LLaDA-MedV：探索用于生物医学图像理解的大语言扩散模型
LLaDA-MedV: Exploring Large Language Diffusion Models for Biomedical Image Understanding
将掩码扩散语言模型（LLaDA）通过视觉指令微调迁移到生物医学图像理解，为视觉-语言建模提供了一条非自回归的替代路径；其关于初始化权重选择、微调策略以及采样步数与响应重复之间权衡的分析，对任何尝试用扩散式语言模型做多模态生成的视觉研究者都有直接参考价值。
原文依据 we introduce LLaDA-MedV, the first large language diffusion model tailored for biomedical image understanding through vision instruction tuning
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论自回归模型（ARMs）长期主导生物医学视觉-语言模型（VLMs）领域。近期，LLaDA 等掩码扩散模型作为有前景的替代方案出现，但其在生物医学领域的应用仍鲜有探索。为填补这一空白，我们提出 LLaDA-MedV，这是首个通过视觉指令微调、面向生物医学图像理解的大语言扩散模型。在开放式生物医学视觉对话任务中，LLaDA-MedV 相较 LLaVA-Med 取得 7.855% 的相对性能提升，相较 LLaDA-V 提升 1.867%，并在三个 VQA 基准的闭式子集上刷新最优准确率：VQA-RAD 84.93%、SLAKE 92.31%、PathVQA 95.15%。此外，与 LLaVA-Med 的详细对比表明，LLaDA-MedV 能够通过显式控制响应长度生成合理更长的回答，从而产生信息量更丰富的输出。我们还对训练与推理阶段进行了深入分析，强调了初始化权重选择、微调策略以及采样步数与响应重复之间相互作用的关键作用。代码与模型权重已发布于 https://github.com/LLM-VLM-GSL/LLaDA-MedV。
英文原文摘要 1275 字符
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. Furthermore, a detailed comparison with LLaVA-Med suggests that LLaDA-MedV is capable of generating reasonably longer responses by explicitly controlling response length, which can lead to more informative outputs. We also conduct an in-depth analysis of both the training and inference stages, highlighting the critical roles of initialization weight selection, fine-tuning strategies, and the interplay between sampling steps and response repetition. The code and model weight is released at https://github.com/LLM-VLM-GSL/LLaDA-MedV.