Microsoft's CARE-X hits 94% on ReXVQA with a 3.8B backbone, and tool use lifts aorta F1 from 39 to 100.

SigLIP2-so400M vision encoder plus Phi-4-mini-instruct via a lightweight adapter, three-stage training, DAPO-based RL refinement, co-trained classification and grounding heads (Microsoft Research). First on the ReXVQA leaderboard at 94% as of August 2026. The number worth staring at: cardiomegaly F1 goes 74.56 → 96.00 and ascending aorta enlargement 39.33 → 100.00 when the model calls measurement tools instead of eyeballing. Research model, not a product, not a medical device. But "give the model a ruler" beating "make the model better at estimating" is a pattern that transfers everywhere.