PathAgentBench: Benchmarking Evidence-Seeking Vision-Language Models on Whole-Slide Pathology Image

TL;DR - PathAgentBench evaluates whether vision-language models can actively find and integrate diagnostic evidence in gigapixel pathology slides. Results expose a major gap between reasoning over curated regions and autonomously locating relevant tissue.

• Covers 1,822 TCGA slides and 17,135 diagnostic paths annotated by ten board-certified pathologists.
• Tests cross-modal matching, diagnostic-region localization, and multi-scale reasoning across 20 models.
• Leading open-weight models exceed 93% multi-scale reasoning accuracy, but the best localization mean IoU remains below 0.09.
• Autonomous evidence discovery deteriorates sharply at higher magnifications, reaching a 0.020 unconditional hit rate.