From the source
- The paper introduces boundary-aware self-distillation for controlled LLM safety refusal, focusing on refusing only a harmful subset of a topic rather than the entire topic.
- Using political persuasion as a testbed, the authors identify weaknesses in standard self-generation safety tuning, including coverage gaps, downside reactions, and inadequate boundary measurement.
- Experiments on Qwen3-8B show that while political refusal rates improve from 9.47% to 84.75%, over-refusal on safe prompts rises from 2.00% to 74.00%, highlighting the trade-off between safety and over-refusal.






