File size: 5,373 Bytes
621894b c1656bf 621894b c1656bf |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 |
---
license: mit
language: en
pipeline_tag: text-generation
tags:
- conversational
- text-generation
- medical
- diagnosis
- agent
- reinforcement-learning
base_model: Qwen3-8B
datasets:
- HealthBench
- MAQuE
- MedQA
- MMLU
paper: 2510.04284
model_name: Doctor-R1
metrics:
- accuracy
---
# Doctor-R1: Mastering Clinical Inquiry with Experiential Agentic Reinforcement Learning
**Doctor-R1** is an AI doctor agent trained to conduct strategic, multi-turn patient inquiries to guide its diagnostic decision-making. Unlike traditional models that excel at static medical QA, Doctor-R1 is designed to master the complete, dynamic consultation process, unifying the two core skills of a human physician: communication and decision-making.
This model is an 8B parameter agent built upon **Qwen3-8B** and fine-tuned using a novel **Experiential Agentic Reinforcement Learning** framework.
## ✨ Key Features
* **Unified Clinical Skills:** The first agent framework to holistically integrate two core clinical skills, **strategic patient inquiry** and **accurate medical decision-making** within a single model.
* **Experiential Reinforcement Learning:** A novel closed-loop framework where the agent learns and improves from an accumulating repository of its own high-quality experiences.
* **Dual-Competency Reward System:** A sophisticated two-tiered reward architecture that separately optimizes for both conversational quality (soft skills) and diagnostic accuracy (hard skills), featuring a "safety-first" veto system.
* **State-of-the-Art Performance:** Outperforms leading open-source models on challenging dynamic benchmarks like HealthBench and MAQuE with high parameter efficiency (8B).
## 🏆 Leaderboards
Doctor-R1 demonstrates state-of-the-art performance among open-source models and surpasses several powerful proprietary models on HealthBench. It demonstrates superior performance on dynamic benchmarks and strong foundational knowledge on static QA tasks.
| Benchmark | Key Metric | Doctor-R1 | Best Open-Source (>=32B) |
| :----------------- | :--------- | :-------: | :----------------------: |
| **HealthBench** | Avg. Score | **36.29** | 33.16 |
| **MAQuE** | Accuracy | **60.00** | 57.00 |
| **MedQA** | Accuracy | **83.50** | 81.50 |
| **MMLU (Medical)** | Accuracy | **85.00** | 84.00 |
The detailed breakdown of **HealthBench Main (Dynamic Consultation)** is as below:
| Model | Avg. Score | Accuracy | Comm. Quality | Context Aware. |
| :------------------------ | :--------: | :-------: | :-----------: | :------------: |
| **GPT-o3** (Proprietary) | 38.91 | 40.31 | 64.78 | 48.09 |
| **Doctor-R1 (8B)** | **36.29** | **37.84** | **64.15** | **49.24** |
| Baichuan-M2-32B | 33.16 | 33.95 | 58.01 | 46.80 |
| Grok-4 (Proprietary) | 33.03 | 37.95 | 61.35 | 45.62 |
| GPT-4.1 (Proprietary) | 31.18 | 34.78 | 60.65 | 44.81 |
| UltraMedical-8B | 22.19 | 25.50 | 57.40 | 40.26 |
| **Base Model (Qwen3-8B)** | 25.13 | 28.57 | 49.35 | 43.00 |
## 👥 Human Evaluation
To validate that our quantitative results align with user experience, we conducted a pairwise human preference evaluation against other leading models. The results show a decisive preference for Doctor-R1, especially in patient-centric metrics.
## 🔬 Ablation Studies
Our ablation studies validate the critical contributions of our framework's key components.
***Impact of Experience Retrieval Mechanism.*** The results show that our full retrieval mechanism with reward and novelty filtering provides a significant performance boost over both a no-experience baseline and a standard similarity-based retrieval, especially in communication skills.
<p align="center">
<img src="assets/radar_exp.jpg" style="width:60%;" />
</p>
***Impact of Patient Agent Scaling.*** We observe a strong, positive correlation between the number of simulated patient interactions during training and the agent's final performance. This validates that our agentic framework effectively learns and improves from a large volume of diverse experiences.
## 📜 Citation
If you find our work useful in your research, please consider citing our paper:
```bibtex
@misc{lai2025doctorr1masteringclinicalinquiry,
title={Doctor-R1: Mastering Clinical Inquiry with Experiential Agentic Reinforcement Learning},
author={Yunghwei Lai and Kaiming Liu and Ziyue Wang and Weizhi Ma and Yang Liu},
year={2025},
eprint={2510.04284},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2510.04284},
}
```
## 💬 Contact & Questions
For collaborations or inquiries, please contact [**[email protected]**](mailto:[email protected]). You’re also welcome to open an issue or join the discussion in this repository, we value your insights and contributions to **Doctor-R1**.
Stay tuned and join our community as we push the boundaries of intelligent healthcare. Together, let’s make medical AI safer, smarter, and more human. 🤝 |