@inproceedings{ostmeier-etal-2026-gradient,
title = "Gradient-Stable Attention Heads Signal {LLM} Correctness",
author = "Ostmeier, Sophie and
Axelrod, Brian and
Varma, Maya and
Aali, Asad and
Zhang, Yabin and
Paschali, Magdalini and
Koyejo, Sanmi and
Langlotz, Curtis and
Chaudhari, Akshay S",
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.2167/",
pages = "46364--46380",
ISBN = "979-8-89176-462-0",
abstract = "Large language models (LLMs) often generate plausible yet incorrect answers, posing risks in safety-critical settings such as medicine. Human evaluation is expensive, and LLM-as-judge approaches risk introducing hidden errors. Recent single-pass white-box methods detect contextual hallucinations using linear probes over model internals, but their generalization to out-of-domain, out-of-task, and training-free settings remains poorly understood. We introduce HeadEntropy, a method that predicts answer correctness by measuring how susceptible each attention head{'}s pattern is to modification under further gradient updates. The key insight is that the trace of the softmax Jacobian is a monotonic function of 2-R{\'e}nyi entropy, linking attention spread to gradient stability during training, even without the reference annotation. With no training, HeadEntropy reaches 0.736 AUROC, outperforms every training-free baseline in paired comparison across 5 models and 5 datasets, and matches a hidden-state probe transferred across domains, at under 1{\%} of inference cost. We evaluate broadly across 5 instruction-tuned LLMs and 5 datasets spanning general knowledge, multi-hop reasoning, mathematical reasoning, claim verification, and medicine. The code is available at https://github.com/StanfordMIMI/HeadEntropy."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ostmeier-etal-2026-gradient">
<titleInfo>
<title>Gradient-Stable Attention Heads Signal LLM Correctness</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sophie</namePart>
<namePart type="family">Ostmeier</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Brian</namePart>
<namePart type="family">Axelrod</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Maya</namePart>
<namePart type="family">Varma</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Asad</namePart>
<namePart type="family">Aali</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yabin</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Magdalini</namePart>
<namePart type="family">Paschali</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sanmi</namePart>
<namePart type="family">Koyejo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Curtis</namePart>
<namePart type="family">Langlotz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Akshay</namePart>
<namePart type="given">S</namePart>
<namePart type="family">Chaudhari</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>Large language models (LLMs) often generate plausible yet incorrect answers, posing risks in safety-critical settings such as medicine. Human evaluation is expensive, and LLM-as-judge approaches risk introducing hidden errors. Recent single-pass white-box methods detect contextual hallucinations using linear probes over model internals, but their generalization to out-of-domain, out-of-task, and training-free settings remains poorly understood. We introduce HeadEntropy, a method that predicts answer correctness by measuring how susceptible each attention head’s pattern is to modification under further gradient updates. The key insight is that the trace of the softmax Jacobian is a monotonic function of 2-Rényi entropy, linking attention spread to gradient stability during training, even without the reference annotation. With no training, HeadEntropy reaches 0.736 AUROC, outperforms every training-free baseline in paired comparison across 5 models and 5 datasets, and matches a hidden-state probe transferred across domains, at under 1% of inference cost. We evaluate broadly across 5 instruction-tuned LLMs and 5 datasets spanning general knowledge, multi-hop reasoning, mathematical reasoning, claim verification, and medicine. The code is available at https://github.com/StanfordMIMI/HeadEntropy.</abstract>
<identifier type="citekey">ostmeier-etal-2026-gradient</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2167/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>46364</start>
<end>46380</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Gradient-Stable Attention Heads Signal LLM Correctness
%A Ostmeier, Sophie
%A Axelrod, Brian
%A Varma, Maya
%A Aali, Asad
%A Zhang, Yabin
%A Paschali, Magdalini
%A Koyejo, Sanmi
%A Langlotz, Curtis
%A Chaudhari, Akshay S.
%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 ostmeier-etal-2026-gradient
%X Large language models (LLMs) often generate plausible yet incorrect answers, posing risks in safety-critical settings such as medicine. Human evaluation is expensive, and LLM-as-judge approaches risk introducing hidden errors. Recent single-pass white-box methods detect contextual hallucinations using linear probes over model internals, but their generalization to out-of-domain, out-of-task, and training-free settings remains poorly understood. We introduce HeadEntropy, a method that predicts answer correctness by measuring how susceptible each attention head’s pattern is to modification under further gradient updates. The key insight is that the trace of the softmax Jacobian is a monotonic function of 2-Rényi entropy, linking attention spread to gradient stability during training, even without the reference annotation. With no training, HeadEntropy reaches 0.736 AUROC, outperforms every training-free baseline in paired comparison across 5 models and 5 datasets, and matches a hidden-state probe transferred across domains, at under 1% of inference cost. We evaluate broadly across 5 instruction-tuned LLMs and 5 datasets spanning general knowledge, multi-hop reasoning, mathematical reasoning, claim verification, and medicine. The code is available at https://github.com/StanfordMIMI/HeadEntropy.
%U https://aclanthology.org/2026.findings-emnlp.2167/
%P 46364-46380
Markdown (Informal)
[Gradient-Stable Attention Heads Signal LLM Correctness](https://aclanthology.org/2026.findings-emnlp.2167/) (Ostmeier et al., Findings 2026)
ACL
- Sophie Ostmeier, Brian Axelrod, Maya Varma, Asad Aali, Yabin Zhang, Magdalini Paschali, Sanmi Koyejo, Curtis Langlotz, and Akshay S Chaudhari. 2026. Gradient-Stable Attention Heads Signal LLM Correctness. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 46364–46380, Budapest, Hungary. Association for Computational Linguistics.