New Model Controls Token Routing Dependencies in Mixture-of-Experts

2026-08-05

Researchers introduce Hierarchical Copula-Gumbel-Top-K (CGA), a novel method for managing dependencies between token routing decisions in Mixture-of-Experts (MoE) models. The approach allows for precise control over how tokens are routed to experts.

Source: arXiv · cs.LG

Reported by VERA Newswire.