Discrete Diffusion Language Models for Interactive Radiology Report Drafting | Max Van Puyvelde et al. | ResearchPod

Transcript: Discrete Diffusion Language Models for Interactive Radiology Report Drafting

Alex: Welcome to another episode of ResearchPod.

Sam: Today we're looking at a study that compares two fundamentally different ways AI models generate text—and the question at the heart of it is whether a newer approach can outperform the method used in almost every AI writing tool available today.

Alex: What's the context? Where does this comparison actually matter?

Sam: The setting is medical reporting—specifically, the kind of written reports that radiologists produce after reviewing a scan. The researchers wanted to know whether a newer generation method, called discrete diffusion, could match or beat the standard approach, which is called autoregressive generation, on tasks that are genuinely useful in a clinical setting.

Alex: Before we get into the results—can you walk me through what those two approaches actually do differently?

Sam: Sure. Most AI writing tools today work the same basic way: they produce text one word at a time, strictly left to right. Each word is chosen based only on what came before it—like a person typing a sentence from start to finish without ever going back. That's the autoregressive approach. The diffusion model works quite differently. Instead of committing to each word in sequence, it starts with a rough, noisy draft of the entire report and gradually refines the whole thing at once—until a clean, coherent text emerges.

Alex: So one is like a typewriter moving left to right, and the other is more like a sculptor refining a block of marble from all sides at once. Why does that distinction matter for a radiologist specifically?

Sam: It matters because of how doctors actually work. A radiologist might draft a report and deliberately leave a gap in the middle of a sentence—to fill in a specific finding after reviewing a scan more carefully. An autoregressive model struggles with that, because it only knows what came before the gap. It's blind to the text that should come after it. The diffusion model, because it treats the whole report as a single canvas, can see the context before and after the blank space at the same time.

Alex: So it can drop into the middle of a half-finished paragraph and fill in a missing phrase, guided by everything surrounding it—not just what came before.

Sam: Exactly. The researchers call this capability "any-order infill," and it's a meaningful practical advantage for clinical drafting work. But the more important question is whether that flexibility comes at a cost to accuracy.

Alex: Right—a flexible approach isn't much use if the output is less reliable. Did the diffusion model hold up?

Sam: That's the central finding. The study shows the diffusion model is competitive with standard autoregressive models on accuracy—it matched or exceeded the standard approach on medical question-answering tasks. And beyond accuracy, it was notably faster at generating the final report.

Discrete Diffusion Language Models for Interactive Radiology Report Drafting

Max Van Puyvelde, H. Ibrahim Gulluk, Wim Van Criekinge, Olivier Gevaert

Jul 1, 2026·3:43·arxiv·Original Paper

Discrete Diffusion Language ModelsRadiology Report Generation (RRG)Any-order InfillAutoregressive (AR) GenerationDiffusion Language ModelAutoregressive Model

Listen

4 min

Briefing

Research Question

Most medical foundation models currently rely on autoregressive (AR) generation, which produces text sequentially from left to right. This study investigates whether discrete diffusion language models—which generate text by iteratively denoising a fixed token canvas bidirectionally—can serve as a viable, more flexible alternative for radiology report drafting and medical visual question answering (VQA).

Approach

The researchers conducted a controlled comparison between a discrete diffusion model (DiffusionGemma-26B) and its same-size autoregressive sibling (Gemma-4-26B). By using an identical LoRA (Low-Rank Adaptation) recipe and training data, they isolated the generative paradigm as the only variable. They evaluated these models on three medical VQA datasets using a verbosity-robust LLM judge to ensure semantic accuracy, and measured inference speed and the ability to perform any-order infill (filling gaps in text based on context from both sides).

Main Findings

The diffusion model proved to be a competitive medical foundation model, matching or exceeding the performance of the autoregressive sibling across all tested VQA datasets. Notably, the diffusion model achieved this while decoding 3.5–4.4x faster than the AR model. Furthermore, the diffusion model demonstrated a unique capability for any-order infill: because it denoises the entire canvas bidirectionally, it can effectively fill gaps in a report by conditioning on text fragments both before and after the gap. In contrast, the AR model failed to utilize post-gap context, even when provided in the prompt, highlighting a fundamental architectural advantage for interactive clinical drafting.

Why It Matters

Radiology reporting often requires editing or normalizing inconsistent text across different institutions. The ability to perform any-order infill allows radiologists to fix specific report fragments while the model intelligently fills in the surrounding text, a workflow that is inherently difficult for standard autoregressive models. This research suggests that shifting toward diffusion-based architectures could improve both the efficiency and the interactive utility of AI-assisted medical documentation.

Key Terms and Definitions

Discrete Diffusion Language Model — A generative model that creates text by iteratively denoising a fixed-length sequence of tokens rather than predicting them one by one.
Autoregressive (AR) Generation — A standard language modeling approach where each token is generated sequentially based only on the preceding tokens.
Any-Order Infill — A drafting operation where a model fills in missing text at arbitrary positions within a document by conditioning on fixed fragments on both the left and right sides.
Token Canvas — The fixed-length window of tokens that a diffusion model processes and denoises in parallel during generation.
Verbosity-Robust LLM Judge — An evaluation method that uses a separate, powerful LLM to determine if an answer is semantically equivalent to the ground truth, avoiding the limitations of exact-match scoring.

Abstract

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.

Discrete Diffusion Language Models for Interactive Radiology Report Drafting: Research Paper Podcast