Cross-Lingual Safety Transfer for LLMs via Self-Distillation

Large language models (LLMs) face significant issues with multilingual safety alignment, showing robust protections in high-resource languages but being particularly susceptible to jailbreak attacks in low-resource languages. The existing methods for safety alignment necessitate high-quality response data for each specific language, which is both costly and challenging to produce. To address this, researchers have introduced a framework called Multilingual Self-Distillation (MSD), which allows for the transfer of an LLM's safety features from high-resource languages like English to low-resource languages such as Javanese, without requiring response data. This adaptable framework can work with various self-distillation techniques. Two specific approaches, on-policy MSD and off-policy MSD, facilitate effective safety transfer across languages using only multilingual queries. The paper is accessible on arXiv under ID 2605.02971.