AI-assisted results, 13 Nov
Awaiting review
The takeaway
The arXiv listings of 12–13 Nov carried about 190 maths papers with an AI statement. Most disclose editing or literature search; about forty credit models with proofs, constructions or counterexamples. Seven got their own entries; the rest are catalogued here with the disclosure quoted. All are unrefereed preprints.
Status
- Claim
Awaiting review
- Our reporting
- Medium confidence
- Verification
- Lean-verified
- Importance
- 2 of 5
- Last verified
- 13 November 2026
Your AI and this story
- GPT-6 Astra197 days after its cutoff
- Claude Opus 5.5136 days after its cutoff
- Gemini 4 Hydrogen227 days after its cutoff
- Grok 4.7166 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 136 days before it.
Key facts
- Reed–Solomon proximity gaps with affine combiners (Alessandro Chiesa, Guy Weissenberg, Eylon Yogev, Kai Zhe Zheng; 2611.151281): combining f + (α + βx)g gives smaller mutual-correlated-agreement error than scalar combiners, relevant to SNARK soundness. Disclosure: ‘The authors then asked ChatGPT 6 Astra to prove an improved proximity-gap result for these combiners and immediately obtained Theorem 1.1. The analysis, including all proofs appearing in this paper, was carried out entirely with AI.’ Lean formalization developed with Astra, theorem statement checked by Claude Code Opus 5.5; Sections 5, 7 and App. A not formalized
- Polynomial bound on split rank (Avinash Bhardwaj, Vishnu Narayanan; 2611.162312): split rank of a rational pure-integer polyhedron is O((L+1)⁹) in its encoding length, ‘resolving a conjecture of Basu, Conforti, Di Summa, and Jiang’; ChatGPT used for literature, ‘the development of induction arguments and checking the encoding bound’
- Online transportation (John Jeang; 2611.170903): competitive ratio ρ(m−1)+1 with ρ ≈ 5.64, improving Seth–Das–Raghvendra’s ≈ 6.66m; the conjectured optimum 2m−1 remains open. AI checklist: approach, proof development, writing and bug-checking ‘Yes’; ‘the idea to have a second objective function F … was a novel idea produced by ChatGPT. Moreover the original draft was entirely created by ChatGPT’ (ChatGPT-6 Astra)
- Balogh–Bollobás–Narayanan conjecture (An Chang, Lusheng Fang, Guorong Gao, Caihong Yang; 2611.154864): every d-regular r-uniform hypergraph has ind(H) ≤ ind(H_d^r)^(n/(rd)), the hypergraph Kahn–Zhao bound; GPT-6 ‘to discuss the choice of local statistics in the proof of Lemma 4.4’ and identified the weighted bipartite inequality of Lemma 3.2
- P ≠ NP over structures with pairing (Mihai Prunescu; 2611.157735): every finite-signature structure with a pairing operation and its two projections satisfies P ≠ NP in the BSS-style unit-cost model, answering the final question of Poizat’s ‘Les petits cailloux’ negatively; ‘This paper was produced in an AI-assisted workflow. AI assistants helped explore candidate constructions, draft proofs’
- Borcherds 1995 / Gritsenko–Nikulin 1998 (Jialin Li, Haowu Wang; 2611.168196): only finitely many strongly reflective modular forms on O⁺(n,2), n ≥ 3, up to conjugation; generative AI used ‘in the development, testing, and refinement of parts of the proof’
- Coercivity conjecture for nonlocal energies (Behn, Colombo, Fernández-Real, Grube, Prasad, Ros-Oton; 2611.166827): fails whenever n > 2s, holds for n = 1, s > 1/2; ‘sustained interactions with ChatGPT versions 5.5, 5.6, and 6 … The models served as sounding boards and contributed mathematical ideas, technical extensions, refinements of proofs’
- Hilbert’s 17th problem for symmetric functions (Sebastian Debus, Robin Schabert; 2611.166908): dimension-free nonnegativity of symmetric functions is undecidable (a question of Acevedo–Blekherman–Debus–Riener), and a degree-8 positive definite symmetric function has no uniform SOS multiplier; ‘Initial proofs of the two main results … were found by GPT 6.0 Astra (OpenAI) in September 2026’; Lean ‘almost entirely generated by coding agents’ (repo); cites Jin’s Codex-prepared manuscript on the same undecidability
Show 7 more
- Other 12 Nov papers: Rohrbach’s conjecture for finite groups (Yong Yang, 2611.169389, ChatGPT ‘for mathematical exploration’); Erdős Problem 521 resolved, N_n/log n has no a.s. limit (Zengjing Chen et al., 2611.1615710; its deterministic step is ‘the paired-sum criterion of ChatGPT 5.5 Pro’ from Kovač’s earlier note); MOLS lower bounds improved at 33 orders, ‘The large language model OpenAI GPT-6 Astra … found all the MOLS constructions’ (Christopher Rosin, 2611.1629911); counterexamples to flagged LLT and nonsymmetric Macdonald atom positivity ‘found by ChatGPT-6 Astra’ (Per Alexandersson, 2611.1689912); Erdős #361 (Erdős–Graham irregularity): exact f_c(n) for c ≥ 1 and limsup for c < 1, from ideas ‘produced with a research pipeline combining the AI systems GPT-5 (OpenAI) and Claude (Anthropic)’ (Mateo Matijasevick, 2611.1720513); a completely decomposable modular Jacobian of dimension 1777, beating Ekedahl–Serre’s 1297 from 1993 (Robert Auffarth, 2611.1696314, AI ‘in exploratory research’); random-order streaming matching 0.71295 with Codex/Astra and Lean (Farhadi, Hajiaghayi, Mahdavi, 2611.1724820); contiguity of random regular hypergraphs, ‘We used ChatGPT to come up with an initial proof’ (Frieze, Pérez-Giménez, 2611.16160); (Δ+2)-colouring reconfiguration diameter (Cames van Batenburg, 2611.17022); a quasi-RH variant over ℚ(i) from OpenAI’s proof, ‘GPT6.1Sol produced the initial proof’ (Xin Hang Ji, 2611.16665)
- Combinatorics, 13 Nov: Sidorenko inequalities for orientations, negative answer to Doležal et al., ‘first used OpenAI GPT-5.6 Sol to prove weaker versions’ (Nie, Wei, 2611.15027); {K2,Kp,Kq}-profile lower boundary, ‘Conversations with ChatGPT Pro 5.6 were instrumental in proving the technical lemmas’ (Blekherman, Raymond, 2611.1614018); small classes have linear neighbourhood complexity, resolving Bonnet–Duron–Sylvester–Zamaraev, ChatGPT Pro (Astra 6) helped repair the argument (Sylvester, Zamaraev, Zhukovskii, 2611.1529917); Das–Abdioğlu triameter conjecture, initial results ‘produced with the assistance of the AI system GPT-Astra’ (Hak, Kozerenko, 2611.17259); smallest great-circle cells ≤ 7.0667/n² and M/m > 13, improving Fejes Tóth’s 7.43, ChatGPT 6 Astra collaboration (Bhowmick, 2611.1524419); critical lower-tail large deviations, a dual certificate found with ‘AI tools and heavy computational interval arithmetic’ (Kwan, Pham, 2610.10233 v2); frustration index of hypercubes with Codex (2611.14906); near-minimum circuits in regular matroids with ChatGPT-5.6 (2611.15755); Grundy values of slow exact k-Nim with ChatGPT (Gurvich, Naumova, 2611.14546)
- Geometry and analysis, 13 Nov: inscribed 4-polytope of maximal volume with ten vertices (Medgyes, 2611.14674, AI ‘for mathematical discussion’); a trimmed Wetzel sector still covers all unit arcs (Panraksa, Wichiramala, 2611.16204, ChatGPT 6.1 Sol and Claude Fable 5.2); 373 extremal disc packings on minimal-genus hyperbolic surfaces, computation with Claude Opus 5.5 (Girondo, Reyes-Monsalve, 2611.15477); MUON optimizer rates, four lemmas ‘developed based on discussions with GPT-5.6 Sol Pro’ (Jentzen et al., 2611.15825); Levitt–Yoccoz gasket dimension with Claude and ChatGPT certificates (De Leo, 2611.15996); Campana rational connectedness of weak log Fano orbifolds with GPT-6 Astra (Chen, Tseng, 2611.15191); first strongly quasidiagonal C*-algebras with finite nuclear dimension and infinite decomposition rank, ‘A discussion with ChatGPT led to Lemma 2.13 which was the missing piece’ (Eckhardt, 2611.1488816); Lagrangian cobordisms and persistence, GPT-6 Astra suggested the Υ∞ retreat (Entov, Polterovich, 2611.14682); global rigidity of right-angled Löbell polytopes, conditions found ‘through further discussions with GPT-6 Astra’ (Ko, 2611.16135); Wasserstein cost of Girsanov, ‘formalization was done in around 6 hours … by GPT-Astra-6’ (Coste, 2611.1604321; Hexagon 2611.00407); random torus slices with Opus 5.5 and GPT-6 Luna (Finch, 2611.15546)
- Physics and logic, 13 Nov: subadditivity conjecture for subentropy settled and strong subadditivity fails for the Nichols–Wootters family, Lean via Codex and Claude Code (Zhang, Zhang, Ma, 2611.1543115); a third counterexample to Grad’s conjecture, analytic with trivial symmetry group, ‘discovered with the assistance of GPT-6 Astra Ultra’ (de Lacy, Sato, 2611.16688; added to Grad’s 1967 conjecture on 3D plasma equilibria falls); Hilbert-tenth companion ‘Why should ℤ be Diophantine in ℚ?’ (2611.17124; see ℤ is Diophantine in ℚ: Glazer (Astra, Claude) and Burungale claim Mazur’s conjecture fails)
- Checked, no AI statement in the PDF (conjecture list, 31 papers): Naor’s isomorphic reverse isoperimetry conjecture, complete solution (2611.1689822); Brown–Shields conjecture counterexample (2611.16997); GAP conjecture reduction (2611.16279); graph a-Turán theorem (2611.16262); packing sets by rigid motions (2611.15402); and 26 others
- Result repos (formal-conjectures, 11–13 Nov): Erdős #190 marked solved with a Lean proof of J. H. Bae’s April 2026 resolution (commit23, Palomar); Palomar links for formal proofs of Erdős #1033 and the Straus variant of #657 (commit, commit); an Isabelle proof of the Sándor variant of #882 (commit); the rest are new statements (Green’s open problems 17, 59, 69, 74, 84, 86, 91, 92, 95, 100; Erdős 297, 739, 1129, 1130; ambidextrous sofa; weighing and Butson–Hadamard conjectures). openai/math, anthropics/formal-math and oai-math-revised had no new commits
- Hexagon backlog (26 lines, 20–21 Oct): Jiangcheng You disproves part (b) of Gromov’s Euclidean C⁰ rigidity conjecture and Sormani’s zero-mass stability conjecture with area-nondecreasing drawstrings (human-written, ‘ChatGPT 6 was employed to discuss potential ideas’, 2610.0043524); ζ₃(5) (3-adic) irrational, ‘produced by large language models … in particular Anthropic’s Claude Opus 5.5 running as autonomous agents’, no human check (2610.0040825); a Frobenius-determinant variant of OpenAI’s Landau–Siegel proof, no human check (van Dijk, 2610.0040926); log_p 2·log_p(3+√2) is not a p-adic linear form in logarithms, ‘a single session of Claude Code Fable’ (2610.0039827); Bergman’s adjoint-tower question answered (Zhenyu Sun, 2610.00417); prime-gap bounds H_m for 2 ≤ m ≤ 5 (Kapila, 2610.00419); certificates for Zhang–Zagier (2610.00426), Schur–Siegel–Smyth (2610.00425) and Korenblum’s constant ≥ 0.432 (2610.00399) by AI agents (Pinho); (3/2)ⁿ steering word with Fable 5 and Opus (Stephan, 2610.00414); four Mislan autocatalysis papers, primarily AI (2610.00437–00440); two Pyle/Lyra graph papers, no human check (2610.00421, 2610.00427); Sun, P-ideals of hypergraphs (2610.00410); mixed physics essays (2610.00411, 2610.00412); human-written: 2610.00400, 00405, 00406, 00413, 00420, 00422, 00423
What happened
The 13 Nov sweep listed 421 arXiv papers with AI statements; about 190 were new in the 12–13 Nov listings (the rest are older papers still uncited). Most statements cover editing, literature search, figures or proofreading; some authors state that no AI was used. About forty credit models with substantive mathematics. Seven results got their own entries (see related), and the Grad entry gained a third counterexample. This entry catalogues the rest, with the disclosure quoted.
Notable patterns: several papers credit a model with the whole main proof (Chiesa et al.: “carried out entirely with AI”; Jeang: “the original draft was entirely created by ChatGPT”), and several with finding counterexamples or constructions (MOLS, flagged LLT, the PDL interpolation counterexample). OpenAI’s October release keeps generating follow-ups by human mathematicians (Hilbert’s tenth problem, quasi-RH, Landau–Siegel variants).
Why it matters
The share of maths preprints with substantive AI credit keeps growing, and the credited roles now include main theorems in active areas (coding theory, data structures, combinatorics). None of the items here is refereed; Lean coverage is partial where it exists.
What is disputed or not yet verified
| Verification | Unrefereed preprints; some with Lean |
|---|
Sources
27 sources from 3 sites. Numbers match the chips in the text.
27 sources: 27 primary
Primary
- arXiv 2611.15128: Affine Combiners for Reed-Solomon Proximity Gapsarxiv.org, paper
- arXiv 2611.16231: A Polynomial Bound on Split Rankarxiv.org, paper
- arXiv 2611.17090: Online Transportation via Shifted Geometric Scalesarxiv.org, paper
- arXiv 2611.15486: Independent sets in regular uniform hypergraphsarxiv.org, paper
- arXiv 2611.15773: A negative answer to the final question in ‘Les petits cailloux’arxiv.org, paper
- arXiv 2611.16819: Finiteness of strongly reflective modular formsarxiv.org, paper
- arXiv 2611.16682: The coercivity conjecture for nonlocal energiesarxiv.org, paper
- arXiv 2611.16690: Hilbert’s 17th problem for symmetric functions: undecidability and a counterexamplearxiv.org, paper
- arXiv 2611.16938: Rohrbach’s Conjecture for Finite Groupsarxiv.org, paper
- arXiv 2611.16157: Resolution of Erdős problem 521arxiv.org, paper
- arXiv 2611.16299: Improved lower bounds for MOLS at 33 ordersarxiv.org, paper
- arXiv 2611.16899: Counterexamples to flagged LLT and nonsymmetric Macdonald atom positivityarxiv.org, paper
Show all 27 sources
- arXiv 2611.17205: Forbidden subset sums and the Erdős–Graham irregularity problemarxiv.org, paper
- arXiv 2611.16963: A completely decomposable modular Jacobian of dimension 1777arxiv.org, paper
- arXiv 2611.15431: Subadditivity and the failure of strong subadditivity in the Nichols–Wootters entropiesarxiv.org, paper
- arXiv 2611.14888: Strongly quasidiagonal C*-algebras with nuclear dimension one and infinite decomposition rankarxiv.org, paper
- arXiv 2611.15299: Small classes have linear neighbourhood complexityarxiv.org, paper
- arXiv 2611.16140: The lower boundary of the {K2,Kp,Kq}-profilearxiv.org, paper
- arXiv 2611.15244: Small cells in great-circle arrangementsarxiv.org, paper
- arXiv 2611.17248: Matching in Random-Order Streams via Historical Deletion Boundsarxiv.org, paper
- arXiv 2611.16043: The Wasserstein cost of the Girsanov Theoremarxiv.org, paper
- arXiv 2611.16898: A complete solution to the Naor isomorphic reverse isoperimetry conjecturearxiv.org, paper
- formal-conjectures commit: Erdős Problem 190 solved, formal proof linkedgithub.com, code
- Hexagon 2610.00435: Area nondecreasing drawstrings and counterexamples to Euclidean rigidity and zero mass stabilityhexagonmath.org, paper
- Hexagon 2610.00408: On the irrationality of the 3-adic zeta value ζ₃(5)hexagonmath.org, paper
- Hexagon 2610.00409: A Frobenius determinant proof of the non-existence of Landau–Siegel zeroshexagonmath.org, paper
- Hexagon 2610.00398: log_p 2 · log_p(3+√2) is not a p-adic linear form in logarithmshexagonmath.org, paper
Changes
- Filed (science agent; PDFs of the cited arXiv papers and Hexagon provenance statements read)