@inproceedings{hacker-bar-2026-saying,
title = "Saying vs. Seeing: Evaluating {LLM} Conflict Awareness and Resolution Preference under Latent Claim-Evidence Disagreement",
author = "Hacker, Daniel and
Bar, Kfir",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.1727/",
pages = "39397--39423",
ISBN = "979-8-89176-459-0",
abstract = "Language models are routinely exposed to contexts where an authoritative claim conflicts with a conclusion derivable from underlying evidence. Unlike explicit contradictions, such conflicts are latent: they emerge only after a full multi-step derivation, a setting current benchmarks largely overlook. We introduce and publicly release CLECK (CLaim Evidence Conflict of Knowledge), a counterfactual benchmark of 5,922 instances (evaluated on a balanced 2,669-instance subset) probing this blind spot across eleven reasoning strategies. Among instruction-tuned models, the capability to recover both competing answers does not translate into recognizing their conflict at inference: models default to the authoritative claim. When the conflict does surface in a model{'}s reasoning, preference shifts by 34 percentage points toward the derived evidence on average. Reasoning strategies and evidence load further modulate this behavior, and while chain-of-thought and claim hedging both improve awareness, they shift resolution in opposite directions: chain-of-thought toward the claim, hedging toward the evidence."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hacker-bar-2026-saying">
<titleInfo>
<title>Saying vs. Seeing: Evaluating LLM Conflict Awareness and Resolution Preference under Latent Claim-Evidence Disagreement</title>
</titleInfo>
<name type="personal">
<namePart type="given">Daniel</namePart>
<namePart type="family">Hacker</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kfir</namePart>
<namePart type="family">Bar</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 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Language models are routinely exposed to contexts where an authoritative claim conflicts with a conclusion derivable from underlying evidence. Unlike explicit contradictions, such conflicts are latent: they emerge only after a full multi-step derivation, a setting current benchmarks largely overlook. We introduce and publicly release CLECK (CLaim Evidence Conflict of Knowledge), a counterfactual benchmark of 5,922 instances (evaluated on a balanced 2,669-instance subset) probing this blind spot across eleven reasoning strategies. Among instruction-tuned models, the capability to recover both competing answers does not translate into recognizing their conflict at inference: models default to the authoritative claim. When the conflict does surface in a model’s reasoning, preference shifts by 34 percentage points toward the derived evidence on average. Reasoning strategies and evidence load further modulate this behavior, and while chain-of-thought and claim hedging both improve awareness, they shift resolution in opposite directions: chain-of-thought toward the claim, hedging toward the evidence.</abstract>
<identifier type="citekey">hacker-bar-2026-saying</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1727/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>39397</start>
<end>39423</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Saying vs. Seeing: Evaluating LLM Conflict Awareness and Resolution Preference under Latent Claim-Evidence Disagreement
%A Hacker, Daniel
%A Bar, Kfir
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F hacker-bar-2026-saying
%X Language models are routinely exposed to contexts where an authoritative claim conflicts with a conclusion derivable from underlying evidence. Unlike explicit contradictions, such conflicts are latent: they emerge only after a full multi-step derivation, a setting current benchmarks largely overlook. We introduce and publicly release CLECK (CLaim Evidence Conflict of Knowledge), a counterfactual benchmark of 5,922 instances (evaluated on a balanced 2,669-instance subset) probing this blind spot across eleven reasoning strategies. Among instruction-tuned models, the capability to recover both competing answers does not translate into recognizing their conflict at inference: models default to the authoritative claim. When the conflict does surface in a model’s reasoning, preference shifts by 34 percentage points toward the derived evidence on average. Reasoning strategies and evidence load further modulate this behavior, and while chain-of-thought and claim hedging both improve awareness, they shift resolution in opposite directions: chain-of-thought toward the claim, hedging toward the evidence.
%U https://aclanthology.org/2026.emnlp-main.1727/
%P 39397-39423
Markdown (Informal)
[Saying vs. Seeing: Evaluating LLM Conflict Awareness and Resolution Preference under Latent Claim-Evidence Disagreement](https://aclanthology.org/2026.emnlp-main.1727/) (Hacker & Bar, EMNLP 2026)
ACL