Counterfactual training makes a small medical model check the image
A 4B-parameter medical vision-language model cut false claims about erased lesions to 9.6%, versus 41-62% for nine frontier models, without a significant accuracy loss on real lesions. Training showed the model an image, an inpainted version with the lesion removed, and a control edit elsewhere, teaching it to answer from visual evidence instead of the question's wording. The result is promising but not deployment-ready: it missed all 48 pneumothorax films in a fresh external test.