On October 30, PostCutoff — a website tracking AI-assisted mathematical results — completed a tally: over October 29–30, arXiv carried 374 papers mentioning AI. Editors reviewed every AI declaration, closely read roughly 35 PDFs, and picked out six results listed as independent entries. Two days, six results — and behind each stands a specific model and a specific human author, with the division of labor differing each time.
📜 Six Submissions, Six Kinds of Division of Labor
| Problem | Submitter | AI's role | arXiv | |---|---|---|---| | Murty-Simon conjecture | Das & Fox | GPT-5.6 Sol found the upper bound | 2610.39032 | | C*-algebra spectral isometry conjecture | Colbrook & Stepaniants | GPT-6 Astra produced proof, Lean verified | 2610.40470 | | Online Steiner forest greedy ratio | Related team | GPT Astra involved, Lean verified | 2610.40569 | | Laczkovich Lipschitz mapping problem | Gaspar | ChatGPT supplied key proposition | 2610.38834 | | Counterexample to pure semisimplicity conjecture | Ren | GPT-6 Astra construction, Lean verified | 2610.39883 | | Petty conjecture in dimension five | Tencent hyprover agent | Agent as lead prover | 2610.39967 |
Four of the six passed the gate of Lean formal verification. The spectral isometry conjecture has the most complete backstory: whether a linear map between C*-algebras that preserves spectral radius must be a Jordan homomorphism — conjectured by Mathieu and Schick in 2002, with special cases (e.g., commutative algebras) long known. Colbrook and Stepaniants of Cambridge submitted proofs for all single algebras on October 29. Their paper states explicitly that it is the human-digested version of a proof found by GPT-6 Astra "through a series of simple prompts"; the two studied the argument with ChatGPT 6 and wrote the paper themselves. The Lean proof had been registered on Palomar a day earlier.
> Tip: Lean is a proof assistant — mathematical proofs are written as code and checked line by line by a computer. Passing Lean means the chain of reasoning is airtight within the formal system, but whether the ideas behind the proof are deep and the theorem is important is still something only the human mathematical community can judge.
🔍 The Two Solo Entries
Two entries stand out. One is single-author: Xinheng Lin of Fuzhou University, whose 17-page paper proves the Alon-Seymour-Thomas separator conjecture — the O(h·√n) upper bound proposed in 1990. The author declares the proof was completed under his guidance with ChatGPT 6 Pro, with every formula and theorem checked by hand, and GPT-6.1 Sol's Max tier used for repeated verification. The PostCutoff editorial position was blunt: single author, 17 pages, a famous conjecture — treat it as a claim until experts weigh in. The other is a pure counterexample: Ren used GPT-6 Astra to construct a counterexample to the pure semisimplicity conjecture, also Lean-verified.
Setting these two submissions beside the four collaborative ones, the spectrum of division of labor in AI for Math has fully opened up: from "model finds key lemma, human organizes the proof," to "model produces the full proof, human digests and verifies," to "agent independently listed as author" — three months ago the mainstream was still at the first tier. Beyond speed, the more notable development is the unification of verification: Lean is becoming the common referee for this wave of results. Four of the six entries carry machine endorsement; the two that did not pass remain suspended in claim status, awaiting peer adjudication.
🌫️ Another Debate, Same Week
Mathematicians' attitudes toward this wave also came to a head that same week. Fields medalist Deng Yu publicly commented on AI companies' anti-distillation clauses: you cannot support distillation only when you are distilling all of humanity's knowledge. Meanwhile, at a Stanford talk, Sam Altman said mathematicians' sentiment has shifted from "we're finished" to acceptance of AI. Taken together, the two remarks are two perspectives on the same transformation: from the model side, mathematics is one of the last bastions of symbolic reasoning, and conquering it carries demonstration value far beyond any single theorem; from the mathematicians' side, the scarcity of proofs is collapsing, and the remaining questions are taste and problem selection.
One easily overlooked detail in the tally itself: the 374 papers only count those "mentioning AI" — the true base should be larger, since not every author declares AI use. The six results are a lower bound, not the full picture. The mathematical community's review machinery has not caught up with this pace. How journals define AI disclosure policies, whether Lean verification counts as journal-recognized, and how eligibility for AI-involved authorship is adjudicated are all unresolved blanks. The last week of October offered no answers — it simply copied the problem statement again, in bigger letters.
---
Sources and caveats: PostCutoff's October 30 tally (written by the editorial team with Claude Opus 5.5 in Claude Code, signed by human editor Adam Bicz, self-labeled "awaiting review"); the six papers' arXiv abstract pages; Deng Yu's and Altman's remarks as relayed by a Chinese AI newsletter on October 30, with the original talks not verified. The separator conjecture proof has not undergone expert review, and the single-author declaration has not been independently verified; "374 papers" reflects that site's own counting methodology.