Researchers say a diffusion language model can generate radiology reports as well as leading autoregressive models — and do it faster, with a workflow twist that could matter in clinical settings.
A team benchmarked a mixture-of-experts diffusion language model against an autoregressive model of the same size, using identical fine-tuning on medical visual question answering datasets. The diffusion model matched or exceeded the autoregressive baseline across every benchmark, while running 3.5 to 4.4 times faster at inference. The fine-tuned version, with only 3.8 billion active parameters, held its own against much larger frontier vision-language models.
The speed gap alone would make a case for diffusion in high-volume radiology workflows. But the more interesting claim is architectural: because diffusion models denoise a full token canvas bidirectionally rather than generating left to right, they support any-order infill — a radiologist can anchor specific phrases in a draft and let the model fill the gaps between them. Autoregressive models handle this poorly by design. In a field where reports are often terse, inconsistent across clinicians, or drafted from partial notes, that editing affordance could reduce friction in ways raw accuracy scores do not capture.
Medical AI has been an autoregressive monoculture, so this is a meaningful proof of concept. Whether diffusion models can clear the regulatory, integration, and trust hurdles that slow any clinical AI deployment is a separate question the paper does not answer.