An AI System That Writes Its Own Research Papers—and Gets Them Accepted

You've probably used an LLM to help draft an email or debug a function. But what happens when you point one at the entire scientific method—hypothesis, experiment, analysis, manuscript—and let it run on its own? That's what SakanaAI's AI Scientist-v2 is built to find out. It's an end-to-end agentic system that autonomously generates research ideas, runs the experiments, and writes up the results as a paper. And according to the project, it produced the first workshop paper written entirely by AI and accepted through peer review.

What It Does

The AI Scientist-v2 is a generalized, end-to-end agentic system for automated scientific discovery in machine learning. You give it a starting point, and it handles the full research loop: generating hypotheses, running experiments, analyzing data, and writing scientific manuscripts.

What separates it from the original AI Scientist is how it explores. V1 leaned on human-authored templates—you'd hand it a solid foundation and it would follow it closely. V2 drops that reliance, generalizes across ML domains, and uses a progressive agentic tree search guided by an experiment manager agent. The tree search is the interesting part: instead of committing to a single line of inquiry, the system branches through possible experiment paths and lets a dedicated manager agent steer the exploration.

The codebase runs on Linux with NVIDIA GPUs and expects CUDA and PyTorch. Installation involves setting up a conda environment, installing PyTorch with CUDA support, pulling in PDF and LaTeX tooling (poppler and chktex), and then the Python requirements. You'll also need API keys—it supports OpenAI models via OPENAI_API_KEY, Gemini via GEMINI_API_KEY, and Claude through AWS Bedrock.

Why It's Cool

• It's honest about its own limits. The README comes with a note that v2 doesn't necessarily produce better papers than v1—especially when you have a strong template. V1's template-following gives it high success rates on well-defined tasks; v2 trades that reliability for open-ended exploration with lower success rates. That kind of candor is rare and useful when you're deciding which tool fits your problem.

• The tree search changes the shape of the problem. Rather than a linear pipeline, the progressive agentic tree search with an experiment manager agent means the system can explore multiple research directions and adapt. That's a more realistic model of how research actually unfolds—messy, branching, and iterative.

• It closes the entire loop. Hypothesis generation, experiment execution, data analysis, and manuscript writing are all in scope. Whether or not every paper is a winner, having the full cycle automated in one system is a genuinely different proposition from tools that help with just one stage.

• It's a real, peer-reviewed data point. The project points to a workshop paper written entirely by the system that passed peer review, plus an ICLR2025 workshop experiment repo you can inspect. That's a concrete artifact to evaluate rather than a demo.

• It's upfront about danger. The README warns that the codebase executes LLM-written code, with risks like dangerous packages, uncontrolled web access, and unintended processes. It tells you to run it in a controlled sandbox like a Docker container. Take that seriously—this isn't a tool you run casually on your workstation.

How to Try It

Getting started requires a Linux machine with NVIDIA GPUs. Here's the install path from the README:

# Create a new 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

The README notes installation usually takes under an hour. From there, you'll want to configure your API keys depending on which models you're using. For Claude via Bedrock, install the extra packages with pip install anthropic[bedrock] and set up your AWS credentials and region.

The workflow then moves in two stages: generating research ideas, followed by running the paper generation experiments. Before you run anything, set up that sandbox—the code will execute model-written code, and the README is clear about the risks.

You can find the full setup details, paper, and blog post at the repository.

Final Thoughts

This is a research tool, not a product, and it's aimed at people who want to experiment with autonomous scientific discovery—ML researchers, agent-system builders, and anyone curious about how far the full research loop can be automated. The honest framing in the README is the most useful thing about it: if you have a well-defined task with a solid template, v1 may serve you better. If you're chasing open-ended exploration and can tolerate a lower hit rate, v2 is the more interesting bet. Either way, it's a compelling look at where agentic systems are heading, and the fact that it produced a peer-reviewed workshop paper makes it worth a closer read.

Follow @githubprojects for more developer tools and open source projects.