META-IR: Meta-Introspective Reasoning for Adaptive Safety Alignment in LLM

Mahabubur Rahman Miraj, Hongyu Huang, Nankun Mu +2

Large Language Models (LLMs) remain vulnerable to jailbreak attacks and harmful-content prompts, particularly under adversarial distribution shifts where static alignment pipelines…