Rota’s 1970 unimodality conjecture for matroid flats disproved, with ChatGPT 5.6 Pro suggesting a key construction
Result confirmed
The takeaway
In July–August 2026 three arXiv papers overturned classic matroid conjectures with AI help.
Status
- Claim
Result confirmed
- Our reporting
- High confidence
- Verification
- Preprint only
- Importance
- 3 of 5
- Last verified
- 30 September 2026
Your AI and this story
- GPT-6.2 Astra54 days after its cutoff
- Claude Opus 5.524 days after its cutoff
- Gemini 4 Hydrogen115 days after its cutoff
- Grok 4.754 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 24 days before it.
Key facts
- Rota’s conjecture (1970): the Whitney numbers of the second kind (number of flats of each rank) of every matroid form a unimodal sequence
- arXiv 2607.02208 (Matt Larson, Jul 2, 2026): counterexamples to Mason’s conjecture (log-concavity of flat counts; e.g. W_74^2 < W_73·W_75 for a graphic matroid of a generalized theta graph with three 26-edge paths and one edge) and to White’s conjecture on symmetric exchanges (a rank-9 binary matroid on 18 elements)
- Larson’s AI use: ‘I prompted ChatGPT 5.5 Pro to look for a counterexample around 20 times’; the paper also credits Claude Opus 4.8
- arXiv 2607.22515 (Alexander Divoux, Chayim Lowen, Shouda Wang, Jul 24, 2026; 8 pages): ‘We give counterexamples to Rota’s 1970 conjecture’; built from Larson’s matroid via Whittle’s q-lift
- AI role in 2607.22515: ‘We used ChatGPT 5.6 Pro to search for a counterexample to this weaker property. Based on a suggestion of the authors, ChatGPT 5.6 Pro returned a non-convex example built using a q-lift’; the paper was ‘written entirely by the authors’
- arXiv 2608.07342 (Divoux, Larson, Lowen, Wang, Aug 7, 2026): flat counts can have arbitrarily many peaks
- Listed in Table 3 of the AI4Math survey arXiv 2608.24961 as a conjecture disproved with ChatGPT 5.6 Pro
What happened
Larson first used ChatGPT 5.5 Pro, prompted around 20 times with hints (for example, look at matroids realizable over F5 and at failures at high indices), to find a graphic matroid that breaks Mason’s log-concavity conjecture. Divoux, Lowen and Wang then asked ChatGPT 5.6 Pro to look for an example of a weaker property; it returned a q-lift construction, which the authors turned into counterexamples to Rota’s unimodality conjecture. All four then showed the flat-count sequence can have many peaks.
Why it matters
Rota’s conjecture was one of the best-known open problems about matroids. The chain is a clear example of the 2026 pattern: AI systems search for counterexamples, and humans generalize and write up the results, with the AI’s role stated precisely in the papers.
Sources
4 sources from 1 site. Numbers match the chips in the text.
4 sources: 4 primary
Primary
- arXiv 2607.22515: Matroid flat counts are not unimodalarxiv.org, paper
- arXiv 2607.02208: Counterexamples to two conjectures about matroids (Larson)arxiv.org, paper
- arXiv 2608.07342: Matroid flat counts can have many peaksarxiv.org, paper
- arXiv 2608.24961: The Gold Rush in AI4Math (Table 3)arxiv.org, paper
Changes
- Filed