Multilingual Safety Alignment via Self-Distillation

Safety alignment of LLMs is predominantly conducted in English, leaving low-resource languages significantly under-aligned and vulnerable to harmful outputs. Traditional approaches rely heavily on high-quality human-annotated safety response data in each target language, which is expensive and hard to scale. This paper proposes an on-policy self-distillation method that transfers the model’s own safety capabilities from high-resource languages to low-resource ones. By leveraging the model’s existing safety knowledge in well-aligned languages as the distillation source, our approach eliminates the dependency on external multilingual safety data. Experiments across multiple models and diverse multilingual scenarios demonstrate significant improvements in both in-distribution and out-of-distribution language safety, while preserving the model’s general reasoning capabilities.

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