Elastic Expert Routing Softens Top-k Boundaries in Sparse MoE
October 9, 2026
Elastic Expert Routing replaces static top-k expert selection with stochastic sampling from a discrete distribution centered at k. This method eliminates the rigid decision boundary between supervised and zero-feedback experts during training while maintaining the same expected computational cost and inference budget.
HOW THIS AFFECTS YOU
●
researcherYou can mitigate training instability in MoE models by smoothing the expert selection threshold.