@inproceedings{ansari-etal-2026-reasoning,
title = "Reasoning Well, Judging Poorly: Theory of Mind Scaffolding Exposes the Limits of {LLM} Persuasion Judges",
author = "Ansari, Baktash and
Dewan, Mouly and
Mashhadi, Afra",
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.529/",
pages = "11088--11109",
ISBN = "979-8-89176-462-0",
abstract = "LLM-as-judge frameworks are widely used for subjective NLP evaluation, yet their reliability on social-cognition tasks remains unclear. We investigate whether Theory-of-Mind (ToM) scaffolding can help LLM judges evaluate persuasion on Reddit ChangeMyView (CMV). We test four judges (GPT-5, GPT-4.1, Claude Sonnet 4.6, Qwen3-30B-A3B-Instruct-2507) under three conditions that share the same ToM annotations but vary how they are presented: No-ToM, Full-ToM (flat), and our proposed PAToM (Persuasion-Aware ToM), which structures reasoning across five sequential belief-stack turns. Our results show that better reasoning does not lead to better verdicts. Although PAToM improves reasoning coherence by 1.6x, semantic coverage by 1.1-1.8x, and our proposed process metric PRISM by 2-3x, all four judges still struggle to reliably distinguish persuasive comments from high-quality non-persuasive ones. Process-level analysis shows that reasoning quality and verdict accuracy are empirically decoupled under this scaffolding, an open challenge for prompt-based approaches to persuasion judging. We release prompts, ToM annotations, the PRISM metric suite, and a 1000-item benchmark."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ansari-etal-2026-reasoning">
<titleInfo>
<title>Reasoning Well, Judging Poorly: Theory of Mind Scaffolding Exposes the Limits of LLM Persuasion Judges</title>
</titleInfo>
<name type="personal">
<namePart type="given">Baktash</namePart>
<namePart type="family">Ansari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mouly</namePart>
<namePart type="family">Dewan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Afra</namePart>
<namePart type="family">Mashhadi</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>LLM-as-judge frameworks are widely used for subjective NLP evaluation, yet their reliability on social-cognition tasks remains unclear. We investigate whether Theory-of-Mind (ToM) scaffolding can help LLM judges evaluate persuasion on Reddit ChangeMyView (CMV). We test four judges (GPT-5, GPT-4.1, Claude Sonnet 4.6, Qwen3-30B-A3B-Instruct-2507) under three conditions that share the same ToM annotations but vary how they are presented: No-ToM, Full-ToM (flat), and our proposed PAToM (Persuasion-Aware ToM), which structures reasoning across five sequential belief-stack turns. Our results show that better reasoning does not lead to better verdicts. Although PAToM improves reasoning coherence by 1.6x, semantic coverage by 1.1-1.8x, and our proposed process metric PRISM by 2-3x, all four judges still struggle to reliably distinguish persuasive comments from high-quality non-persuasive ones. Process-level analysis shows that reasoning quality and verdict accuracy are empirically decoupled under this scaffolding, an open challenge for prompt-based approaches to persuasion judging. We release prompts, ToM annotations, the PRISM metric suite, and a 1000-item benchmark.</abstract>
<identifier type="citekey">ansari-etal-2026-reasoning</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.529/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>11088</start>
<end>11109</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Reasoning Well, Judging Poorly: Theory of Mind Scaffolding Exposes the Limits of LLM Persuasion Judges
%A Ansari, Baktash
%A Dewan, Mouly
%A Mashhadi, Afra
%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 ansari-etal-2026-reasoning
%X LLM-as-judge frameworks are widely used for subjective NLP evaluation, yet their reliability on social-cognition tasks remains unclear. We investigate whether Theory-of-Mind (ToM) scaffolding can help LLM judges evaluate persuasion on Reddit ChangeMyView (CMV). We test four judges (GPT-5, GPT-4.1, Claude Sonnet 4.6, Qwen3-30B-A3B-Instruct-2507) under three conditions that share the same ToM annotations but vary how they are presented: No-ToM, Full-ToM (flat), and our proposed PAToM (Persuasion-Aware ToM), which structures reasoning across five sequential belief-stack turns. Our results show that better reasoning does not lead to better verdicts. Although PAToM improves reasoning coherence by 1.6x, semantic coverage by 1.1-1.8x, and our proposed process metric PRISM by 2-3x, all four judges still struggle to reliably distinguish persuasive comments from high-quality non-persuasive ones. Process-level analysis shows that reasoning quality and verdict accuracy are empirically decoupled under this scaffolding, an open challenge for prompt-based approaches to persuasion judging. We release prompts, ToM annotations, the PRISM metric suite, and a 1000-item benchmark.
%U https://aclanthology.org/2026.findings-emnlp.529/
%P 11088-11109
Markdown (Informal)
[Reasoning Well, Judging Poorly: Theory of Mind Scaffolding Exposes the Limits of LLM Persuasion Judges](https://aclanthology.org/2026.findings-emnlp.529/) (Ansari et al., Findings 2026)
ACL