Block-diffusion language models denoise a whole block of token positions in parallel, so each forward pass can emit many tokens. This within-block acceleration is a clear efficiency gain, but the abstract notes that blocks are still denoised in strict sequence—a cost that the parallel approach leaves unaddressed.
The paper, titled "One Block Ahead: Promoting Sampler State in Diffusion Language Models," appears to focus on this limitation. Based on the abstract, the proposed remedy is to promote sampler state, presumably to allow blocks to be processed less rigidly or to carry information forward more effectively.
Because the abstract is truncated, the exact mechanism is not described in the available text. The source provides only the problem statement and the title's hint at a solution, so the article cannot go beyond that without speculating.