🤖 AI

Discrete Diffusion Language Models for Interactive Radiology Report Drafting

This paper introduces a discrete diffusion language model adapted for radiology report drafting that not only matches or exceeds the performance of autoregressive models on medical benchmarks with faster decoding but also uniquely enables any-order text infill, allowing clinicians to seamlessly edit and complete report fragments in a bidirectional manner.

Original authors: Max Van Puyvelde, Halil Ibrahim Gulluk, Wim Van Criekinge, Olivier Gevaert

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Max Van Puyvelde, Halil Ibrahim Gulluk, Wim Van Criekinge, Olivier Gevaert

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to write a medical report based on an X-ray. Usually, computers do this like a person writing a story from left to right: they write the first word, then the second, then the third, never looking back or changing what they've already written. This is called Autoregressive (AR) generation. It's like a train on a single track; once the wheels pass a point, they can't go back.

This paper introduces a new way for computers to write reports using Discrete Diffusion. Instead of writing word-by-word, imagine the computer starts with a blank page full of "noise" (random gibberish). It then looks at the whole page at once, cleans up the noise, and gradually refines the text until it makes sense. It can look at the beginning, middle, and end of the sentence simultaneously to figure out what fits best.

Here is a breakdown of what the researchers found, using simple analogies:

1. The Race: Two Siblings, One Goal

The researchers took two "siblings" from the same family of AI models (both based on a model called Gemma).

Sibling A (AR): The traditional writer who writes left-to-right.
Sibling B (Diffusion): The new writer who cleans up a whole page at once.

They gave both siblings the exact same training data (medical images and questions) and the exact same "study plan" (using a technique called LoRA to teach them). They wanted to see if the new Diffusion method could keep up with the old AR method in the medical field.

The Result: The Diffusion sibling didn't just keep up; it often performed better or equal to the AR sibling on medical quizzes. Even more impressively, it was 3.5 to 4.4 times faster.

Analogy: If the AR model is a painter who carefully adds one brushstroke at a time, the Diffusion model is like a sculptor who starts with a block of stone and chips away the excess all at once to reveal the shape. The sculptor finished the job much faster.

2. The Superpower: "Any-Order Infill"

This is the paper's biggest claim. The Diffusion model has a special ability that the AR model lacks: Any-Order Infill.

The AR Limit: Imagine you are writing a report and you stop in the middle to fix a sentence. If you use the AR model, it can only write what comes after your fix. It can't look at what you wrote after the gap to help it fix what comes before the gap. It's like trying to solve a puzzle while only seeing the pieces to the left of your hand.
The Diffusion Superpower: Because the Diffusion model looks at the whole page at once, a radiologist can leave a blank space in the middle of a draft, type a few words before it, and a few words after it. The model can then fill in the blank space, using the context from both sides to get it right.
Analogy: If you are editing a paragraph and you delete a sentence in the middle, the AR model is like a writer who only reads the text before the cursor and guesses what should come next. The Diffusion model is like a writer who reads the sentence before the gap and the sentence after the gap to perfectly reconstruct the missing middle.

The researchers tested this by hiding a sentence in a real medical report and asking the models to fill it back in.

The Diffusion model used the clues from both sides of the gap and got the sentence right.
The AR model, even when it was "told" what was on the right side, still struggled to use that information to fix the left side. It just couldn't "look back" effectively.

3. Why This Matters for Doctors

The paper suggests that medical reports are often messy. Different doctors might write different things, or leave sections blank to fill in later.

The AR model is rigid; it expects a strict order.
The Diffusion model is flexible. It allows a doctor to draft a report in a "non-linear" way—typing the conclusion first, then the findings, then filling in the middle later. The model can fill in the gaps based on the context surrounding them, making the drafting process feel more like a conversation and less like a rigid form-filling exercise.

Summary of Claims

Accuracy: The new Diffusion model is just as good (or better) at answering medical questions than the traditional model.
Speed: It is significantly faster (up to 4x).
Flexibility: It is the first model of its kind to successfully "fill in the blanks" in the middle of a report using context from both the left and right sides, a task the traditional model fails at.
Availability: The researchers have released their code and the trained models for others to use.

The paper does not claim this is currently being used in hospitals to diagnose patients, but rather that it is a powerful new tool for drafting reports that is faster and more flexible than the current standard.

Technical Summary: Discrete Diffusion Language Models for Interactive Radiology Report Drafting

Problem Statement

Autoregressive (AR) generation, which produces text token-by-token from left to right, currently underpins nearly all large language models (LLMs) and vision-language models (VLMs) in the medical domain, including radiology report generation (RRG). While AR models are effective for sequential generation, they lack inherent capabilities for any-order infill—the ability to generate text in the middle of a sequence conditioned on fixed content on both the left and right sides. This limitation hinders interactive drafting workflows where radiologists may fix specific report fragments and require the model to fill gaps between them. Furthermore, existing medical foundation models are almost exclusively AR, leaving the viability of discrete diffusion models in clinical settings largely untested.

Methodology

The authors conduct a controlled comparison between discrete diffusion and autoregressive paradigms using matched backbones to isolate the generative paradigm as the sole variable.

1. Matched Backbones and Adaptation

Models: The study compares DiffusionGemma-26B (a discrete diffusion language model) against its autoregressive sibling, Gemma-4-26B. Both are 25.2B parameter models with 3.8B active parameters (Mixture-of-Experts) and share a SigLIP-lineage vision encoder.
Adaptation: Both models are fine-tuned using an identical Low-Rank Adaptation (LoRA) recipe. The adaptation targets attention and shared-MLP projections (rank-64, $\alpha=128$ ) while freezing the experts, router, and vision tower.
Optimization: The only deliberate difference is the generative objective and optimizer settings:
- Diffusion: Uses the uniform-state dLLM objective, where a random fraction of target tokens is replaced with uniform vocabulary draws, and the model learns to recover them.
- Autoregressive: Uses standard next-token cross-entropy.
- Optimizers differ (AdamW with different $\beta$ values) to align with established settings for each objective.

2. Training and Evaluation Tasks

Medical Visual Question Answering (VQA): The models are evaluated on three datasets (VQA-RAD, SLAKE, VQA-Med-2019) using a verbosity-robust LLM judge (Claude Sonnet 4.6) to score semantic equivalence rather than exact match.
Inference Speed: Latency and throughput are measured on a single H100 GPU for $\sim$ 256-token generations.
Any-Order Infill: The authors define infill as sampling the conditional distribution $p_\theta(x_{\bar{F}} | x_F = a, c)$ $p_{θ} (x_{\overset{ˉ}{F}} ∣ x_{F} = a, c)$ , where $F$ $F$ represents fixed positions and $\bar{F}$ $\overset{ˉ}{F}$ represents free positions.
- Diffusion: Naturally supports this by bidirectionally denoising the entire canvas. Fixed tokens are re-imposed at every denoising step.
- Autoregressive: Cannot natively sample this conditional. To simulate bidirectional context, the right-side text is provided in the prompt, but the model still generates left-to-right.

Key Contributions

Competitive Medical Foundation Model: The study demonstrates that a discrete diffusion model (DiffusionGemma-26B) matches or exceeds its autoregressive sibling on medical VQA benchmarks and rivals frontier VLMs, while decoding 3.5–4.4 $\times$ faster.
Any-Order Infill Capability: The paper establishes that diffusion offers a unique conditioning capability absent in AR: filling gaps using context from both sides simultaneously. This is inherent to the bidirectional denoising process.
Open Release: The authors release code and fine-tuned checkpoints for the diffusion and AR models.

Results

Medical VQA Performance

Accuracy: Fine-tuned diffusion matches or exceeds fine-tuned AR across all three datasets. On the SLAKE dataset, fine-tuned diffusion achieved 0.863 accuracy compared to 0.817 for AR.
Significance: The difference was statistically significant on SLAKE (fine-tuned) and VQA-RAD (base). Diffusion performed particularly well on closed-ended (yes/no) questions.
Frontier Comparison: The fine-tuned 26B model (3.8B active) is competitive with frontier non-reasoning VLMs (Gemini-3.5-Flash, GPT-4.1-mini, Claude-Sonnet-4.6), outperforming GPT-4.1-mini on all datasets.

Inference Speed

DiffusionGemma drafts reports 3.5–4.4 $\times$ faster than the AR sibling (Gemma-4) on an H100 GPU.
Throughput for diffusion ranges from 139.4 to 175.3 tokens/second, compared to 24.6 tokens/second for AR.
This speed advantage arises because diffusion cost is determined by the denoising-step budget over a fixed canvas size, whereas AR latency scales linearly with report length.

Any-Order Infill Performance

Context Exploitation: On MIMIC-CXR reports with masked sentences, the diffusion model significantly improved when given context from both sides of the gap:
- Token-F1: Increased by +0.109 ( $p < 10^{-10}$ ).
- LLM-Judge Accuracy: Increased by +0.129 ( $p = 2 \times 10^{-5}$ ).
AR Limitation: The autoregressive model showed no significant improvement (+0.031 Token-F1, non-significant) even when the right-side context was provided in the prompt. The interaction between model type and context direction was significant, indicating diffusion benefits roughly 3.5 $\times$ more from bidirectional context.

Significance and Claims

The paper claims that discrete diffusion is a viable substrate for medical foundation models, offering accuracy comparable to autoregressive models with the added benefits of significantly faster inference and unique interactive drafting capabilities.

Specifically, the authors argue that the any-order infill capability is a critical differentiator for clinical workflows. Real-world radiology reports are often terse, inconsistent, or drafted in fragments. A diffusion-based tool allows a radiologist to fix specific report sections and have the model fill the gaps between them, leveraging context from both directions. This capability is described as "inherent to diffusion but not to autoregression," positioning diffusion as a superior paradigm for interactive, collaborative report drafting rather than just sequential generation.

The study concludes that while AR remains dominant, diffusion models offer a competitive alternative that is not only accurate but also more efficient and better suited for non-sequential, interactive editing tasks in radiology.