Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models | Read Paper on Bytez