Researchers introduced ScientistTwo, a multi-agent AI framework for autonomous science. It takes a problem, sets baselines, forms hypotheses and runs experiments. It validates findings with a simulated peer-review rebuttal engine. ScientistTwo was benchmarked against papers from ICLR, ICML and NeurIPS. It produced expert-level papers and codebases that outperformed human state-of-the-art. Its papers received higher review ratings than human work under AI reviewers. The authors claim it can push the frontiers of human discovery. The paper was submitted to arXiv on 17 September 2026.