AI-Scientist-v2: Automated Scientific Discovery via Agentic Tree Search

Introduction

The AI Scientist-v2, developed by SakanaAI, represents a groundbreaking step in automated scientific discovery. This generalized, end-to-end agentic system is designed to autonomously conduct scientific research, from generating novel hypotheses to writing peer-reviewed papers. It distinguishes itself from its predecessor by removing reliance on human-authored templates, generalizing across various Machine Learning (ML) domains, and employing a progressive agentic tree search guided by an experiment manager agent.

This system has already achieved a notable milestone, generating the first workshop paper written entirely by AI and accepted through peer review, marking a significant step towards fully autonomous research. While AI Scientist-v2 takes a broader, more exploratory approach compared to v1, it is ideal for open-ended scientific exploration, pushing the boundaries of what AI can achieve in research.

Installation

To get started with AI Scientist-v2, you'll need a Linux environment with NVIDIA GPUs, CUDA, and PyTorch. The installation process involves setting up a Conda environment and installing necessary dependencies.

Create a Conda environment:

conda create -n ai_scientist python=3.11
conda activate ai_scientist

Install PyTorch with CUDA support: (Adjust pytorch-cuda version for your setup)

conda install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia

Install PDF and LaTeX tools:

conda install anaconda::poppler
conda install conda-forge::chktex

Install Python package requirements:
```
pip install -r requirements.txt
```

Ensure you set up API keys for supported models (OpenAI, Gemini, Claude via AWS Bedrock) and optionally for Semantic Scholar for enhanced literature search. Refer to the official repository for detailed instructions on API key configuration.

Examples

The AI Scientist-v2 workflow typically involves two main stages: generating research ideas and then running paper generation experiments.

1. Generate Research Ideas

First, you use the perform_ideation_temp_free.py script to brainstorm and refine research ideas based on a high-level topic description you provide. This script leverages LLMs and tools like Semantic Scholar to check for novelty.

Example Command:

python ai_scientist/perform_ideation_temp_free.py \
 --workshop-file "ai_scientist/ideas/my_research_topic.md" \
 --model gpt-4o-2024-05-13 \
 --max-num-generations 20 \
 --num-reflections 5

This will generate a JSON file containing structured research ideas, which will be used in the next step.

2. Run AI Scientist-v2 Paper Generation Experiments

Once you have your research ideas, you can launch the main pipeline to run experiments via agentic tree search, analyze results, and generate a paper draft.

Example Command:

python launch_scientist_bfts.py \
 --load_ideas "ai_scientist/ideas/my_research_topic.json" \
 --load_code \
 --add_dataset_ref \
 --model_writeup o1-preview-2024-09-12 \
 --model_citation gpt-4o-2024-11-20 \
 --model_review gpt-4o-2024-11-20 \
 --model_agg_plots o3-mini-2025-01-31 \
 --num_cite_rounds 20

After completion, you will find a timestamped log folder in experiments/ containing the tree visualization and, eventually, the generated PDF paper.

Why Use It

AI Scientist-v2 offers a powerful platform for accelerating scientific discovery and exploring new research frontiers. Its key advantages include:

Full Autonomy: It automates the entire research lifecycle, from hypothesis generation to paper writing, significantly reducing human effort and time.
Generalization: Unlike previous versions, it can generalize across various ML domains, making it a versatile tool for diverse research areas.
Agentic Tree Search: The progressive agentic tree search, guided by an experiment manager, allows for more exploratory and less template-dependent research.
Pioneering AI Research: It represents a significant milestone in AI's capability to contribute to scientific literature, having produced a peer-reviewed paper entirely by AI.

Caution: This codebase will execute Large Language Model (LLM)-written code. It is crucial to run this within a controlled sandbox environment (e.g., a Docker container) due to potential risks like dangerous packages or unintended processes.