META-IR: Meta-Introspective Reasoning for Adaptive Safety Alignment in LLM
Large Language Models (LLMs) remain vulnerable to jailbreak attacks and harmful-content prompts, particularly under adversarial distribution shifts where static alignment pipelines fail to generalize.
arXiv - cs.AI2 min read