Psyche-R1: Towards Reliable Psychological LLMs through Unified Empathy, Expertise, and Reasoning
[ English | 简体中文 ]
🔥[2025.10.17] We've developed an APP for our model which is available for Windows, Linux, and Android users. Download it here: Psyche-R1-APP.
🔥[2025.8.16] We have released the Chinese Psychological Reasoning LLM Psyche-R1! For model downloads, please click here: MindIntLab/Psyche-R1
Psyche-R1 is a Chinese psychological reasoning LLM that unifies empathy, expertise, and reasoning capabilities. We propose a novel data synthesis pipeline, as illustrated in the figure below. Through processes including data cleaning, question generation, rationale iteration, and empathetic dialogue synthesis, we generated psychology QA pairs with detailed rationales, as well as empathetic dialogue data. Based on this, we utilize multi-LLM selection to filter out "challenging questions" to enhance the model's complex reasoning capabilities, while the remaining data is divided into "non-challenging questions."
We select Qwen2.5-7B-Instruct as the base model. The model first undergoes SFT on massive "non-challenging questions" (including psychological questions and empathetic dialogues) to infuse extensive expertise and empathetic capabilities. Subsequently, the model undergoes GRPO reinforcement learning based on "challenging questions" to further improve the model's complex reasoning abilities.
We present the results on the Psychological Counselor Examination Benchmark (PCEB) which are shown below.
| Model | Case | Moral | Theory | Avg. | Case (QA) | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SMCQ | MMCQ | SMCQ | MMCQ | SMCQ | MMCQ | R-1 | R-L | B-4 | ||||||
| Qwen2.5-7B-Instruct | 47.57 | 31.64 | 47.49 | 87.83 | 59.50 | 71.02 | 78.46 | 42.45 | 55.17 | 57.91 | (64.59) | 20.94 | 11.28 | 1.28 |
| Qwen2.5-72B-Instruct | 46.91 | 40.34 | 53.11 | 90.79 | 70.25 | 78.48 | 82.63 | 47.63 | 59.74 | 63.09 | (68.61) | 21.43 | 12.02 | 1.16 |
| DeepSeek-R1 | 79.25 | 44.25 | 60.86 | 95.39 | 68.99 | 77.95 | 92.19 | 57.60 | 69.41 | 72.95 | (79.18) | 17.65 | 9.19 | 0.94 |
| DeepSeek-R1-70B | 56.30 | 30.72 | 46.95 | 88.16 | 52.53 | 65.66 | 68.01 | 25.64 | 45.63 | 53.56 | (61.79) | 22.77 | 13.23 | 1.16 |
| QwQ-32B | 56.51 | 23.35 | 41.27 | 88.82 | 41.14 | 53.06 | 82.12 | 32.69 | 49.90 | 54.11 | (61.95) | 18.39 | 7.48 | 0.84 |
| Qwen3-235B-A22B | 68.58 | 41.91 | 57.24 | 93.42 | 69.62 | 78.90 | 88.36 | 56.70 | 68.64 | 69.77 | (75.86) | 18.96 | 11.14 | 1.11 |
| GPT-4o | 65.63 | 13.67 | 34.53 | 88.15 | 33.54 | 54.79 | 74.65 | 24.10 | 45.07 | 49.96 | (60.47) | 23.45 | 12.75 | 1.18 |
| Claude3.7-Sonnet | 63.39 | 19.40 | 34.23 | 90.13 | 60.13 | 70.04 | 76.73 | 37.37 | 48.99 | 57.86 | (63.92) | 21.59 | 11.11 | 1.23 |
| EmoLLM | 46.93 | 21.87 | 40.02 | 84.21 | 34.17 | 51.05 | 71.72 | 26.18 | 44.49 | 47.51 | (56.40) | 22.15 | 11.69 | 1.20 |
| PsycoLLM | 55.58 | 35.07 | 42.89 | 88.81 | 69.62 | 74.20 | 72.63 | 48.59 | 54.12 | 61.72 | (64.71) | 24.45 | 17.45 | 2.04 |
| Psyche-R1 | 63.31 | 56.26 | 66.21 | 92.76 | 79.62 | 82.54 | 87.70 | 66.54 | 73.34 | 74.37 | (77.64) | 27.31 | 15.33 | 2.40 |
We also conducted more detailed and comprehensive experiments, including experiments on CPsyExam and PsyDT test sets, demonstrating Psyche-R1's capabilities in psychological examinations and counseling. For detailed experimental results, please refer to the original paper.
- Clone this project locally
git clone https://github.com/MindIntLab-HFUT/Psyche-R1.git- Set up the environment
conda create -n psycher1 python=3.10
conda activate psycher1
pip install -r requirements.txt- Run the Python file run.py
deepspeed --num_gpus=1 run.py- Start interacting Note that for multi-turn dialogue, we recommend you to prompt the model to output the reasoning process enclosed with and tags explicitly (as in run.py).
Model training is based on the LLaMA-Factory and VeRL frameworks.
We also thank the following students for their help with this project, including but not limited to data collection, data processing , and so on (in no particular order): Yuhang Deng, Yiduo Jin, Xiang Li, Yue Liu, Yan Luo, Weidong Wang, Jinming Yu. We also thank Weidong Wang for developing an impressive APP!
If this work is helpful, please kindly cite as:
@article{dai2025psyche,
title={Psyche-R1: Towards Reliable Psychological LLMs through Unified Empathy, Expertise, and Reasoning},
author={Dai, Chongyuan and Hu, Jinpeng and Shi, Hongchang and Li, Zhuo and Yang, Xun and Wang, Meng},
journal={arXiv preprint arXiv:2508.10848},
year={2025}
}