SMELT paper: looping middle layers twice in MoE saves 6.8-18% training FLOPs

iScienceLuvr · x · 2026-09-02

New paper SMELT (Sparse MoE Transformer, middle layers Loop Twice) studies looping on Mixture-of-Experts Transformers while closely matching per-token FLOPs, total non-embedding parameters, and KV cache.

The recipe loops the middle half of layers twice, matching an unlooped baseline on all three budgets. The authors scale across four sizes up to 54B non-embedding parameters and fit a separate Chinchilla-style scaling law per architecture.

Result: SMELT's loss drops faster with compute, saving 6.8–18.0% of training FLOPs on the compute-optimal frontier.

Related event: ByteDance's SMELT: Looping MoE Layers Cuts Training Compute Up to 18%(4 posts)→

Original post →

More from Research

Research channel →