{"ID":2858704,"CreatedAt":"2026-06-01T04:54:23.091178241Z","UpdatedAt":"2026-06-01T04:54:23.091178241Z","DeletedAt":null,"paper_url":"https://arxiv.org/abs/2510.06871","arxiv_id":"2510.06871","title":"SaFeR-VLM: Toward Safety-aware Fine-grained Reasoning in Multimodal Models","abstract":"Multimodal Large Reasoning Models (MLRMs) demonstrate impressive cross-modal reasoning but often amplify safety risks under adversarial or unsafe prompts, a phenomenon we call the \\textit{Reasoning Tax}. Existing defenses mainly act at the output level and do not constrain the reasoning process, leaving models exposed to implicit risks. In this paper, we propose SaFeR-VLM, a safety-aligned reinforcement learning framework that embeds safety directly into multimodal reasoning. The framework integrates four components: (I) QI-Safe-10K, a curated dataset emphasizing safety-critical and reasoning-sensitive cases; (II) safety-aware rollout, where unsafe generations undergo reflection and correction instead of being discarded; (III) structured reward modeling with multi-dimensional weighted criteria and explicit penalties for hallucinations and contradictions; and (IV) GRPO optimization, which reinforces both safe and corrected trajectories. This unified design shifts safety from a passive safeguard to an active driver of reasoning, enabling scalable and generalizable safety-aware reasoning. SaFeR-VLM further demonstrates robustness against both explicit and implicit risks, supporting dynamic and interpretable safety decisions beyond surface-level filtering. SaFeR-VLM-3B achieves average performance $70.13$ and $78.97$ on safety and helpfulness across six benchmarks, surpassing both same-scale and $\u003e10\\times$ larger models such as Skywork-R1V3-38B, Qwen2.5VL-72B, and GLM4.5V-106B. Remarkably, SaFeR-VLM-7B benefits from its increased scale to surpass GPT-5-mini and Gemini-2.5-Flash by \\num{6.47} and \\num{16.76} points respectively on safety metrics, achieving this improvement without any degradation in helpfulness performance. Our codes are available at https://github.com/HarveyYi/SaFeR-VLM.","short_abstract":"Multimodal Large Reasoning Models (MLRMs) demonstrate impressive cross-modal reasoning but often amplify safety risks under adversarial or unsafe prompts, a phenomenon we call the \\textit{Reasoning Tax}. Existing defenses mainly act at the output level and do not constrain the reasoning process, leaving models exposed...","url_abs":"https://arxiv.org/abs/2510.06871","url_pdf":"https://arxiv.org/pdf/2510.06871v2","authors":"[\"Huahui Yi\",\"Kun Wang\",\"Qiankun Li\",\"Miao Yu\",\"Liang Lin\",\"Gongli Xi\",\"Hao Wu\",\"Xuming Hu\",\"Kang Li\",\"Yang Liu\"]","published":"2025-10-08T10:39:12Z","proceeding":"cs.LG","tasks":"[\"cs.LG\",\"cs.CV\"]","methods":"[\"Reinforcement Learning\"]","has_code":false,"code_links":[{"ID":608572,"CreatedAt":"2026-06-01T04:54:23.091178241Z","UpdatedAt":"2026-06-01T04:54:23.091178241Z","DeletedAt":null,"paper_id":2858704,"paper_url":"https://arxiv.org/abs/2510.06871","paper_title":"SaFeR-VLM: Toward Safety-aware Fine-grained Reasoning in Multimodal Models","repo_url":"https://github.com/HarveyYi/SaFeR-VLM","is_official":false,"mentioned_in_paper":false,"mentioned_in_github":true,"github_stars":0}]}