arXiv · Computer Vision· Ruochen Zhang, Yao Huang, Yitong Sun, Jiahe Xie, Jin Yan, Jifan Ma, Yuanfang Guo, Xingxing Wei·· 6 小时前AI 评分34
ThinkingGuard：通过逐步风险归因解码多模态大语言模型中的隐式风险
ThinkingGuard: Decoding Implicit Hazards via Step-by-Step Risk Attribution in Multimodal Large Language Models
AI 导读
ThinkingGuard 是针对多模态大语言模型隐式风险检测的训练框架，结合 Step-Supervised Structured Reasoning 与 step-reward Monte Carlo Tree Search 探索推理轨迹，并通过 Dual-Constraint Preference Alignment 蒸馏到模型中。
来源：arXiv · Computer Vision · arxiv.org