AI Scientist — Automated Research Paper Generation

What it is

AI Scientist is a fully automated research system by Sakana AI that takes a research template and compute budget, then autonomously generates novel ideas, designs experiments, runs code, analyzes results, and produces complete scientific papers in LaTeX. The output includes literature review, methodology, results, and discussion sections.

This tool targets researchers exploring AI-assisted scientific discovery, labs looking to accelerate hypothesis exploration, and anyone studying the frontier of automated research. It works with Claude, GPT-4, and Gemini as the underlying LLM.

How it saves time or tokens

The traditional research cycle from idea to manuscript takes weeks or months of manual work. AI Scientist compresses the entire pipeline into a single command. It generates multiple candidate ideas, implements each as code, runs experiments, and writes up findings -- all without human intervention between steps.

The token_estimate for this workflow is approximately 2,800 tokens per run. The system supports batching multiple ideas in a single launch, amortizing setup overhead.

How to use

1. Clone and install dependencies:

git clone https://github.com/SakanaAI/AI-Scientist.git
cd AI-Scientist
pip install -r requirements.txt

1. Run the full pipeline with your chosen model and experiment template:

python launch_scientist.py \
  --model claude-sonnet-4-20250514 \
  --experiment nanoGPT \
  --num-ideas 5

1. Collect the output LaTeX manuscripts from the results directory.

Example

The system generates structured research ideas before implementing them:

Input: 'Improve training efficiency of small language models'

Generated Ideas:
1. Adaptive learning rate scheduling based on gradient noise
2. Curriculum learning with dynamic difficulty assessment
3. Sparse attention patterns for resource-constrained training

For each idea, AI Scientist writes experiment code, runs it, and produces a full paper:

# Check generated papers
ls results/nanoGPT/
# idea_1_paper.pdf  idea_2_paper.pdf  idea_3_paper.pdf

Related on TokRepo

• AI Tools for Research -- Research automation and discovery tools
• Prompt Library -- Reusable prompts for various AI workflows

Common pitfalls

• Generated papers should be treated as drafts requiring human review. The system may produce plausible-sounding but incorrect analysis.
• Experiment templates constrain the research scope. Without a well-designed template, the system may explore unproductive directions.
• API costs can accumulate quickly when generating multiple ideas with large models. Set --num-ideas conservatively for initial runs.