Multilingual Safety Alignment via Self-Distillation
Published in NeurIPS, 2026
Recommended citation: Qin, R., Wang, Q., Liu, D., Li, Q., Wei, Z., & Shen, W. Multilingual Safety Alignment via Self-Distillation. In NeurIPS 2026. https://arxiv.org/abs/2605.02971
Abstract. This paper proposes Multilingual Self-Distillation, a cross-lingual safeguard transfer framework that improves multilingual safety alignment of large language models without requiring response data for each target language.
Authors: Ruiyang Qin*, Qingzhuo Wang*, Dongrui Liu, Qiang Li, Zhihua Wei, Wen Shen†.
* Equal contribution; † Corresponding author.