稀疏谱 LoRA：面向医学视觉语言模型的路由专家
Sparse Spectral LoRA: Routed Experts for Medical VLMs
把预训练权重按 SVD 分段初始化 MoE 专家，并配一个与全秩全量微调对齐的缩放规则，为视觉语言模型在异构医学数据上做参数高效适配提供了可复用的低秩专家路由方案，可直接迁移到其他多数据集 VLM 微调场景。
原文依据 we initialize each expert from non-overlapping singular value decomposition (SVD) segments of the pretrained weight
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论大型视觉语言模型在通用基准上表现优异，但在医学影像中往往缺乏鲁棒性：异构监督会引发跨数据集干扰，并对数据机制（即监督信号如何混合）敏感。在真实临床流程中，数据与任务按顺序到达，因此朴素的持续训练会进一步导致灾难性遗忘。为应对这些挑战，我们提出 MedQwen，一种参数高效的医学视觉语言模型，它将谱路由混合专家（MoE）与一个有理论依据的缩放规则相结合，在不改变基础架构的前提下，使低秩更新与全秩、全量微调的 MoE 对齐。具体而言，我们用预训练权重中互不重叠的奇异值分解（SVD）片段初始化每个专家，并引入残差补偿与缩放方案，以在分布偏移下实现稳定的专家专门化与一致的路由。在涵盖视觉问答、报告生成、放射影像分类与幻觉缓解的 23 个医学数据集上，MedQwen 取得了强劲而可靠的性能：在零样本分类上以少 339 倍的可训练参数逼近全量微调，并将顺序遗忘降至约 5%，而强基线则退化超过 20%–50%。
英文原文摘要 1323 字符
Large vision-language models (VLMs) excel on general benchmarks but often lack robustness in medical imaging, where heterogeneous supervision induces cross-dataset interference and sensitivity to data regime (i.e., how the supervisory signals are mixed). In realistic clinical workflows, data and tasks arrive sequentially, so naive continual training further leads to catastrophic forgetting. To address these challenges, we propose MedQwen, a parameter-efficient medical VLM that couples a spectrally routed Mixture-of-Experts (MoE) with a theoretically grounded scaling rule that aligns low-rank updates with a full-rank, fully fine-tuned MoE, without changing the base architecture. Concretely, we initialize each expert from non-overlapping singular value decomposition (SVD) segments of the pretrained weight and introduce a residual compensation and scaling scheme to enable stable expert specialization and consistent routing under distribution shift.Across 23 medical datasets covering visual question answering, report generation, radiology classification, and hallucination mitigation, MedQwen achieves strong, reliable performance: it approaches full fine-tuning on zero-shot classification with 339xfewer trainable parameters, and reduces sequential forgetting to ~5% where strong baselines degrade by >20-50%.