AI RESEARCH

MESA: Improving MoE Safety Alignment via Decentralized Expertise

arXiv CS.AI

ArXi:2606.00651v1 Announce Type: cross Mixture-of-Experts (MoE) architectures scale Large Language Models (LLMs) efficiently, enabling greater capacity with reduced computational cost by dynamically routing inputs to relevant experts, yet