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

TL;DR

This paper introduces a resource-efficient multilingual safety alignment method for large language models, enhancing cross-lingual consistency with minimal supervision by improving multilingual representation collinearity.

Contribution

The paper proposes a plug-and-play Multi-Lingual Consistency loss that enforces semantic alignment across languages in a single update, reducing resource requirements.

Findings

Improves multilingual safety alignment across diverse languages.

Enhances cross-lingual generalization with limited supervision.

Effective across different model architectures and alignment paradigms.

Abstract

The widespread deployment of large language models (LLMs) across linguistic communities necessitates reliable multilingual safety alignment. However, recent efforts to extend alignment to other languages often require substantial resources, either through large-scale, high-quality supervision in the target language or through pairwise alignment with high-resource languages, which limits scalability. In this work, we propose a resource-efficient method for improving multilingual safety alignment. We introduce a plug-and-play Multi-Lingual Consistency (MLC) loss that can be integrated into existing monolingual alignment pipelines. By improving collinearity between multilingual representation vectors, our method encourages directional consistency at the multilingual semantic level in a single update. This allows simultaneous alignment across multiple languages using only multilingual prompt variants without requiring additional response-level supervision in low-resource languages. We validate the proposed method across different model architectures and alignment paradigms, and demonstrate its effectiveness in enhancing multilingual safety with limited impact on general model utility. Further evaluation across languages and tasks indicates improved cross-lingual generalization, suggesting the proposed approach as a practical solution for multilingual consistency alignment under limited supervision.