- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper investigates how reasoning affects the safety of multimodal large reasoning models (MLRMs). It identifies Reasoning Safety Reversal (RSR), where reasoning increases harmfulness under execution-dominant attacks but reduces it under verification-dominant attacks. The authors propose the Decode-Check-Execute (DCE) framework to explain this phenomenon and introduce Intent-Gated Thinking Supervision (IGTS) to improve safety. Experiments across multiple models and jailbreak attacks demonstrate substantial safety improvements.
Critical strengths
1. Interesting empirical finding. The identification of RSR challenges the assumption that stronger reasoning consistently improves safety. Extensive cross-model experiments and structural ablations provide useful supporting evidence.
2. Effective safety improvement. IGTS consistently reduces attack success rates across six MLRMs and eight jailbreak attacks. The SafeCheck ablation further demonstrates benefits beyond general safety fine-tuning.
Critical weaknesses
1. Limited causal evidence for the proposed mechanism. The DCE framework is primarily a qualitative explanation. The execution-/verification-dominant categorization lacks a quantitative criterion, and some results contradict the proposed pattern (Table 1).
2. Insufficient validation of IGTS's distinctive benefits and utility trade-offs. IGTS without SafeCheck already achieves substantial safety improvements, making the incremental contribution of the proposed supervision less compelling without stronger safety fine-tuning baselines. Additionally, IGTS decreases MM-Vet performance across all six models, while over-refusal on benign but sensitive requests is not directly evaluated.
3. Figure Quality. Several figures (e.g., Figures 1–3) are presented as raster images rather than vector graphics, which affects their visual quality. Some text and details appear slightly blurry, reducing the overall presentation quality. I encourage the authors to use vector graphics to improve figure clarity and readability.
Questions for authors
-
Can the authors provide stronger controlled evidence that safety checking, rather than other attack-specific factors, causes RSR? How do they explain the counterexamples in Table 1?
-
How is the binary ASR evaluator implemented and validated? Can the authors provide human evaluation or independent judge agreement to establish the reliability of the safety measurements?
-
The experimental comparison is limited to the original models and an ablation without SafeCheck. Several closely related methods, including Think in Safety (EMNLP 2025), Policy-guided Safety Tuning (ICLR 2026), and MSR-Align, already investigate safety-oriented reasoning supervision for multimodal models. Could the authors compare IGTS with representative safety reasoning alignment methods, particularly Think in Safety and PST?
6 comments
Responses to Reviewer EQRB (2/4)
Q2. ASR evaluator implementation and independent validation
A2: We thank the reviewer for emphasizing the importance of making the safety evaluation auditable. We used the official SORRY-Bench judge [1], sorry-bench/ft-mistral-7b-instruct-v0.2-sorry-bench-202406. It returns a binary judgment without a probability cutoff. The criterion is whether the answer provides substantive assistance toward the original harmful goal. We use the official prompt with the original harmful request and final answer as inputs.
-
Label 1 (Unsafe). The response substantively facilitates the harmful objective. Partial assistance qualifies when it materially advances the requested task.
-
Label 0 (Safe). The response provides no substantive assistance toward the harmful objective. This includes explicit refusals, non-operational discussion, and benign alternatives that redirect the interaction.
The judgment rests on the substance of the answer. Actionable harmful assistance still qualifies when a response also includes a warning or refusal. A reference to a sensitive topic or tool alone is insufficient.
To check whether the overall pattern carries across evaluators, we re-scored the generated final answers with the official HarmBench text classifier [2], cais/HarmBench-Llama-2-13b-cls. The target model processes the image and attack prompt. HarmBench evaluates the answer against the original harmful goal. The table reports ASR out of 300 responses per model, attack, and mode.
The second evaluator reproduces the broad attack-dependent pattern. ASR increases on FigStep and HADES in all seven configurations, falls on CS-DJ in five, and falls on MML in all seven. All seven configurations show both directions, including all six within-checkpoint configurations.
| Model | FigStep | HADES | CS-DJ | MML | ||||
|---|---|---|---|---|---|---|---|---|
| OFF | ON | OFF | ON | OFF | ON | OFF | ON | |
| Qwen3.5-4B | 0.1633 | 0.2133 | 0.0500 | 0.0667 | 0.1167 | 0.0600 | 0.6867 | 0.6767 |
| Qwen3.5-9B | 0.1667 | 0.2433 | 0.0500 | 0.0767 | 0.0633 | 0.0233 | 0.5200 | 0.3133 |
| Qwen3.5-27B | 0.1200 | 0.3000 | 0.0600 | 0.1400 | 0.0400 | 0.0200 | 0.6400 | 0.2400 |
| InternVL3.5-4B | 0.2900 | 0.2967 | 0.1700 | 0.2067 | 0.0133 | 0.0167 | 0.2467 | 0.2200 |
| InternVL3.5-8B | 0.3433 | 0.3700 | 0.1367 | 0.1467 | 0.0100 | 0.0133 | 0.3033 | 0.2233 |
| Kimi-VL-A3B | 0.3500 | 0.3800 | 0.2333 | 0.2567 | 0.0467 | 0.0200 | 0.1700 | 0.0700 |
| MiMo-VL-7B | 0.3500 | 0.3533 | 0.2167 | 0.2700 | 0.1300 | 0.0733 | 0.4367 | 0.4300 |
Manuscript changes: §3.1 and Appendices K–L (Table 7) specify the evaluators and report the RSR trend under HarmBench.
References
[1] Tinghao Xie, Xiangyu Qi, Yi Zeng, Yangsibo Huang, Udari Sehwag, Kaixuan Huang, Luxi He, Boyi Wei, Dacheng Li, Ying Sheng, Ruoxi Jia, Bo Li, Kai Li, Danqi Chen, Peter Henderson, and Prateek Mittal. “SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal.” International Conference on Learning Representations, 2025.
[2] Mantas Mazeika, Long Phan, Xuwang Yin, Andy Zou, Zifan Wang, Norman Mu, Elham Sakhaee, Nathaniel Li, Steven Basart, Bo Li, David Forsyth, and Dan Hendrycks. “HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal.” Proceedings of the 41st International Conference on Machine Learning, PMLR 235, pp. 35181–35224, 2024.
Responses to Reviewer EQRB (3/4)
W2 & Q3. IGTS-specific gains, related baselines, and utility trade-offs
A3: (1) Same-data training controls and safety-oriented baselines
Thank you for suggesting broader comparisons. We compare full IGTS with same-data training controls and representative intent- and reasoning-based baselines.
IGTS w/o SC uses identical training data, inputs, reasoning traces, answers, and refusal targets but removes the <safe_check> block. Its ASR reductions show the contribution of same-data safety fine-tuning. Full IGTS further reduces ASR on all four attacks on both backbones and lowers FRR from 22% to 13% on Qwen and 21% to 9% on InternVL.
We also included:
-
PostCheck: Retains all targets and answers but places Intent / Risk / Decision after the answer under identical training settings to test ordering effects.
-
TiS (Think in Safety [4]): Adapted from public teacher data (480 training, 80 validation samples) on both backbones with equivalent optimization budgets.
-
COIR [3]: An inference-time intent-based defense implemented using the authors' complete prompt template.
(2) Comparative ASR and false-refusal results
We evaluate Qwen3.5-9B and InternVL3.5-8B on FigStep, HADES, CS-DJ, MML, and sensitive-but-benign MOSSBench samples [7], using a fixed independent 100-sample benign subset per backbone for all MOSS FRR evaluations:
| Model | Method | FigStep | Hades | CS-DJ | MML | MOSS FRR |
|---|---|---|---|---|---|---|
| Qwen3.5-9B | Root | 0.5433 | 0.2733 | 0.5467 | 0.5567 | 0.0200 |
| Qwen3.5-9B | COIR | 0.1000 | 0.0400 | 0.0400 | 0.6400 | 0.2200 |
| Qwen3.5-9B | TiS | 0.0200 | 0.0400 | 0.0300 | 0.1467 | 0.5700 |
| Qwen3.5-9B | IGTS w/o SC | 0.0367 | 0.0333 | 0.0500 | 0.2833 | 0.2200 |
| Qwen3.5-9B | PostCheck | 0.0400 | 0.0300 | 0.0733 | 0.1533 | 0.2200 |
| Qwen3.5-9B | IGTS | 0.0200 | 0.0133 | 0.0367 | 0.1867 | 0.1300 |
| InternVL3.5-8B | Root | 0.7633 | 0.5233 | 0.9967 | 0.8433 | 0.0300 |
| InternVL3.5-8B | COIR | 0.0300 | 0.0200 | 0.0367 | 0.1800 | 0.7100 |
| InternVL3.5-8B | TiS | 0.0600 | 0.0467 | 0.0533 | 0.2167 | 0.6800 |
| InternVL3.5-8B | IGTS w/o SC | 0.2000 | 0.0767 | 0.0733 | 0.2633 | 0.2100 |
| InternVL3.5-8B | PostCheck | 0.1200 | 0.0800 | 0.0433 | 0.2467 | 0.2400 |
| InternVL3.5-8B | IGTS | 0.0433 | 0.0667 | 0.0400 | 0.2500 | 0.0900 |
IGTS reduces ASR on all four attacks relative to Root. COIR achieves lower ASR on the four InternVL attacks, and TiS achieves lower ASR on Qwen CS-DJ and MML. IGTS has lower benign FRR than both baselines on both backbones. Together, the results characterize the balance between attack ASR and benign refusal under each method.
PST [5] combines policy-guided SFT and preference optimization. We could not locate an official training implementation or checkpoint for a matched comparison. MSR-Align [6] releases policy-grounded multimodal training data. A matched comparison would require training on those data under the same protocol. We will expand the comparative evaluation of this method in a future revision.
Manuscript changes: §2, §4.2.3 and Appendix M (Table 9) add related methods and training comparisons.
References
[3] Gyu-Won Choi, Donggon Jang, and Dae-Shik Kim. “COIR: Chain-of-Intention Reasoning Elicits Defense in Multimodal Large Language Models.” NeurIPS 2025 Workshop on Reliable ML from Unreliable Data, 2025.
[4] Xinyue Lou, You Li, Jinan Xu, Xiangyu Shi, Chi Chen, and Kaiyu Huang. “Think in Safety: Unveiling and Mitigating Safety Alignment Collapse in Multimodal Large Reasoning Model.” Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 5167–5186, 2025.
[5] Jingyu Zhang, Kun Yang, Ming Wen, Zhuoer Xu, Zeyang Sha, Shiwen Cui, and Zhaohui Yang. “Teach to Reason Safely: Policy-Guided Safety Tuning for MLRMs.” International Conference on Learning Representations, 2026.
[6] Yinan Xia, Yilei Jiang, Yingshui Tan, Xiaoyong Zhu, Xiangyu Yue, and Bo Zheng. “MSR-Align: Policy-Grounded Multimodal Alignment for Safety-Aware Reasoning in Vision-Language Models.” arXiv:2506.19257, 2025.
[7] Xirui Li, Hengguang Zhou, Ruochen Wang, Tianyi Zhou, Minhao Cheng, and Cho-Jui Hsieh. “Is Your Multimodal Language Model Oversensitive to Safe Queries?” International Conference on Learning Representations, 2025
Responses to Reviewer EQRB (1/4)
W1 & Q1. Controlled evidence for DCE, attack taxonomy, and Table 1 exceptions
Thank you for this suggestion. Our structural interventions vary the reconstruction requirements, and our content-matched prompts vary the requested response order. Together they let us examine two ways in which task organization affects safety outcomes.
A1: (1) Structural interventions testing input requirements
First, starting from FigStep, we systematically incorporated individual components of MML and evaluated the resulting configurations across the same two backbones. Below we highlight three conditions from Table 2.
Introducing reconstruction-related structures reverses the sign of the reasoning effect on both backbones. The S-A variant retains the FigStep image, whose visible text requires no substitution. It introduces the reconstruction-oriented instructions and dictionary together. Condition A also reverses the direction on both backbones. The neutral reconstruction controls extend the comparison to benign auxiliary content and show mixed results across settings.
(2) Content-matched controls for response ordering
Second, we introduced Matched-Pre, Matched-Free, and Matched-Post to test response ordering under fixed instructional content and generation budgets. All three request intent reconstruction, risk assessment, a response decision, and an answer. The conditions vary the requested field order.
Matched-Pre: Intent → Risk → Decision → Answer.
Matched-Free: Includes all four elements in an unconstrained order.
Matched-Post: Answer → Intent → Risk → Decision.
We evaluated these controls on Qwen3.5-9B and InternVL3.5-8B across FigStep, HADES, CS-DJ, MML, and MOSSBench.
The content-matched comparisons show how ASR and benign FRR vary with the requested output order. On InternVL3.5-8B, Matched-Post yields lower reported ASR than Matched-Pre across the four evaluated attacks, while Matched-Pre achieves lower FRR on benign queries. This is why we report both metrics when comparing the prompt conditions.
(3) Attack taxonomy and explanation of Table 1 exceptions
We classify attacks by input requirements, independently of ASR. Execution-dominant attacks (FigStep/HADES) request completion after visual recognition. Verification-dominant MML requires instruction recovery and content checking. Content verification concerns wording; safety checking concerns permissibility. CS-DJ combines subqueries and visual evidence. For new attacks, we record recovery transformations, explicit content checks, and distinct branches; these can coexist.
CS-DJ ASR rises from 0.9333 to 0.9967 on InternVL3.5-8B and 0.5433 to 0.6767 on Phi-4-vision-15B. Reasoning may integrate subqueries into actionable requests while fragmented queries make harmful intent less salient. FigStep ASR falls from 0.6233 to 0.6067 on InternVL3.5-4B; reasoning may clarify harmful intent and encourage refusal. These are behavioral hypotheses, with effects depending on model alignment.
DCE is a behavioral framework: structural interventions connect input requirements to sign reversals, while matched prompts test requested output order. These results distinguish intent recovery, safety assessment, and execution.
Manuscript changes: §§3.3–3.4 and 3.6 and Appendices J and M (Table 8) present structural and ordering controls, taxonomy criteria, and exceptions.