@inproceedings{zhou-etal-2026-ird,
title = "{IRD}: Steering Reasoning Models with Intent, Risk, and Decision",
author = "Zhou, Kang and
Zhou, Guanghao and
Chen, Cen and
Liu, Zhaoyang and
Gao, Ming and
Li, Yaliang",
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.613/",
pages = "14103--14121",
ISBN = "979-8-89176-459-0",
abstract = "Large reasoning models (LRMs) are strong multi-step problem solvers, yet their long reasoning traces may gradually drift into unsafe compliance when exposed to harmful or adversarial prompts. Existing safety alignment methods can reduce harmful outputs; however, they often rely on broad refusal or supervision over full safe reasoning traces, which may weaken the reasoning behavior that makes LRMs useful. We introduce \textbf{I}ntent{--}\textbf{R}isk{--}\textbf{D}ecision (IRD), a lightweight prefix-tuning method that steers only the beginning of the reasoning trace: the model first emits a prompt-intent tag, a risk statement, and a compliance decision, and then continues under its original reasoning policy. Across three LRMs and evaluations covering harmfulness, jailbreak robustness, over-refusal, and reasoning ability, IRD substantially reduces unsafe outputs while preserving reasoning performance. In particular, IRD lowers average harmfulness to 1.2{--}7.4{\%} across the evaluated backbones and keeps near-lossless reasoning performance, {\%}within 0.2 average points of the base models, showing that targeted early steering can improve safety without supervising complete safe traces or full model fine-tuning. Code and data are available at \url{https://github.com/ReXT9527/IRD}."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhou-etal-2026-ird">
<titleInfo>
<title>IRD: Steering Reasoning Models with Intent, Risk, and Decision</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kang</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Guanghao</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Cen</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhaoyang</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ming</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yaliang</namePart>
<namePart type="family">Li</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) are strong multi-step problem solvers, yet their long reasoning traces may gradually drift into unsafe compliance when exposed to harmful or adversarial prompts. Existing safety alignment methods can reduce harmful outputs; however, they often rely on broad refusal or supervision over full safe reasoning traces, which may weaken the reasoning behavior that makes LRMs useful. We introduce Intent–Risk–Decision (IRD), a lightweight prefix-tuning method that steers only the beginning of the reasoning trace: the model first emits a prompt-intent tag, a risk statement, and a compliance decision, and then continues under its original reasoning policy. Across three LRMs and evaluations covering harmfulness, jailbreak robustness, over-refusal, and reasoning ability, IRD substantially reduces unsafe outputs while preserving reasoning performance. In particular, IRD lowers average harmfulness to 1.2–7.4% across the evaluated backbones and keeps near-lossless reasoning performance, %within 0.2 average points of the base models, showing that targeted early steering can improve safety without supervising complete safe traces or full model fine-tuning. Code and data are available at https://github.com/ReXT9527/IRD.</abstract>
<identifier type="citekey">zhou-etal-2026-ird</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.613/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>14103</start>
<end>14121</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T IRD: Steering Reasoning Models with Intent, Risk, and Decision
%A Zhou, Kang
%A Zhou, Guanghao
%A Chen, Cen
%A Liu, Zhaoyang
%A Gao, Ming
%A Li, Yaliang
%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 zhou-etal-2026-ird
%X Large reasoning models (LRMs) are strong multi-step problem solvers, yet their long reasoning traces may gradually drift into unsafe compliance when exposed to harmful or adversarial prompts. Existing safety alignment methods can reduce harmful outputs; however, they often rely on broad refusal or supervision over full safe reasoning traces, which may weaken the reasoning behavior that makes LRMs useful. We introduce Intent–Risk–Decision (IRD), a lightweight prefix-tuning method that steers only the beginning of the reasoning trace: the model first emits a prompt-intent tag, a risk statement, and a compliance decision, and then continues under its original reasoning policy. Across three LRMs and evaluations covering harmfulness, jailbreak robustness, over-refusal, and reasoning ability, IRD substantially reduces unsafe outputs while preserving reasoning performance. In particular, IRD lowers average harmfulness to 1.2–7.4% across the evaluated backbones and keeps near-lossless reasoning performance, %within 0.2 average points of the base models, showing that targeted early steering can improve safety without supervising complete safe traces or full model fine-tuning. Code and data are available at https://github.com/ReXT9527/IRD.
%U https://aclanthology.org/2026.emnlp-main.613/
%P 14103-14121
Markdown (Informal)
[IRD: Steering Reasoning Models with Intent, Risk, and Decision](https://aclanthology.org/2026.emnlp-main.613/) (Zhou et al., EMNLP 2026)
ACL
- Kang Zhou, Guanghao Zhou, Cen Chen, Zhaoyang Liu, Ming Gao, and Yaliang Li. 2026. IRD: Steering Reasoning Models with Intent, Risk, and Decision. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 14103–14121, Budapest, Hungary. Association for Computational Linguistics.