Study uses counterfactual lesion erasure to curb medical VLM hallucinations

A paper on arXiv proposes counterfactual lesion erasure training so small medical vision-language models report only findings actually present in images, addressing bias where models name findings because the question mentions them.

Read the full story at arXiv cs.AI ↗

Timeline · 1 report

1. 2026-11-11 05:00 · arXiv cs.AI
   Seeing Is Not Believing: Counterfactual Lesion Erasure Teaches a Small Medical Vision-Language Agent to Report Only What the Image Shows