Jul 2026
Rethinking the Generation Order of Block Diffusion Language Models
This work proposes Parallel Autoregressive Decoding (PARD), a simple training-free sampling method that preserves left-to-right unmasking structure while allowing parallel token commitment and consistently outperforms existing parallel samplers in generation quality.
Kai-Nuo Hou, James T. Kwok
· arXiv.org · 0 citations