This repository contains the code for our paper Many Heads Are Better Than One: Improved Scientific Idea Generation by A LLM-Based Multi-Agent System.

• To the best of our knowledge, we propose the first multi-agent system with a scientific research ecosystem for conducting and benchmarking scientific collaborations, named VirSci, where real data is used for role-playing and objective evaluation.
• To simulate a reliable scientific collaboration process, we propose an end-to-end pipeline that spans team organization to idea generation. A novel inter- and intra-team discussion mechanism is introduced to promote communication topology and enhance the simulation realism.
• Extensive experiments demonstrate that multi-agent collaboration improves outcome quality, surpassing existing methods. We also conduct systematic experiments to investigate factors influencing the system, with findings aligning with established principles from the Science of Science, such as fresh teams tend to generate more innovative research, underscoring VirSci's reliability as a powerful tool for autonomous scientific discovery.

Our project website is Website.

1. We update our paper with Appendix on Arxiv.

1. We propose the VirSci, a multi-agent system has the potential to improve scientific idea generation.
2. Watch demo video for our project at YouTube.
3. Full paper with Appendix is available on Arxiv.
4. VirSci code and data are available for Research community.

We tested our codebase with PyTorch 2.3.1 and CUDA 12.1. Please install the corresponding versions of PyTorch and CUDA based on your computational resources.

To install the required packages, run:

pip install -r requirements.txt

Note

If you encounter any errors while setting up the environment, do not panic, as our environment is deployed on the ARM architecture, which may cause some package versions to be unavailable. The most important thing is to install agentscope in editable mode 😀, which can be easily installed using the command:

cd agentscope-main
pip install -e .

Another kind advice is that our code uses faiss during execution, and we highly recommend installing faiss-gpu.

The raw data is based on the AMiner Computer Science Dataset.

After preprocessing, the used data is publicly available at Google Drive.

• Past paper database is put in the Papers/papers.tar.gz, which is used in paper_folder_path of Line 34 in sci_platform/sci_platform.py. The corresponding embedding database is put in the Embeddings/faiss_index.index, which is used in cpu_index of Line 135 in sci_platform/sci_platform.py.
• Contemporary paper database is put in the Papers/papers_future.tar.gz, which is used in future_paper_folder_path of Line 35 in sci_platform/sci_platform.py. The corresponding embedding database is put in the Embeddings/faiss_index_future.index, which is used in cpu_future_index of Line 139 in sci_platform/sci_platform.py.
• Author knowledge bank is put in the Authors/books.tar, which is used in in input_dir of Line 13 in sci_platform/configs/knowledge_config.json and author_info_dir of Line 36 in sci_platform/sci_platform.py.
• Adjacency matrix is put in the adjacency.txt, which is used in adjacency_matrix_dir of Line 37 in sci_platform/sci_platform.py.

Note

Please replace all paths in sci_platform/sci_platform.py with your own settings after download the data.

Here we explain the roles of several critial files.

• agentscope-main/src/agentscope/agents/sci_agent.py defines the customized scientist agent in this project.
• sci_platform/run.py is the main execution file.
• sci_platform/sci_platform.py defines the platform for the initialization of our multi-agent system.
• sci_platform/utils/prompt.py contains all the prompts used.
• sci_platform/utils/scientist_utils.py contains all the common functions used.
• sci_platform/sci_team/SciTeam.py defines the execution mechanism of each scientist team.

Ollama

In our experiments, we use ollama to deploy the llama3.1-8b and llama3.1-70b language models and mxbai-embed-large embedding model. The details of deployment could refer to URL. Here we show some key steps:

1. Ollama should be installed. The linux version:

curl -fsSL https://ollama.com/install.sh | sh

2. Run ollama in the path where ollama is installed:

./ollama serve

3. Pull a model to use with the library:

./ollama pull llama3.1
./ollama pull llama3.1:70b
./ollama pull mxbai-embed-large

4. Install the ollama python library in your environment:

pip install ollama

5. Complete the installation and close the terminal.

Run

After pull all models, you need to open the ollama server before running our codes:

./ollama serve

Then, run our codes:

cd sci_platform/
python run.py

Our code support different collaboration settings. The commonly used arguments:

--runs: how many times does the program run

--team_limit: the max number of teams for a scientist

--max_discuss_iteration: the max discussion iterations for a team in a step

--max_team_member: the max team member of a team (including the team leader)

--epochs: the allowed time steps for one program run (default value is 6, which is enough for a scientist to finish all steps)

• {info_dir}/{current_time}_{self.team_name}_dialogue.json saves the team information: all team members, selected topic, generated idea and abstract, where info_dir denotes the storage path, current_time denotes the start running time, and self.team_name is the name of the team.
• {log_dir}/{current_time}_{self.team_name}_dialogue.log saves the log record of the team, where log_dir denotes the storage path.

This project is supported by Shanghai Artificial Intelligence Laboratory.

The multi-agent framework in this work is based on the AgentScope.

The raw data is based on the AMiner Computer Science Dataset.

If you have any questions, please contact at [chenrenqi@pjlab.org.cn, suhaoyang@pjlab.org.cn].

Welcome to join our discussion group on collective intelligence technology!

Wechat

If you find that the WeChat QR code has expired, please send us an email with your WeChat ID, and we will manually invite you to the group.

This repository is licensed under the Apache-2.0 License.

If you find this code useful, please consider citing our work:

@article{su2024two,
  title={Two Heads Are Better Than One: A Multi-Agent System Has the Potential to Improve Scientific Idea Generation},
  author={Su, Haoyang and Chen, Renqi and Tang, Shixiang and Zheng, Xinzhe and Li, Jingzhe and Yin, Zhenfei and Ouyang, Wanli and Dong, Nanqing},
  journal={arXiv preprint arXiv:2410.09403},
  year={2024}
}