Diffusion Language Models Enhance Radiology Report Drafting

While most medical foundation models rely on autoregressive (AR) text generation, this research explores the potential of discrete diffusion language models. The study adapts DiffusionGemma-26B, a mixture-of-experts diffusion model, and benchmarks it against its AR counterpart, Gemma-4-26B, using an identical LoRA fine-tuning recipe on medical visual question answering datasets. An LLM judge, robust to verbosity, was used for scoring. The findings indicate that the diffusion model performs comparably to or even surpasses AR models across all evaluated datasets. Furthermore, the fine-tuned diffusion model (with 3.8B active parameters) proves competitive with frontier vision-language models and offers significantly faster decoding, at 3.5-4.4 times the speed. Beyond performance parity, the key advantage of the diffusion model lies in its inherent ability for any-order infill. Unlike AR models, which generate text left-to-right, diffusion models denoise a token canvas bidirectionally. This allows radiologists to correct specific report fragments and have the model intelligently fill in the missing text between them, a capability crucial for improving the efficiency and consistency of real-world radiology report drafting, which often involves terse or inconsistent initial inputs.