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

From Given to Gathered Evidence. Agentic Learning for Longitudinal Medical Reasoning. By Minye Shao and five others, from Durham University, DAMO Academy at Alibaba Group, and two other institutions. Published on the thirtieth of September. Its code repository is worked out: about fifty-five source files, with a readme.

Foundation models can act as clinical agents when they are given a harness of tools. Yet conventional medical benchmarks test reasoning over evidence that has already been selected for the model. They do not test whether a model can go and find the evidence itself, across clinical records and imaging taken over time. Being handed the right scan and being able to locate it are different skills, and only the first one gets measured.

CASE, short for Clinical Agents for Seeking Evidence, is a series of role-specific agents built around that second skill. It comes with a tool-use harness and an agentic post-training framework for compact vision-language policy models. To train and test it, the authors also introduce a longitudinal multimodal benchmark built on UK Biobank. It holds fifty thousand four hundred and one clinical questions, derived from real-world ICD-10-coded diagnoses of four thousand seven hundred and thirty-nine participants.

Each question opens onto a patient-specific environment, containing clinical context and multi-sequence MRI from a baseline visit and a follow-up visit. The agent decides for itself which visits, organs, modalities, slices and specialist tools to inspect and compare.

Training runs in stages. Supervised fine-tuning first transfers evidence-seeking workflows from fourteen thousand seven hundred and thirty-four interaction trajectories produced by frontier models. Agentic reinforcement learning then follows, on the learner's own interactions with the environment. Finally, privileged on-policy self-distillation and rubric-based feedback from a language model refine how the agent gets from evidence to conclusion, without prescribing which sequence of tools it should use.

The authors report that CASE moves beyond imitating question-answer pairs toward investigation policies that transfer, and that it strengthens evidence-grounded longitudinal reasoning. Under matched evaluation conditions, their agent based on Qwen3-VL-8B achieves relative improvements in answer accuracy of over sixteen percent compared with GPT-5.4, and over ten percent compared with Claude Opus 4.8. Code is to be released.

The shift named in the title is the whole point. A model that is given the evidence is being tested on interpretation. A model that has to gather it is being tested on investigation too, and the benchmark and training pipeline here are built so that a small open model can be trained, and measured, on both.