AI contribution section placement system

The wiki is no longer updated. The latest data is as of Jun 30, 2026.

---

We divide AI contributions into two top-level categories:

• Section 1 contains contributions to Erdős problems in which AI systems play a primary role.

• Section 2 contains contributions to Erdős problems in which AI systems play a secondary role.

What is a primary/secondary contribution? Roughly, a contribution is primary if it proposes ideas, advances understanding, or makes progress in a way that, for a human, qualifies for authorship in a potential work. Admittedly this is subjective.

What is a contribution? For the purposes of this wiki, some contributions are not recorded. See the end of this page for details.

Provisional status. The section placement may be updated if new information comes to light, per disclaimer 10.

Section 1 is further subdivided into subsections 1(a), 1(b), 1(c), 1(d) for organization purposes.

Notice. Sections 1(a)-1(d) are equally important. We do not intend section 1(a) to be more important than other sections.

Indeed, there are currently original AI solutions to Erdős problems in all four subsections.

Do not treat the tables in this wiki as a leaderboard.

The subsections are organized by (1) presence of literature and (2) involvement of humans.

These depend on two subjective concepts:

1. Comparable literature means works that make contributions on roughly the same levels as the AI contributions.

2. Significant human involvement means the human contributions are at levels usually qualified for authorship.

Literature presence: comparable literature unknown.
Human involvement: non-significant.

Be mindful of a strong reporting bias: a failure of an AI system to make progress on a problem may not be reported. We do not advise inferring success rates of AI systems just from this data.

Literature presence: comparable literature discovered afterwards.
Human involvement: non-significant.

The literature might be included in the AI training data. But even if it were, we cannot conclude from this whether the AI contributions were influenced by this literature. The AI system may have independently arrived at a comparable solution by itself. However, we do our best to extract any literature connections we can find according to professional standards. You can also compare the AI contributions with literature and draw your own tentative conclusions. (We do not endorse the view that plagiarism is automatic just from presence in training data, as then by analogy a human who has seen a result, does not remember it, and later re-derives it also plagiarizes, which is absurd.)

Literature presence: comparable literature known beforehand and not discovered afterwards.
Human involvement: non-significant.

The literature may have been provided as input to the AI system itself. In the case where it was not provided, the literature might still be included in the AI training data. The exact nature of such interaction varies with the problem; please visit the problem pages for more details.

Literature presence: any.
Human involvement: significant.

It is expected that projects involving significant human input rely on or are inspired by previous literature; and during literature review or refereeing, it is common that other relevant literature is discovered. We do not report these routine linkages. Please refer to the discussion pages for the Erdős problems and/or the writeups of these results for literature connections.

Section 2 is further subdivided into:

Literature search is to look for works and results that are relevant to the problem.

This section includes both (i) results of an AI system being explicitly instructed to perform a literature search, and (ii) results of an AI system that was not explicitly instructed to perform a literature search but performed one, and for which this was an essential component of the final results.

Formalization is to convert a mathematical proof into a formal language such as Lean.

Rewriting is to produce a document based on a previous document either directly or via a critique.

Uses of AI to perform minor edits such as spelling or grammar fixes, translate from one language to another, extract or reformat data, or summarize an existing document, are not recorded. The standard is that the generated revision or critique led to substantive mathematical contributions to the argument, for instance by identifying gaps, simplifying proofs, or strengthening results.

Computation is to carry out a well-defined calculation.

Q: Isn't AI automatically computing by its nature?
A: Unless the desired task is directly performed by AI, the AI must write code to perform the requested computation. The writing of code and other informal reasoning is usually not a well-defined task, even though the code itself performs a well-defined task. So in practice this correctly defines what you'd expect regardless of the AI's nature. See also here if curious.

Only computation which materially helped a research project on the problem in question are recorded.

• Minor observations.

• Works in progress or otherwise not ready for review.

• Verification and critique of other works.

• Auxiliary writing tasks; see section 2(c).

• Non-significant computation; see section 2(d).