Google's DiffusionGemma-26B drafts radiology reports 4x faster
Radiology reports get 3.5-4.4x faster drafting with diffusion models...
Deep Dive
A new study adapts a mixture-of-experts diffusion language model, DiffusionGemma-26B, and benchmarks it against its same-size autoregressive sibling Gemma-4-26B on medical visual question answering datasets. Diffusion matches or exceeds AR on all benchmarks, and the finetuned model is competitive with frontier vision-language models, with 3.5-4.4x faster decoding. The diffusion model also offers any-order infill for fixing report fragments, a capability autoregressive models lack.
Key Points
- DiffusionGemma-26B outperforms autoregressive Gemma-4-26B in medical benchmarks while decoding 3.5–4.4x faster
- The model enables any-order infill, letting radiologists edit report fragments and have AI fill gaps bidirectionally
- Fine-tuned DiffusionGemma-26B (3.8B active params) is competitive with frontier vision-language models
Why It Matters
Could slash radiology report drafting time by 75% while improving accuracy in fragmented clinical notes.