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

За 30 секунд

Что произошло

PathAgentBench benchmark evaluates vision-language models on whole-slide pathology images across evidence-seeking tasks using 1,822 annotated TCGA slides.

Почему это важно

Matters for engineers building AI systems for digital pathology who need to assess model performance on realistic gigapixel image analysis workflows.

На что обратить внимание

Models excel at reasoning over curated evidence but struggle with finding relevant regions directly in slides, with hit rates dropping sharply at higher magnifications.