@inproceedings{yang-etal-2026-harmmetric,
title = "{H}arm{M}etric Eval: Benchmarking Metrics and Judges for {LLM} Harmfulness Assessment",
author = "Yang, Langqi and
Chen, Yixuan and
Xiu, Kedong and
Zheng, Tianhang and
Zhou, Hao and
Ni, Wangze and
Chen, Lei and
Qin, Zhan and
Ren, Kui",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1598/",
pages = "34046--34064",
ISBN = "979-8-89176-462-0",
abstract = "Automatic harmfulness metrics and LLM-based judges are central to LLM safety assessment, yet differences in evaluation criteria, methodologies, and output formats may yield inconsistent judgments for the same model response. The lack of systematic research on these evaluators undermines the credibility and consistency of LLM safety assessments. To bridge this gap, we introduce HarmMetric Eval, a benchmark for both overall and fine-grained assessment of harmfulness evaluators. We construct a dataset of representative harmful prompts paired with diverse harmful responses and non-harmful counterparts across multiple categories. We propose a flexible scoring mechanism that rewards metrics for correctly ranking harmful responses above non-harmful ones, making it applicable to nearly all metrics and judges with varying output formats and scoring scales. Using HarmMetric Eval, we find that conventional reference-based metrics such as ROUGE and METEOR can outperform existing LLM-based judges in fine-grained harmfulness evaluation, challenging prevailing assumptions about LLM superiority. Fine-grained analysis reveals that LLM judges tend to overestimate irrelevant and useless responses. Furthermore, we build a new harmfulness judge by incorporating fine-grained criteria into its prompt template and leveraging reference-based metrics to fine-tune its base LLM. The resulting judge demonstrates leading performance on HarmMetric Eval and two external human-labeled datasets."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yang-etal-2026-harmmetric">
<titleInfo>
<title>HarmMetric Eval: Benchmarking Metrics and Judges for LLM Harmfulness Assessment</title>
</titleInfo>
<name type="personal">
<namePart type="given">Langqi</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yixuan</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kedong</namePart>
<namePart type="family">Xiu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tianhang</namePart>
<namePart type="family">Zheng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hao</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wangze</namePart>
<namePart type="family">Ni</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lei</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhan</namePart>
<namePart type="family">Qin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kui</namePart>
<namePart type="family">Ren</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>Findings of the Association for Computational Linguistics: EMNLP 2026</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">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Automatic harmfulness metrics and LLM-based judges are central to LLM safety assessment, yet differences in evaluation criteria, methodologies, and output formats may yield inconsistent judgments for the same model response. The lack of systematic research on these evaluators undermines the credibility and consistency of LLM safety assessments. To bridge this gap, we introduce HarmMetric Eval, a benchmark for both overall and fine-grained assessment of harmfulness evaluators. We construct a dataset of representative harmful prompts paired with diverse harmful responses and non-harmful counterparts across multiple categories. We propose a flexible scoring mechanism that rewards metrics for correctly ranking harmful responses above non-harmful ones, making it applicable to nearly all metrics and judges with varying output formats and scoring scales. Using HarmMetric Eval, we find that conventional reference-based metrics such as ROUGE and METEOR can outperform existing LLM-based judges in fine-grained harmfulness evaluation, challenging prevailing assumptions about LLM superiority. Fine-grained analysis reveals that LLM judges tend to overestimate irrelevant and useless responses. Furthermore, we build a new harmfulness judge by incorporating fine-grained criteria into its prompt template and leveraging reference-based metrics to fine-tune its base LLM. The resulting judge demonstrates leading performance on HarmMetric Eval and two external human-labeled datasets.</abstract>
<identifier type="citekey">yang-etal-2026-harmmetric</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1598/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>34046</start>
<end>34064</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T HarmMetric Eval: Benchmarking Metrics and Judges for LLM Harmfulness Assessment
%A Yang, Langqi
%A Chen, Yixuan
%A Xiu, Kedong
%A Zheng, Tianhang
%A Zhou, Hao
%A Ni, Wangze
%A Chen, Lei
%A Qin, Zhan
%A Ren, Kui
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F yang-etal-2026-harmmetric
%X Automatic harmfulness metrics and LLM-based judges are central to LLM safety assessment, yet differences in evaluation criteria, methodologies, and output formats may yield inconsistent judgments for the same model response. The lack of systematic research on these evaluators undermines the credibility and consistency of LLM safety assessments. To bridge this gap, we introduce HarmMetric Eval, a benchmark for both overall and fine-grained assessment of harmfulness evaluators. We construct a dataset of representative harmful prompts paired with diverse harmful responses and non-harmful counterparts across multiple categories. We propose a flexible scoring mechanism that rewards metrics for correctly ranking harmful responses above non-harmful ones, making it applicable to nearly all metrics and judges with varying output formats and scoring scales. Using HarmMetric Eval, we find that conventional reference-based metrics such as ROUGE and METEOR can outperform existing LLM-based judges in fine-grained harmfulness evaluation, challenging prevailing assumptions about LLM superiority. Fine-grained analysis reveals that LLM judges tend to overestimate irrelevant and useless responses. Furthermore, we build a new harmfulness judge by incorporating fine-grained criteria into its prompt template and leveraging reference-based metrics to fine-tune its base LLM. The resulting judge demonstrates leading performance on HarmMetric Eval and two external human-labeled datasets.
%U https://aclanthology.org/2026.findings-emnlp.1598/
%P 34046-34064
Markdown (Informal)
[HarmMetric Eval: Benchmarking Metrics and Judges for LLM Harmfulness Assessment](https://aclanthology.org/2026.findings-emnlp.1598/) (Yang et al., Findings 2026)
ACL
- Langqi Yang, Yixuan Chen, Kedong Xiu, Tianhang Zheng, Hao Zhou, Wangze Ni, Lei Chen, Zhan Qin, and Kui Ren. 2026. HarmMetric Eval: Benchmarking Metrics and Judges for LLM Harmfulness Assessment. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 34046–34064, Budapest, Hungary. Association for Computational Linguistics.