Preprint
Aug 2026
Router Sensitivity Under Lightweight Fine-Tuning Identifies Prunable Experts in Mixture-of-Experts Models
Lightweight router sensitivity makes provably motivated, task-conditioned expert pruning practical at scale and ranks experts by the induced router change, and prune the least-changed experts in one shot.
Ali Janati, K. El Maghraoui, Cheng-Ke Zou et al.
· 0 citations