{"id":"https://openalex.org/W4412887960","doi":"https://doi.org/10.18653/v1/2025.findings-acl.932","title":"ShieldHead: Decoding-time Safeguard for Large Language Models","display_name":"ShieldHead: Decoding-time Safeguard for Large Language Models","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4412887960","doi":"https://doi.org/10.18653/v1/2025.findings-acl.932"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2025.findings-acl.932","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-acl.932","pdf_url":"https://aclanthology.org/2025.findings-acl.932.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: ACL 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-acl.932.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Zitao Xuan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zitao Xuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5111968616","display_name":"Xiaofeng Mao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiaofeng Mao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100392734","display_name":"Da Chen","orcid":"https://orcid.org/0000-0003-2508-5461"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Da Chen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100327449","display_name":"Xin Zhang","orcid":"https://orcid.org/0000-0002-4086-1943"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xin Zhang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5108047157","display_name":"Yuhan Dong","orcid":"https://orcid.org/0000-0001-5275-1787"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yuhan Dong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5050300353","display_name":"Zhou Jun","orcid":"https://orcid.org/0000-0001-5810-5153"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jun Zhou","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":"18129","last_page":"18143"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.8036999702453613,"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":0.8036999702453613,"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/T10764","display_name":"Privacy-Preserving Technologies in Data","score":0.7473000288009644,"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":0.7243000268936157,"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/safeguard","display_name":"Safeguard","score":0.8776860237121582},{"id":"https://openalex.org/keywords/decoding-methods","display_name":"Decoding methods","score":0.7643694877624512},{"id":"https://openalex.org/keywords/computer-science","display_name":"Computer science","score":0.744757890701294},{"id":"https://openalex.org/keywords/programming-language","display_name":"Programming language","score":0.35492706298828125},{"id":"https://openalex.org/keywords/algorithm","display_name":"Algorithm","score":0.13476967811584473},{"id":"https://openalex.org/keywords/business","display_name":"Business","score":0.07038739323616028}],"concepts":[{"id":"https://openalex.org/C2780771206","wikidata":"https://www.wikidata.org/wiki/Q3271761","display_name":"Safeguard","level":2,"score":0.8776860237121582},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.7643694877624512},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.744757890701294},{"id":"https://openalex.org/C199360897","wikidata":"https://www.wikidata.org/wiki/Q9143","display_name":"Programming language","level":1,"score":0.35492706298828125},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.13476967811584473},{"id":"https://openalex.org/C144133560","wikidata":"https://www.wikidata.org/wiki/Q4830453","display_name":"Business","level":0,"score":0.07038739323616028},{"id":"https://openalex.org/C155202549","wikidata":"https://www.wikidata.org/wiki/Q178803","display_name":"International trade","level":1,"score":0.0}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.findings-acl.932","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-acl.932","pdf_url":"https://aclanthology.org/2025.findings-acl.932.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: ACL 2025","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.findings-acl.932","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-acl.932","pdf_url":"https://aclanthology.org/2025.findings-acl.932.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: ACL 2025","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4412887960.pdf","grobid_xml":"https://content.openalex.org/works/W4412887960.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/W3121667336","https://openalex.org/W2492640494","https://openalex.org/W4379054516","https://openalex.org/W2384877859","https://openalex.org/W2392961223","https://openalex.org/W2386423843","https://openalex.org/W1484464798"],"abstract_inverted_index":{"In":[0],"light":[1],"of":[2,6,66,92,123,174,205],"the":[3,11,88,93,121,164,172,175,182],"widespread":[4],"deployment":[5],"Large":[7],"Language":[8],"Models":[9],"(LLMs),":[10],"responsibility":[12],"for":[13],"safeguarding":[14],"and":[15,43,73,142,184],"regulating":[16],"LLMgenerated":[17],"content":[18],"has":[19],"taken":[20],"on":[21,87,181],"heightened":[22],"significance.Recent":[23],"advancements":[24],"in":[25,35,45,126],"LLM-based":[26,49,198],"moderation":[27,62,157,199],"methods,":[28],"e.g.,":[29],"LlamaGuard,":[30],"have":[31],"demonstrated":[32],"remarkable":[33,151],"promise":[34],"identifying":[36],"safety":[37],"risks":[38,125],"associated":[39],"with":[40,64,116,139,201],"both":[41,140],"inputs":[42],"outputs":[44],"human-AI":[46],"interactions.However,":[47],"integrating":[48],"safeguards":[50],"into":[51],"a":[52,61,84,97,130,178,190],"chatbot":[53,165],"system":[54],"requires":[55],"an":[56,112],"additional":[57],"inference":[58],"stage":[59],"involving":[60],"LLM":[63],"billions":[65],"parameters,":[67],"which":[68,145],"significantly":[69],"increases":[70],"computational":[71],"costs":[72],"reduces":[74],"overall":[75],"efficiency.In":[76],"this":[77],"paper,":[78],"we":[79],"demonstrate":[80,171],"that":[81],"simply":[82],"learning":[83],"classification":[85,104],"head":[86],"last-layer":[89],"hidden":[90],"states":[91],"dialogue":[94],"model":[95],"provides":[96],"strong":[98],"capability":[99],"to":[100,107,136],"identify":[101],"harmful":[102],"contents.The":[103],"head,":[105,119],"referred":[106],"as":[108,111],"Shield-Head,":[109],"serves":[110],"auxiliary":[113],"branch":[114],"paralleled":[115],"next-token-prediction":[117],"LM":[118],"enabling":[120],"detection":[122],"potential":[124],"past":[127],"text":[128],"sequences.Additionally,":[129],"label":[131],"disambiguation":[132],"technique":[133],"is":[134],"employed":[135],"supervise":[137],"ShieldHead":[138],"token-level":[141],"sentence-level":[143],"labels,":[144],"further":[146],"enhances":[147],"its":[148],"performance.Shield-Head":[149],"exhibits":[150],"efficiency":[152],"during":[153,163],"inference,":[154],"providing":[155],"real-time":[156],"results":[158,170],"alongside":[159],"token-wise":[160],"streaming":[161],"output":[162],"system's":[166],"decoding":[167],"phase.Extensive":[168],"experimental":[169],"superiority":[173],"proposed":[176],"framework:":[177],"state-of-theart":[179],"performance":[180],"XSTest":[183],"SafeRLHF":[185],"datasets":[186],"while":[187],"running":[188],"at":[189],"speed":[191],"about":[192],"300":[193],"faster":[194],"(<1ms)":[195],"than":[196],"previous":[197],"models":[200],"99%":[202],"less":[203],"parameters":[204],"LlamaGuard.":[206]},"counts_by_year":[],"updated_date":"2026-07-14T23:27:15.235271","created_date":"2025-10-10T00:00:00"}