MENTOR Framework: Metacognition-Driven Self-Evolution for Uncovering and Mitigating Implicit Domain Risks in LLMs
ORIGINAL / MENTOR: A Metacognition-Driven Self-Evolution Framework for Uncovering and Mitigating Implicit Domain Risks in LLMs
The study introduces a dataset of 3,000 annotated queries across education, finance, and management, finds an average jailbreak success rate of 57.8% across 14 leading LLMs, and proposes MENTOR, a metacognition-driven self-evolution framework that reduces attack success rates via self-assessment, rule-guided inference, and a selective metacognitive evolution cycle.
01 ABSTRACT
The authors built a 3,000-query domain-annotated dataset and evaluated 14 LLMs, finding an average jailbreak success rate of 57.8%. They propose MENTOR, which couples single-pass rule-guided inference with a selectively invoked metacognitive evolution cycle to revise unsafe responses and distill rules into activation-level steering signals. Experiments show MENTOR reduces attack success rates across all tested domains and outperforms existing safety alignment methods.
02 KEY FINDINGS
- A dataset of 3,000 annotated queries spanning education, finance, and management was constructed to evaluate implicit domain risks.
- Across 14 leading LLMs, the average jailbreak success rate reached 57.8%, exposing gaps in current safety measures.
- MENTOR achieves self-evolving safety alignment through metacognitive self-assessment, rule-graph distillation, and activation-level steering signals.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.CL