SakanaAI has unveiled AI Scientist-v2, a generalized end-to-end agentic system that has generated the first workshop paper written entirely by AI and accepted through peer review. The system autonomously generates hypotheses, runs experiments, analyzes data, and writes scientific manuscripts.
From Template-Based to Autonomous Discovery
Unlike its predecessor AI Scientist-v1, which relied on human-authored templates, v2 removes this limitation entirely. It generalizes across machine learning domains and employs a progressive agentic tree search guided by an experiment manager agent.
How It Works
The system operates in three stages. First, it generates research ideas by brainstorming and refining topics based on high-level descriptions, checking novelty against existing literature via Semantic Scholar. Second, it runs experiments via agentic tree search, exploring multiple paths simultaneously with parallel workers. Third, it analyzes results and writes scientific papers in LaTeX format.
Technical Requirements
The code is designed to run on Linux with NVIDIA GPUs using CUDA and PyTorch. It supports multiple LLM backends including GPT-4o, Claude through Amazon Bedrock, and Gemini. A typical run generates ideas in minutes, runs experiments over several hours, and produces a final paper draft in 20-30 minutes.
Important Caveats
The creators emphasize this codebase will execute LLM-written code with various risks including potential use of dangerous packages, uncontrolled web access, and unintended processes. They strongly advise running within a controlled sandbox environment such as Docker.
The Future of Scientific Discovery
AI Scientist-v2 represents a significant step toward fully autonomous scientific research. While it does not necessarily produce better papers than v1—especially when a strong starting template is available—it excels at open-ended scientific exploration where no clear path exists. The system is available on GitHub with full documentation for researchers interested in exploring automated scientific discovery.