Who Transfers Safety? Identifying and Targeting Cross-Lingual Shared Safety Neurons

The paper finds that large language models contain a set of cross-lingual shared safety neurons (SS-Neurons), a minimal subset of neurons that jointly regulate safety behavior across languages. The authors first identify monolingual safety neurons (MS-Neurons) and verify their causal role in safety refusal behavior through targeted activation and suppression; cross-lingual analysis shows that SS-Neurons are a subset shared between high-resource and low-resource languages, and suppressing them causes safety to decline synchronously across multiple low-resource languages, while strengthening them improves cross-lingual defense consistency. Based on this, the paper proposes a neuron training strategy that targets SS-Neurons according to language resource distribution and model architecture. Experiments claim that fine-tuning this minimal subset of neurons outperforms existing methods, improving low-resource language safety while maintaining general model capabilities. Code and dataset will be released on GitHub.