Token Superposition Training Cuts Pretraining Compute 2.5x on 10B MoE

gordic_aleksa · x · 2026-09-19

Aleksa Gordić breaks down Nous Research's token superposition training (TST): during early pretraining, group every s neighboring tokens, average their embeddings into one latent token, and use it to predict the next non-overlapping bag of s tokens. This cuts sequence positions 4x, exposing the model to more raw text per FLOP, before switching back to standard next-token training to recover exact ordering. Their 10B MoE experiment reaches comparable/better loss with 2.5x less compute. Intuition: a blurred, low-resolution version of text first, then sharpening — akin to a Laplacian pyramid. Big caveat: it assumes a compute-bound, not data-bound regime, which contradicts today's consensus.

Related event: Token Superposition Training Could Cut MoE Pretraining Compute by 2.5x(2 posts)→

Original post →

More from Infra

Infra channel →