ScientistTwo is claimed to run experiments and write papers after a researcher sets the problem

A post says Google researchers built ScientistTwo to form hypotheses, run experiments, write papers and simulate peer review after a human sets the problem. It reportedly beat human results on 86 of 107 problems from accepted ICLR, ICML and NeurIPS papers. The post cites a 25.2% average gain but a 7.7% median. Nine reviewers reportedly rated 33 AI papers tied with accepted human papers overall, with a small methodological rigor edge for the human papers.