Saved in:
Bibliographic Details
Main Authors: Shan, Liang, Shen, Kaicheng, Wu, Wen, Ying, Zhenyu, Lu, Chaochao, Teng, Yan, Huang, Jingqi, Ye, Guangze, Wang, Guoqing, He, Liang
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2511.07107
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918277457903616
author Shan, Liang
Shen, Kaicheng
Wu, Wen
Ying, Zhenyu
Lu, Chaochao
Teng, Yan
Huang, Jingqi
Ye, Guangze
Wang, Guoqing
He, Liang
author_facet Shan, Liang
Shen, Kaicheng
Wu, Wen
Ying, Zhenyu
Lu, Chaochao
Teng, Yan
Huang, Jingqi
Ye, Guangze
Wang, Guoqing
He, Liang
contents 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 first performs structured self-assessment through simulated critical thinking, such as perspective-taking and consequential reasoning to uncover latent model misalignments. These reflections are formalized into dynamic rule-based knowledge graphs that evolve with emerging risk patterns. To enforce these rules at inference time, we introduce activation steering, a method that directly modulates the model's internal representations to ensure compliance. Experiments demonstrate that MENTOR substantially reduces attack success rates across all tested domains and achieves risk analysis performance comparable to human experts. Our work offers a scalable and adaptive pathway toward robust domain-specific alignment of LLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2511_07107
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle MENTOR: A Metacognition-Driven Self-Evolution Framework for Uncovering and Mitigating Implicit Domain Risks in LLMs
Shan, Liang
Shen, Kaicheng
Wu, Wen
Ying, Zhenyu
Lu, Chaochao
Teng, Yan
Huang, Jingqi
Ye, Guangze
Wang, Guoqing
He, Liang
Artificial Intelligence
Computation and Language
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 first performs structured self-assessment through simulated critical thinking, such as perspective-taking and consequential reasoning to uncover latent model misalignments. These reflections are formalized into dynamic rule-based knowledge graphs that evolve with emerging risk patterns. To enforce these rules at inference time, we introduce activation steering, a method that directly modulates the model's internal representations to ensure compliance. Experiments demonstrate that MENTOR substantially reduces attack success rates across all tested domains and achieves risk analysis performance comparable to human experts. Our work offers a scalable and adaptive pathway toward robust domain-specific alignment of LLMs.
title MENTOR: A Metacognition-Driven Self-Evolution Framework for Uncovering and Mitigating Implicit Domain Risks in LLMs
topic Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2511.07107