跳转至

文章背景与核心概要

罕见病的诊断高度依赖稀缺且难以传承的医学专家推理能力，而开箱即用的大语言模型（LLM）在传统基准测试中，仅能在 35.4% 的病例中将正确疾病排在第一位。为了攻克这一难题，本文研究团队推出了名为 liteOdyssey 的智能诊断系统。该系统通过“人类反馈策略迭代”（Policy Iteration with Human Feedback, PIHF）开发，这是一种借鉴自强化学习中广义策略迭代的上下文策略学习方法。

通过将模型的诊断失败案例与医生的纠正意见整合为明确的、受临床医生把关的策略，liteOdyssey 成功将标准大模型转化为高精度的智能体诊断系统，同时确保整个过程受控于临床医生。实验结果表明，在涵盖 722 种罕见病的 1,243 个公开基准测试病例中，liteOdyssey 将正确疾病的排首位准确率从无策略时的 26.5% 大幅提升至 59.3%，并且在未参与策略开发的病例和疾病上表现出极强的泛化与跨模型迁移能力。

Teaching Agentic AI to Learn Expert Reasoning for Rare Disease Diagnosis

arXiv ID: 2606.16149 [cs.AI]
Submitted: 15 June 2026 (v1), last revised 12 August 2026 (v3)

license icon (view license)

arXiv ID: 2606.16149 [cs.AI]
提交时间: 2026年6月15日 (v1)，最后修订于 2026年8月12日 (v3)

license icon (查看许可协议)

Authors

Minh-Ha Nguyen, Erica Gray, Bryce A. Schuler, Kevin W. Byram, Chih-Ting Yang, Fan Ma, Hua Xu, Wu-Chen Su, Chao Yan, Wei-Qi Wei, Adam Wright, Lisa Bastarache, Josh Peterson, Lingyao Li, Siyuan Ma, Undiagnosed Diseases Network, Rizwan Hamid, Thomas A. Cassini, Cathy Shyr

作者

Minh-Ha Nguyen, Erica Gray, Bryce A. Schuler, Kevin W. Byram, Chih-Ting Yang, Fan Ma, Hua Xu, Wu-Chen Su, Chao Yan, Wei-Qi Wei, Adam Wright, Lisa Bastarache, Josh Peterson, Lingyao Li, Siyuan Ma, Undiagnosed Diseases Network, Rizwan Hamid, Thomas A. Cassini, Cathy Shyr

Executive Summary

Diagnosing rare diseases requires specialized expert reasoning that is both scarce and difficult to transfer; off-the-shelf Large Language Models (LLMs) traditionally rank the correct disease first in only 35.4% of benchmark cases.

This paper introduces liteOdyssey, developed through Policy Iteration with Human Feedback (PIHF)—an in-context policy-learning method adapted from generalized policy iteration in reinforcement learning. By consolidating model failures and expert corrections into an explicit, clinician-gated policy, liteOdyssey transforms standard LLMs into highly accurate, agentic diagnostic systems that remain under strict clinician control.

Key Highlights & Results:

  • Improved Accuracy: Across 1,243 public benchmark cases spanning 722 rare diseases, liteOdyssey ranked the correct disease first in 59.3% of cases (compared to 26.5% without the policy).
  • Generalization & Transferability: Achieved nearly identical gains on 1,193 cases and 679 diseases explicitly excluded from policy development. Furthermore, the policy transferred seamlessly without modification across both closed- and open-weight models.
  • Clinical Validation: Tested on 515 patients from the Undiagnosed Diseases Network (UDN), liteOdyssey improved accuracy, with blinded physicians rating its differential diagnoses more often exact and less often unhelpful.

执行摘要

罕见病诊断需要专门的专家推理能力，这种能力既稀缺又难以迁移；在传统基准测试中，开箱即用的大语言模型（LLM）仅在 35.4% 的病例中能将正确的疾病排在首位。

本文介绍了 liteOdyssey，它是通过“人类反馈策略迭代”（Policy Iteration with Human Feedback, PIHF）开发出来的——这是一种借鉴自强化学习中广义策略迭代的上下文策略学习方法。通过将模型的失误和专家的修正整合为明确的、受临床医生把关的策略，liteOdyssey 将标准 LLM 转化为高精度、受严格临床控制的智能体诊断系统。

核心亮点与结果：

  • 准确率提升： 在涵盖 722 种罕见病的 1,243 个公共基准测试病例中，liteOdyssey 将正确疾病排首位的比例提升至 59.3%（无策略时为 26.5%）。
  • 泛化与迁移能力： 在被明确排除在策略开发之外的 1,193 个病例和 679 种疾病上，获得了几乎相同的性能增益。此外，该策略在闭源和开源模型之间实现了无缝、无需修改的迁移。
  • 临床验证： 在来自未诊断疾病网络（UDN）的 515 名患者身上进行了测试，liteOdyssey 提高了准确率，盲法评估的医师对其鉴别诊断的评价更频繁地表现为“精确”、更少表现为“无帮助”。

Abstract

Rare disease diagnosis depends on expert reasoning that is scarce and difficult to transfer; off-the-shelf large language models (LLMs) rank the correct disease first in only 35.4% of benchmark cases. Here we show that this expert reasoning can be converted into a scalable AI capability through a governed learning process rather than model training alone. We developed liteOdyssey through Policy Iteration with Human Feedback (PIHF), an in-context policy-learning method adapted from generalized policy iteration in reinforcement learning, in which model failures and expert corrections consolidate into an explicit, clinician-gated policy that turns an off-the-shelf LLM into an agentic diagnostic system.

We demonstrated that such a policy improved diagnostic accuracy to match the best published systems at a fraction of their deployment footprint, generalized to unseen diseases, transferred across models, and remained under clinician control. Across 1,243 public benchmark cases spanning 722 rare diseases, liteOdyssey ranked the correct disease first in 59.3% of cases versus 26.5% without the policy, with nearly identical gains on the 1,193 cases and 679 diseases excluded from policy development. Ablations showed that gains exceeded automated prompting improvement or source access alone, and the policy transferred without modification across closed- and open-weight models. In 515 Undiagnosed Diseases Network patients, liteOdyssey again improved accuracy, and blinded physicians rated its differentials more often exact and less often unhelpful. Through PIHF, expert reasoning becomes an LLM capability that experts can inspect, revise, and transfer across models.

摘要

罕见病诊断依赖于稀缺且难以传承的专家推理能力；开箱即用的大语言模型（LLM）在基准测试中仅有 35.4% 的概率将正确的疾病排在首位。在此，我们表明这种专家推理能力可以通过受控的学习过程（而不仅仅依赖模型训练）转化为可扩展的 AI 能力。我们通过人类反馈策略迭代（PIHF）开发了 liteOdyssey，这是一种改编自强化学习中广义策略迭代的上下文策略学习方法。在该方法中，模型的失误和专家的修正被汇集成一个明确的、受临床医生把关的策略，从而将开箱即用的 LLM 变成了智能体诊断系统。

我们证明，这种策略将诊断准确率提升至与目前发表的最佳系统相媲美的水准，而其部署开销却只是后者的冰山一角。它能够泛化到未见过的疾病，跨模型迁移，并始终处于临床医生的控制之下。在跨越 722 种罕见病的 1,243 个公共基准测试病例中，liteOdyssey 将正确疾病排首位的比例提高到了 59.3%（不使用该策略时为 26.5%），在被排除在策略开发之外的 1,193 个病例和 679 种疾病上也获得了几乎相同的增益。消融实验表明，其增益超越了单纯的自动化提示词优化或仅依赖源访问，且该策略在闭源和开源模型之间实现了无修改迁移。在 515 名未诊断疾病网络（UDN）患者中，liteOdyssey 再次提高了准确率，盲测医师对其鉴别诊断的评价更常被评为精准，较少被评为无帮助。通过 PIHF，专家推理变成了一种 LLM 能力，专家可以对其进行检查、修改并在不同模型间进行迁移。

Submission History

  • [v1] Mon, 15 Jun 2026, 03:10:17 UTC (233 KB)
  • [v2] Thu, 9 Jul 2026, 08:06:50 UTC (1,632 KB)
  • [v3] Wed, 12 Aug 2026, 04:32:43 UTC (594 KB, current version)

提交历史

  • [v1] 2026年6月15日 周一 03:10:17 UTC (233 KB)
  • [v2] 2026年7月9日 周四 08:06:50 UTC (1,632 KB)
  • [v3] 2026年8月12日 周三 04:32:43 UTC (594 KB，当前版本)

Access & Resources

References & Academic Tools

访问与资源

参考文献与学术工具