The AI Scientist: Towards Fully Automated
Open-Ended Scientific Discovery 🧑‍🔬

📚 [Paper] | 📝 [Blog Post] | 📂 [Drive Folder]

One of the grand challenges of artificial intelligence is developing agents capable of conducting scientific research and discovering new knowledge. While frontier models have already been used to aid human scientists, e.g. for brainstorming ideas or writing code, they still require extensive manual supervision or are heavily constrained to a specific task.

We're excited to introduce The AI Scientist, the first comprehensive system for fully automatic scientific discovery, enabling Foundation Models such as Large Language Models (LLMs) to perform research independently.

We further provide all runs and data from our paper here, where we run each base model on each template for ~50 ideas. We highly recommend reading through some of the Claude papers, (especially the diffusion ones), to get a sense of its strengths and weaknesses. Here are some example papers generated by The AI Scientist 📝:

1. DualScale Diffusion: Adaptive Feature Balancing for Low-Dimensional Generative Models
2. Multi-scale Grid Noise Adaptation: Enhancing Diffusion Models For Low-dimensional Data
3. GAN-Enhanced Diffusion: Boosting Sample Quality and Diversity
4. DualDiff: Enhancing Mode Capture in Low-dimensional Diffusion Models via Dual-expert Denoising
5. StyleFusion: Adaptive Multi-style Generation in Character-Level Language Models
6. Adaptive Learning Rates for Transformers via Q-Learning
7. Unlocking Grokking: A Comparative Study of Weight Initialization Strategies in Transformer Models
8. Grokking Accelerated: Layer-wise Learning Rates for Transformer Generalization
9. Grokking Through Compression: Unveiling Sudden Generalization via Minimal Description Length
10. Accelerating Mathematical Insight: Boosting Grokking Through Strategic Data Augmentation

Note: Caution! This codebase will execute LLM-written code. There are various risks and challenges associated with this autonomy. This includes e.g. the use of potentially dangerous packages, web access, and potential spawning of processes. Use at your own discretion. Please make sure to containerize and restrict web access appropriately.

Table of Contents

1. Requirements
2. Run AI Scientist Paper Generation Experiments
3. Getting an LLM Generated Paper Review
4. Making your own Template
5. Template Resources
6. Citing The AI Scientist
7. Frequently Asked Questions

Requirements

Installation

conda create -n ai_scientist python=3.11
conda activate ai_scientist

# LLM APIs
pip install anthropic aider-chat backoff openai
# Viz
pip install matplotlib pypdf pymupdf4llm
# Install pdflatex
sudo apt-get install texlive-full

# Common Requirements
pip install torch numpy transformers datasets tiktoken wandb tqdm

We use the following environment variables for the different API providers for different models:

OPENAI_API_KEY, ANTHROPIC_API_KEY, DEEPSEEK_API_KEY, OPENROUTER_API_KEY

Our code can also optionally use a Semantic Scholar API Key (S2_API_KEY) for higher throughput if you have one, though in principle it should work without it.

Be sure to provide the key for the model used for your runs, e.g.

export OPENAI_API_KEY="YOUR KEY HERE"
export S2_API_KEY="YOUR KEY HERE"

Setup NanoGPT

# Prepare NanoGPT data
python data/enwik8/prepare.py
python data/shakespeare_char/prepare.py
python data/text8/prepare.py

Create baseline runs (machine dependent)

# Set up NanoGPT baseline run
# NOTE: YOU MUST FIRST RUN THE PREPARE SCRIPTS ABOVE!
cd templates/nanoGPT && python experiment.py --out_dir run_0 && python plot.py

Create NanoGPT_lite baseline run. We use this for sanity-checking

# NOTE: YOU MUST FIRST RUN THE PREPARE SCRIPTS ABOVE!
cd templates/nanoGPT_lite && python experiment.py --out_dir run_0 && python plot.py

Setup 2D Diffusion

# Set up 2D Diffusion
git clone https://github.com/gregversteeg/NPEET.git
cd NPEET
pip install .
pip install scikit-learn

# Set up 2D Diffusion baseline run
cd templates/2d_diffusion && python experiment.py --out_dir run_0 && python plot.py

Setup Grokking

# Set up Grokking baseline run
cd templates/grokking && python experiment.py --out_dir run_0 && python plot.py

Run AI Scientist Paper Generation Experiments

Note: please ensure the setup steps above are completed.

conda activate ai_scientist
# Run the paper generation.
python launch_scientist.py --model "gpt-4o-2024-05-13" --experiment nanoGPT_lite --num-ideas 2
python launch_scientist.py --model "claude-3-5-sonnet-20240620" --experiment nanoGPT_lite --num-ideas 2

Getting an LLM Generated Paper Review

import openai
from ai_scientist.perform_review import load_paper, perform_review

client = openai.OpenAI()
model = "gpt-4o-2024-05-13"

# Load paper from pdf file (raw text)
paper_txt = load_paper("report.pdf")
# Get the review dict of the review
review = perform_review(
    paper_txt,
    model,
    client,
    num_reflections=5,
    num_fs_examples=1,
    num_reviews_ensemble=5,
    temperature=0.1,
)

# Inspect review results
review["Overall"]  # overall score 1-10
review["Decision"]  # ['Accept', 'Reject']
review["Weaknesses"]  # List of weaknesses (str)

To run batch analysis:

cd review_iclr_bench
python iclr_analysis.py --num_reviews 500  --batch_size 100 --num_fs_examples 1 --num_reflections 5 --temperature 0.1 --num_reviews_ensemble 5

Making your own Template

If there is an area of study you would like The AI Scientist to explore, it should be very easy to create your own templates. In general, follow the structure of the existing templates, which consists of:

• experiment.py -- This is a single file where the 'meat' of the content is. It takes in an argument for out_dir, which is where it should create the folder and save the relevant information from the run.
• plot.py -- This should take in the information from the run folders and create plots. The code should be clear and easy to edit.
• prompt.json -- Put information about your template here.
• seed_ideas.json -- Put example ideas here. You can also try to generate ideas without any examples, and then pick the best one or two to put here.
• latex/template.tex -- We recommend using our latex folder, but be sure to replace the pre-loaded citations with ones that you would expect to be more relevant.

Template Resources

We provide 3 templates, which heavily use code from other repositories, which we credit below. (Normally, we would do this in the files themselves, but it's unclear how this would affect The AI Scientist since it would be visible).

The NanoGPT template used code from NanoGPT and this PR.

The 2D Diffusion template used code from tiny-diffusion and ema-pytorch.

The Grokking template used code from Sea-Snell/grokking and danielmamay/grokking.

We would like to thank the developers of the open-source models and packages for their contributions and for making their work available.

Citing The AI Scientist

If you use The AI Scientist in your research, please cite it as follows:

@article{lu2024aiscientist,
  title={The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery},
  author={Chris Lu and Cong Lu and Robert Tjarko Lange and Jakob Foerster and Jeff Clune and David Ha},
  journal={arXiv preprint arXiv:2408.06292},
  year={2024}
}

FAQ

We recommend reading our paper in the first instance for any questions you have on The AI Scientist.

Why am I missing files when running The AI Scientist?

Make sure you have completed all the setup and preparation steps before the main experiment script.

Why has a PDF not been generated?

The AI Scientist finishes an idea with a success rate that depends on both the template, the base foundation model, and the complexity of the idea. We advise referring to our main paper. The highest success rates are observed with Claude Sonnet 3.5.

What is the cost of each idea generated?

Typically less than $15 per paper with Claude Sonnet 3.5.

How do I change the base conference format associated with the write-ups?

Change the base template.tex files contained within each template.

How do I run The AI Scientist for different subject fields?

Please refer to the instructions for different templates. In this current iteration, this is restricted to ideas that can be expressed in code. However, lifting this restriction would represent exciting future work! :)