Whole-slide images (WSIs) pose severe challenges for vision‑language reasoning because diagnostically relevant morphology is sparse, heterogeneous, and spread across gigapixel‑scale images at multiple spatial resolutions. Existing WSI models either aggregate slide‑level features or actively acquire evidence, yet the retained information is often hard to access semantically while preserving its link to the original visual evidence. SlideBank introduces a training‑free framework that represents each WSI as a persistent, concept‑indexed, and spatially grounded evidence bank. The system first performs question‑independent coarse‑to‑fine exploration to locate informative regions and generate multi‑scale views, then converts these views into explicit morphological observations and grounds pathology signals to their supporting patches and WSI coordinates. During inference, questions are routed to relevant signals and evidence scales, and the linked global, regional, and patch evidence is integrated through confidence‑based cross‑level consensus. Experiments on WSI‑VQA and SlideBench‑BCNB demonstrate that with Patho‑R1 SlideBank reaches 52.77% on WSI‑VQA, and with Quilt‑LLaVA it achieves 50.92% average accuracy on SlideBench‑BCNB, while structured signal‑guided retrieval consistently outperforms random evidence sampling. Reusing the same bank across repeated queries yields over 99% rephrasing consistency and substantially reduces amortized inference cost via persistent evidence reuse.
Review