aiqbench analysis
MENTOR addresses a critical gap in LLM safety by using metacognitive self-assessment to detect and mitigate implicit, domain-specific risks that traditional alignment methods miss. The framework converts model reflections into dynamic rule-based knowledge graphs, which generate activation-level steering signals to effectively reduce jailbreak success rates across diverse sectors like finance and education. This approach offers a promising direction for enhancing robustness against adversarial attacks without relying solely on static fine-tuning.
MENTOR: A Metacognition-Driven Self-Evolution Framework for Uncovering and Mitigating Implicit Domain Risks in LLMs
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. 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: https://anonymous.4open.science/r/MENTOR-Evo.
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. 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.
Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
arXiv:2511.07107 [cs.AI]
(or arXiv:2511.07107v3 [cs.AI] for this version)
Full text under the publisher's open licence · 155 words
Related coverage
MENTOR: A Metacognition-Driven Self-Evolution Framework for Uncovering and Mitigating Implicit Domain Risks in LLMs
arXiv · Aug 5, 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · Sep 18, 2026
.webp%3Fssl%3D1&w=320)
Top 50 Best Penetration Testing Companies in 2026
Cybersecurity News · Jul 31, 2026
Rethinking Self-Evolution: A Constrained Exploration-Exploitation Process for Mitigating Skill Overfitting
arXiv · Jul 31, 2026