Block Diffusion Interpolating Between Autoregressive And Diffusion Language Models 2025 03 12

Block Diffusion generates text in successive chunks while refining multiple text units together inside each chunk. Once a chunk is complete, the model retains computations from it to help generate the next one. The researchers' technical explanation shows how this combines parallel generation within blocks with the ability to keep extending a document.

The publisher abstract reports improved language-modeling results among diffusion models. The contribution is a way to make diffusion-based text generation more flexible; the project's perplexity tables still favor the conventional autoregressive baseline.