Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal

The article discusses the limitations of current safety alignment methods in AI, particularly how they treat harmful prompts as a property of a topic rather than recognizing the need for nuanced boundaries within those topics. It introduces a new approach that focuses on refusing only harmful subsets of topics while allowing benign prompts to be answered, specifically in the context of political prompts.