{"id":"https://openalex.org/W7166856062","doi":"https://doi.org/10.18653/v1/2026.acl-long.1572","title":"More Thinking, Less Talking: Internalizing Deliberative Safety into LLM Parameters","display_name":"More Thinking, Less Talking: Internalizing Deliberative Safety into LLM Parameters","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166856062","doi":"https://doi.org/10.18653/v1/2026.acl-long.1572"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.acl-long.1572","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1572","pdf_url":"https://aclanthology.org/2026.acl-long.1572.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 64th 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/2026.acl-long.1572.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5139842138","display_name":"Guan Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Guan Wang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139759210","display_name":"Xuehai Tang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xuehai Tang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5064648779","display_name":"Biyu Zhou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Biyu Zhou","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139753539","display_name":"Jizhong Han","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jizhong Han","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5139848109","display_name":"Songlin Hu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Songlin Hu","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":0.87116564,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"34079","last_page":"34094"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T13295","display_name":"Safety Systems Engineering in Autonomy","score":0.09870000183582306,"subfield":{"id":"https://openalex.org/subfields/2213","display_name":"Safety, Risk, Reliability and Quality"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T13295","display_name":"Safety Systems Engineering in Autonomy","score":0.09870000183582306,"subfield":{"id":"https://openalex.org/subfields/2213","display_name":"Safety, Risk, Reliability and Quality"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"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":0.07909999787807465,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.03460000082850456,"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/action","display_name":"Action (physics)","score":0.3440000116825104},{"id":"https://openalex.org/keywords/perspective","display_name":"Perspective (graphical)","score":0.2728999853134155},{"id":"https://openalex.org/keywords/agency","display_name":"Agency (philosophy)","score":0.27160000801086426},{"id":"https://openalex.org/keywords/context","display_name":"Context (archaeology)","score":0.26019999384880066},{"id":"https://openalex.org/keywords/matching","display_name":"Matching (statistics)","score":0.25540000200271606}],"concepts":[{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.5670999884605408},{"id":"https://openalex.org/C77805123","wikidata":"https://www.wikidata.org/wiki/Q161272","display_name":"Social psychology","level":1,"score":0.42890000343322754},{"id":"https://openalex.org/C2780791683","wikidata":"https://www.wikidata.org/wiki/Q846785","display_name":"Action (physics)","level":2,"score":0.3440000116825104},{"id":"https://openalex.org/C75630572","wikidata":"https://www.wikidata.org/wiki/Q538904","display_name":"Applied psychology","level":1,"score":0.31150001287460327},{"id":"https://openalex.org/C12713177","wikidata":"https://www.wikidata.org/wiki/Q1900281","display_name":"Perspective (graphical)","level":2,"score":0.2728999853134155},{"id":"https://openalex.org/C108170787","wikidata":"https://www.wikidata.org/wiki/Q3951828","display_name":"Agency (philosophy)","level":2,"score":0.27160000801086426},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.26019999384880066},{"id":"https://openalex.org/C165064840","wikidata":"https://www.wikidata.org/wiki/Q1321061","display_name":"Matching (statistics)","level":2,"score":0.25540000200271606},{"id":"https://openalex.org/C26760741","wikidata":"https://www.wikidata.org/wiki/Q160402","display_name":"Perception","level":2,"score":0.2547999918460846},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.24459999799728394}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.1572","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1572","pdf_url":"https://aclanthology.org/2026.acl-long.1572.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 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.acl-long.1572","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1572","pdf_url":"https://aclanthology.org/2026.acl-long.1572.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 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[{"id":"https://openalex.org/G5208670511","display_name":null,"funder_award_id":"U24A20335","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"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166856062.pdf","grobid_xml":"https://content.openalex.org/works/W7166856062.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Prevailing":[0],"safety":[1,55,73,137,142],"alignment":[2,138],"methods":[3],"still":[4],"leave":[5],"Large":[6],"Language":[7],"Models":[8],"(LLMs)":[9],"vulnerable":[10,172],"to":[11,57,85,119,156,170,195],"sophisticated":[12],"jailbreak":[13,192],"attacks.To":[14],"bolster":[15],"defenses,":[16],"explicit":[17,78],"reasoning":[18,37,79,143],"mechanisms":[19],"like":[20],"Safety-oriented":[21],"Chain-of-Thought":[22],"(SCoT)":[23],"have":[24],"emerged,":[25],"significantly":[26],"enhancing":[27],"robustness.However,":[28],"this":[29,62,127],"transparency":[30],"introduces":[31],"a":[32,41,100,115,135,158,162,183],"critical":[33],"trade-off:":[34],"the":[35,46,53,68,75,87,104,120,154,168],"exposed":[36],"process":[38,89],"itself":[39],"becomes":[40],"new":[42],"attack":[43],"surface,":[44],"risking":[45],"leakage":[47],"of":[48,71,77,107,131],"harmful":[49],"information":[50],"and":[51],"revealing":[52],"model's":[54,121],"logic":[56],"adversaries.This":[58],"paper":[59],"directly":[60],"confronts":[61],"dilemma,":[63],"asking:":[64],"Can":[65],"we":[66],"achieve":[67],"full":[69],"benefits":[70],"deliberative":[72,88],"without":[74],"costs":[76],"generation?We":[80],"propose":[81],"Safety":[82],"Reasoning":[83,133],"Internalization":[84,130],"make":[86],"in":[90,99],"SCoT":[91,109,173],"\"available":[92],"but":[93],"not":[94],"visible\".This":[95],"approach":[96],"is":[97],"grounded":[98],"key":[101],"theoretical":[102],"insight:":[103],"corrective":[105],"influence":[106],"an":[108,145],"can":[110],"be":[111],"effectively":[112],"approximated":[113],"by":[114],"targeted,":[116],"low-rank":[117],"update":[118],"Feed-Forward":[122],"Network":[123],"(FFN)":[124],"layers.We":[125],"operationalize":[126],"through":[128],"Hierarchical":[129],"Adversarially-Guided":[132],"(HIAR),":[134],"layer-wise":[136],"framework":[139],"that":[140,180],"internalizes":[141],"into":[144],"implicit":[146],"computational":[147],"pathway":[148],"using":[149],"Low-Rank":[150],"Adaptation":[151],"(LoRA).HIAR":[152],"enables":[153],"model":[155],"reach":[157],"safe":[159],"conclusion":[160],"within":[161],"single":[163],"forward":[164],"pass,":[165],"entirely":[166],"eliminating":[167],"need":[169],"generate":[171],"text.Extensive":[174],"experiments":[175],"on":[176],"various":[177],"LLMs":[178],"demonstrate":[179],"HIAR":[181],"achieves":[182],"43%":[184],"lower":[185],"Attack":[186],"Success":[187],"Rate":[188],"(ASR)":[189],"against":[190],"distinct":[191],"attacks":[193],"compared":[194],"strong":[196],"baselines.":[197]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}