代理訓練與醫療 AI
BaT Schedules Reinforcement Learning Around Agent Failure Stages, Raising a 9B Policy’s AutoMedBench-Lite Score From 19.9 to 53.4
Benchmark-as-Teacher turns stage-level evaluation scores directly into the curriculum for the next training round, while using isolated synthetic data to prevent test answers from leaking into updates. The paper reports that its 9B policy substantially outperforms standard GRPO, but the full training dataset has not yet been released, and cross-system rankings mix different execution frameworks.
Medical research agents typically need to plan, configure tools, validate data, run inference, and submit results in sequence. If an entire trajectory receives only a single terminal reward, it is difficult for the model to determine which step needs improvement. Benchmark-as-Teacher (BaT) turns this kind of structured evaluation into a training controller: a fixed held-out test set first identifies weaknesses at each stage, an outer-loop BiCuRL curriculum selects the target for the next round, and an inner loop updates the policy with GRPO. Candidate checkpoints are retained only if they pass reevaluation.
Training data is supplied by an asynchronous Stage Bank. The team created 20,299 candidate states that can be scored against rubrics, divided into weakness-targeted S-target data, S-mix data covering the remaining stages, and E2E data for full-workflow agents. Task identities, answers, paths, reports, and original trajectories are prohibited from crossing the evaluation boundary. Each round also records the data mixture, reward version, and checkpoint, allowing rollbacks when performance regresses. This approach is closer to designing a curriculum around defects in a software workflow than simply increasing the number of rollouts.
Across seven AutoMedBench-Lite task types, with ten reruns per task, Qwen3.5-4B’s Overall score increased from 6.1 to 22.9, while the 9B model rose from 19.9 to 53.4. Terminal-reward GRPO reached 31.9 for the same 9B model. The full BaT-9B Agent achieved 79.6 when paired with a fixed OpenHands environment, but that figure is a system-level comparison and is not directly interchangeable with the standalone policy score.
Engineers can currently inspect the MIT-licensed code, VeRL/Slime training interfaces, and vLLM/SGLang inference interfaces. However, the repository was only recently created, and the full training data, synthetic prompts, and leakage checker are still marked for future release. The models compared in the paper also use different agent harnesses. Key next steps are to determine whether the data can be fully reconstructed, whether the gains persist in fixed-compute ablations, and whether repeatedly training against evaluation weaknesses gradually leads to overfitting to the scorer.