@inproceedings{liu-etal-2026-position-bias,
title = "From Position Bias to Position-Stable Agreement: Testing What {LLM} Judges Evaluate",
author = "Liu, Lixin and
Lin, Xiankun and
Hao, Zhezheng",
editor = "Wang, Yimu and
Choi, Yee Man and
Wu, Di and
Ouyang, Siqi and
Azadani, Mozhgan Nasr and
Fung, Yi R. (May)",
booktitle = "Proceedings of the 1st Workshop on Grounding Language Models: Learning Faithfully and Efficiently ({G}round{LM} 2026)",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.groundlm-1.61/",
pages = "839--856",
ISBN = "979-8-89176-478-1",
abstract = "Pairwise LLM-as-judge evaluation should preserve response identity when presentation order changes. Across four judge models, short A/B prompts yield 70{--}100{\%} second-position selection (72/80 comparisons overall). Targeted prompt ablations show format sensitivity: label changes alter parsed choices, and a neutral buffer produces complete first-position selection on 59 fixed-order comparisons, reversing rather than removing position preference. We use a position-swapped evaluation protocol, retaining a directional verdict only when both orders select the same response. In a new 16-condition GPT-4.1 factorial audit, analysis-first and a verdict prefix raise directional coverage by 5.6 and 6.5 percentage points, respectively. On 40 Arena-Hard queries with three generators, the 120 pairs comprise 62 directional results, no double ties, 14 swap failures, and 44 invalid pairs. Cross-Model Agreement (CMA) yields 25 unique query winners. A retrospective domain-routing check predicts 17 of these 25 labels using rules fitted outside each test fold. These labels and any oracle score remain judge-relative; human-preference validity and downstream training gains are not established."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liu-etal-2026-position-bias">
<titleInfo>
<title>From Position Bias to Position-Stable Agreement: Testing What LLM Judges Evaluate</title>
</titleInfo>
<name type="personal">
<namePart type="given">Lixin</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiankun</namePart>
<namePart type="family">Lin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhezheng</namePart>
<namePart type="family">Hao</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 1st Workshop on Grounding Language Models: Learning Faithfully and Efficiently (GroundLM 2026)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yimu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yee</namePart>
<namePart type="given">Man</namePart>
<namePart type="family">Choi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Di</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Siqi</namePart>
<namePart type="family">Ouyang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mozhgan</namePart>
<namePart type="given">Nasr</namePart>
<namePart type="family">Azadani</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yi</namePart>
<namePart type="given">R</namePart>
<namePart type="given">(May)</namePart>
<namePart type="family">Fung</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">979-8-89176-478-1</identifier>
</relatedItem>
<abstract>Pairwise LLM-as-judge evaluation should preserve response identity when presentation order changes. Across four judge models, short A/B prompts yield 70–100% second-position selection (72/80 comparisons overall). Targeted prompt ablations show format sensitivity: label changes alter parsed choices, and a neutral buffer produces complete first-position selection on 59 fixed-order comparisons, reversing rather than removing position preference. We use a position-swapped evaluation protocol, retaining a directional verdict only when both orders select the same response. In a new 16-condition GPT-4.1 factorial audit, analysis-first and a verdict prefix raise directional coverage by 5.6 and 6.5 percentage points, respectively. On 40 Arena-Hard queries with three generators, the 120 pairs comprise 62 directional results, no double ties, 14 swap failures, and 44 invalid pairs. Cross-Model Agreement (CMA) yields 25 unique query winners. A retrospective domain-routing check predicts 17 of these 25 labels using rules fitted outside each test fold. These labels and any oracle score remain judge-relative; human-preference validity and downstream training gains are not established.</abstract>
<identifier type="citekey">liu-etal-2026-position-bias</identifier>
<location>
<url>https://aclanthology.org/2026.groundlm-1.61/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>839</start>
<end>856</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T From Position Bias to Position-Stable Agreement: Testing What LLM Judges Evaluate
%A Liu, Lixin
%A Lin, Xiankun
%A Hao, Zhezheng
%Y Wang, Yimu
%Y Choi, Yee Man
%Y Wu, Di
%Y Ouyang, Siqi
%Y Azadani, Mozhgan Nasr
%Y Fung, Yi R. (May)
%S Proceedings of the 1st Workshop on Grounding Language Models: Learning Faithfully and Efficiently (GroundLM 2026)
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-478-1
%F liu-etal-2026-position-bias
%X Pairwise LLM-as-judge evaluation should preserve response identity when presentation order changes. Across four judge models, short A/B prompts yield 70–100% second-position selection (72/80 comparisons overall). Targeted prompt ablations show format sensitivity: label changes alter parsed choices, and a neutral buffer produces complete first-position selection on 59 fixed-order comparisons, reversing rather than removing position preference. We use a position-swapped evaluation protocol, retaining a directional verdict only when both orders select the same response. In a new 16-condition GPT-4.1 factorial audit, analysis-first and a verdict prefix raise directional coverage by 5.6 and 6.5 percentage points, respectively. On 40 Arena-Hard queries with three generators, the 120 pairs comprise 62 directional results, no double ties, 14 swap failures, and 44 invalid pairs. Cross-Model Agreement (CMA) yields 25 unique query winners. A retrospective domain-routing check predicts 17 of these 25 labels using rules fitted outside each test fold. These labels and any oracle score remain judge-relative; human-preference validity and downstream training gains are not established.
%U https://aclanthology.org/2026.groundlm-1.61/
%P 839-856
Markdown (Informal)
[From Position Bias to Position-Stable Agreement: Testing What LLM Judges Evaluate](https://aclanthology.org/2026.groundlm-1.61/) (Liu et al., GroundLM 2026)
ACL