To address the issue of excessive rejection of harmless content when large language models are deployed in specific domains, Hugging Face’s team proposed the “Boundary-Sensitive Self-Distillation” method. The study found that focusing solely on the lowest harmful response rate would result in up to 74% of harmless prompts being wrongly rejected. To address this, researchers developed an offline self-generation framework combining controlled topic generation, coverage repair, and data pairs of harmful and benign content, and trained it on the Qwen3-8B model. Experiments showed that by adding boundary-related data, this false rejection rate was significantly reduced from 32.94% to 4.16%, while maintaining a basically unchanged real rejection rate. Comparative experiments further verified that replacing external responses with the responses of the validated target model could significantly reduce excessive rejection rates. The study emphasized that safe alignment should not be based solely on the harmful response rate; it must be evaluated on both sides of the boundary between intention to reject and usability, in order to balance safety and availability.
Regarding the issue of needing to set different security boundaries for different deployment scenarios under the same theme, researchers proposed an offline self-generating framework that combines controlled theme generation, coverage repair, and harmful-healthy pairing data. After training on the Qwen3-8B model, the target domain’s refusal rate increased from 9.47% to 84.75%, and the average unsafe response rate for the three broad harmfulness benchmarks decreased to 0.14%. However, the XSTest over-refusal rate rose to 74.00%. Comparative experiments showed that replacing external responses with the validated target model responses could reduce the over-refusal rate from 15.20% to 5.20%; boundary pairing data reduced the over-refusal rate on the retained side from 32.94% to 4.16%, while the harmful-side refusal rate decreased slightly from 91.88% to 87.72%. The results indicate that data composition can adjust the balance between security and usability, and security alignment needs to be evaluated on both sides of the intention-to-refuse boundary.