Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment

The researchers propose a plug-and-play multilingual consistency (MLC) loss that can be embedded into existing monolingual alignment pipelines to improve LLM multilingual safety alignment at low cost. By increasing the collinearity of multilingual representation vectors, the method can align multiple languages simultaneously in a single update using only multilingual prompt variants, without requiring response-level supervision in low-resource languages. Validation across multiple model architectures and alignment paradigms shows that it enhances multilingual safety, has limited impact on general capabilities, and improves cross-lingual generalization.