Soteria: Language-Specific Functional Parameter Steering for Multilingual Safety Alignment

TL;DR

Soteria is a novel method that minimally adjusts language-specific functional parameters in multilingual LLMs to improve safety and reduce harmful content generation across diverse languages.

Contribution

It introduces Soteria, a lightweight approach for language-specific safety tuning by adjusting functional heads, and presents XThreatBench, a new multilingual dataset for evaluating harmful behaviors.

Findings

Soteria significantly reduces policy violations across multiple languages.

The approach maintains overall model performance in low-resource settings.

Experiments demonstrate improved safety metrics in open-source LLMs.

Abstract

Ensuring consistent safety across multiple languages remains a significant challenge for large language models (LLMs). We introduce Soteria, a lightweight yet powerful strategy that locates and minimally adjusts the "functional heads" most responsible for harmful content generation in each language. By altering only a fraction of parameters, Soteria drastically reduces policy violations without sacrificing overall model performance, even in low-resource settings. To rigorously evaluate our approach, we also present XThreatBench, a specialized multilingual dataset capturing fine-grained harmful behaviors drawn from real policy guidelines. Experiments with leading open-source LLMs (e.g., Llama, Qwen, Mistral) show that Soteria consistently improves safety metrics across high-, mid-, and low-resource languages. These findings highlight a promising path toward scalable, linguistically attuned, and ethically aligned LLMs worldwide.