MultiMDM Lets Masked Diffusion Draft Before Writing

Yinuo Ren, a Kempner Research Fellow at Harvard, posted about a COLM2026 paper on MultiMDM. The approach gives masked diffusion models typed masks that carry a rough plan for each word. These masks first sketch the sentence structure, then resolve into tokens. The post states this enables few-step generation. An attached animated video shows the process with a blue header. The tweet links to an arXiv abstract and forms the first part of a thread.