Flow Map Language Models: One-Step Generation Beats Multi-Step Diffusion

A new paper proposes Flow Map language models that generate text in one continuous-flow step, outperforming 8-step discrete diffusion. The recipe: represent vocabulary as one-hot vectors, learn flows in continuous space, and decode at simplex vertices.

2026-09-18 ~ 2026-09-18 · 2 related posts

• A simple recipe for language generation with continuous flow models over one-hot tokens — alec_helbling · 2026-09-18
• Flow Map Language Models: continuous flows enable one-step LM generation beating 8-step diffusion — alec_helbling · 2026-09-18