@inproceedings{teng-etal-2026-analyzing,
title = "Analyzing and Steering Reflection in Large Reasoning Models: A Difficulty-Aware Geometric Approach",
author = "Teng, Ge and
Shen, Chen and
Mao, Ting and
Tian, Xiang and
Liu, Xuesong and
Jiang, Rongxin and
Chen, Yaowu and
Ye, Jieping",
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.326/",
pages = "6891--6915",
ISBN = "979-8-89176-462-0",
abstract = "Large Reasoning Models (LRMs) have demonstrated exceptional capabilities in complex problem-solving, largely driven by emergent reflection and backtracking mechanisms. However, understanding these mechanisms through mechanistic interpretability remains challenging. Current sparsity-based approaches rely heavily on heuristic keywords (e.g., ``wait'') and focus on isolated features, lacking a holistic view of how these features collaborate. In this paper, we propose a novel interpretability framework that bridges the microscopic and macroscopic analysis of Sparse Auto-Encoder (SAE) features. By leveraging an auxiliary LRM to robustly annotate reasoning trajectories, we introduce a comprehensive ``Reflection Score'' to quantify a feature{'}s relevance to reflection without keyword bias. Our geometric analysis of these features reveals that LRMs encode reflection not as discrete, isolated vectors, but within continuous regions of the hidden state space. Furthermore, we uncover an intrinsic, difficulty-aware reflection representation mechanism: the model dynamically activates sub-regions based on problem complexity. Building upon these insights, we introduce a static, difficulty-informed steering strategy that suppresses superficial, simple-problem oriented reflection features during inference. Comprehensive evaluations across three LRMs and challenging reasoning benchmarks demonstrate that our intervention simultaneously improves reasoning accuracy by an average of 2.25{\%} and reduces token consumption by 5.3{\%}. Our findings provide a deeper geometric understanding of LRM reflection and offer an interpretable, scalable approach to optimizing LRM reasoning."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="teng-etal-2026-analyzing">
<titleInfo>
<title>Analyzing and Steering Reflection in Large Reasoning Models: A Difficulty-Aware Geometric Approach</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ge</namePart>
<namePart type="family">Teng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chen</namePart>
<namePart type="family">Shen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ting</namePart>
<namePart type="family">Mao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiang</namePart>
<namePart type="family">Tian</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuesong</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rongxin</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yaowu</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jieping</namePart>
<namePart type="family">Ye</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 demonstrated exceptional capabilities in complex problem-solving, largely driven by emergent reflection and backtracking mechanisms. However, understanding these mechanisms through mechanistic interpretability remains challenging. Current sparsity-based approaches rely heavily on heuristic keywords (e.g., “wait”) and focus on isolated features, lacking a holistic view of how these features collaborate. In this paper, we propose a novel interpretability framework that bridges the microscopic and macroscopic analysis of Sparse Auto-Encoder (SAE) features. By leveraging an auxiliary LRM to robustly annotate reasoning trajectories, we introduce a comprehensive “Reflection Score” to quantify a feature’s relevance to reflection without keyword bias. Our geometric analysis of these features reveals that LRMs encode reflection not as discrete, isolated vectors, but within continuous regions of the hidden state space. Furthermore, we uncover an intrinsic, difficulty-aware reflection representation mechanism: the model dynamically activates sub-regions based on problem complexity. Building upon these insights, we introduce a static, difficulty-informed steering strategy that suppresses superficial, simple-problem oriented reflection features during inference. Comprehensive evaluations across three LRMs and challenging reasoning benchmarks demonstrate that our intervention simultaneously improves reasoning accuracy by an average of 2.25% and reduces token consumption by 5.3%. Our findings provide a deeper geometric understanding of LRM reflection and offer an interpretable, scalable approach to optimizing LRM reasoning.</abstract>
<identifier type="citekey">teng-etal-2026-analyzing</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.326/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>6891</start>
<end>6915</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Analyzing and Steering Reflection in Large Reasoning Models: A Difficulty-Aware Geometric Approach
%A Teng, Ge
%A Shen, Chen
%A Mao, Ting
%A Tian, Xiang
%A Liu, Xuesong
%A Jiang, Rongxin
%A Chen, Yaowu
%A Ye, Jieping
%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 teng-etal-2026-analyzing
%X Large Reasoning Models (LRMs) have demonstrated exceptional capabilities in complex problem-solving, largely driven by emergent reflection and backtracking mechanisms. However, understanding these mechanisms through mechanistic interpretability remains challenging. Current sparsity-based approaches rely heavily on heuristic keywords (e.g., “wait”) and focus on isolated features, lacking a holistic view of how these features collaborate. In this paper, we propose a novel interpretability framework that bridges the microscopic and macroscopic analysis of Sparse Auto-Encoder (SAE) features. By leveraging an auxiliary LRM to robustly annotate reasoning trajectories, we introduce a comprehensive “Reflection Score” to quantify a feature’s relevance to reflection without keyword bias. Our geometric analysis of these features reveals that LRMs encode reflection not as discrete, isolated vectors, but within continuous regions of the hidden state space. Furthermore, we uncover an intrinsic, difficulty-aware reflection representation mechanism: the model dynamically activates sub-regions based on problem complexity. Building upon these insights, we introduce a static, difficulty-informed steering strategy that suppresses superficial, simple-problem oriented reflection features during inference. Comprehensive evaluations across three LRMs and challenging reasoning benchmarks demonstrate that our intervention simultaneously improves reasoning accuracy by an average of 2.25% and reduces token consumption by 5.3%. Our findings provide a deeper geometric understanding of LRM reflection and offer an interpretable, scalable approach to optimizing LRM reasoning.
%U https://aclanthology.org/2026.findings-emnlp.326/
%P 6891-6915
Markdown (Informal)
[Analyzing and Steering Reflection in Large Reasoning Models: A Difficulty-Aware Geometric Approach](https://aclanthology.org/2026.findings-emnlp.326/) (Teng et al., Findings 2026)
ACL
- Ge Teng, Chen Shen, Ting Mao, Xiang Tian, Xuesong Liu, Rongxin Jiang, Yaowu Chen, and Jieping Ye. 2026. Analyzing and Steering Reflection in Large Reasoning Models: A Difficulty-Aware Geometric Approach. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 6891–6915, Budapest, Hungary. Association for Computational Linguistics.