@inproceedings{zhu-etal-2026-large-reasoning,
title = "When Can Large Reasoning Models Save Thinking? Mechanistic Analysis of Behavioral Divergence in Reasoning",
author = "Zhu, Rongzhi and
Liu, Yi and
Wang, Jiancheng and
Liu, Xiangyu and
Sun, Zequn and
Wang, Yiwei and
Deng, Yu and
Zhou, Zijian and
Hu, Wei",
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.761/",
doi = "10.18653/2026.findings-emnlp.761",
pages = "16028--16040",
ISBN = "979-8-89176-462-0",
abstract = "Large reasoning models (LRMs) have achieved remarkable success on complex tasks, yet their tendency to ``overthink'' leads to inefficiencies. Although ``save-thinking'' prompts are intended to mitigate this issue, we find that LRMs still frequently enter the ``Still-thinking'' mode instead of the expected ``No-thinking'' mode, especially when facing difficult queries. First, we examine the LRM{'}s confidence when terminating thinking process, finding that high perplexity signals internal uncertainty and precedes the thinking process. Second, we analyze the divergence in internal attention distributions across different LRMs. Third, we investigate where the LRM allocates its attention within the prompt, identifying a critical ``re-reading'' pattern. Based on these analyses, we propose an attention intervention method to regulate this behavior. While this intervention successfully suppresses thinking, we find it causes a drop in accuracy, confirming that the thinking process is functional. Our work provides a mechanistic explanation for this behavior, highlighting the trade-off between instruction following and reasoning correctness, and offers insights for developing more controllable LRMs."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhu-etal-2026-large-reasoning">
<titleInfo>
<title>When Can Large Reasoning Models Save Thinking? Mechanistic Analysis of Behavioral Divergence in Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Rongzhi</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yi</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiancheng</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiangyu</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zequn</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yiwei</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yu</namePart>
<namePart type="family">Deng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zijian</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</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 (LRMs) have achieved remarkable success on complex tasks, yet their tendency to “overthink” leads to inefficiencies. Although “save-thinking” prompts are intended to mitigate this issue, we find that LRMs still frequently enter the “Still-thinking” mode instead of the expected “No-thinking” mode, especially when facing difficult queries. First, we examine the LRM’s confidence when terminating thinking process, finding that high perplexity signals internal uncertainty and precedes the thinking process. Second, we analyze the divergence in internal attention distributions across different LRMs. Third, we investigate where the LRM allocates its attention within the prompt, identifying a critical “re-reading” pattern. Based on these analyses, we propose an attention intervention method to regulate this behavior. While this intervention successfully suppresses thinking, we find it causes a drop in accuracy, confirming that the thinking process is functional. Our work provides a mechanistic explanation for this behavior, highlighting the trade-off between instruction following and reasoning correctness, and offers insights for developing more controllable LRMs.</abstract>
<identifier type="citekey">zhu-etal-2026-large-reasoning</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.761</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.761/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>16028</start>
<end>16040</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T When Can Large Reasoning Models Save Thinking? Mechanistic Analysis of Behavioral Divergence in Reasoning
%A Zhu, Rongzhi
%A Liu, Yi
%A Wang, Jiancheng
%A Liu, Xiangyu
%A Sun, Zequn
%A Wang, Yiwei
%A Deng, Yu
%A Zhou, Zijian
%A Hu, Wei
%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-large-reasoning
%X Large reasoning models (LRMs) have achieved remarkable success on complex tasks, yet their tendency to “overthink” leads to inefficiencies. Although “save-thinking” prompts are intended to mitigate this issue, we find that LRMs still frequently enter the “Still-thinking” mode instead of the expected “No-thinking” mode, especially when facing difficult queries. First, we examine the LRM’s confidence when terminating thinking process, finding that high perplexity signals internal uncertainty and precedes the thinking process. Second, we analyze the divergence in internal attention distributions across different LRMs. Third, we investigate where the LRM allocates its attention within the prompt, identifying a critical “re-reading” pattern. Based on these analyses, we propose an attention intervention method to regulate this behavior. While this intervention successfully suppresses thinking, we find it causes a drop in accuracy, confirming that the thinking process is functional. Our work provides a mechanistic explanation for this behavior, highlighting the trade-off between instruction following and reasoning correctness, and offers insights for developing more controllable LRMs.
%R 10.18653/2026.findings-emnlp.761
%U https://aclanthology.org/2026.findings-emnlp.761/
%U https://doi.org/10.18653/2026.findings-emnlp.761
%P 16028-16040
Markdown (Informal)
[When Can Large Reasoning Models Save Thinking? Mechanistic Analysis of Behavioral Divergence in Reasoning](https://aclanthology.org/2026.findings-emnlp.761/) (Zhu et al., Findings 2026)
ACL
- Rongzhi Zhu, Yi Liu, Jiancheng Wang, Xiangyu Liu, Zequn Sun, Yiwei Wang, Yu Deng, Zijian Zhou, and Wei Hu. 2026. When Can Large Reasoning Models Save Thinking? Mechanistic Analysis of Behavioral Divergence in Reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 16028–16040, Budapest, Hungary. Association for Computational Linguistics.