AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Teach to Reason Safely: Policy-Guided Safety Tuning for MLRMs

Jingyu Zhang, Kun Yang, Ming Wen, Zhuoer Xu, Zeyang Sha, shiwen cui, Zhaohui Yang

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Analyzes the safety-reasoning trade-off in multimodal large reasoning models (MLRMs) and proposes policy-guided safety tuning to restore safety without sacrificing reasoning.

Problem

Reasoning-based multimodal models are more prone to generating harmful content, so reasoning gains come with safety degradation.

Approach

Large-scale analysis identifying two mechanisms of safety degradation in MLRMs (the first being visual-related; the abstract is truncated) and a policy-guided safety tuning approach to teach models to reason safely. Details beyond the abstract are unknown.

Key results

Measured effects reported in the paper

What was changedWhereMeasuredEffect
reasoning-oriented tuning of base MLLMsQwen2.5-VL-7B to R1-Onevision · BeaverTails-VHarmful Rate (%): 58.40 -> 78.61Hurt
add explicit multi-step reasoning generation (ablation)multimodal models · safety benchmarksHarmful RateHurt
apply PST (policy-guided SFT + SRPO)R1-Onevision · BeaverTails-VHarmful Rate (%): 78.61 -> 9.00Improved
add SRPO preference stage after PST-SFT (ablation)R1-Onevision + PST-SFT · MM-SafetyBenchHarmful Rate (%): 5.48 -> 2.68Improved
add SRPO preference stage after PST-SFT (ablation)R1-Onevision + PST-SFT · SIUOHarmful Rate (%): 10.18 -> 12.57Hurt
add SRPO preference stage after PST-SFT (ablation)LLaMA-CoT + PST-SFT · MMSafetyAware (over-sensitivity)Refusal Rate (%): 54.91 -> 46.26Improved
use PST data instead of MSR-Align for alignmentR1-Onevision · MMSafetyAware (over-sensitivity)Refusal Rate (%): 86.45 -> 69.39Improved
use PST instead of MSR-Align on general reasoningR1-Onevision · VQAv2accuracy: 74.40 -> 80.87Improved
use PST instead of Think-in-Safety on LLaMA-CoTLLaMA-CoT · ScienceQAaccuracy: 33.00 -> 84.10Improved
apply PST to LLaMA-CoTLLaMA-CoT · GQAaccuracy: 60.00 -> 48.00Hurt
increase PST-SFT data from 1k to 2k-4k samples (ablation)PST-SFT · BeaverTails-V, MM-SafetyBench, SPA-VLHarmful Rate (%): 10.70 -> 11.54 (R1-Onevision, 1k -> 2k)No clear effect
apply PSTR1-Onevision · BeaverTails-V failure typesFlawed Reasoning Initiation failure count: 331 -> 27Improved

Surprising findings

Limitations

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy