Chvátal’s 1972 conjecture proved (Chang–Liu–Liu, ChatGPT-assisted), then a GPT-6 Astra ‘proof from The Book’ and a Codex-built Lean formalization
Result confirmed
Importance: major (4 of 5)The takeaway
On Sept 16, 2026 Fan Chang, Hong Liu and Miao Liu posted a proof of Chvátal’s conjecture (1972): every hereditary family of sets has a largest intersecting subfamily that is a star.
Status
- Claim
Result confirmed
- Our reporting
- High confidence
- Verification
- Lean-verified
- Importance
- Major (4 of 5)
- Last verified
- 7 October 2026
Your AI and this story
- GPT-6.2 Astra108 days after its cutoff
- Claude Opus 5.578 days after its cutoff
- Gemini 4 Hydrogen169 days after its cutoff
- Grok 4.7108 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 78 days before it.
Key facts
- arXiv 2609.19123 (Sept 16, 2026), Chang, Liu & Liu: proves Chvátal’s conjecture and the Friedgut–Kahn–Kalai–Keller correlation form Cov(f,g) ≥ ¼ min_i Inf_i[f] for antipodal g
- AI disclosure (Chang–Liu–Liu): ‘We used ChatGPT to test candidate inequalities… ChatGPT proved the special case of the inequality in Theorem 1.2 when f is a symmetric threshold function, inspiring the authors to pursue and ultimately prove the stronger inequality’; all proofs written and checked by the authors
- Lean: boonsuan/chvatal formalizes the Chang–Liu–Liu proof, ‘generated with GPT-6 through Codex’; registered with Palomar as PALOMAR-
2026-09-17-000004; no sorry or extra axioms in the library; also imported into Lean Pool (PR #443) - arXiv 2609.28404 (Sept 23), Ellis, Filmus & Friedgut, ‘Chvátal’s conjecture: a proof from The Book’: a short direct spectral proof plus a projection-packing-number strengthening
- Their AI chronology: weeks of failed attempts with ChatGPT-6 Astra and Claude Fable 5.1 on their own conjectures (H, I); after they fed in the Chang–Liu–Liu paper, Astra announced a proof that the projection packing number is tight for every downset, then ‘stripped [it] down to the spectral bare bones’; ChatGPT used only for proofreading of the final paper
What happened
Chvátal’s conjecture, posed in 1972, is one of the best-known open problems in extremal set theory. Chang, Liu and Liu (IBS) proved it through a sharp correlation inequality for increasing Boolean functions, and they credit ChatGPT with proving a special case that showed them the way. A Lean formalization made with GPT-6 in Codex appeared the next day in the Palomar registry. A week later, Ellis, Filmus and Friedgut described how GPT-6 Astra had failed for weeks on their related conjectures. Once given the new paper, it produced a short spectral proof, which they present as “a proof from The Book”.
Why it matters
It is a clear example of the late-2026 pattern in mathematics. A human-led breakthrough with model assistance is followed within days by AI simplification and AI formalization. Disclosures are now detailed enough to separate the contribution of each model.
Sources
4 sources from 2 sites. Numbers match the chips in the text.
4 sources: 4 primary
Primary
- arXiv 2609.19123: A proof of Chvátal’s conjecture via a sharp correlation inequalityarxiv.org, paper
- arXiv 2609.28404: Chvátal’s conjecture, a proof from The Book (Ellis, Filmus, Friedgut)arxiv.org, paper
- GitHub boonsuan/chvatal: Lean formalization generated with GPT-6 through Codexgithub.com, code
- Lean Pool PR #443: import Chvátal’s conjecture and sharp Boolean correlationgithub.com, code
Changes
- Science block: short label for the open-problems table
- Filed