Can AI systems generate research papers that pass peer review?

The AI Scientist-v2 is an end-to-end agentic system that formulates hypotheses, runs experiments, analyzes data and writes manuscripts. Its central claim is a test result reported by its builders: of three manuscripts generated entirely by the system and submitted to a peer-reviewed ICLR workshop, one averaged 6.33 across reviewers, "roughly in the top 45% of submissions," and "would have been accepted after meta-review were it human-generated." The authors call this "the first fully AI-generated manuscript to successfully pass a peer-review process." The scores are their figures. The reviewers had been told beforehand that some submissions could be AI-generated and could opt out.

The excerpt credits the result to three changes from the predecessor, The AI Scientist-v1. Idea generation starts at a higher level of abstraction and queries Semantic Scholar during formulation, rather than relying on post-hoc checks. An experiment progress manager agent moves work through staged experimentation, and an agentic tree search explores the experiments. A vision-language model feedback loop refines the figures, and manuscript writing becomes a single pass followed by a reflection stage run by reasoning models. The excerpt names four stages for the manager but lists only three, and it reports no ablation showing how much each change contributed.

Set against the nearby notes, this is a different kind of evidence. Can automated review loops handle AI-generated research at scale? argues that AI research needs a venue built for automated review. This excerpt instead sends its output to an existing human workshop and lets human reviewers decide, which tests whether the current gate can be passed, not whether a new gate is needed. PaperOrchestra's win rates are margins in human evaluation against autonomous baselines, so they measure writing quality, not acceptance. Can separating judgment from verification improve research paper reliability? requires evidence to be checkable before results are seen. This excerpt shows the failure mode such checks target: citations the authors say were sometimes inaccurate, "similar to the well-known 'hallucination' issue." How should AI agents and humans divide research tasks? describes the same human-retained division of labor. The AI Scientist-v2 authors draw that line too: they withdrew the accepted paper before publication to avoid putting purely AI-generated work into the record without wider community discussion.

The excerpt establishes less than the headline. One accepted paper out of three is too small a sample to estimate how often the system clears review. The acceptance-rate context is the authors' own: they cite workshop acceptance "typically 60-80%" against 20-30% at ICLR, ICML and NeurIPS, with no source given in the excerpt. They also say the system "does not yet consistently reach" top-tier standards, "nor does it even reach workshop-level consistently." The reviewers' verdict was "an interesting and technically sound workshop contribution that needs further development." The supportable implication is narrow: an autonomous pipeline got one manuscript past a lenient human review once. That is a capability result, not evidence that its output is reliable. The authors' expectation that AI "will likely generate papers that match or exceed human quality" is a prediction the excerpt does not test.