Microsoft Research's CARE-X hits 94% on ReXVQA with a 3.8B backbone — and tool-augmented measurement lifts aorta F1 from 39 to 100
Published 2026-08-11, CARE-X is a unified chest X-ray VLM pairing a SigLIP2-so400M vision encoder with a Phi-4-mini-instruct 3.8B language backbone via a lightweight adapter, trained in three stages (vision pre-training, adapter/head training, LoRA) then refined with DAPO-based reinforcement learning and co-trained classification/grounding heads. It ranks first on the ReXVQA leaderboard at 94% overall accuracy as of August 2026 and posts the highest CRIMSON scores across MIMIC-CXR, IU-Xray, CheXpert-Plus and ReXGradient. The tool-augmented measurement result is the striking part: cardiomegaly F1 rises 74.56 → 96.00 and ascending aorta enlargement 39.33 → 100.00 when the model calls measurement tools instead of eyeballing. Microsoft stresses it is a research model, not a product or medical device.
↳ Follow the thread