Feed

Pathology vision-language models (VLMs) are promising for building the clinical decision support systems. However, a key barrier to real-world clinical deployment lies in the lack of rigorous and clinically meaningful model evaluation. Existing pathology visual question answering (VQA) benchmarks exhibit the weak alignment of vision-language information, insufficient support for knowledge-intensive reasoning, and a narrow disease spectrum, thereby undermining the ability to provide clinical-level evaluation of model performance. In this work, we introduce OmniPathoVQA, a new pathology VQA benchmark offering a broad disease coverage, spanning all human anatomical systems, and thousands of disease entities. To enhance well-organized visual-language alignment, we leverage pathology educational materials and design a fine-grained extraction pipeline for linking pathology images with the correct knowledge. To examine the knowledge depth for pathological reasoning, the hard-version questions are designed based on microanatomic features underlying similar diseases. A comprehensive evaluation of eighteen VLMs reveals that closed-source VLMs attain higher scores compared to open-source counterparts, yet their performance drops dramatically in difficult questions. We find that strong general reasoning ability is a crucial catalyst for the successful fine-tuning and pathology feature interpretation. Our study establishes a rigorous standard for assessing high-level reasoning and guiding the rapid development of clinical pathology VLMs.