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

Problem

Existing medical benchmarks only evaluate reasoning over evidence that has been selected in advance. In real clinical practice, clinicians must decide for themselves which records and images to look at across multiple visits and gather the evidence, but environments for evaluating and training this ability have been lacking.

Method

Using UK Biobank, the authors built a longitudinal multimodal benchmark of 50,401 multiple-choice clinical questions created from the ICD-10 diagnoses of 4,739 participants. Each question is linked to a patient-specific environment that contains clinical context and multi-sequence MRI (5 organ regions) from baseline and follow-up visits. The agent chooses the visit, organ, modality, slice and specialist tools (Vista3D measurements, MedGemma reports) on its own. Three roles, Radiologist, Anatomist and Consultant, are implemented in a single policy model. First, Teacher (frontier model) trajectories are quality-filtered and used for rejection-sampling SFT of Qwen3-VL-8B-Instruct. Next, GRPO-based agentic reinforcement learning is run through self-interaction. On top of the outcome reward, this stage adds on-policy self-distillation (OPSD), which penalizes divergence from a privileged view that knows the correct answer, and rubric feedback from an external LLM judge.

Results

• The authors report that under the same evaluation conditions, the answer accuracy of the Qwen3-VL-8B-based CASE agent was more than 16% higher than GPT-5.4 and more than 10% higher than Claude Opus 4.8, in relative terms.
• The benchmark consists of 50,401 questions (of which 4,850 questions from 455 participants are held out for evaluation), 7 clinical families and 10 task/format strata.
• The authors report that OPSD reduces a pattern of repeatedly delaying conclusions that appears in the middle and later stages of reinforcement learning. They also state that controlled ablations isolate the contribution of each component.

Limitations

The data comes from a single cohort, UK Biobank, and the ground-truth labels come from ICD-10 diagnosis records. The questions are multiple-choice. During reinforcement learning, tool outputs are replayed from precomputed results. The judge used to select SFT data sees only text, so it cannot verify whether visual findings are valid. Within the excerpted text, per-stratum accuracy and ablation numbers could not be confirmed.

Why it matters

The work shifts medical AI evaluation from "answering from given evidence" to "finding the needed evidence and then answering." It shows that a small open model can surpass frontier models through interactive training, which has implications for developing lightweight agents suited to clinical settings.

Key terms

Longitudinal reasoning

Reasoning that compares a patient's records and images from multiple time points and makes judgments based on changes over time.

GRPO

A reinforcement learning algorithm that samples multiple trajectories from the same prompt and updates the policy using each trajectory's reward relative to the group average.

Privileged on-policy self-distillation (OPSD)

A technique that uses the output distribution of the same model, when it also sees the correct answer, as a reference, and reflects in the reward how well the policy reaches conclusions in trajectories it generates itself.

RLAIF (rubric-based LLM feedback)

An approach in which an external LLM judge, instead of a human, scores reasoning quality according to a rubric, and the score is used as the reinforcement learning reward.

Original

• arXiv abstract
• PDF
• HTML full text
• Code
• Project
• Hugging Face

Authors: Minye Shao, Chaohui Yu, Yixuan Wu, Fan Wang, Ling Shao, Yang Long · Posted 2026-09-30 · License arXiv 비독점 배포 라이선스 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.