The authors propose CASE, a set of role-specific clinical agents with a tool-use harness and agentic post-training framework for compact vision-language models. They also introduce a longitudinal multimodal benchmark from UK Biobank with 50,401 clinical questions, where agents autonomously select which visits, organs, modalities, and tools to inspect. Using supervised fine-tuning on frontier-model trajectories followed by with self-distillation and rubric feedback, a Qwen3-VL-8B based agent achieved over 16% and 10% relative accuracy gains over GPT-5.4 and Claude Opus 4.8 under matched conditions. This shows compact models can learn effective evidence-seeking policies beyond simple question-answer imitation.
Clinical agents learn to gather evidence across longitudinal MRI records
CASE trains a compact vision-language model to select and synthesize relevant clinical evidence rather than answer from preselected inputs.
Chinese Tech
Minye Shao · Chaohui Yu · Yixuan Wu · Fan Wang · Ling Shao · Yang Long
Durham University · DAMO Academy, Alibaba Group · Hupan Laboratory · University of the Chinese Academy of Sciences
Research Digest··2 min read
In one line
CASE agents achieve over 16% and 10% relative improvements in answer accuracy over GPT-5.4 and Claude Opus 4.8 on longitudinal medical reasoning.
Before you read this
Three ideas this paper builds onevidence retrievalPulling specific, relevant information from a large corpus to support an answer, decision, or reasoning step.115 papers →agentic reinforcement learningTraining approach that lets LLM-based agents learn multi-step tool use and planning via reinforcement learning, often with dense or per-tool rewards.124 papers →agent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.186 papers →
Read it as
Shao and colleagues built an interactive benchmark in which agents must investigate patient records and MRI scans from multiple visits before answering clinical questions.
The authors created 50,401 questions from ICD-10-coded diagnoses for 4,739 UK Biobank participants.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.