Diffusion language models differ from standard autoregressive models in a key way: they can produce multiple tokens in one forward pass. They also decide during generation how many passes to spend on a given output. As a result, two outputs of the same token length can vary substantially in computational cost.
The paper, posted on arXiv, sets out to understand the underlying reasons for this variability. Its title points to speed changing across tasks and languages, suggesting the cost differences are not random but tied to what is being generated. The abstract, however, is truncated before presenting specific findings.
Based on the available text, the study's contribution at this stage is framing the problem: measuring and explaining why diffusion LMs allocate compute unevenly. Readers should treat any specific results as not yet visible in the abstract.