AI and the World's Hard Problems · Medicine & drug discovery · arXiv · 2026-10-20
What Language Adds, and at What Cost: Four Controlled Probes of Pathology Vision-Language Models
This study tests how much language adds to the vision backbone in pathology vision-language models, using released checkpoints and public data with one input varied at a time. Adding a stage-redacted report embedding improved prognostic prediction slightly, while prompts gave no measurable accuracy gain and slide-level question answering was barely grounded in the image. The authors conclude that language adds real but narrow value and propose minimum controls for future evaluations.