@inproceedings{tsunefuji-etal-2026-two,
title = "Two-Stage {LLM}-as-a-Judge Using Required and Bonus Criteria",
author = "Tsunefuji, Kohei and
Tamura, Akihiro and
Nakatsuji, Daichi and
Oobayashi, Hiroaki and
Hamada, Mitsuo and
Kato, Tsuneo",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.40/",
pages = "768--777",
ISBN = "979-8-89176-444-6",
abstract = "This study proposes a new LLM-as-a-judge method for relative evaluation of LLM responses based on both required criteria (i.e., minimal conditions that responses must satisfy) and bonus criteria (i.e., desirable additional qualities). Recent LLM-as-a-judge methods have improved the stability and accuracy of relative evaluation through the use of automatically generated criteria. However, existing methods rely solely on the required criteria, which may limit their ability to distinguish closely matched responses. To address this issue, the proposed method first evaluates responses based on required criteria and, in the case of a tie, further evaluates them using bonus criteria, enabling more fine-grained comparisons. Experiments on the Writing, Roleplay, and Humanities categories of the MT-Bench dataset demonstrate that the proposed method achieves higher agreement with human judgments than existing methods."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="tsunefuji-etal-2026-two">
<titleInfo>
<title>Two-Stage LLM-as-a-Judge Using Required and Bonus Criteria</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kohei</namePart>
<namePart type="family">Tsunefuji</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Akihiro</namePart>
<namePart type="family">Tamura</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Daichi</namePart>
<namePart type="family">Nakatsuji</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hiroaki</namePart>
<namePart type="family">Oobayashi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mitsuo</namePart>
<namePart type="family">Hamada</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tsuneo</namePart>
<namePart type="family">Kato</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>This study proposes a new LLM-as-a-judge method for relative evaluation of LLM responses based on both required criteria (i.e., minimal conditions that responses must satisfy) and bonus criteria (i.e., desirable additional qualities). Recent LLM-as-a-judge methods have improved the stability and accuracy of relative evaluation through the use of automatically generated criteria. However, existing methods rely solely on the required criteria, which may limit their ability to distinguish closely matched responses. To address this issue, the proposed method first evaluates responses based on required criteria and, in the case of a tie, further evaluates them using bonus criteria, enabling more fine-grained comparisons. Experiments on the Writing, Roleplay, and Humanities categories of the MT-Bench dataset demonstrate that the proposed method achieves higher agreement with human judgments than existing methods.</abstract>
<identifier type="citekey">tsunefuji-etal-2026-two</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.40/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>768</start>
<end>777</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Two-Stage LLM-as-a-Judge Using Required and Bonus Criteria
%A Tsunefuji, Kohei
%A Tamura, Akihiro
%A Nakatsuji, Daichi
%A Oobayashi, Hiroaki
%A Hamada, Mitsuo
%A Kato, Tsuneo
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F tsunefuji-etal-2026-two
%X This study proposes a new LLM-as-a-judge method for relative evaluation of LLM responses based on both required criteria (i.e., minimal conditions that responses must satisfy) and bonus criteria (i.e., desirable additional qualities). Recent LLM-as-a-judge methods have improved the stability and accuracy of relative evaluation through the use of automatically generated criteria. However, existing methods rely solely on the required criteria, which may limit their ability to distinguish closely matched responses. To address this issue, the proposed method first evaluates responses based on required criteria and, in the case of a tie, further evaluates them using bonus criteria, enabling more fine-grained comparisons. Experiments on the Writing, Roleplay, and Humanities categories of the MT-Bench dataset demonstrate that the proposed method achieves higher agreement with human judgments than existing methods.
%U https://aclanthology.org/2026.findings-aacl.40/
%P 768-777
Markdown (Informal)
[Two-Stage LLM-as-a-Judge Using Required and Bonus Criteria](https://aclanthology.org/2026.findings-aacl.40/) (Tsunefuji et al., Findings 2026)
ACL
- Kohei Tsunefuji, Akihiro Tamura, Daichi Nakatsuji, Hiroaki Oobayashi, Mitsuo Hamada, and Tsuneo Kato. 2026. Two-Stage LLM-as-a-Judge Using Required and Bonus Criteria. In Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026, pages 768–777, Hengqin, China. Association for Computational Linguistics.