@inproceedings{zhu-etal-2026-demystifying-entropy,
title = "Demystifying Entropy Conflict: Why Reasoning Compression Stalls in Large Reasoning Models",
author = "Zhu, Hourun and
Gao, Yang and
Fei, Wenlong and
Li, Jiawei and
Sun, Huashan and
Ren, Bowen and
Hu, Chengzhi",
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.837/",
pages = "17734--17747",
ISBN = "979-8-89176-462-0",
abstract = "Large reasoning models achieve strong problem-solving performance, but often generate excessively long chain-of-thought outputs, which substantially increase inference cost and limit practical deployment. Existing reasoning compression methods reduce length only partially and often suffer from unstable optimization or accuracy degradation. We trace this limitation to an underlying trajectory-level objective conflict: when the compression objective and the accuracy objective are optimized jointly, compression favors short-and-correct trajectories while accuracy optimization continues to preserve long-but-correct ones that remain useful on hard problems. This conflict creates opposing gradients on the same behaviors and can trap training in a local trade-off where further compression harms accuracy while accuracy recovery reintroduces verbose reasoning. At the model-decision level, we refer to the resulting failure mode as entropy conflict: the policy fails to fully collapse onto concise trajectories because accuracy-oriented updates keep restoring broader support over long-correct reasoning paths. To address this issue, we propose an entropy-guided training framework. Instead of forcing a mixed objective to follow an unstable path toward the final efficient-output set, we first contract the policy to a short-output set and then expand from this compact regime to recover accuracy and robustness. Experiments on multiple mathematical reasoning benchmarks show that our method achieves performance comparable to the original model using only about 25{\%} of its reasoning tokens."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhu-etal-2026-demystifying-entropy">
<titleInfo>
<title>Demystifying Entropy Conflict: Why Reasoning Compression Stalls in Large Reasoning Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hourun</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yang</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenlong</namePart>
<namePart type="family">Fei</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiawei</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Huashan</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bowen</namePart>
<namePart type="family">Ren</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chengzhi</namePart>
<namePart type="family">Hu</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 reasoning models achieve strong problem-solving performance, but often generate excessively long chain-of-thought outputs, which substantially increase inference cost and limit practical deployment. Existing reasoning compression methods reduce length only partially and often suffer from unstable optimization or accuracy degradation. We trace this limitation to an underlying trajectory-level objective conflict: when the compression objective and the accuracy objective are optimized jointly, compression favors short-and-correct trajectories while accuracy optimization continues to preserve long-but-correct ones that remain useful on hard problems. This conflict creates opposing gradients on the same behaviors and can trap training in a local trade-off where further compression harms accuracy while accuracy recovery reintroduces verbose reasoning. At the model-decision level, we refer to the resulting failure mode as entropy conflict: the policy fails to fully collapse onto concise trajectories because accuracy-oriented updates keep restoring broader support over long-correct reasoning paths. To address this issue, we propose an entropy-guided training framework. Instead of forcing a mixed objective to follow an unstable path toward the final efficient-output set, we first contract the policy to a short-output set and then expand from this compact regime to recover accuracy and robustness. Experiments on multiple mathematical reasoning benchmarks show that our method achieves performance comparable to the original model using only about 25% of its reasoning tokens.</abstract>
<identifier type="citekey">zhu-etal-2026-demystifying-entropy</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.837/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>17734</start>
<end>17747</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Demystifying Entropy Conflict: Why Reasoning Compression Stalls in Large Reasoning Models
%A Zhu, Hourun
%A Gao, Yang
%A Fei, Wenlong
%A Li, Jiawei
%A Sun, Huashan
%A Ren, Bowen
%A Hu, Chengzhi
%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 zhu-etal-2026-demystifying-entropy
%X Large reasoning models achieve strong problem-solving performance, but often generate excessively long chain-of-thought outputs, which substantially increase inference cost and limit practical deployment. Existing reasoning compression methods reduce length only partially and often suffer from unstable optimization or accuracy degradation. We trace this limitation to an underlying trajectory-level objective conflict: when the compression objective and the accuracy objective are optimized jointly, compression favors short-and-correct trajectories while accuracy optimization continues to preserve long-but-correct ones that remain useful on hard problems. This conflict creates opposing gradients on the same behaviors and can trap training in a local trade-off where further compression harms accuracy while accuracy recovery reintroduces verbose reasoning. At the model-decision level, we refer to the resulting failure mode as entropy conflict: the policy fails to fully collapse onto concise trajectories because accuracy-oriented updates keep restoring broader support over long-correct reasoning paths. To address this issue, we propose an entropy-guided training framework. Instead of forcing a mixed objective to follow an unstable path toward the final efficient-output set, we first contract the policy to a short-output set and then expand from this compact regime to recover accuracy and robustness. Experiments on multiple mathematical reasoning benchmarks show that our method achieves performance comparable to the original model using only about 25% of its reasoning tokens.
%U https://aclanthology.org/2026.findings-emnlp.837/
%P 17734-17747
Markdown (Informal)
[Demystifying Entropy Conflict: Why Reasoning Compression Stalls in Large Reasoning Models](https://aclanthology.org/2026.findings-emnlp.837/) (Zhu et al., Findings 2026)
ACL