Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic

The Hugging Face blog post describes a research paper titled 'Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal.' It argues that current safety alignment treats harm as a topic-level property, but real deployments need to refuse only a subset of a topic (e.g., political manipulation) while answering benign prompts within the same topic. The paper formalizes this narrow-boundary safety setting and proposes a training method to achieve a sharp refusal step.