From Given to Gathered Evidence: Agentic Learning for Longitudinal Medical Reasoning

CASE introduces role-specific clinical agents and an agentic post-training framework designed to improve evidence retrieval in vision-language models. The framework utilizes a longitudinal multimodal benchmark from the UK Biobank, featuring 50,401 questions tied to real-world ICD-10 diagnoses and multi-sequence MRI sequences across baseline and follow-up visits.