Grouter: Decoupling Routing from Representation for Accelerated MoE Training 事件

Grouter: Decoupling Routing from Representation for Accelerated MoE Training arXiv:2603.06626v2 Announce Type: replace-cross Abstract: Traditional Mixture-of-Experts (MoE) training typically proceeds without any structural priors, effectively requiring the model to simultaneously train expert weights while searching for an optimal routing policy within a vast combinatorial space. This entanglement often leads to sluggish convergence and training instabilities. This paper introduces Grouter, a p