Medicine

Discrete Diffusion Language Models for Interactive Radiology Report Drafting

Max Van Puyvelde, Halil Ibrahim Gulluk, Wim Van Criekinge, Olivier Gevaert

Featured July 7, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Imagine filling in a crossword puzzle where you can write anywhere and use clues from all directions; this paper shows a new AI that does just that for radiology reports, making it faster and better than older AIs that could only write from left to right, especially with any-order infill.

In depth

The paper demonstrates that discrete diffusion language models can serve as highly effective medical foundation models, outperforming or matching autoregressive counterparts in medical visual question answering and offering significantly faster decoding. Crucially, these diffusion models introduce a unique any-order infill capability, allowing radiologists to interactively complete report sections by fixing text fragments at arbitrary positions, a task where autoregressive models struggle due to their left-to-right generation constraint.

Key Takeaways

1
Diffusion models are competitive with or superior to autoregressive models for medical VQA, achieving comparable accuracy and 3.5-4.4x faster decoding.
2
The any-order infill capability, inherent to diffusion models, enables interactive radiology report drafting by filling gaps using bidirectional context.
3
A rigorous matched comparison methodology, using identical LoRA recipes and an LLM judge, validates the diffusion paradigm's viability in medical imaging.

Conceptual Flow

HIGH LEVEL

Methodology: The Logic of Bidirectional Text Generation

The AI learns to write medical reports by cleaning up noisy text, using clues from both the image and any words already written, no matter where they are.

Medical Image

Noisy Report Text

Clean Up Text

Clear Report Text

Results: Faster and More Flexible Report Drafting

This new AI writes reports faster and can fill in missing parts using clues from both sides, which old AIs couldn't do well.

Old AI

New AI

Compare Performance

Faster Writing

Fill Gaps Easily

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.