Eight extremal combinatorialists write up GPT-6 Astra’s proof of the Erdős–Simonovits rational exponents conjecture (Conlon, Janzer, Liu et al.)
Result confirmed
Importance: major (4 of 5)The takeaway
The rational exponents proof that GPT-6 Astra found during Epoch AI’s FrontierMath Erdős runs now has a human-written paper by leading experts in the area, who explain and situate it and call it one of ‘the recent AI-driven breakthroughs’.
Status
- Claim
Result confirmed
- Our reporting
- High confidence
- Verification
- Lean-verified
- Importance
- Major (4 of 5)
- Last verified
- 14 October 2026
Your AI and this story
- GPT-6 Astra166 days after its cutoff
- Claude Opus 5.5105 days after its cutoff
- Gemini 4 Argon196 days after its cutoff
- Grok 4.7135 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 105 days before it.
Key facts
- Paper: arXiv 2610.18713 [math.CO], 13 Oct 2026, 13 pages; authors from Caltech (Conlon), EPFL (Janzer), Miami University (Jiang), KAIST (Kim), KIAS (Lee), IBS Daejeon (Liu) and Emory (Longbrake, Yepremyan)
- Theorem 1.2: for every rational r in [1, 2] there is a graph H with ex(n, H) = Θ(n^r), the Erdős–Simonovits rational exponents conjecture
- Framing: ‘we wish to explain and situate a proof of the conjecture for all exponents claimed by Epoch AI [1], obtained during benchmarking for the recently released model GPT-6 Astra.’ Reference [1] is Thomas Bloom’s erdosproblems.com/571 page
- The authors name four sources the proof builds on: the Bukh–Conlon lower bound, an Erdős–Simonovits result on powers of rooted graphs, the Kang–Kim–Liu subdivision conjecture, and a technique of Conlon, Janzer and Lee
- Key technical result (Theorem 1.3): a variant of the Kang–Kim–Liu subdivision conjecture for the graphs H_k (each edge replaced by a path of length k plus two apex vertices), which lets one ‘percolate’ from the exponent 1 to all rational exponents
- Before the AI proof, the conjecture was known for exponents 1 + a/b with b > a² and 2 − a/b with b ≥ (a − 1)², among other ranges
- Still open, per the concluding remarks: the Bukh–Conlon conjecture on rooted powers of trees and the Kang–Kim–Liu subdivision conjecture, both of which would imply the result
- AI use in this paper: ‘We used ChatGPT Pro to draw the figures in this paper and to proofread an early draft.’
What happened
In September, Epoch AI and Thomas Bloom reported that GPT-6 Astra, in larger-budget runs outside the FrontierMath Erdős benchmark, had proved the Erdős–Simonovits rational exponents conjecture (#571) with a Lean proof (see the FrontierMath Erdős entry). The benchmark report called the informal write-ups placeholders until experts prepared proper papers.
On 13 Oct 2026 eight extremal combinatorialists posted that paper, arXiv 2610.187131. Several of them wrote the earlier partial results the proof builds on, including the Bukh–Conlon lower bound, the Kang–Kim–Liu subdivision conjecture and the Conlon–Janzer–Lee technique. The abstract says: “We present a proof of this conjecture claimed by Epoch AI, explaining how it builds upon ideas and methods developed by many researchers.” The introduction adds: “As with many of the recent AI-driven breakthroughs, this advance draws heavily on ideas and methods developed by many researchers over several years.”
The core is Theorem 1.3, a variant of the Kang–Kim–Liu subdivision conjecture for graphs H_k (each edge of H replaced by a path of length k, plus two vertices joined to the two sides of the bipartition). Combined with the Erdős–Simonovits result on powers of balanced rooted graphs and the Bukh–Conlon lower bound, it reaches every rational exponent starting from the exponent 1. The two stronger conjectures (Bukh–Conlon on rooted tree powers, Kang–Kim–Liu on 1-subdivisions) remain open.
The paper’s own AI statement covers only figures and proofreading by ChatGPT Pro. No Hacker News thread was found on 14 Oct (Algolia search); X was not searched directly.
Why it matters
It is the second wave of one of the strongest AI-found results of 2026. Experts in the area accept the proof, rewrite it for humans and place it in the literature, while making the point that the model built on their programme. Its verification moves from a Lean certificate plus placeholder write-up to an expert-written preprint.
Sources
3 sources from 3 sites. Numbers match the chips in the text.
3 sources: 2 primary, 1 reaction
Primary
- arXiv 2610.18713: A proof of the rational exponents conjecture (Conlon, Janzer, Jiang, Kim, Lee, Liu, Longbrake, Yepremyan)arxiv.org, paper
- GitHub: tadamcz/erdos571 (Lean proof from the Epoch AI run)github.com, code
Reactions
- erdosproblems.com: Problem #571erdosproblems.com, discussion
Changes
- Filed from the arXiv paper (arXiv AI-disclosure list)