From Given to Gathered Evidence: Agentic Learning for Longitudinal Medical Reasoning
CASE trains compact medical agents to gather longitudinal evidence, with an 8B model outperforming larger frontier systems.
The paper proposes CASE, role-specific clinical agents that autonomously gather evidence from longitudinal records and multi-sequence MRI rather than reasoning over preselected inputs. A UK Biobank benchmark contains 50,401 questions tied to ICD-10 diagnoses of 4,739 participants, with patient-specific visits, organs, modalities, and slices. Supervised fine-tuning on 14,734 frontier-model trajectories is followed by agentic reinforcement learning, privileged on-policy self-distillation, and rubric-based feedback. Under matched evaluation, a Qwen3-VL-8B agent reports more than 16% and 10% relative answer-accuracy gains over GPT-5.4 and Claude Opus 4.8.
- CASE trains compact vision-language agents to seek evidence across records and MRI.
- Benchmark has 50,401 questions from 4,739 UK Biobank participants.
- SFT uses 14,734 frontier trajectories, then agentic reinforcement learning.
- Qwen3-VL-8B agent gains over 16% and 10% relative accuracy versus GPT-5.4 and Claude Opus 4.8.
- Organizations
- UK Biobank
- AI models
- Qwen3-VL-8BGPT-5.4Claude Opus 4.8
From the original article
1 min read at huggingface.co
Foundation models can serve as clinical agents through tool-use harnesses. However, conventional medical benchmarks assess reasoning over preselected evidence rather than the ability to seek it across clinical records and longitudinal imaging.
By Minye Shao, Chaohui Yu, Yixuan Wu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.