IPAB Workshop - 12/11/2026

Title: Can Vision–Language Models Understand Physical Scale?

Abstract: Vision–language models can describe images and videos, but how well do they understand physical scale? This talk introduces MedVision (https://medvision-vlm.github.io) and its findings on quantitative medical image analysis, then reviews recent work on spatial measurement and temporal reasoning. We examine how models encode and learn to use scale, their current limitations, and what reliable measurement requires.