Solution update
This problem has been solved! A solution was first elicited by Kevin Barreto and Liam Price, using GPT-5.4 Pro. This solution was confirmed by problem contributor Will Brian, and will be written up for publication. A full transcript of the original conversation with GPT-5.4 Pro can be found here and GPT-5.4 Pro’s write-up from the end of that transcript can be found here.
Brian’s comments: “This is an exciting solution to a problem I find very interesting. I had previously wondered if the AI’s approach might be possible, but it seemed hard to work out. Now I see that it works out perfectly. It eliminates an inefficiency in our lower-bound construction and in some sense mirrors the intricacy of our upper-bound construction. The matching lower and upper bounds are quite good for Ramsey-theoretic problems, and I’m interested in further understanding why this works out so well.”
Brian plans to write up the solution for publication, possibly including follow-on work spurred by the AI’s ideas. Barreto and Price have the option of being coauthors on any resulting papers. We will update this page with links to future work.
Subsequent to this solve, we finished developing our general scaffold for testing models on FrontierMath: Open Problems. In this scaffold, several other models were able to solve the problem: Opus 4.6 (max), Gemini 3.1 Pro, and GPT-5.4 (xhigh). We sampled each problem four times: Opus 4.6 solved it once, and Gemini 3.1 Pro and GPT-5.4 solved it twice. We did not check whether these models could then write a self-contained proof of the general result.
To isolate the first model able to solve the problem, we also ran GPT-5.2 (xhigh), Opus 4.5 (max), and Kimi K2.5 Thinking on the problem using the same scaffold, again sampling four times. None of them solved it. Unfortunately, Gemini 3 Pro is deprecated so we could not run it.
About the problem
Update (2026-07-31): We have removed this problem from the benchmark. It was solved by AI during the pilot phase. We determined that the problem did not meet the minimum notability bar we wanted to hold for the benchmark, and that this was true regardless of the content of the AI solution.
This problem is about improving lower bounds on the values of a sequence, \(H(n)\), that arises in the study of simultaneous convergence of sets of infinite series, defined as follows.
A hypergraph \((V,\mathcal H)\) is said to contain a partition of size \(n\) if there is some \(D \subseteq V\) and \(\mathcal P \subseteq \mathcal H\) such that \(\|D\| = n\) and every member of \(D\) is contained in exactly one member of \(\mathcal P\). \(H(n)\) is the greatest \(k \in \mathbb{N}\) such that there is a hypergraph \((V,\mathcal H)\) with \(\|V\| = k\) having no isolated vertices and containing no partitions of size greater than \(n\).
It is believed that the best-known lower bounds for \(H(n)\) are suboptimal, even asymptotically, and that they can be improved by finding new constructions of hypergraphs. The goal of this problem is to find such a construction.
We ask for a general construction for all \(n\).
Read more here.
Attempts by AI
We have evaluated the following models on this problem.
Prompt
A hypergraph (V, H) is said to contain a partition of size n if there is some D ⊆ V and P ⊆ H such that |D| = n and every member of D is contained in exactly one member of P. Define H(n) to be the largest integer k such that there is a hypergraph (V, H) with |V| = k having no isolated vertices and containing no partitions of size greater than n.
It is known that H(n) ≥ k_n, where k_n is defined recursively by the formula k_1 = 1 and k_n = ⌊n/2⌋ + k_⌊n/2⌋ + k_⌊(n+1)/2⌋.
Your task is to improve this lower bound by a constant factor, i.e. show that H(n) ≥ c*k_n for some c > 1. It is acceptable if this improvement does not work for small n, but it must already be "in effect" for n=15. You must demonstrate this improvement by providing an algorithm that takes n as input and produces a hypergraph witnessing H(n) ≥ c * k_n.
Please provide an algorithm that takes n as input and outputs the witness hypergraph as a string where vertices are labeled, 1, ..., |V|, and edges are denoted with curly braces. Example: {1,2,3},{2,4},{3,4,5},{1,5}
Solution format:
* Write a Python script defining a function `solution(n: int) -> str`.
* Do not include any code at the file level. You may include a `main` block for testing, but it will not be executed by the verifier.
* For n ≤ 100, the algorithm must complete within 10 minutes when run on a typical laptop.