Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic
Hugging Face 1 month ago 9
A research paper argues that LLM safety work should refuse only the harmful subset of a topic, not the whole topic, because topic-level guards can mis-handle deployment-specific boundaries. It reports that single-shot self-generation for harmful political prompts drops 19.88% of prompts (8,009) but an escalating retry strategy reduces residual failures to 0.20% (79). It introduces boundary-aware self-distillation and corresponding measurements so training data composition can be tuned to reduce over-refusals on benign prompts near the boundary while keeping harmful refusals high.