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

Abstract:

The deployment of Large Language Models (LLMs) in practical applications hinges on their safety. Nevertheless, existing protective mechanisms frequently overlook implicit risks that are specific to certain domains. To bridge this oversight, the authors present a curated dataset comprising 3,000 annotated queries drawn from the fields of education, finance, and management. Comprehensive evaluations involving 14 prominent LLMs highlight a significant vulnerability, revealing an average jailbreak success rate of 57.8%.

To address this challenge, the study introduces MENTOR, a novel framework driven by metacognitive self-evolution. MENTOR leverages techniques like consequential reasoning and perspective-taking to conduct metacognitive self-assessments, thereby exposing latent misalignments within the models. These insights are then distilled into dynamic, rule-based knowledge graphs. During inference, the framework retrieves specific rules from these graphs and transforms them into activation-level steering signals to direct the model’s internal representations. Empirical results show that MENTOR significantly lowers attack success rates across all evaluated domains, surpassing current safety alignment approaches. The associated code and dataset are accessible at: https://anonymous.4open.science/r/MENTOR-Evo.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC