@inproceedings{hu-kao-2026-decision,
title = "Decision Rigidity in {LLM} Disinformation Detection: Behavioral and Representational Evidence",
author = "Hu, Ming-Jhe and
Kao, Hung-Yu",
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.1515/",
doi = "10.18653/2026.findings-emnlp.1515",
pages = "32248--32286",
ISBN = "979-8-89176-462-0",
abstract = "LLMs deployed as judges in high-stakes classification are commonly assumed to let their generated rationale shape the verdict; we show this assumption fails in disinformation detection. We introduce the Multi-Bias Injection Protocol (MBIP), which perturbs the reasoning stage along four cognitive axes (confirmation, anchoring, framing, authority) while holding the input fixed, and quantifies induced rationale variation with technique-count metrics anchored on the number of persuasion techniques the model reports. Across five disinformation datasets, MBIP produces substantial rationale variation but near-zero verdict flipping; an explicit rationale-deference prompt raises the flip rate only modestly, to 0.118{--}0.300 across three open-weight models, with API-only gpt-4o-mini in the same range. We call this verdict-level invariance Decision Rigidity. At the representational level, our residual-stream probe localizes a candidate commitment band late in the network, where biased and neutral final-class trajectories converge. The band{'}s location is architectural, replicated by a neutral-vs-neutral null; a within-band alignment-rate elevation and a random-lexicon control gap co-occur with the behavioral invariance and attenuate in Gemma-2-9B. All representational evidence is correlational. Generalization to other high-prior, evidence-dependent classifications and attenuation on reasoning-intensive tasks remain falsifiable hypotheses."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hu-kao-2026-decision">
<titleInfo>
<title>Decision Rigidity in LLM Disinformation Detection: Behavioral and Representational Evidence</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ming-Jhe</namePart>
<namePart type="family">Hu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hung-Yu</namePart>
<namePart type="family">Kao</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>LLMs deployed as judges in high-stakes classification are commonly assumed to let their generated rationale shape the verdict; we show this assumption fails in disinformation detection. We introduce the Multi-Bias Injection Protocol (MBIP), which perturbs the reasoning stage along four cognitive axes (confirmation, anchoring, framing, authority) while holding the input fixed, and quantifies induced rationale variation with technique-count metrics anchored on the number of persuasion techniques the model reports. Across five disinformation datasets, MBIP produces substantial rationale variation but near-zero verdict flipping; an explicit rationale-deference prompt raises the flip rate only modestly, to 0.118–0.300 across three open-weight models, with API-only gpt-4o-mini in the same range. We call this verdict-level invariance Decision Rigidity. At the representational level, our residual-stream probe localizes a candidate commitment band late in the network, where biased and neutral final-class trajectories converge. The band’s location is architectural, replicated by a neutral-vs-neutral null; a within-band alignment-rate elevation and a random-lexicon control gap co-occur with the behavioral invariance and attenuate in Gemma-2-9B. All representational evidence is correlational. Generalization to other high-prior, evidence-dependent classifications and attenuation on reasoning-intensive tasks remain falsifiable hypotheses.</abstract>
<identifier type="citekey">hu-kao-2026-decision</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1515</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1515/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>32248</start>
<end>32286</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Decision Rigidity in LLM Disinformation Detection: Behavioral and Representational Evidence
%A Hu, Ming-Jhe
%A Kao, Hung-Yu
%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 hu-kao-2026-decision
%X LLMs deployed as judges in high-stakes classification are commonly assumed to let their generated rationale shape the verdict; we show this assumption fails in disinformation detection. We introduce the Multi-Bias Injection Protocol (MBIP), which perturbs the reasoning stage along four cognitive axes (confirmation, anchoring, framing, authority) while holding the input fixed, and quantifies induced rationale variation with technique-count metrics anchored on the number of persuasion techniques the model reports. Across five disinformation datasets, MBIP produces substantial rationale variation but near-zero verdict flipping; an explicit rationale-deference prompt raises the flip rate only modestly, to 0.118–0.300 across three open-weight models, with API-only gpt-4o-mini in the same range. We call this verdict-level invariance Decision Rigidity. At the representational level, our residual-stream probe localizes a candidate commitment band late in the network, where biased and neutral final-class trajectories converge. The band’s location is architectural, replicated by a neutral-vs-neutral null; a within-band alignment-rate elevation and a random-lexicon control gap co-occur with the behavioral invariance and attenuate in Gemma-2-9B. All representational evidence is correlational. Generalization to other high-prior, evidence-dependent classifications and attenuation on reasoning-intensive tasks remain falsifiable hypotheses.
%R 10.18653/2026.findings-emnlp.1515
%U https://aclanthology.org/2026.findings-emnlp.1515/
%U https://doi.org/10.18653/2026.findings-emnlp.1515
%P 32248-32286
Markdown (Informal)
[Decision Rigidity in LLM Disinformation Detection: Behavioral and Representational Evidence](https://aclanthology.org/2026.findings-emnlp.1515/) (Hu & Kao, Findings 2026)
ACL