Study Shows Quadratic Gap Favoring Masked Diffusion in Parallel Generation

Researchers sitannch and Liye Wang published the first systematic comparison of discrete diffusion paradigms, proving a quadratic gap in parallel generation capability between uniform/Gaussian diffusion and unmasking, explained by masked diffusion's quadratically narrower critical decision window.

2026-09-21 ~ 2026-09-21 · 2 related posts