MENTOR: A Metacognition-Driven Self-Evolution Framework for Uncovering and Mitigating Implicit Domain Risks in LLMs 文章

arXiv:2511.07107v3 Announce Type: replace-cross Abstract: Ensuring the safety of Large Language Models (LLMs) is critical for real-world deployment. However, current safety measures often fail to address implicit, domain-specific risks. To investigate this gap, we introduce a dataset of 3,000 annotated queries spanning education, finance, and management. Evaluations across 14 leading LLMs reveal a concerning vulnerability: an average jailbreak success rate of 57.8\%. In response, we propose MENTOR, a metacognition-driven self-evolution framework. MENTOR performs metacognitive self-assessment, using strategies such as perspective-taking and consequential reasoning to uncover latent model misalignments. The resulting reflections are distilled into dynamic rule-based knowledge graphs, from which retrieved rules are converted into activation-level steering signals to guide internal representations during inference.