Talagrand's Ultimate Matching Conjecture
Let be independent uniform points in and let be evenly spread (each in its own rectangle of area and diameter ). Talagrand's Ultimate Matching Conjecture (Research Problem 17.1.2 of his 2021 monograph) asks: for exponents with , is there a constant such that with high probability one permutation satisfies for both coordinates , with read as ?
- Result
- Proved(see note)
- Status
- Candidate (review pending)
- AI contribution
- AI co-developed
- Method
- Argument
- Field
- probability theory, random matching, optimal transport
- Posed by
- Michel Talagrand
- Year posed
- 2021
- Years open
- 5y
- Solved
- 2026-11-11
- Model
- GPT-6, Claude Fable
- Vendor
- OpenAI; Anthropic
- Collaborators
- Michael Goldman, Francesco Grotto, Dario Trevisan
- Verification
- Unreviewed
- Publication
- Preprint
- Significance
- 34 / 100
- Disclosed cost
- —
- Wikipedia
- No dedicated article
What was actually shown
Theorem 1.1: for finite with (including the quartic case ), with probability tending to one a single permutation satisfies the conjectured bound in both coordinates, for an evenly spread deterministic configuration and also for two independent samples; subcritical exponents follow. Theorem 20.1: at the hard endpoints and the normalized optimal matching cost diverges in probability, so the conjecture fails there. The paper thus answers the conjecture in every case: true for finite critical exponents, false at the endpoints. Exact endpoint asymptotics are claimed only in the repository's endpoints.pdf, not in the paper.
What the AI did
Per the paper's 'Use of AI tools' section and the Hexagon provenance statement: several weeks of interaction with LLMs produced a first proof in the subcritical case; a later round produced the critical finite-exponent case (finite-exponents.pdf in the GitHub repository), and GPT-6 then obtained the endpoint asymptotics (endpoints.pdf). The Hexagon manuscript itself was mostly generated and proof-checked by LLMs while the authors worked to understand those versions; the authors say they have an overall understanding and plan a fully digested version. Hexagon lists OpenAI ChatGPT and Anthropic Claude Fable as tools.
Verification
No independent mathematician has checked this yet. Checked here on 15 November 2026: the Hexagon PDF (2611.00294v1, dated 10 November 2026, three named authors at CNRS/Universite Gustave Eiffel and Pisa) and its abstract page; Theorem 1.1, Definition 10.1 and Theorem 20.1 read against the paper's own restatement of Research Problem 17.1.2 (Talagrand's book itself was not opened); the AI disclosure is in the paper and on the Hexagon page, which labels the text 'Primarily AI-generated' with 'Human understanding: some parts'; the GitHub repository README lists the undigested AI-generated derivations. The proof (about 70 pages) was not checked. Hexagon is a moderated but not peer-reviewed repository; acceptance there is not a judgement of correctness.
Sources
Submitted by MellowGander562 on