Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Think in Safety: Unveiling and Mitigating Safety Alignment Collapse in Multimodal Large Reasoning Model

Xinyue Lou, You Li, Jinan Xu, Xiangyu Shi, Chi Chen, Kaiyu Huang

TL;DR

This work systematically evaluates the safety of Multimodal Large Reasoning Models across jailbreak robustness and safety-awareness benchmarks, revealing pervasive safety degradation that worsens with enhanced reasoning in many models. It highlights a nuanced pattern: jailbreak robustness suffers more than safety-awareness, while longer reasoning can sometimes improve safety by better detecting unsafe intent. To address safety collapse, the authors introduce TiS, a safety oriented thinking dataset that preserves reasoning chains and enables supervised fine-tuning to align multimodal models with safety objectives. Experimental results show TiS markedly improves safety performance on several benchmarks while maintaining reasoning capabilities, offering a new direction for safe multimodal reasoning and practical pathways for deployment.

Abstract

The rapid development of Multimodal Large Reasoning Models (MLRMs) has demonstrated broad application potential, yet their safety and reliability remain critical concerns that require systematic exploration. To address this gap, we conduct a comprehensive and systematic safety evaluation of 11 MLRMs across 5 benchmarks and unveil prevalent safety degradation phenomena in most advanced models. Moreover, our analysis reveals distinct safety patterns across different benchmarks: significant safety degradation is observed across jailbreak robustness benchmarks, whereas safety-awareness benchmarks demonstrate less pronounced degradation. In particular, the long thought process in some scenarios even enhances safety performance. Therefore, it is a potential approach to address safety issues in MLRMs by leveraging the intrinsic reasoning capabilities of the model to detect unsafe intent. To operationalize this insight, we construct a multimodal tuning dataset that incorporates a safety-oriented thought process. Experimental results from fine-tuning existing MLRMs with this dataset effectively enhances the safety on both jailbreak robustness and safety-awareness benchmarks. This study provides a new perspective for developing safe MLRMs. Our dataset is available at https://github.com/xinyuelou/Think-in-Safety.

Think in Safety: Unveiling and Mitigating Safety Alignment Collapse in Multimodal Large Reasoning Model

TL;DR

This work systematically evaluates the safety of Multimodal Large Reasoning Models across jailbreak robustness and safety-awareness benchmarks, revealing pervasive safety degradation that worsens with enhanced reasoning in many models. It highlights a nuanced pattern: jailbreak robustness suffers more than safety-awareness, while longer reasoning can sometimes improve safety by better detecting unsafe intent. To address safety collapse, the authors introduce TiS, a safety oriented thinking dataset that preserves reasoning chains and enables supervised fine-tuning to align multimodal models with safety objectives. Experimental results show TiS markedly improves safety performance on several benchmarks while maintaining reasoning capabilities, offering a new direction for safe multimodal reasoning and practical pathways for deployment.

Abstract

The rapid development of Multimodal Large Reasoning Models (MLRMs) has demonstrated broad application potential, yet their safety and reliability remain critical concerns that require systematic exploration. To address this gap, we conduct a comprehensive and systematic safety evaluation of 11 MLRMs across 5 benchmarks and unveil prevalent safety degradation phenomena in most advanced models. Moreover, our analysis reveals distinct safety patterns across different benchmarks: significant safety degradation is observed across jailbreak robustness benchmarks, whereas safety-awareness benchmarks demonstrate less pronounced degradation. In particular, the long thought process in some scenarios even enhances safety performance. Therefore, it is a potential approach to address safety issues in MLRMs by leveraging the intrinsic reasoning capabilities of the model to detect unsafe intent. To operationalize this insight, we construct a multimodal tuning dataset that incorporates a safety-oriented thought process. Experimental results from fine-tuning existing MLRMs with this dataset effectively enhances the safety on both jailbreak robustness and safety-awareness benchmarks. This study provides a new perspective for developing safe MLRMs. Our dataset is available at https://github.com/xinyuelou/Think-in-Safety.
Paper Structure (55 sections, 12 figures, 11 tables)

This paper contains 55 sections, 12 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Safety Evaluation of MLRMs
  3. Evaluation Settings
  4. Datasets.
  5. Models and Configurations.
  6. Metrics and Evaluator.
  7. Safety Results and Inspection
  8. Overall Performance.
  9. Safety Degradation.
  10. Performance on Safety-Awareness.
  11. Modality Ablation.
  12. Performance of Thought Process.
  13. Comparative Effect between Thought and Answer.
  14. Data Construction
  15. Data Preparation.
  16. ...and 40 more sections

Figures (12)

  • Figure 1: Examples of multimodal safety benchmarks and their corresponding responses on different models.
  • Figure 2: Case study of the better safety consideration on safety-awareness tasks. Kimi-VL directly outputs the answer that ignores the potential risk, while Kimi-VL-Thinking dives deeper into the insidious safety issue with stronger reasoning abilities. The red indicates the unsafe parts, while the green indicates the content related to potential risks identified by reasoning models.
  • Figure 3: ASR scores of the thought process and the final answer generated by MLRMs on FigStep respectively.
  • Figure 4: Examples of self-deception in responses generated by MLRMs. The content of reframing intent of users is highlighted with HTML]FFD966yellow background, and harmful content is marked with the red font.
  • Figure 5: Examples of the safety inconsistency between the thought process and the final answer. The harmful content is marked with the red font.
  • ...and 7 more figures