[ACL 2026] Repositories for our paper: Psyche-r1: Towards reliable psychological llms through unified empathy, expertise, and reasoning
We propose the first Chinese psychological reasoning LLM that unifies empathy, expertise, and reasoning.
This model is a fine-tuned version of Qwen/Qwen2.5-7B-Instruct on our proposed dataset encompassing psychological questions paired with detailed rationales, and empathetic single-turn dialogues.
We conduct a hybrid training strategy, including SFT and GRPO training. We will present detailed training hyperparameters later.
It achieves comparable performance to DeepSeek-R1 on several psychology benchmarks, including psychology counselor examination benchmark (PCEB) proposed by Hu et al. (2024), and CPsyExam test set proposed by Zhao et al. (2024). It also demonstates better performance in empathy on SoulChat2.0 test set (Xie et al. 2025).
Training procedure
SFT Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-05
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- gradient_accumulation_steps: 16
- total_train_batch_size: 256
- total_eval_batch_size: 8
- optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 2.0
GRPO Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-06
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- total_train_batch_size: 128
- total_eval_batch_size: 8
- ppo_mini_batch_size: 32
- ppo_micro_batch_size_per_gpu: 20
- kl_loss_coef: 0.001
- lr_scheduler_warmup_steps: 10
- num_epochs: 2.0
Usage
For quick start, please see MindIntLab-HFUT/Psyche-R1 on GitHub.
Citation
If this work is helpful, please kindly cite as:
@misc{dai2025psycher1reliablepsychologicalllms,
title={Psyche-R1: Towards Reliable Psychological LLMs through Unified Empathy, Expertise, and Reasoning},
author={Chongyuan Dai and Jinpeng Hu and Hongchang Shi and Zhuo Li and Xun Yang and Meng Wang},
year={2025},
eprint={2508.10848},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2508.10848},
}