LLaDA-MedV: Exploring Large Language Diffusion Models for Biomedical Image Understanding

📰 News

• [2026-06] LLaDA-MedV has been accepted to CVPR 2026.
• [2025-08-06] We uploaded our paper to arXiv and released model weights.

🧠 Introduction

Autoregressive models (ARMs) have long dominated the landscape of biomedical vision-language models (VLMs). Recently, masked diffusion models such as LLaDA have emerged as promising alternatives, yet their application in the biomedical domain remains largely underexplored. To bridge this gap, we introduce LLaDA-MedV, the first large language diffusion model tailored for biomedical image understanding through vision instruction tuning.

LLaDA-MedV achieves relative performance gains of 7.855% over LLaVA-Med and 1.867% over LLaDA-V in the open-ended biomedical visual conversation task, and sets new state-of-the-art accuracy on the closed-form subset of three VQA benchmarks: 84.93% on VQA-RAD, 92.31% on SLAKE, and 95.15% on PathVQA. We further analyze both training and inference behaviors, highlighting the importance of initialization weight selection, fine-tuning strategies, and the interplay between sampling steps and response repetition.

✨ Highlights

• Biomedical diffusion VLM: LLaDA-MedV explores large language diffusion models for biomedical image understanding instead of the more common autoregressive paradigm.
• Strong open-ended chat performance: The model improves biomedical visual conversation quality over LLaVA-Med and LLaDA-V.
• Strong downstream VQA results: LLaDA-MedV achieves strong performance on VQA-RAD, SLAKE, and PathVQA after task-specific fine-tuning.
• Controllable long-form responses: The model can generate longer answers when response length is explicitly controlled, improving informativeness in biomedical dialogue.

🗂️ Repository Overview

This repository currently contains:

• the core multimodal model implementation under llava/
• LLaDA-based biomedical vision-language model extensions
• a single-image VQA inference demo in test_lladamedv_vqa.py
• evaluation helpers such as eval_vlm_chat_gpt_score.py and openai_api.py
• result figures and tables used in the paper

The current repository is strongest on model/inference code and result presentation. The README in its current form indicates that the full training and evaluation release is still being finalized.

🏗️ Repository Structure

LLaDA-MedV/
├── llava/                         # Core multimodal model, serving, and training code
├── images/                        # Figures used in the README and paper presentation
├── test_lladamedv_vqa.py          # Single-image biomedical VQA demo
├── eval_vlm_chat_gpt_score.py     # GPT-based evaluation helper
├── openai_api.py                  # Async OpenAI API wrapper for evaluation
├── ASU Non-Commercial License
└── README.md

📊 Model Performance

Open-End Biomedical Conversation

We adopt the Biomedical Visual Chatbot benchmark to evaluate LLaDA-MedV in a realistic open-ended biomedical conversation setting. As shown below, LLaDA-MedV demonstrates superior performance compared to several baselines, and we provide qualitative generation results for visualization.

Downstream Biomedical VQA

We also evaluate LLaDA-MedV on three biomedical visual question answering benchmarks after fine-tuning on the training sets of VQA-RAD, SLAKE, and PathVQA, following their official splits.

⚙️ Model Details

Training

Please refer to the training implementation of LLaDA-V.

Evaluation

Please refer to our GPT-based evaluation helper in eval_vlm_chat_gpt_score.py. For more detailed evaluation protocols, please refer to LLaVA-Med evaluation.

Weights

We release our model weights to support future research in the community.

🚀 Quick Start

The most direct runnable entry point in this repository is test_lladamedv_vqa.py, which provides a single-image biomedical VQA demo.

python test_lladamedv_vqa.py \
  --image /path/to/your/image.png \
  --question "Please describe this biomedical image in detail." \
  --model-path XZDong123/LLaDA-MedV \
  --llada-v-codebase /path/to/LLaDA-V

Notes:

• --llada-v-codebase is required so the script can import the llava codebase correctly, and it can point to this repository root.
• The default vision tower is google/siglip2-so400m-patch14-384.
• The default model path is XZDong123/LLaDA-MedV.
• CUDA is expected for typical use.

🙏 Acknowledgements

We gratefully acknowledge the authors of the following open-source repositories, which served as valuable references during our implementation:

• LLaDA
• LLaDA-V
• LLaVA-Med

We deeply appreciate their contributions to the research community.

📚 Citation

If you find our work useful in your research, please cite:

@article{dong2025llada,
  title={LLaDA-MedV: Exploring Large Language Diffusion Models for Biomedical Image Understanding},
  author={Dong, Xuanzhao and Zhu, Wenhui and Chen, Xiwen and Wang, Zhipeng and Qiu, Peijie and Tang, Shao and Li, Xin and Wang, Yalin},
  journal={arXiv preprint arXiv:2508.01617},
  year={2025}
}