MIT's ModaLens finds report availability sharply reduces medical VLM image sensitivity

MIT · hf · 2026-09-15

MIT released ModaLens on Hugging Face, a tool for measuring image sensitivity in report-conditioned medical VLMs.

A paired image-swap audit shows that when reports are available in radiology QA, VLMs become substantially less sensitive to image changes—suggesting overreliance on text over the actual imaging.

Original post →

More from Research

Research channel →