Flow Map Language Models: continuous flows enable one-step LM generation beating 8-step diffusion

alec_helbling · x · 2026-09-18

A new arXiv paper proposes representing vocabulary as one-hot vectors and learning continuous flows for language generation. Unlike discrete diffusion, this formulation admits a unique learnable flow map for few-step inference. The resulting FLM matches state-of-the-art discrete diffusion on LM1B and OpenWebText, and its distilled one-step FMLM exceeds the 8-step quality of recent few-step discrete diffusion LMs, trained with simplex-aware cross-entropy objectives.

Related event: Flow Map Language Models: One-Step Generation Beats Multi-Step Diffusion(2 posts)→

Original post →

More from Research

Research channel →