BaT: Towards Self-Evolving Medical Research Agent with Stage Rubrics
| Source: arXiv AI
Tags: medical AI, reinforcement learning, self-improvement, medical imaging, AutoMedBench, BiCuRL, autonomous agents
BaT-9B, trained with BiCuRL self-improvement on medical imaging tasks, scores 79.6 on AutoMedBench-Lite — outperforming Claude Opus 4.6 with Claude Code (77.5) at a fraction of the model size, via recursive curriculum-based RL post-training without new expert annotations.
Details
Benchmark-as-Teacher (BaT) is a recursive self-improvement system for medical imaging AI agents. It combines Stage Bank — an asynchronous data pipeline that synthesizes training states while keeping content isolated from the policy update loop — with BiCuRL (Bilevel Curriculum Reinforcement Learning), which uses a fixed held-out evaluation to select the next curriculum stage, 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 reaches 79.6 Overall — exceeding Claude Opus 4.6 with Claude Code at 77.5. A 9B specialized model trained with self-improvement outperforming a frontier closed model with agentic tools on a medical benchmark is a notable result. The Stage Bank design isolates training data synthesis from the policy update loop, a key feature for long-horizon medical workflows where expert trajectories are scarce and data-sensitive. The framework supports recursive improvement without requiring new expert annotations per round. As with all benchmark results, clinical deployment would require separate validation.