A simple recipe for language generation with continuous flow models over one-hot tokens

alec_helbling · x · 2026-09-18

The author summarizes a growing line of work applying continuous flow models to discrete problems like language generation: represent vocabulary items as one-hot vectors, learn a flow in the continuous space, then decode at simplex corners — enabling few-step, even one-step, high-quality generation.

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

Original post →

More from Research

Research channel →