Discrete Diffusion Language Models for Interactive Radiology Report Drafting

2026-07-01

AI Summary

This work proposes the first application of a discrete diffusion language model with a Mixture-of-Experts architecture—DiffusionGemma-26B—to medical imaging report generation. By incorporating LoRA-based fine-tuning and activating only 3.8B parameters, the model achieves performance on par with or superior to autoregressive counterparts of comparable scale across multiple medical visual question answering benchmarks. Unlike conventional autoregressive models, which constrain editing to sequential generation, this approach enables any-order text infilling, substantially enhancing human-AI collaborative report drafting efficiency. Moreover, it attains a 3.5–4.4× speedup in inference latency while demonstrating competitive capabilities relative to state-of-the-art vision-language models.

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.

Problem

• diffusion language models
• interactive radiology report drafting
• any-order infill
• autoregressive generation
• medical foundation models

Innovation

• Discrete Diffusion Language Models
• Any-order Infill
• Interactive Radiology Report Drafting
• Mixture-of-Experts
• Non-autoregressive Generation