TL;DR
GPT-6 Astra has solved a nine-year-old open problem in election mathematics, proving a perfectly fair voting committee always exists and inventing an entirely new voting rule to do it.
What happened
- GPT-6 Astra, working with three human researchers over several days, solved a major open problem in social choice theory that has stood since 2017.
- The problem, listed on the FrontierMath benchmark as one of its hardest challenges, asked solvers to find a counterexample showing a fair voting committee (the "core") could be empty under some preference configurations.
- Astra flipped the premise entirely: it proved no counterexample exists and that a core committee is always guaranteed to exist.
- To do so, Astra invented a new objective function called "harmonic entropy" and proved any local optimum under it falls inside the core, yielding a polynomial-time algorithm for computing a fair committee.
- The 20-page arXiv paper (arxiv.org/pdf/2609.11912) credits the discovery explicitly: "The voting rule we present and the proof that it satisfies core+ were found by GPT-6 Astra."
Why it matters
- This is the first time an AI has solved a problem rated as a major mathematical breakthrough, crossing from tool to discoverer of new mathematical structure.
- The result is not just theoretical: the polynomial-time algorithm can be directly implemented in real-world ranked-choice and approval-voting systems, with immediate engineering relevance.
- Oxford researcher Dominik Peters called the proof "elegant," noting it avoided brute-force case analysis and that "the techniques used here may well be applicable to other models as well," signaling broad downstream impact.
- The harmonic entropy framework improves on prior approaches (Shannon entropy, Lindahl equilibrium) that only handled fractional committees, making this the first clean solution for integer committee selection.
- The result reframes what frontier AI benchmarks are for: FrontierMath set this as a find-the-counterexample task, and Astra invalidated the question itself, a qualitatively different capability than benchmark optimization.
What to watch next
- Whether peer review confirms the harmonic entropy proof holds, and whether the technique generalizes to other open problems in cooperative game theory and social choice.
- How FrontierMath and similar benchmarks respond: if AI can dissolve benchmark problems rather than answer them, the benchmarks themselves need redesign.
- Whether OpenAI or collaborating institutions publish follow-on work applying harmonic entropy to adjacent voting and mechanism-design problems, which Peters explicitly flagged as likely.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.