Multilingual Safety Alignment via Self-Distillation

Multilingual Self-Distillation (MSD) transfers an LLM’s existing safety capabilities from high-resource languages to low-resource languages. It uses multilingual queries without requiring external response data in any language. The framework supports both on-policy and off-policy self-distillation, using student-sampled and teacher-sampled responses, respectively.

We introduce Dual-Perspective Safety Weighting (DPSW) to adapt the distillation objective at the token level. DPSW uses teacher and student confidence to increase the weights of safety-critical tokens and reduce those of non-critical tokens. Experiments on multilingual jailbreak and utility benchmarks show improved safety, including transfer to unseen languages and more challenging datasets, while preserving general capabilities.

Recommended citation: Ruiyang Qin*, Qingzhuo Wang*, Dongrui Liu, Qiang Li, Zhihua Wei, Wen Shen. (2026). "Multilingual Safety Alignment via Self-Distillation." NeurIPS 2026.
Download Paper