MESA: Improving MoE Safety Alignment via Decentralized Expert Division of Labor

The researchers propose MESA, a safety alignment framework for large language models with MoE architectures, which enhances defenses by decentralizing safety responsibilities across multiple experts while minimizing harm to model usefulness. The paper points out that MoE architectures have a Safety Sparsity vulnerability: safety capability is concentrated in a few experts and is easily bypassed by adversarial means; conventional alignment methods uniformly adjust all parameters, ignoring functional differences among experts, which incidentally causes performance degradation. MESA is based on optimal transport (OT) theory and includes two mechanisms: expert capacity reallocation uses a transport cost matrix to assign safety responsibilities to the most cost-effective experts, and dynamic routing refinement constrains the router to precisely activate these decentralized modules. Experiments show that MESA achieves robust defense performance on multiple harmful benchmarks while maintaining usefulness. The code is open-sourced at https://github.com/lorraine021/MESA。