LatentLoop scales looped MoE models with asymmetric compute

LatentLoop is a method for scaling looped Mixture-of-Experts models by allocating compute asymmetrically across loop visits, improving performance at matched budgets.

Read it at HF Papers →

Coverage: 1 sources, in order

1. 14 Oct 00:00 HF Papers: LatentLoop: Scaling Looped MoE Models with Asymmetric Compute Allocation

← Front page