CASE

CASE: From Given to Gathered Evidence

Clinical reasoning is not merely about interpreting evidence, but also about knowing what evidence to seek. Existing medical AI benchmarks primarily evaluate reasoning over preselected information, overlooking the ability to autonomously gather, compare, and synthesize evidence across clinical records and longitudinal imaging.

We introduce CASE (Clinical Agents for Seeking Evidence), shifting medical AI from passive question answering to active, evidence-grounded clinical investigation.

Key Contributions

• Longitudinal Multimodal Benchmark. A UK Biobank-based benchmark comprising 50,401 clinical questions from 4,739 participants, integrating clinical records and multi-sequence MRI across baseline and follow-up visits.

• Autonomous Evidence-Seeking Agents. A role-conditioned, multi-tool framework enabling compact vision-language models to autonomously select visits, modalities, anatomical structures, image slices, and specialist tools for evidence collection and cross-visit reasoning.

• Beyond Teacher Imitation. An agentic post-training framework integrating Agentic RL, On-Policy Self-Distillation (OPSD), and RLAIF to improve evidence-to-conclusion reasoning while preserving autonomous tool-use decisions.