The paper defines the reported routed-expert budget as the average retained routed experts over evaluated tokens and MoE layers, with the retained fraction equal to that average divided by native K. Shared experts are excluded. · CiteArk
The paper defines the reported routed-expert budget as the average retained routed experts over evaluated tokens and MoE layers, with the retained fraction equal to that average divided by native K. Shared experts are excluded.
来源：fixed-paper:PDF pp. 4 and 8, Equations 1–3 and Sections 3.1 and 5.1
