference":[98],"importance":[101],"for":[102],"measured":[106],"corresponding":[109],"projection.The":[111],"top":[112],"scored":[113],"vectors":[115],"are":[116,121],"located":[117],"used":[122],"initialize":[124],"LoRA":[126],"existing":[129],"methods":[132,163],"in":[133,170],"plug-and-play":[135],"manner,":[136],"thereby":[137],"constraining":[138],"updates":[141],"parameters.Additionally,":[144],"dynamic":[148],"rank":[149],"number":[150],"determination":[151],"strategy":[152],"further":[154],"reduce":[155],"overhead.Experiments":[157],"HarmBench":[159],"multiple":[161],"validate":[164],"effectiveness":[166],"our":[168,181],"approach":[169],"training,":[172],"while":[173],"evaluations":[174],"several":[176],"benchmarks":[178],"demonstrate":[179],"successfully":[183],"adverse":[186],"utility,":[193],"enhancing":[194],"performance":[197],"evaluated":[200],"baselines.":[203]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2025-10-10T00:00:00"}