Discrete Diffusion Language Models for Interactive Radiology Report Drafting
06:00 · July 3, 2026 · arXiv cs.AI RSS
Diffusion language models, which generate text by denoising a token canvas bidirectionally instead of emitting tokens left to right, have become competitive with autoregressive (AR) generation. Medical foundation models, however, remain almost entirely autoregressive. We adapt a mixture-of-experts diffusion language model, DiffusionGemma-26B, and benchmark it against its same-size AR sibling Gemma-4-26B under an identical LoRA recipe on medical visual question answering datasets, scored by a verbosity-robust LLM judge. Diffusion matches or exceeds AR on all of them, and the finetuned model (3.8B active) is competitive with frontier vision-language models; its decoding is also 3.5-4.4x faster. Beyond this parity, the diffusion model offers a drafting capability AR lacks: any-order infill. Because the canvas is denoised bidirectionally, a radiologist can fix report fragments and have the model fill the text between them, an operation inherent to diffusion but not to autoregression, which is subpar at it. This suits real reports, which are often terse or inconsistent across clinicians and institutions.
Summary
Diffusion language models generate text by iteratively denoising a fixed token canvas, allowing every position to attend to the entire sequence rather than restricting attention to preceding tokens. Researchers have adapted one such model, DiffusionGemma-26B, a 26-billion-parameter mixture-of-experts architecture with 3.8 billion active parameters, for medical visual question answering and radiology report drafting. They compared it directly to its autoregressive counterpart, Gemma-4-26B, using identical low-rank adaptation on paired image-text data from standard medical VQA benchmarks.
In these matched experiments the diffusion model matched or exceeded the autoregressive baseline across all evaluated datasets. After fine-tuning, its performance remained competitive with frontier vision-language models while delivering 3.5- to 4.4-fold faster decoding. The same bidirectional denoising process also enables a capability absent from left-to-right generation: any-order infill. A radiologist can anchor arbitrary fragments of a report and have the model complete the intervening text, conditioning on context from both sides of each gap.
This property aligns with clinical practice, where reports are frequently terse, omit normal findings, or follow inconsistent sectional conventions across institutions. The approach therefore supports interactive drafting workflows that current autoregressive medical models handle poorly.
Why it matters
This research is highly relevant for Dutch AI researchers and MedTech enterprises focusing on clinical workflow automation. The introduction of diffusion models for text generation offers a novel, faster, and more flexible alternative to autoregressive models in healthcare applications.
