(ACL-2025) Dolphin: Moving Towards Closed-loop Auto-research through Thinking, Practice, and Feedback

[ Paper 📓 ] [ Project Page 🚀 ]

🔥 News

• 2025.06:  🎉🎉 Codes have been released.

• 2025.05:  🎉🎉 Please checkout our latest work NovelSeek on arXiv, Github, and Huggingface.

• 2025.05:  🎉🎉 Congratulations: Dolphin was accepted by ACL-2025 main conference.

• 2025.04:  🎉🎉 We release the paper of Dolphin, a closed-loop LLM-driven framework to enhance the automation level of scientific research

Introduction

Recent works demonstrate that various AI-assisted research methods can largely improve research efficiency by improving data analysis, accelerating computation, and fostering novel ideageneration. To further move towards the ultimate goal (i.e., automatic scientific research), we introduce DOLPHIN, a closed-loop LLM-driven framework to enhance the automation level of scientific research. DOLPHIN first generates novel ideas based on feedback from previous experiments and relevant papers ranked by the topic and task attributes. Then, the generated ideas can be implemented using a code template refined and debugged with the designed exception-traceback-guided local code structure. Finally, DOLPHIN automatically analyzes the results of each idea and feeds the results back to the next round of idea generation.

The usage of Dolphin 🐬

• Propose novel ideas and implement code automatically

  Task	Point Classfication	Image Classification	Sentiment Classification
  Baseline	91.0%	81.2%	91.0%
  Dolphin	93.9%	82.0%	92.5%

• Perform version updates for technologies or code

  Task	Code Source	Previous Score	Current Score
  Detecting insults in social commentary	AIDE	81.0	84.7
  Tabular playground series dec 2021	Kaggle	95.3	96.2
  Jigsaw toxic comment classification	Kaggle	94.7	97.2

🚀 How to use Dolphin

Installation

conda create -n dolphin python=3.11
conda activate dolphin

# Install PyPI requirements
pip install -r requirements.txt

Start Auto-Research using Dolphin

bash launch_dolphin.sh

# modify launch_dolphin.py line # line 189 if round > 0
# exp_base_file_list = [List your exp dir] 

• Note that you need to add api_key and specify the model and topic in launch_dolphin.sh. You can refer to the doc if you want to use self-deployed model.
• Data for Point Classfication, Image Classification, and Sentiment Classification tasks can be downloaded here.

🤔 TODO List

• Release code of Dolphin
• Support more tasks

Citations

@article{yuan2025dolphin,
  title={Dolphin: Closed-loop Open-ended Auto-research through Thinking, Practice, and Feedback},
  author={Yuan, Jiakang and Yan, Xiangchao and Feng, Shiyang and Zhang, Bo and Chen, Tao and Shi, Botian and Ouyang, Wanli and Qiao, Yu and Bai, Lei and Zhou, Bowen},
  journal={arXiv preprint arXiv:2501.03916},
  year={2025}
}