{"id":"https://openalex.org/W7104869295","doi":"https://doi.org/10.18653/2026.emnlp-main.76","title":"Efficient LLM Safety Evaluation through Multi-Agent Debate","display_name":"Efficient LLM Safety Evaluation through Multi-Agent Debate","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7104869295","doi":"https://doi.org/10.18653/2026.emnlp-main.76"},"language":"en","primary_location":{"id":"doi:10.18653/2026.emnlp-main.76","doi":"https://doi.org/10.18653/2026.emnlp-main.76","is_oa":true,"landing_page_url":"https://doi.org/10.18653/2026.emnlp-main.76","pdf_url":"https://aclanthology.org/2026.emnlp-main.76.pdf","source":{"id":"https://openalex.org/S4363608991","display_name":"Conference on Empirical Methods in Natural Language Processing (EMNLP)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","is_traditional":true,"indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"bronze","oa_url":"https://aclanthology.org/2026.emnlp-main.76.pdf","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Dachuan Lin","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Dachuan Lin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Guobin Shen","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I19820366","display_name":"Chinese Academy of Sciences","ror":"https://ror.org/034t30j35","country_code":"CN","type":"government","lineage":["https://openalex.org/I19820366"]},{"id":"https://openalex.org/I4210112150","display_name":"Institute of Automation","ror":"https://ror.org/022c3hy66","country_code":"CN","type":"facility","lineage":["https://openalex.org/I19820366","https://openalex.org/I4210112150"]},{"id":"https://openalex.org/I4210165038","display_name":"University of Chinese Academy of Sciences","ror":"https://ror.org/05qbk4x57","country_code":"CN","type":"education","lineage":["https://openalex.org/I19820366","https://openalex.org/I4210165038"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Guobin Shen","raw_affiliation_strings":["Institute of Automation , Chinese Academy of Sciences ,","University of Chinese Academy of Sciences ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Institute of Automation , Chinese Academy of Sciences ,","institution_ids":["https://openalex.org/I19820366","https://openalex.org/I4210112150"]},{"raw_affiliation_string":"University of Chinese Academy of Sciences ,","institution_ids":["https://openalex.org/I4210165038"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Zihao Yang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I177725633","display_name":"Chinese University of Hong Kong","ror":"https://ror.org/00t33hh48","country_code":"HK","type":"education","lineage":["https://openalex.org/I177725633"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Zihao Yang","raw_affiliation_strings":["The Chinese University of Hong Kong ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"The Chinese University of Hong Kong ,","institution_ids":["https://openalex.org/I177725633"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Tianrong Liu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I177725633","display_name":"Chinese University of Hong Kong","ror":"https://ror.org/00t33hh48","country_code":"HK","type":"education","lineage":["https://openalex.org/I177725633"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Tianrong Liu","raw_affiliation_strings":["The Chinese University of Hong Kong ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"The Chinese University of Hong Kong ,","institution_ids":["https://openalex.org/I177725633"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Dongcheng Zhao","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I19820366","display_name":"Chinese Academy of Sciences","ror":"https://ror.org/034t30j35","country_code":"CN","type":"government","lineage":["https://openalex.org/I19820366"]},{"id":"https://openalex.org/I4210112150","display_name":"Institute of Automation","ror":"https://ror.org/022c3hy66","country_code":"CN","type":"facility","lineage":["https://openalex.org/I19820366","https://openalex.org/I4210112150"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Dongcheng Zhao","raw_affiliation_strings":["Institute of Automation , Chinese Academy of Sciences ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Institute of Automation , Chinese Academy of Sciences ,","institution_ids":["https://openalex.org/I19820366","https://openalex.org/I4210112150"]}]},{"author_position":"last","author":{"id":null,"display_name":"Yi Zeng","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yi Zeng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":4,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":8.2459,"has_fulltext":true,"cited_by_count":7,"citation_normalized_percentile":{"value":0.97238926,"is_in_top_1_percent":false,"is_in_top_10_percent":true},"cited_by_percentile_year":{"min":89,"max":99},"biblio":{"volume":null,"issue":null,"first_page":"1739","last_page":"1755"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.9976000189781189,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.9976000189781189,"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":0.0005000000237487257,"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/T12262","display_name":"Hate Speech and Cyberbullying Detection","score":0.00039999998989515007,"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/llm-safety-evaluation","display_name":"LLM safety evaluation","score":0.8769999742507935},{"id":"https://openalex.org/keywords/jailbreak-attacks","display_name":"jailbreak attacks","score":0.652999997138977},{"id":"https://openalex.org/keywords/multi-agent-debate","display_name":"multi-agent debate","score":0.9309999942779541},{"id":"https://openalex.org/keywords/llm-as-a-judge","display_name":"LLM-as-a-judge","score":0.8610000014305115},{"id":"https://openalex.org/keywords/small-language-models","display_name":"small language models","score":0.9200000166893005},{"id":"https://openalex.org/keywords/manual-annotation","display_name":"manual annotation","score":0.5699999928474426},{"id":"https://openalex.org/keywords/model-robustness","display_name":"model robustness","score":0.6039999723434448},{"id":"https://openalex.org/keywords/cost-efficiency","display_name":"cost efficiency","score":0.5299999713897705}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6323000192642212},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6301000118255615},{"id":"https://openalex.org/C37736160","wikidata":"https://www.wikidata.org/wiki/Q1801315","display_name":"Adversarial system","level":2,"score":0.5954999923706055},{"id":"https://openalex.org/C48044578","wikidata":"https://www.wikidata.org/wiki/Q727490","display_name":"Scalability","level":2,"score":0.5900999903678894},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.553600013256073},{"id":"https://openalex.org/C2780801425","wikidata":"https://www.wikidata.org/wiki/Q5164392","display_name":"Construct (python library)","level":2,"score":0.48829999566078186},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.3889000117778778},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.38040000200271606},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.35010001063346863},{"id":"https://openalex.org/C146849305","wikidata":"https://www.wikidata.org/wiki/Q370766","display_name":"Ground truth","level":2,"score":0.33649998903274536},{"id":"https://openalex.org/C38652104","wikidata":"https://www.wikidata.org/wiki/Q3510521","display_name":"Computer security","level":1,"score":0.3319999873638153},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3122999966144562},{"id":"https://openalex.org/C2780966255","wikidata":"https://www.wikidata.org/wiki/Q5474306","display_name":"Foundation (evidence)","level":2,"score":0.296099990606308},{"id":"https://openalex.org/C2778571376","wikidata":"https://www.wikidata.org/wiki/Q1355821","display_name":"Frontier","level":2,"score":0.2727999985218048},{"id":"https://openalex.org/C2777877512","wikidata":"https://www.wikidata.org/wiki/Q1116097","display_name":"Common ground","level":2,"score":0.26840001344680786},{"id":"https://openalex.org/C2522767166","wikidata":"https://www.wikidata.org/wiki/Q2374463","display_name":"Data science","level":1,"score":0.26600000262260437},{"id":"https://openalex.org/C3017944768","wikidata":"https://www.wikidata.org/wiki/Q1450463","display_name":"Poison control","level":2,"score":0.26570001244544983},{"id":"https://openalex.org/C134121241","wikidata":"https://www.wikidata.org/wiki/Q899301","display_name":"Yield (engineering)","level":2,"score":0.26269999146461487},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.25529998540878296},{"id":"https://openalex.org/C184356942","wikidata":"https://www.wikidata.org/wiki/Q830382","display_name":"Best practice","level":2,"score":0.251800000667572}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.18653/2026.emnlp-main.76","doi":"https://doi.org/10.18653/2026.emnlp-main.76","is_oa":true,"landing_page_url":"https://doi.org/10.18653/2026.emnlp-main.76","pdf_url":"https://aclanthology.org/2026.emnlp-main.76.pdf","source":{"id":"https://openalex.org/S4363608991","display_name":"Conference on Empirical Methods in Natural Language Processing (EMNLP)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2511.06396","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2511.06396","pdf_url":"https://arxiv.org/pdf/2511.06396","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by-sa","license_id":"https://openalex.org/licenses/cc-by-sa","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2511.06396","doi":"https://doi.org/10.48550/arxiv.2511.06396","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2511.06396","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.18653/2026.emnlp-main.76","doi":"https://doi.org/10.18653/2026.emnlp-main.76","is_oa":true,"landing_page_url":"https://doi.org/10.18653/2026.emnlp-main.76","pdf_url":"https://aclanthology.org/2026.emnlp-main.76.pdf","source":{"id":"https://openalex.org/S4363608991","display_name":"Conference on Empirical Methods in Natural Language Processing (EMNLP)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320321133","display_name":"Chinese Academy of Sciences","ror":"https://ror.org/034t30j35"},{"id":"https://openalex.org/F4320335639","display_name":"Institute of Automation, Chinese Academy of Sciences","ror":"https://ror.org/022c3hy66"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7104869295.pdf","grobid_xml":"https://content.openalex.org/works/W7104869295.grobid-xml"},"referenced_works_count":22,"referenced_works":["https://openalex.org/W4385374425","https://openalex.org/W4385849309","https://openalex.org/W4386907292","https://openalex.org/W4387559087","https://openalex.org/W4387994739","https://openalex.org/W4389713766","https://openalex.org/W4391832496","https://openalex.org/W4393160208","https://openalex.org/W4393946427","https://openalex.org/W4400434624","https://openalex.org/W4401042350","https://openalex.org/W4401306838","https://openalex.org/W4401657794","https://openalex.org/W4410641004","https://openalex.org/W4412945143","https://openalex.org/W4414281281","https://openalex.org/W4415689913","https://openalex.org/W4416596100","https://openalex.org/W4417292349","https://openalex.org/W6885068669","https://openalex.org/W6966618268","https://openalex.org/W7120088840"],"related_works":["https://openalex.org/W4416113418","https://openalex.org/W7131133318","https://openalex.org/W7140112949","https://openalex.org/W4414891524","https://openalex.org/W4414580537","https://openalex.org/W4416047745","https://openalex.org/W4415068033","https://openalex.org/W4412944925","https://openalex.org/W4402670748","https://openalex.org/W7167874103"],"abstract_inverted_index":{"Safety":[0],"evaluation":[1],"of":[2],"large":[3],"language":[4,30],"models":[5,31],"(LLMs)":[6],"increasingly":[7],"relies":[8],"on":[9],"LLM-as-a-judge":[10],"pipelines,":[11],"but":[12],"strong":[13],"judges":[14],"can":[15],"still":[16],"be":[17],"costly":[18],"to":[19,127],"deploy":[20],"at":[21,137],"scale.We":[22],"present":[23],"a":[24,51,124],"multi-agent":[25,91],"judge":[26,40,77],"that":[27,104,120],"uses":[28],"small":[29],"(SLMs)":[32],"in":[33],"structured":[34,121],"debates":[35],"among":[36,113],"critic,":[37],"defender,":[38],"and":[39,64,76,89],"agents.To":[41],"evaluate":[42],"these":[43],"systems,":[44],"we":[45],"introduce":[46],"ForesightSafety":[47],"Human-Annotated":[48],"JailBench":[49],"(FS-HAJailBench),":[50],"human":[52],"annotated":[53],"jailbreak":[54],"benchmark":[55],"with":[56,97],"11,100":[57],"labeled":[58],"interactions":[59],"spanning":[60],"diverse":[61],"attack":[62],"methods":[63],"target":[65],"models.FS-HAJailBench":[66],"provides":[67],"fine-grained":[68],"ground":[69],"truth":[70],"for":[71],"measuring":[72],"both":[73],"target-model":[74],"robustness":[75],"reliability.On":[78],"this":[79],"benchmark,":[80],"our":[81],"SLM-based":[82],"framework":[83],"improves":[84],"over":[85],"matched":[86],"small-model":[87],"baselines":[88],"prior":[90],"judges,":[92],"while":[93],"approaching":[94],"GPT-4o-level":[95],"agreement":[96],"lower":[98],"estimated":[99],"unit":[100],"cost.Ablation":[101],"results":[102,118],"show":[103],"three":[105],"debate":[106,122],"rounds":[107],"provide":[108],"the":[109,114],"best":[110],"accuracy-efficiency":[111],"trade-off":[112],"current":[115],"Qwen3-14B":[116],"variants.These":[117],"indicate":[119],"is":[123,135],"practical":[125],"way":[126],"improve":[128],"scalable":[129],"LLM":[130],"safety":[131],"evaluation.The":[132],"source":[133],"code":[134],"available":[136],"https://github.com/Beijing-AISI/":[138],"ForesightSafety-HAJailBench.":[139]},"counts_by_year":[{"year":2026,"cited_by_count":6},{"year":2025,"cited_by_count":1}],"updated_date":"2026-11-08T12:12:28.789561","created_date":"2025-11-12T00:00:00"}