BaT: Towards Self-Evolving Medical Research Agent with Stage Rubrics

2026-08-17

AI Summary

This study addresses the scarcity of expert trajectories and the loss of stage-wise diagnostic information during post-training for medical imaging agents. We propose Benchmark-as-Teacher (BaT), a recursive self-improvement framework that transforms benchmarks into teacher signals via an asynchronous Stage Bank data pipeline and a BiCuRL bi-level curriculum reinforcement learning algorithm. This approach leverages stage-specific scoring criteria to validate and guide continuous policy evolution. Experimental results demonstrate that BaT-9B achieves a total score of 79.6 on AutoMedBench-Lite, doubling baseline performance and surpassing Claude Opus 4.6. These findings confirm that BaT effectively enables long-horizon workflow automation and self-improvement in medical research agents by mitigating data limitations and preserving critical diagnostic context throughout the training process.

Abstract

Long-horizon agents are beginning to automate complete workflows that produce code, reports, and research artifacts. Medical imaging workflows are multi-stage and data-sensitive, while expert trajectories remain scarce and difficult to share. Structured benchmarks can localize failures through stage-level rubrics, but standard post-training discards these diagnostics before the next training round. We present Benchmark-as-Teacher (BaT), a recursive self-improvement system for agent post-training. BaT contains two linked components: the asynchronous Stage Bank data pipeline and BiCuRL (Bilevel Curriculum Reinforcement Learning), its self-improving post-training method. Stage Bank synthesizes content-isolated training states outside the policy-update loop. BiCuRL uses a fixed held-out evaluation to select the next stage curriculum, verifies rollouts with task rubrics, updates the policy with GRPO, and returns the candidate checkpoint to evaluation. On AutoMedBench-Lite, BaT-4B and BaT-9B more than double the Overall scores of their Qwen Instruct baselines. BaT-9B Agent reaches 79.6 Overall, exceeding Claude Opus 4.6 with Claude Code at 77.5.

Problem

• Medical imaging workflows
• Long-horizon agents
• Stage rubrics
• Self-evolving agent
• Post-training

Innovation

• Benchmark-as-Teacher
• BiCuRL
• Stage Bank
• Self-Evolving Agent
• GRPO