Multilingual Safety Alignment Via Sparse Weight Editing

TL;DR

This paper introduces a training-free, sparse weight editing method to improve safety alignment of large language models across multiple languages, especially low-resource ones, with minimal impact on their general capabilities.

Contribution

It presents a novel, closed-form solution for cross-lingual safety alignment by sparsely editing model weights, avoiding costly retraining or data requirements.

Findings

Significantly reduces attack success rates in low-resource languages.

Maintains general reasoning abilities with negligible performance loss.

Effective across multiple languages and model architectures.

Abstract

Large Language Models (LLMs) exhibit significant safety disparities across languages, with low-resource languages (LRLs) often bypassing safety guardrails established for high-resource languages (HRLs) like English. Existing solutions, such as multilingual supervised fine-tuning (SFT) or Reinforcement Learning from Human Feedback (RLHF), are computationally expensive and dependent on scarce multilingual safety data. In this work, we propose a novel, training-free alignment framework based on Sparse Weight Editing. Identifying that safety capabilities are localized within a sparse set of safety neurons, we formulate the cross-lingual alignment problem as a constrained linear transformation. We derive a closed-form solution to optimally map the harmful representations of LRLs to the robust safety subspaces of HRLs, while preserving general utility via a null-space projection constraint. Extensive experiments across 8 languages and multiple model families (Llama-3, Qwen-2.5) demonstrate that our method substantially reduces Attack Success Rate (ASR) in LRLs with negligible impact on general reasoning capabilities, all achieved with a single, data-efficient calculation.