MPO: Multilingual Safety Alignment via Reward Gap Optimization

TL;DR

This paper introduces MPO, a novel method for multilingual safety alignment of large language models that transfers safety capabilities from English to other languages by minimizing reward gaps, validated through extensive experiments.

Contribution

MPO is a new approach that effectively aligns safety across multiple languages by directly minimizing reward gaps, addressing limitations of existing monolingual preference learning methods.

Findings

MPO improves multilingual safety alignment without reducing utility.

It outperforms existing methods on three large language models.

MPO maintains safety capabilities while preserving multilingual performance.

Abstract

Large language models (LLMs) have become increasingly central to AI applications worldwide, necessitating robust multilingual safety alignment to ensure secure deployment across diverse linguistic contexts. Existing preference learning methods for safety alignment, such as RLHF and DPO, are primarily monolingual and struggle with noisy multilingual data. To address these limitations, we introduce Multilingual reward gaP Optimization (MPO), a novel approach that leverages the well-aligned safety capabilities of the dominant language (English) to improve safety alignment across multiple languages. MPO directly minimizes the reward gap difference between the dominant language and target languages, effectively transferring safety capabilities while preserving the original strengths of the dominant language. Extensive experiments on three LLMs, LLaMA-3.1, Gemma-2 and Qwen2.5, validate MPO's efficacy in multilingual safety alignment without degrading general multilingual utility.