AI-Scientist-v2

Why we included this project

Researchers and ML engineers curious about how far autonomous experimentation can go will find this the most ambitious open attempt yet. It hands the entire research loop, from idea generation and experiment design to data analysis and manuscript writing, to a set of coordinated LLM agents, and it does not rely on human-authored paper templates. The progressive agentic tree search, guided by an experiment manager agent that explores multiple research branches and prunes unpromising paths, is what lets the system generalize across machine-learning domains instead of locking onto one benchmark. It also produced the first fully AI-authored workshop paper accepted through peer review, a rare real-world check on whether AI-generated science can survive scrutiny. Teams should treat it as research software rather than a plug-and-play pipeline: the codebase deliberately executes LLM-written code, so the authors require a sandboxed container and supervised runs, and that trade-off between open-ended exploration and safety overhead is the thing to weigh before adopting it.