Source: arXiv 2609.19644, shared by @HowToPrompt__
The ScientistTwo paper describes a set of AI agents that takes a research problem, sets baselines, forms hypotheses, runs experiments and ablations, and argues with a simulated peer reviewer before writing up the result. The authors compare its output with papers accepted at ICLR, ICML and NeurIPS and say it beats the human state of the art. In the same abstract they report that its papers got higher average ratings than human ones “under automated AI review agents”, so the reviewers were AI too.