FOLTMed reaches 85.4% accuracy across 42 medical image benchmarks

Researchers introduced FOLTMed, a medical image model trained on more than one million visual question-answer pairs assembled from de-identified images and clinician commentary. In preprint tests, the model achieved 85.4% macro accuracy across 42 medical visual-question-answering benchmarks. The researchers used clinician-in-the-loop verification to build the ThoughtMed-1M dataset and reported that FOLTMed exceeded other state-of-the-art models by 3% to 5% on factuality and similarity measures for its test set. The results remain benchmark findings from a preprint, not evidence of clinical deployment or patient outcomes.