MENTOR：用元认知自我进化揭示并缓解 LLM 隐式领域风险
MENTOR: A Metacognition-Driven Self-Evolution Framework for Uncovering and Mitigating Implicit Domain Risks in LLMs
研究者提出元认知驱动的自我进化框架 MENTOR，用于发现并缓解 LLM 在特定领域的隐式安全风险。团队构建了覆盖教育、金融、管理三个领域的 3000 条标注查询数据集，测试 14 个主流 LLM 发现平均越狱成功率达 57.8%。MENTOR 通过视角转换与后果推理进行元认知自评，将修正经验蒸馏为动态规则图并编译为激活层引导信号，在所有测试领域显著降低攻击成功率并优于现有安全对齐方法。
Authors:Liang Shan, Kaicheng Shen, Wen Wu, Zhenyu Ying, Chaochao Lu, Yan Teng, Jingqi Huang, Qingshan Liu, Guoqing Wang, Jie Zhou, Liang He
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. MENTOR couples single-pass rule-guided inference for routine requests with a selectively invoked metacognitive evolution cycle that revises residual unsafe responses, distills successful corrections into a dynamic rule graph, and compiles validated rules into activation-level steering signals for future inference. Experiments demonstrate that MENTOR substantially reduces attack success rates across all tested domains and outperforms existing safety alignment methods. The code and dataset for MENTOR are available at: this https URL.
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL) |
| Cite as: | arXiv:2511.07107 [cs.AI] |
| (or arXiv:2511.07107v5 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2511.07107 arXiv-issued DOI via DataCite |
Submission history
From: Liang Shan [view email]
[v1]
Mon, 10 Nov 2025 13:51:51 UTC (290 KB)
[v2]
Thu, 8 Jan 2026 04:30:15 UTC (428 KB)
[v3]
Wed, 3 Jun 2026 14:35:36 UTC (395 KB)
[v4]
Fri, 31 Jul 2026 18:29:13 UTC (573 KB)
[v5]
Wed, 21 Oct 2026 20:32:48 UTC (629 KB)
来源：arXiv cs.CL · arxiv.org