SaFeR-VLM：迈向多模态模型中的安全感知细粒度推理
SaFeR-VLM: Toward Safety-aware Fine-grained Reasoning in Multimodal Models
AI总结：
提出SaFeR-VLM，一种将安全嵌入多模态推理的强化学习框架，通过安全数据集、反思展开、结构化奖励和GRPO优化，在保持有用性的同时显著提升安全性能，超越更大规模模型。
摘要：
多模态大型推理模型（MLRMs）展示了令人印象深刻的跨模态推理能力，但在对抗性或不安全提示下往往会放大安全风险，我们将这一现象称为“推理税”（Reasoning Tax）。现有防御措施主要作用于输出层面，并未约束推理过程，使模型暴露于隐含风险之中。在本文中，我们提出了SaFeR-VLM，一种安全对齐的强化学习框架，将安全直接嵌入多模态推理中。该框架整合了四个组成部分：（I）QI-Safe-10K，一个强调安全关键和推理敏感案例的精选数据集；（II）安全感知的展开（safety-aware rollout），其中不安全的生成结果经过反思和纠正，而非被丢弃；（III）结构化奖励建模，采用多维加权标准和针对幻觉与矛盾的明确惩罚；（IV）GRPO优化，同时强化安全和纠正后的轨迹。这种统一设计将安全从被动保障转变为主动推理驱动力，实现了可扩展且可泛化的安全感知推理。SaFeR-VLM进一步展示了对显性和隐性风险的鲁棒性，支持超越表面过滤的动态和可解释的安全决策。SaFeR-VLM-3B在六个基准上平均安全性和有用性得分分别为70.13和78.97，超越了同规模以及大于10倍规模的模型，如Skywork-R1V3-38B、Qwen2.5VL-72B和GLM4.5V-106B。值得注意的是，SaFeR-VLM-7B得益于其更大的规模，在安全指标上分别超越GPT-5-mini和Gemini-2.5-Flash 6.47和16.76个百分点，且在有用性性能上没有任何下降。我们的代码可在https://github.com/HarveyYi/SaFeR-VLM获取。
英文摘要：
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 $>10\times$ 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.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
