From Given to Gathered Evidence: Agentic Learning for Longitudinal Medical Reasoning
CASE trains compact clinical agents to find their own evidence across patient history and MRI, not just answer from supplied context.
The paper introduces a UK Biobank-based benchmark with 50,401 questions tied to ICD-10-coded diagnoses from 4,739 participants. Each question gives the agent a patient-specific environment spanning clinical context plus baseline and follow-up multi-sequence MRI. The authors train agents from frontier-model evidence-seeking trajectories, then refine them through reinforcement learning, self-distillation, and rubric-based LLM feedback. In matched tests, their Qwen3-VL-8B agent reports relative accuracy gains of more than 16% over GPT-5.4 and 10% over Claude Opus 4.8. HF Daily Papers' note
The paper introduces a UK Biobank-based benchmark with 50,401 questions tied to ICD-10-coded diagnoses from 4,739 participants. Each question gives the agent a patient-specific environment spanning clinical context plus baseline and follow-up multi-sequence MRI. The authors train agents from frontier-model evidence-seeking trajectories, then refine them through reinforcement learning, self-distillation, and rubric-based LLM feedback. In matched tests, their Qwen3-VL-8B agent reports relative accuracy gains of more than 16% over GPT-5.4 and 10% over Claude Opus 4.8. HF Daily Papers' note
score 6