Curie: A Research Experimentation Agent

Curie is the first AI-agent framework designed for automated and rigorous scientific experimentation. Curie helps answer your curiosity through end-to-end experimentation automation, ensuring that every step—from hypothesis formulation to result interpretation—is conducted with precision, reliability, and reproducibility. Our mission is to empower scientists to move research at the speed of thought.

Curie’s Role in the Scientific Research Lifecycle

🗞️ News

• [2025/06] We published EXP-Bench: Can AI Conduct AI Research Experiments? → 📄 Paper | 🗂️ Dataset | 📰 Blog
• [2025/05] We launched an AutoML feature to help researcher find the optimal ML solution → 📢 Blog
• [2025/02] We published Curie: Toward Rigorous and Automated Scientific Experimentation with AI Agents → 📄 Paper | 📰 Blog

Key Features

• 🚀 Automated Experimentation – From hypothesis formulation, experiment implementation, experiment execution, result analysis and finding reflection.
• 📊 Rigor Enhancement - Built-in verification modules enforce methodical procedure, agent reliability and reproducibility.
• 🔬 Broad Applicability – Supports ML Engineering, system analysis, and scientific discovery.
• 💻 Use Your Starter Code – Supports working on arbitrary user's starter code.
• 📂 Bring Your Own Dataset – Supports working on arbitrary user's datasets.
• 🧾 Automatic, Insightful Reporting - See a sample report here

Table of Contents

• ⚙️ Installation
• ⚡ Quick Start
• 📚 Tutorial
• 🎬 Demo

⚙️ Installation

Prerequisite: Install Docker from here

sudo chmod 666 /var/run/docker.sock
docker ps  # Verify Docker installation

Option 1: Quick Install via pip

pip install curie-ai

Option 2: Manual Installation for Developers

⚡ Quick Start

• It's recommended to use tmux or a similar terminal multiplexer before running Curie, as experiments can take several minutes depending on the task and budget.

• Do not use Jupyter Notebook.

(Simple) Example 1: You Have a Single Question that Needs to be Verified.

👩‍🎓: I want to understand the Sorting Algorithm Efficiency.

import curie
# Set up your API keys, refer to curie/setup/env.sh.example
key_dict = {
    "MODEL": "claude-3-7-sonnet-20250219",
    "ANTHROPIC_API_KEY": "your-anthropic-key"
}

result = curie.experiment(api_keys=key_dict, 
                          question="How does the choice of sorting algorithm impact runtime performance across different input distributions?",
                          max_global_steps=10)

• 🧾 Auto-Generated Experiment Report: logs/research_<ID>.md.

• 📊 Experiment Result Notebook: logs/research_*_all_results.txt.

• 🪵 The Experimentation Process (generated script generated code to reproduce experiment results): workspace/research_<ID>/.

Example 2: Find Optimal ML Strategies for Noisy Cancer Data.

👩‍🎓: I want to find the most robust ML methods for my noisy data.

result = curie.experiment(api_keys=key_dict, 
                          question="Are ensemble methods (e.g., Random Forests, Gradient Boosting) more robust to added noise in the Breast Cancer Wisconsin dataset compared to linear models like Logistic Regression for a binary classification task?")

• 🧾 Auto-generated Experiment Report

• 📊 Experiment Result Notebook

(Advanced) Example 3: You Have a Dataset and Want to Gain Insight from It

👨‍🎓: I have a dataset and some starter code,and I want to train/deloy ML models to achieve specific goals. (GPU is recommended for ML training tasks.)

result = curie.experiment(
    api_keys=key_dict,
    question="E.g. How to improve my prediction accuracy on my dataset.",
    dataset_dir="/abs/path/to/your/dataset",
    codebase_dir="[Optional] /abs/path/to/your/code",
    env_requirements="[Optional] /abs/path/to/requirements.txt",
)

Curie AutoML Feature Overview.

• Check out how Curie is able to find optimal ML soltuions in these examples from MLE-Bench.
  • Predict the dog breed
  • Identify melanoma in images of skin lesions
  • Predict the severity level of diabetic retinopathy based on retinal images
  • Histopathologic Cancer Detection
  • Predict the stock price ranking
• Sample Curie-Generated Experiment Report:

Check out more Machine Learning Use Cases here.

📚 Tutorial

• How to use Curie to answer your curiosity (with a concrete example)?

🎬 Demo Video

Curie Overview & Demo.

📜 Citation

If you use Curie in a research paper, please cite our work:

@article{kon2025expbenchaiconductai,
      title={EXP-Bench: Can AI Conduct AI Research Experiments?}, 
      author={Patrick Tser Jern Kon and Jiachen Liu and Xinyi Zhu and Qiuyi Ding and Jingjia Peng and Jiarong Xing and Yibo Huang and Yiming Qiu and Jayanth Srinivasa and Myungjin Lee and Mosharaf Chowdhury and Matei Zaharia and Ang Chen},
      journal={arXiv preprint 2505.24785}
      year={2025},
}

@article{kon2025curie,
  title={Curie: Toward rigorous and automated scientific experimentation with ai agents},
  author={Kon, Patrick Tser Jern and Liu, Jiachen and Ding, Qiuyi and Qiu, Yiming and Yang, Zhenning and Huang, Yibo and Srinivasa, Jayanth and Lee, Myungjin and Chowdhury, Mosharaf and Chen, Ang},
  journal={arXiv preprint arXiv:2502.16069},
  year={2025}
}

Community and Support

• GitHub Issues - Report bugs or request features
• Schedule a Meeting with Us - Get help from our team
• Join our Slack Community

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.