Tag
The post describes how diffusion language models face KV-cache reuse limitations due to iterative unmasking, and introduces Block Diffusion as a solution that decodes blocks left-to-right for efficient caching while generating in parallel.