Live page · Day archive

LatentLoop: Scaling Looped MoE Models with Asymmetric Compute Allocation

Read the original at HF Daily Papers

Summary

LatentLoop is a method for looped Mixture-of-Experts models that allocates compute asymmetrically by compressing the first visit and expanding the hidden size. It is evaluated across 15 configurations ranging from 1.3B to 7B parameters.

Carried by: HF Daily Papers. First seen: .