CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement
CARE-X tries to make chest X-ray VLMs useful beyond report prose by adding classification, localization, and measurement workflows.
The paper says CARE-X pairs a generative radiology model with auxiliary heads for finding classification and spatial grounding. Its reward-aligned training targets report generation, VQA, and grounding metrics, with reported state-of-the-art results on most report-generation measures and 94.0% accuracy on ReXVQA. For measurement-dependent diagnoses, the authors connect Qwen3-VL-4B-Instruct to deterministic measurement tools, reporting a 43.6-point average F1 gain over perception-only baselines.
ArXiv · AI/CL/LG's note
The paper says CARE-X pairs a generative radiology model with auxiliary heads for finding classification and spatial grounding. Its reward-aligned training targets report generation, VQA, and grounding metrics, with reported state-of-the-art results on most report-generation measures and 94.0% accuracy on ReXVQA. For measurement-dependent diagnoses, the authors connect Qwen3-VL-4B-Instruct to deterministic measurement tools, reporting a 43.6-point average F1 gain over perception-only baselines.
ArXiv · AI/CL/LG's note
score 5