@inproceedings{lin-etal-2026-efficient,
title = "Efficient {LLM} Safety Evaluation through Multi-Agent Debate",
author = "Lin, Dachuan and
Shen, Guobin and
Yang, Zihao and
Liu, Tianrong and
Zhao, Dongcheng and
Zeng, Yi",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.76/",
pages = "1739--1755",
ISBN = "979-8-89176-459-0",
abstract = "Safety evaluation of large language models (LLMs) increasingly relies on LLM-as-a-judge pipelines, but strong judges can still be costly to deploy at scale. We present a multi-agent judge that uses small language models (SLMs) in structured debates among critic, defender, and judge agents. To evaluate these systems, we introduce ForesightSafety Human-Annotated JailBench (FS-HAJailBench), a human annotated jailbreak benchmark with 11,100 labeled interactions spanning diverse attack methods and target models. FS-HAJailBench provides fine-grained ground truth for measuring both target-model robustness and judge reliability. On this benchmark, our SLM-based framework improves over matched small-model baselines and prior multi-agent judges, while approaching GPT-4o-level agreement with lower estimated unit cost. Ablation results show that three debate rounds provide the best accuracy-efficiency trade-off among the current Qwen3-14B variants. These results indicate that structured debate is a practical way to improve scalable LLM safety evaluation."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="lin-etal-2026-efficient">
<titleInfo>
<title>Efficient LLM Safety Evaluation through Multi-Agent Debate</title>
</titleInfo>
<name type="personal">
<namePart type="given">Dachuan</namePart>
<namePart type="family">Lin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Guobin</namePart>
<namePart type="family">Shen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zihao</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tianrong</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dongcheng</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yi</namePart>
<namePart type="family">Zeng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Safety evaluation of large language models (LLMs) increasingly relies on LLM-as-a-judge pipelines, but strong judges can still be costly to deploy at scale. We present a multi-agent judge that uses small language models (SLMs) in structured debates among critic, defender, and judge agents. To evaluate these systems, we introduce ForesightSafety Human-Annotated JailBench (FS-HAJailBench), a human annotated jailbreak benchmark with 11,100 labeled interactions spanning diverse attack methods and target models. FS-HAJailBench provides fine-grained ground truth for measuring both target-model robustness and judge reliability. On this benchmark, our SLM-based framework improves over matched small-model baselines and prior multi-agent judges, while approaching GPT-4o-level agreement with lower estimated unit cost. Ablation results show that three debate rounds provide the best accuracy-efficiency trade-off among the current Qwen3-14B variants. These results indicate that structured debate is a practical way to improve scalable LLM safety evaluation.</abstract>
<identifier type="citekey">lin-etal-2026-efficient</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.76/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>1739</start>
<end>1755</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Efficient LLM Safety Evaluation through Multi-Agent Debate
%A Lin, Dachuan
%A Shen, Guobin
%A Yang, Zihao
%A Liu, Tianrong
%A Zhao, Dongcheng
%A Zeng, Yi
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F lin-etal-2026-efficient
%X Safety evaluation of large language models (LLMs) increasingly relies on LLM-as-a-judge pipelines, but strong judges can still be costly to deploy at scale. We present a multi-agent judge that uses small language models (SLMs) in structured debates among critic, defender, and judge agents. To evaluate these systems, we introduce ForesightSafety Human-Annotated JailBench (FS-HAJailBench), a human annotated jailbreak benchmark with 11,100 labeled interactions spanning diverse attack methods and target models. FS-HAJailBench provides fine-grained ground truth for measuring both target-model robustness and judge reliability. On this benchmark, our SLM-based framework improves over matched small-model baselines and prior multi-agent judges, while approaching GPT-4o-level agreement with lower estimated unit cost. Ablation results show that three debate rounds provide the best accuracy-efficiency trade-off among the current Qwen3-14B variants. These results indicate that structured debate is a practical way to improve scalable LLM safety evaluation.
%U https://aclanthology.org/2026.emnlp-main.76/
%P 1739-1755
Markdown (Informal)
[Efficient LLM Safety Evaluation through Multi-Agent Debate](https://aclanthology.org/2026.emnlp-main.76/) (Lin et al., EMNLP 2026)
ACL
- Dachuan Lin, Guobin Shen, Zihao Yang, Tianrong Liu, Dongcheng Zhao, and Yi Zeng. 2026. Efficient LLM Safety Evaluation through Multi-Agent Debate. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 1739–1755, Budapest, Hungary. Association for Computational Linguistics.