Microsoft Research Unveils CARE-X: A Radiology VLM with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement

Microsoft Research has developed CARE-X, a research model for chest X-ray interpretation that combines generative and discriminative capabilities to support a wide range of radiology tasks, from report generation to disease classification and anatomical grounding. Built on a SigLIP2 vision encoder and Phi-4-mini language model, CARE-X adds task-specific auxiliary heads that provide calibrated confidence scores, allowing clinicians to adjust sensitivity-specificity trade-offs. The model is trained in three stages of supervised fine-tuning followed by DAPO reinforcement learning, which optimizes task-specific rewards for clinical reporting, diagnostic accuracy, and spatial grounding. On benchmarks such as MIMIC-CXR, IU-Xray, CheXpert-Plus, and ReXGradient, CARE-X achieves strong performance, and it reaches 94% accuracy on the ReXVQA benchmark, surpassing other publicly reported models. In a separate experiment, the team paired Qwen3-VL-4B-Instruct with deterministic measurement tools to handle quantitative assessments like cardiomegaly and aortic enlargement, which significantly outperformed visual approximation alone; for instance, F1 for cardiomegaly improved from 74.56 to 96.00. The model was validated on real-world Indian clinical data from Narayana Health, including rare ICU pathologies, but it is a research model not approved for clinical use.