@inproceedings{song-etal-2026-corrsafe,
title = "{C}o{RRS}afe: Improving {LRM} Safety via Co-optimized Fine-Grained Reasoning and Response Alignment",
author = "Song, Yangxiao and
Xu, Bing and
Xu, Bufan and
Yang, Muyun and
Zhu, Conghui and
Wu, Huanqin and
Hu, Bojie and
Guan, Hongjiao and
Lu, Wenpeng",
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.1217/",
pages = "27790--27814",
ISBN = "979-8-89176-459-0",
abstract = "Large reasoning models (LRMs) with explicit Chain-of-Thought (CoT) face significant safety risks, where unsafe content may appear in intermediate reasoning even when the final response is safe. Existing safety alignment methods often ignore the safety of the reasoning and frequently suffer from over-alignment issues. To address these, we propose \textbf{CoRRSafe} (\textbf{C}o-\textbf{o}ptimized \textbf{R}easoning and \textbf{R}esponse \textbf{Safe}), a two-stage framework designed to comprehensively enhance LRM safety while balancing safety and over-refusal. The method begins with a response cold-start stage using filtered high-quality data for initial alignment on responses. Then we implement a GRPO-based training strategy with fine-grained rewards to jointly evaluate segmented reasoning steps and final answers. This approach effectively guides the model to autonomously learn safe reasoning and response behaviors. Experiments on multiple LRMs show that CoRRSafe achieves state of the art performance in reasoning safety and achieves the best trade-off between safety and false refusal. On DeepSeek-Distill-R1-8B, it raises the Reasoning Safety Rate from 16.86{\%} to 89.31{\%} and reduces the Attack Success Rate from 78.74{\%} to 4.06{\%}, with only a 16.56{\%} increase in false refusal. Further analysis confirms that co-optimizing reasoning and response is essential, as optimizing either alone fails to overcome safety performance ceilings."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="song-etal-2026-corrsafe">
<titleInfo>
<title>CoRRSafe: Improving LRM Safety via Co-optimized Fine-Grained Reasoning and Response Alignment</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yangxiao</namePart>
<namePart type="family">Song</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bing</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bufan</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Muyun</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Conghui</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Huanqin</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bojie</namePart>
<namePart type="family">Hu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hongjiao</namePart>
<namePart type="family">Guan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenpeng</namePart>
<namePart type="family">Lu</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) with explicit Chain-of-Thought (CoT) face significant safety risks, where unsafe content may appear in intermediate reasoning even when the final response is safe. Existing safety alignment methods often ignore the safety of the reasoning and frequently suffer from over-alignment issues. To address these, we propose CoRRSafe (Co-optimized Reasoning and Response Safe), a two-stage framework designed to comprehensively enhance LRM safety while balancing safety and over-refusal. The method begins with a response cold-start stage using filtered high-quality data for initial alignment on responses. Then we implement a GRPO-based training strategy with fine-grained rewards to jointly evaluate segmented reasoning steps and final answers. This approach effectively guides the model to autonomously learn safe reasoning and response behaviors. Experiments on multiple LRMs show that CoRRSafe achieves state of the art performance in reasoning safety and achieves the best trade-off between safety and false refusal. On DeepSeek-Distill-R1-8B, it raises the Reasoning Safety Rate from 16.86% to 89.31% and reduces the Attack Success Rate from 78.74% to 4.06%, with only a 16.56% increase in false refusal. Further analysis confirms that co-optimizing reasoning and response is essential, as optimizing either alone fails to overcome safety performance ceilings.</abstract>
<identifier type="citekey">song-etal-2026-corrsafe</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1217/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>27790</start>
<end>27814</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T CoRRSafe: Improving LRM Safety via Co-optimized Fine-Grained Reasoning and Response Alignment
%A Song, Yangxiao
%A Xu, Bing
%A Xu, Bufan
%A Yang, Muyun
%A Zhu, Conghui
%A Wu, Huanqin
%A Hu, Bojie
%A Guan, Hongjiao
%A Lu, Wenpeng
%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 song-etal-2026-corrsafe
%X Large reasoning models (LRMs) with explicit Chain-of-Thought (CoT) face significant safety risks, where unsafe content may appear in intermediate reasoning even when the final response is safe. Existing safety alignment methods often ignore the safety of the reasoning and frequently suffer from over-alignment issues. To address these, we propose CoRRSafe (Co-optimized Reasoning and Response Safe), a two-stage framework designed to comprehensively enhance LRM safety while balancing safety and over-refusal. The method begins with a response cold-start stage using filtered high-quality data for initial alignment on responses. Then we implement a GRPO-based training strategy with fine-grained rewards to jointly evaluate segmented reasoning steps and final answers. This approach effectively guides the model to autonomously learn safe reasoning and response behaviors. Experiments on multiple LRMs show that CoRRSafe achieves state of the art performance in reasoning safety and achieves the best trade-off between safety and false refusal. On DeepSeek-Distill-R1-8B, it raises the Reasoning Safety Rate from 16.86% to 89.31% and reduces the Attack Success Rate from 78.74% to 4.06%, with only a 16.56% increase in false refusal. Further analysis confirms that co-optimizing reasoning and response is essential, as optimizing either alone fails to overcome safety performance ceilings.
%U https://aclanthology.org/2026.emnlp-main.1217/
%P 27790-27814
Markdown (Informal)
[CoRRSafe: Improving LRM Safety via Co-optimized Fine-Grained Reasoning and Response Alignment](https://aclanthology.org/2026.emnlp-main.1217/) (Song et al., EMNLP 2026)
ACL
- Yangxiao Song, Bing Xu, Bufan Xu, Muyun Yang, Conghui Zhu, Huanqin Wu, Bojie Hu, Hongjiao Guan, and Wenpeng Lu. 2026. CoRRSafe: Improving LRM Safety via Co-optimized Fine-Grained Reasoning and Response Alignment. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 27790–27814, Budapest, Hungary. Association for Computational Linguistics.