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)→
More from Research
- Sequential test-time scaling beats parallel multi-agent scaling at equal budget, data shows — eliebakouch · 2026-09-18
- Genomic language model gLM2 controllably generates new biosynthetic domain architectures — BrianHie · 2026-09-18
- TDiMS: fragment distances beat large pretrained models at chromophore property prediction — bravo_abad · 2026-09-18
- UVFaceFusion open-sourced: topology-consistent 3D face reconstruction from multi-view images in under 3s — rsasaki0109 · 2026-09-18
- ProgramAsWeights: compile English-described AI functions into tiny CPU-only neural nets — yuntiandeng · 2026-09-18
- TMLR dumps Medium, launches its own official blog with all posts migrated — thegautamkamath · 2026-09-18