Microsoft Research has introduced CARE-X, an experimental radiology vision-language model designed to tackle a fundamental problem with medical AI: producing language that sounds clinically convincing is not the same thing as producing information clinicians can trust.
CARE-X combines generative AI with structured predictions, confidence scores, visual grounding and reinforcement learning designed around clinical correctness.
On the ReXVQA benchmark containing 41,007 question-and-answer pairs, Microsoft reports that CARE-X achieved 94% overall accuracy, six percentage points above the next-best publicly reported model as of August 2026.
But the most interesting part of the research may be what happens when the AI stops trying to estimate everything itself.
Microsoft separately tested a vision-language model connected to deterministic measurement tools and found substantial improvements on diagnoses requiring precise measurements.
Radiology Requires More Than Good Text
A radiologist may need AI to perform very different tasks from the same chest X-ray.
One request might involve generating findings for a report. Another might require determining whether a pathology is present, locating an abnormality or checking whether a medical device has been positioned incorrectly.
Those tasks also require different types of certainty.
CARE-X addresses this by combining free-text generation with auxiliary prediction heads capable of producing structured outputs and confidence scores.
That means a single inference pass can generate an explanation while also producing a calibrated prediction that clinicians could theoretically adjust around different sensitivity requirements.
Reinforcement Learning Targets Clinical Accuracy
Microsoft trained CARE-X using supervised fine-tuning followed by DAPO-based reinforcement learning.
Instead of optimizing solely for plausible text, the system uses task-specific rewards covering clinical reporting, diagnostic accuracy and spatial grounding.
CARE-X uses a SigLIP2 vision encoder and Phi-4-mini-instruct language model, with additional classification and grounding components.
Microsoft found that jointly training those structured components also improved the model’s generative performance.
Sometimes AI Should Use a Calculator
A separate experiment may have even broader implications for enterprise AI.
Some radiological findings require actual measurements rather than visual interpretation.
Instead of asking a VLM to approximate those measurements, researchers connected Qwen3-VL-4B-Instruct to deterministic tools capable of identifying landmarks, calculating measurements and evaluating diagnostic thresholds.
Across five measurement-dependent conditions, the tool-assisted approach improved F1 scores by an average of 43.6 percentage points compared with perception alone.
For ascending aorta enlargement, F1 increased from 39.33 to 100 in the reported experiment. For descending aorta enlargement, it increased from 28.57 to 100.
The lesson extends well beyond medicine.
AI models do not necessarily need to perform every task internally. Giving them reliable external tools can sometimes be considerably more effective.
Real-World Data Adds Another Test
CARE-X was also evaluated using retrospective clinical data from Narayana Health in India.
Researchers tested the model on 1,047 de-identified chest radiographs involving five relatively rare high-acuity conditions.
A separate tool-assisted study involving 122 CT-confirmed positive cases reported 94.26% recall, 10.65 percentage points higher than the best perception-only baseline.
Microsoft emphasizes that CARE-X remains a research model. It is not a Microsoft product or medical device, has not received regulatory approval and is not intended for clinical diagnosis, screening or patient care.
The Bottom Line
CARE-X is interesting not simply because Microsoft built another medical AI model.
The research explores a larger question facing AI developers everywhere: when should an AI model reason, and when should it call a specialized tool?
Radiology makes that distinction unusually clear. A model may be very good at understanding an image while still being worse than deterministic software at calculating a measurement.
Combining generative reasoning, structured predictions and external tools could therefore prove more valuable than simply building larger models and expecting them to do everything themselves.
CARE-X remains research, and its retrospective results should not be confused with clinical validation. But the architecture offers a compelling glimpse at how specialized AI systems may evolve: not as all-knowing models, but as orchestrators capable of combining learned intelligence with tools built for precision.