Perhaps something similar to speculative decoding.

Speculating Experts Accelerates Inference for Mixture-of-Experts: https://arxiv.org/abs/2603.19289