1

LatentLoop: Scaling Looped MoE Models with Asymmetric Compute Allocation

huggingface.co

LatentLoop is a method for scaling looped Mixture-of-Experts models using asymmetric compute allocation, compressing the first loop visit's sequence while retaining full resolution in the second, and it reportedly lowers validation NLL across 15 configurations.

Sign in to suggest

Key sources

  1. SOURCEhuggingface.co
Markdown