{"id":"https://openalex.org/W4416034407","doi":"https://doi.org/10.18653/v1/2025.findings-emnlp.1130","title":"Pruning Weights but Not Truth: Safeguarding Truthfulness While Pruning LLMs","display_name":"Pruning Weights but Not Truth: Safeguarding Truthfulness While Pruning LLMs","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4416034407","doi":"https://doi.org/10.18653/v1/2025.findings-emnlp.1130"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2025.findings-emnlp.1130","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-emnlp.1130","pdf_url":"https://aclanthology.org/2025.findings-emnlp.1130.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":"Findings of the Association for Computational Linguistics: EMNLP 2025","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2025.findings-emnlp.1130.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5105267000","display_name":"Y. Fu","orcid":"https://orcid.org/0000-0002-7370-7395"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yao Fu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5111937833","display_name":"Ruiteng Li","orcid":"https://orcid.org/0009-0002-4970-2145"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Runchao Li","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101797600","display_name":"Xianxuan Long","orcid":"https://orcid.org/0009-0006-4534-1775"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xianxuan Long","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5085239567","display_name":"Haoran Yu","orcid":"https://orcid.org/0000-0001-7304-2840"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Haotian Yu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5116337235","display_name":"Xiaotian Han","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiaotian Han","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5051574489","display_name":"Yu Yin","orcid":"https://orcid.org/0000-0002-5098-6109"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu Yin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5037159442","display_name":"Li Pan","orcid":"https://orcid.org/0000-0001-6157-3740"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Pan Li","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":null,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"20750","last_page":"20768"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.04529999941587448,"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":0.04529999941587448,"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":0.040300000458955765,"subfield":{"id":"https://openalex.org/subfields/3311","display_name":"Safety Research"},"field":{"id":"https://openalex.org/fields/33","display_name":"Social Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}},{"id":"https://openalex.org/T11492","display_name":"Academic integrity and plagiarism","score":0.03420000150799751,"subfield":{"id":"https://openalex.org/subfields/3311","display_name":"Safety Research"},"field":{"id":"https://openalex.org/fields/33","display_name":"Social Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/pruning","display_name":"Pruning","score":0.8082000017166138},{"id":"https://openalex.org/keywords/safeguarding","display_name":"Safeguarding","score":0.6304000020027161},{"id":"https://openalex.org/keywords/ranking","display_name":"Ranking (information retrieval)","score":0.2906999886035919}],"concepts":[{"id":"https://openalex.org/C108010975","wikidata":"https://www.wikidata.org/wiki/Q500094","display_name":"Pruning","level":2,"score":0.8082000017166138},{"id":"https://openalex.org/C2776743756","wikidata":"https://www.wikidata.org/wiki/Q5097921","display_name":"Safeguarding","level":2,"score":0.6304000020027161},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.44620001316070557},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.42320001125335693},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.32989999651908875},{"id":"https://openalex.org/C105795698","wikidata":"https://www.wikidata.org/wiki/Q12483","display_name":"Statistics","level":1,"score":0.2948000133037567},{"id":"https://openalex.org/C189430467","wikidata":"https://www.wikidata.org/wiki/Q7293293","display_name":"Ranking (information retrieval)","level":2,"score":0.2906999886035919},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.2802000045776367},{"id":"https://openalex.org/C162118730","wikidata":"https://www.wikidata.org/wiki/Q1128453","display_name":"Actuarial science","level":1,"score":0.26080000400543213},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.25699999928474426}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.findings-emnlp.1130","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-emnlp.1130","pdf_url":"https://aclanthology.org/2025.findings-emnlp.1130.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":"Findings of the Association for Computational Linguistics: EMNLP 2025","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.findings-emnlp.1130","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-emnlp.1130","pdf_url":"https://aclanthology.org/2025.findings-emnlp.1130.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":"Findings of the Association for Computational Linguistics: EMNLP 2025","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416034407.pdf","grobid_xml":"https://content.openalex.org/works/W4416034407.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Neural":[0],"network":[1],"pruning":[2,28,81],"has":[3],"emerged":[4],"as":[5],"a":[6,58,150],"promising":[7],"approach":[8,167],"for":[9,20,35,144,158,173],"deploying":[10],"LLMs":[11,66,175],"in":[12],"low-resource":[13],"scenarios":[14],"while":[15,136],"preserving":[16],"downstream":[17],"task":[18],"performance.However,":[19],"the":[21,51,100,155,169],"first":[22],"time,":[23],"we":[24,64,108,148],"reveal":[25],"that":[26,77,165],"such":[27],"disrupts":[29],"LLMs'":[30,133],"internal":[31],"activation":[32,125],"features":[33,49,130,139],"crucial":[34,59,88,102],"lie":[36,71,93,146],"detection,":[37],"where":[38],"probing":[39],"classifiers":[40],"(typically":[41],"small":[42],"logistic":[43],"regression":[44],"models)":[45],"trained":[46],"on":[47,84,99,121,186],"these":[48,69],"assess":[50],"truthfulness":[52],"of":[53,140],"LLM-generated":[54],"statements.This":[55],"discovery":[56],"raises":[57],"open":[60],"question:":[61],"how":[62],"can":[63],"prune":[65],"without":[67],"sacrificing":[68],"critical":[70,138],"detection":[72,94,171],"capabilities?Our":[73],"investigation":[74],"further":[75],"reveals":[76],"naively":[78],"adjusting":[79],"layer-wise":[80],"sparsity":[82],"based":[83],"importance":[85],"inadvertently":[86],"removes":[87],"weights,":[89],"failing":[90],"to":[91,153],"improve":[92],"performance":[95,135,185],"despite":[96],"its":[97],"reliance":[98],"most":[101],"LLM":[103,161],"layer.To":[104],"address":[105],"this":[106],"issue,":[107],"propose":[109],"Truthful":[110],"Pruning":[111],"aligned":[112],"by":[113],"Layer-wise":[114],"Outliers":[115],"(TPLO),":[116],"which":[117],"places":[118],"greater":[119],"emphasis":[120],"layers":[122],"with":[123],"more":[124],"outliers":[126],"and":[127,182,188],"stronger":[128],"discriminative":[129],"simultaneously.This":[131],"preserves":[132],"original":[134],"retaining":[137],"inner":[141],"states":[142],"needed":[143],"robust":[145],"detection.Moreover,":[147],"introduce":[149],"prompting":[151],"rule":[152],"enrich":[154],"TruthfulQA":[156],"benchmark":[157],"better":[159],"calibrating":[160],"pruning.Empirical":[162],"results":[163],"show":[164],"our":[166],"improves":[168],"hallucination":[170],"1":[172],"pruned":[174],"(achieving":[176],"88%":[177],"accuracy":[178],"at":[179],"50%":[180],"sparsity)":[181],"enhances":[183],"their":[184],"TruthfulQA.Codes":[187],"data":[189],"are":[190],"available":[191],"here":[192],"2":[193],".":[194]},"counts_by_year":[],"updated_date":"2026-07-14T23:27:15.235271","created_date":"2025-11-08T00:00:00"}