跳转至

文章背景与核心概要

长周期人工智能智能体在自动化复杂的、多阶段工作流（如生成代码、研究报告和科学成果）方面展现出越来越强的能力。然而，医学影像工作流具有独特的挑战：它们是多阶段的、高度数据敏感的，并且严重依赖稀缺且难以共享的专家操作轨迹。

尽管结构化基准可以通过阶段级评分准则来定位故障，但标准的后训练协议却在训练周期之间丢弃了这些关键的诊断信息。为了弥合这一差距，本文引入了 Benchmark-as-Teacher (BaT)，这是一种专为智能体后训练设计的递归式自我提升系统。BaT 实现了最先进的性能，其中 BaT-9B 模型使基线指令模型的性能翻倍有余，甚至超越了 Claude Opus 4.6 等先进的专有模型。

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

View PDF | HTML Version | arXiv:2608.16211 [cs.AI]

Executive Summary

长周期 AI 智能体正日益具备自动化完整工作流的能力，能够产出代码、报告和研究成果。然而，医学影像工作流具有多阶段、数据敏感的特点，同时专家轨迹十分稀缺且难以共享。结构化基准可以通过阶段级评分准则定位失败原因，但标准的后训练过程却在下一轮训练开始前丢弃了这些诊断信息。

Long-horizon AI agents are increasingly capable of automating complex, multi-stage workflows such as generating code, research reports, and scientific artifacts. However, medical imaging workflows present distinct challenges: they are multi-stage, highly data-sensitive, and rely on expert trajectories that are scarce and difficult to share.

While structured benchmarks can localize failures using stage-level rubrics, standard post-training protocols discard these critical diagnostics between training cycles. To bridge this gap, this paper introduces Benchmark-as-Teacher (BaT), a recursive self-improvement system designed for agent post-training. BaT achieves state-of-the-art performance, with BaT-9B more than doubling baseline instruct models and outperforming advanced proprietary models like Claude Opus 4.6.

为了弥补这一差距，本文提出了 Benchmark-as-Teacher (BaT)，这是一种用于智能体后训练的递归式自我提升系统。BaT 包含两个相互链接的组件： 1. 异步阶段库（Stage Bank）数据流 2. 双层课程强化学习（BiCuRL, Bilevel Curriculum Reinforcement Learning），即其自我提升的后训练方法。

We present Benchmark-as-Teacher (BaT), a recursive self-improvement system for agent post-training. BaT contains two linked components: 1. The asynchronous Stage Bank data pipeline 2. BiCuRL (Bilevel Curriculum Reinforcement Learning), its self-improving post-training method.

BaT 的工作原理

  • Stage Bank 在策略更新循环之外合成内容隔离的训练状态。
  • BiCuRL 利用固定的留出评估来选择下一个阶段课程，通过任务评分准则验证模型输出，使用 GRPO（群组相对策略优化，Group Relative Policy Optimization） 更新策略，并将候选检查点返回给评估环节。

How BaT Works

  • 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 (Group Relative Policy Optimization), and returns the candidate checkpoint to evaluation.

核心结果

  • 在 AutoMedBench-Lite 上进行测试，BaT-4B 和 BaT-9B 的总体得分较其 Qwen Instruct 基线翻倍有余。
  • BaT-9B 智能体取得了 79.6 的总体得分，超越了 Claude Opus 4.6 (配合 Claude Code)（得分为 77.5）。

Key Results

  • Tested on AutoMedBench-Lite, BaT-4B and BaT-9B more than double the Overall scores of their Qwen Instruct baselines.
  • BaT-9B Agent achieves an Overall score of 79.6, outperforming Claude Opus 4.6 with Claude Code (which scores 77.5).

Metadata

  • Title: BaT: Towards Self-Evolving Medical Research Agent with Stage Rubrics
  • Authors: Junqi Liu, Yufan He, Yexiao He, Pengfei Guo, Dong Yang, Andriy Myronenko, Can Zhao, Hanrong Ye, Tianhao Qi, Yuyin Zhou, Daguang Xu, Yucheng Tang
  • Submitted On: August 17, 2026
  • Primary Subject: Artificial Intelligence (cs.AI)
  • DOI: 10.48550/arXiv.2608.16211

Abstract

长周期智能体开始自动化生成代码、报告和研究成果的完整工作流。医学影像工作流具有多阶段和数据敏感的特点，而专家轨迹仍然稀缺且难以共享。结构化基准可以通过阶段级评分准则定位失败原因，但标准的后训练会在下一轮训练之前丢弃这些诊断信息。

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.

我们提出了 Benchmark-as-Teacher (BaT)，这是一个用于智能体后训练的递归自我提升系统。BaT 包含两个相互链接的组件： 1. 异步 Stage Bank 数据管道 2. BiCuRL（双层课程强化学习），即其自我提升的后训练方法。

We present Benchmark-as-Teacher (BaT), a recursive self-improvement system for agent post-training. BaT contains two linked components: 1. The asynchronous Stage Bank data pipeline 2. BiCuRL (Bilevel Curriculum Reinforcement Learning), its self-improving post-training method.

BaT 的工作原理

  • Stage Bank 在策略更新循环之外合成内容隔离的训练状态。
  • BiCuRL 使用固定的留出评估来选择下一个阶段课程，用任务评分准则验证输出，通过 GRPO 更新策略，并将候选检查点返回给评估。

How BaT Works

  • 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 (Group Relative Policy Optimization), and returns the candidate checkpoint to evaluation.

关键结果

  • 在 AutoMedBench-Lite 上测试显示，BaT-4B 和 BaT-9B 的总体得分超过了其 Qwen Instruct 基线的两倍。
  • BaT-9B 智能体 的总体得分达到了 79.6，超越了 Claude Opus 4.6 (配合 Claude Code)（得分为 77.5）。

Key Results

  • Tested on AutoMedBench-Lite, BaT-4B and BaT-9B more than double the Overall scores of their Qwen Instruct baselines.
  • BaT-9B Agent achieves an Overall score of 79.6, outperforming Claude Opus 4.6 with Claude Code (which scores 77.5).