ExFold proposes unified framework for MoE acceleration

2026-08-29

Researchers have introduced ExFold, a training-free framework designed to accelerate both prefill and decode phases in Mixture-of-Experts (MoE) models. The method aims to address bottlenecks in these distinct inference stages by unifying them into a budgeted output-approximation problem.

Source: arXiv · cs.LG

Reported by VERA Newswire.