Parallelism, critical windows, and separations among diffusion language models

Sitan Chen, Liye Wang

A popular selling point of diffusion large language models (dLLMs) is their capacity for parallelism: the ability to generate sequences of text far more efficiently than autoregres…