D. Roos

A novel $\mathrm{SemiCat}$ architecture achieves discrete diffusion-level performance on TinyGSM and outperforms one-hot flows at scale by addressing self-conditioning and embedding geometry in continuous diffusion language models.

Achieve state-of-the-art few-step generation on images, molecular graphs, and text with Categorical Flow Maps, a self-distillation approach that unlocks continuous-time inference for categorical data.