4 ms·Diffusion‑based LLMs generate many parallel tokens rather than one‑by‑one1 points by binyu 4mo ago