Chris Lu, Cong Lu, R. T. Lange, Yutaro Yamada, Shengran Hu, Jakob Foerster, David Ha, Jeff Clune
This paper presents a pipeline that automates the entire scientific research process, demonstrating that AI-generated manuscripts can pass peer review.
Scientific automation has been limited to individual components; a system that autonomously navigates the entire research lifecycle from conception to publication has remained out of reach.
The authors developed 'The AI Scientist', a complex agentic system leveraging modern foundation models. It automates idea generation, code writing, experiment execution, data analysis and plotting, manuscript writing, and its own peer review. The system was evaluated in a focused mode using human-provided code templates and a template-free, open-ended mode using agentic search.
The manuscript generated by the AI system passed the first round of peer review for a workshop at a top-tier machine learning conference (70% acceptance rate). This achievement demonstrates AI's growing capacity for scientific contribution and signifies a potential paradigm shift in research conduct.