Multilingual Safety Alignment Via Sparse Weight Editing

Jiaming Liang, Zhaoxin Wang, Handing Wang

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 de-
pendent on scarce multilingual safety data. In this paper, 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 significantly reduces Attack Success Rate (ASR) in LRLs with negligible impact on general reasoning capabilities, all achieved with a single, data-efficient calculation.

Lay Summary:

Large language models are now used by people around the world, but their safety protections do not work equally well in every language. A model that refuses a harmful request in English may still answer the same kind of request in a lower-resource language, where less safety training data is available.

Our work studies how to transfer safety behavior across languages without expensive retraining. We find that some internal parts of a model are especially related to safety, and we edit only a small set of these parts. This edit moves harmful requests in low-resource languages closer to the safer internal patterns the model already uses in English, while trying to preserve its general abilities.

Across several model families and languages, our method reduces unsafe responses while keeping performance on general reasoning tasks largely unchanged. This suggests that multilingual AI safety can be improved with lightweight model edits, rather than requiring large new multilingual safety datasets.

Link To Code: https://github.com/HandingWangXDGroup/MultilingualSparseWeightEdit

Primary Area: Deep Learning->Large Language Models

Keywords: Safety Alignment, Multilingual Safety

Submission Number: 9993