@inproceedings{mercorio-etal-2026-diagnosing,
title = "Diagnosing {LLM} Fragility: An Empirical Study on Reasoning Entropy and Malicious Outputs",
author = "Mercorio, Fabio and
Potert{\`i}, Daniele and
Seveso, Andrea",
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.1324/",
pages = "30166--30184",
ISBN = "979-8-89176-459-0",
abstract = "Chain-of-Thought prompting is increasingly relied upon as a safety monitoring surface: if a model reasons before acting, its trace can be inspected before harm occurs. Yet reasoning is a stochastic branching process rather than a deterministic, auditable path. We introduce Bifurcation Entropy ($H_{BE}$), a step-level diagnostic metric that extends Kernel Language Entropy to intermediate reasoning steps, measuring the semantic breadth of a model{'}s next-step distribution at each point in the Chain-of-Thought. We deploy $H_{BE}$ via In-Distribution Attacks, a controlled probing instrument that injects the model{'}s own high-entropy continuations to test whether $H_{BE}$ predicts exploitability. Across five reasoning models in high-stakes blackmail scenarios, phrases in the top $H_{BE}$ quartile are exploited 6.2$\times$ more often than the lowest, with $H_{BE}$-targeted probing inducing significantly more alignment failures than random selection. These results establish $H_{BE}$ as a principled probing signal that concentrates safety monitoring effort on the minority of reasoning steps that carry the majority of alignment risk."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="mercorio-etal-2026-diagnosing">
<titleInfo>
<title>Diagnosing LLM Fragility: An Empirical Study on Reasoning Entropy and Malicious Outputs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Fabio</namePart>
<namePart type="family">Mercorio</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Daniele</namePart>
<namePart type="family">Potertì</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Andrea</namePart>
<namePart type="family">Seveso</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>Chain-of-Thought prompting is increasingly relied upon as a safety monitoring surface: if a model reasons before acting, its trace can be inspected before harm occurs. Yet reasoning is a stochastic branching process rather than a deterministic, auditable path. We introduce Bifurcation Entropy (H_BE), a step-level diagnostic metric that extends Kernel Language Entropy to intermediate reasoning steps, measuring the semantic breadth of a model’s next-step distribution at each point in the Chain-of-Thought. We deploy H_BE via In-Distribution Attacks, a controlled probing instrument that injects the model’s own high-entropy continuations to test whether H_BE predicts exploitability. Across five reasoning models in high-stakes blackmail scenarios, phrases in the top H_BE quartile are exploited 6.2\times more often than the lowest, with H_BE-targeted probing inducing significantly more alignment failures than random selection. These results establish H_BE as a principled probing signal that concentrates safety monitoring effort on the minority of reasoning steps that carry the majority of alignment risk.</abstract>
<identifier type="citekey">mercorio-etal-2026-diagnosing</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1324/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>30166</start>
<end>30184</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Diagnosing LLM Fragility: An Empirical Study on Reasoning Entropy and Malicious Outputs
%A Mercorio, Fabio
%A Potertì, Daniele
%A Seveso, Andrea
%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 mercorio-etal-2026-diagnosing
%X Chain-of-Thought prompting is increasingly relied upon as a safety monitoring surface: if a model reasons before acting, its trace can be inspected before harm occurs. Yet reasoning is a stochastic branching process rather than a deterministic, auditable path. We introduce Bifurcation Entropy (H_BE), a step-level diagnostic metric that extends Kernel Language Entropy to intermediate reasoning steps, measuring the semantic breadth of a model’s next-step distribution at each point in the Chain-of-Thought. We deploy H_BE via In-Distribution Attacks, a controlled probing instrument that injects the model’s own high-entropy continuations to test whether H_BE predicts exploitability. Across five reasoning models in high-stakes blackmail scenarios, phrases in the top H_BE quartile are exploited 6.2\times more often than the lowest, with H_BE-targeted probing inducing significantly more alignment failures than random selection. These results establish H_BE as a principled probing signal that concentrates safety monitoring effort on the minority of reasoning steps that carry the majority of alignment risk.
%U https://aclanthology.org/2026.emnlp-main.1324/
%P 30166-30184
Markdown (Informal)
[Diagnosing LLM Fragility: An Empirical Study on Reasoning Entropy and Malicious Outputs](https://aclanthology.org/2026.emnlp-main.1324/) (Mercorio et al., EMNLP 2026)
ACL