Soteria introduces a lightweight approach to enhance LLM safety across multiple languages by targeting only the specific parameters responsible for harmful outputs in each language.

• Identifies and adjusts only the functional heads most responsible for generating harmful content
• Achieves significant safety improvements while modifying just a fraction of parameters
• Maintains model performance across languages, even in low-resource settings
• Introduces XThreatBench as a specialized benchmark for multilingual safety evaluation

This research addresses a critical security challenge for deploying LLMs globally, enabling organizations to implement targeted safety controls without compromising overall model utility or requiring complete retraining.

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