Block Diffusion Model

A generative modeling paradigm that partitions a sequence into blocks; within each block, masked diffusion generates tokens in parallel through denoising, while blocks are linked autoregressively or in parallel, trading off flexibility in generation order against long-range consistency.