Mathematical papers, audited and connected.

Generative AI has become exceptionally strong at mathematics, and we believe that this ability can and should be used to assist human understanding rather than replace it. One of its strongest abilities is finding small and hidden issues in mathematical texts; we have found it incredibly useful in our own work. We believe the whole mathematical community will benefit from having a place where new arXiv submissions are timely and responsibly AI-reviewed.

• It helps the authors. A preprint audit can identify errors, typos and flaws early in public circulation and before publication. Authors can question a finding and have revisions checked. We are aware of cases when authors used MathAudit reports to correct typos in accepted papers before publication.
• It helps the readers. It is useful to know whether a result one wants to understand and use has issues requiring attention. Sometimes a check catches a real mistake before it produces a chain of faulty arguments; sometimes it clarifies what the authors meant and helps the reader understand the proof.
• It helps the journals. The majority of papers submitted to journals appear on arXiv first. With MathAudit, journals can see at an early stage whether something in a paper needs attention and make a more informed decision. In turn, this could improve the quality and speed of publication.

To serve the mathematical community, we believe that the AI reviewer should follow two strict rules:

1. Same standard for all papers. MathAudit reviews all covered papers against the same standardized prompt, based on our experience reviewing papers. We apply an equal standard across fields while maximizing the review's usefulness for human readers.
2. No judgments of importance, scope or style. Evaluating a result's importance or interest is a task for human experts, and writing style should reflect the author's decision. AI review should identify formal issues in mathematical arguments and nothing else. MathAudit therefore does not rank a suspected fatal flaw above a small formal issue: the importance of an issue is subjective, and our goal is only to point out parts of the text requiring additional attention. Human mathematicians decide how to use that information.