Tag
This paper introduces Attention-Aware Routing (AAR), a method that enhances Mixture-of-Experts language models by incorporating attention weights into the router, improving mathematical reasoning performance and revealing coupled dynamics between routing and attention.