@inproceedings{peng-etal-2026-large,
title = "Large Reasoning Models Learn Better Alignment from Flawed Thinking",
author = "Peng, ShengYun and
Chen, Pin-Yu and
Smith, Eric Michael and
Jiang, Song and
Zhan, Hongyuan and
Wang, Haozhu and
Pasupuleti, Mahesh and
Chau, Duen Horng and
Chi, Jianfeng",
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.1651/",
pages = "37538--37566",
ISBN = "979-8-89176-459-0",
abstract = "Large reasoning models (LRMs) ``think'' by generating structured chain-of-thought (CoT) before producing a final answer, yet they still lack the ability to reason critically about safety alignment and are easily biased when a flawed premise is injected into their thought process. We propose RECAP (Robust Safety Alignment via Counter-Aligned Prefilling), a principled reinforcement learning (RL) method for post-training that explicitly teaches models to override flawed reasoning trajectories and reroute to safe and helpful responses. RECAP trains on a mixture of synthetically generated counter-aligned CoT prefills and standard prompts, integrates seamlessly into standard reinforcement learning from human feedback (RLHF) pipelines without additional RL optimization steps or architectural modifications, and substantially improves safety and jailbreak robustness, reduces overrefusal, and preserves core reasoning capability {---} all while maintaining inference token budget. Extensive analysis shows that RECAP-trained models engage in self-reflection more frequently and remain robust under adaptive attacks, preserving safety even after repeated attempts to override their reasoning. Our code is publicly available at https://github.com/poloclub/recap."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="peng-etal-2026-large">
<titleInfo>
<title>Large Reasoning Models Learn Better Alignment from Flawed Thinking</title>
</titleInfo>
<name type="personal">
<namePart type="given">ShengYun</namePart>
<namePart type="family">Peng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pin-Yu</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Eric</namePart>
<namePart type="given">Michael</namePart>
<namePart type="family">Smith</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Song</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hongyuan</namePart>
<namePart type="family">Zhan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haozhu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mahesh</namePart>
<namePart type="family">Pasupuleti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Duen</namePart>
<namePart type="given">Horng</namePart>
<namePart type="family">Chau</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jianfeng</namePart>
<namePart type="family">Chi</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>Large reasoning models (LRMs) “think” by generating structured chain-of-thought (CoT) before producing a final answer, yet they still lack the ability to reason critically about safety alignment and are easily biased when a flawed premise is injected into their thought process. We propose RECAP (Robust Safety Alignment via Counter-Aligned Prefilling), a principled reinforcement learning (RL) method for post-training that explicitly teaches models to override flawed reasoning trajectories and reroute to safe and helpful responses. RECAP trains on a mixture of synthetically generated counter-aligned CoT prefills and standard prompts, integrates seamlessly into standard reinforcement learning from human feedback (RLHF) pipelines without additional RL optimization steps or architectural modifications, and substantially improves safety and jailbreak robustness, reduces overrefusal, and preserves core reasoning capability — all while maintaining inference token budget. Extensive analysis shows that RECAP-trained models engage in self-reflection more frequently and remain robust under adaptive attacks, preserving safety even after repeated attempts to override their reasoning. Our code is publicly available at https://github.com/poloclub/recap.</abstract>
<identifier type="citekey">peng-etal-2026-large</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1651/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>37538</start>
<end>37566</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Large Reasoning Models Learn Better Alignment from Flawed Thinking
%A Peng, ShengYun
%A Chen, Pin-Yu
%A Smith, Eric Michael
%A Jiang, Song
%A Zhan, Hongyuan
%A Wang, Haozhu
%A Pasupuleti, Mahesh
%A Chau, Duen Horng
%A Chi, Jianfeng
%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 peng-etal-2026-large
%X Large reasoning models (LRMs) “think” by generating structured chain-of-thought (CoT) before producing a final answer, yet they still lack the ability to reason critically about safety alignment and are easily biased when a flawed premise is injected into their thought process. We propose RECAP (Robust Safety Alignment via Counter-Aligned Prefilling), a principled reinforcement learning (RL) method for post-training that explicitly teaches models to override flawed reasoning trajectories and reroute to safe and helpful responses. RECAP trains on a mixture of synthetically generated counter-aligned CoT prefills and standard prompts, integrates seamlessly into standard reinforcement learning from human feedback (RLHF) pipelines without additional RL optimization steps or architectural modifications, and substantially improves safety and jailbreak robustness, reduces overrefusal, and preserves core reasoning capability — all while maintaining inference token budget. Extensive analysis shows that RECAP-trained models engage in self-reflection more frequently and remain robust under adaptive attacks, preserving safety even after repeated attempts to override their reasoning. Our code is publicly available at https://github.com/poloclub/recap.
%U https://aclanthology.org/2026.emnlp-main.1651/
%P 37538-37566
Markdown (Informal)
[Large Reasoning Models Learn Better Alignment from Flawed Thinking](https://aclanthology.org/2026.emnlp-main.1651/) (Peng et al., EMNLP 2026)
ACL
- ShengYun Peng, Pin-Yu Chen, Eric Michael Smith, Song Jiang, Hongyuan Zhan, Haozhu Wang, Mahesh Pasupuleti, Duen Horng Chau, and Jianfeng Chi. 2026. Large Reasoning Models Learn Better Alignment from Flawed Thinking. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 37538–37566, Budapest, Hungary. Association for Computational Linguistics.