AI-assisted results, 18 Nov
Partly confirmed
The takeaway
The sweep listed 323 arXiv papers with AI statements (208 new on 18 Nov), 68 conjecture papers without one (16 new), 54 Hexagon works, 18 result-repo commits and 1 watched repo. Every new line was read and its AI statement checked in the PDF. Five results got their own entries, six existing entries were extended, and 96 arXiv papers and 31 Hexagon works where AI supplied proofs, constructions, code or Lean are listed here. All are unrefereed.
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 2 of 5
- Last verified
- 18 November 2026
Your AI and this story
- GPT-6.2 Astra171 days after its cutoff
- Claude Opus 5.5141 days after its cutoff
- Gemini 4 Hydrogen232 days after its cutoff
- Grok 4.7171 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 141 days before it.
Key facts
- Method: 208 new lines of ‘arXiv papers that disclose or mention AI use’ (dated 18 Nov) were read and each PDF scanned for its AI statement, since about 70 quoted only a heading; the 115 lines dated 10–17 Nov are re-listings already read and dropped in earlier runs and were not re-read. 3 own entries and 1 extension from this list; 96 listed below; 108 dropped as editorial: about 62 disclose only editing, literature search, proofreading, figures or routine code, about 15 state that no AI was used, and about 30 mention AI only in references, funding or the topic (e.g. LLM quantization, PINNs)
- Conjecture papers without a sweep disclosure: the 16 new PDFs were scanned for model names; none credits AI with mathematics (two mention only proofreading: Abelian varieties with homotheties 2611.22466, toric Frobenius 2611.22992). Boundedness of a Schur multiplier on S_p not implying complete boundedness (2611.22111) goes to the off-scope file. The other 52 lines are re-listings
- Algebra, number theory and geometry: Kisin’s second and Kronheimer–Mrowka’s second deposits (see their entries); tilings of integer sets of prime-power cardinality, ‘the proofs had not been completed until OpenAI’s ChatGPT 6 assisted’ (Hu, Wang, Wang, Zhu; 2611.22849); Flip Combinatorial Invariance for small residues, GPT-6 Astra and Opus 5.5 ‘working in parallel and coordinating themselves’ suggested Theorem 5.1 ‘both its statement and its proof’ (2611.22528); distance algebras of distributive lattices are Gorenstein, key Lemma 3.7 ‘first shown directly by ChatGPT’ (2611.22513) and its companion on distance algebras of graphs (2611.22478); Morse resolutions and lcm powers, ‘Theorem 5.4 was proved by GPT 5.6’ (2611.22173); the Nygaard graded piece, an LLM found the fiber sequence of Theorem 3.7 (2611.22214); totient bound for cyclotomic polynomials ‘written … in collaboration with GPT 6 Astra’ (2611.22795); obstructions to Hadamard deformations of Fourier matrices, Astra with Lean (2611.21871); actions of quantum groups, arguments from ChatGPT 5.6 Sol and 6 Astra (2611.21776); Morse index of homogeneous fibrations, Astra 6.1 calculations (2611.22090); shifted Hankel determinants, an LLM ‘used to help find some of the proofs’ (Guo-Niu Han, Lihong Yang; 2611.21968); Tubbenhauer’s positivity in symmetric webs, old notes turned into a paper and Lean by ChatGPT-6.0 Astra (2611.22058); reflection character varieties, initial ideas for Proposition 6.3 (2611.21700); integral tensor isomorphism, Astra on Lemma 8.1 (2611.23186); associative log rings, an LLM correctly predicted the unconditional Proposition 4.8 (2611.23064); invariant flags and integral matrices (2611.23031); inversion-sequence classes 152 and 207 with Lean by Codex and Claude (2611.21657); resistance-regular graphs need not be regular, Fable 5.1 search scripts (2611.21670)
- Analysis, PDE and geometry: local Mizohata–Takeuchi conjectures for the paraboloid and sphere disproved with power-type lower bounds; ‘An early version of the proof of a special case of Theorem 1.3 was generated by OpenAI’s ChatGPT-5 Sol and ChatGPT-6 Astra’, the final proof differs (Cairo, Fenves, Kovač, Oliveira, Zakharov, R. Zhang; 2611.224651); regularity for complex Monge–Ampère in big classes, ‘The proof was ultimately completed by ChatGPT Sol on September 8, 2026’ (2611.230944); large diameter growth for vortex patches, ChatGPT 6.1 Sol ‘first discovered a four-point-vortex system’ (2611.23059); waist inequality for mean dimension, GPT-5.6 Sol ‘generated key ideas’ (Huo, Ouyang, Shi, Tsukamoto; 2611.22390); TDDFT inverse problem on the torus, ‘crucial ideas of proof … suggested by’ Opus 5.5 and GPT-6.1 Sol (2611.22942); Helmholtz FEM matrix invertible for almost all meshes, Astra proposed the mesh deformation (2611.22676); Yamabe isolated singularities II, GPT-6.1 Sol built the counterexamples (2611.22818); Steklov extremal metrics, Proposition 5.4 and the Section 7 construction from ChatGPT (2611.22168); sharp Gaussian Brunn–Minkowski, Theorem 3.14 from a commercial AI tool (2611.22749); contractibility of geometric joins, Lemma 2 by ChatGPT 6-Pro (2611.22182); pinned simplex volumes, GPT-6 led to improved thresholds (2611.21833); p-harmonic polynomials (2611.23136); symphonic Hodge bounds with Codex (2611.22865); Moser–Trudinger and Monge–Ampère mean field (2611.22876); tangent-point energies (2611.22920); Stokes flow in a wedge, Claude drafted the routine estimates (2611.22008); RCD spaces with Opus 5.5 and Gemini 4 Argon (2611.22383); translation surfaces with log-growing systole, Astra’s proof strategy (2611.21853); compressible 2D transport (2611.23073); crossing limit cycles, Claude on technical lemmas (2611.23230); Hill’s lunar equation, Claude-written computer-assisted proof code (2611.21818); Bak–Sneppen avalanche sharpness on ℤ^d, ‘Human was used at the level of an annoyingly slow co-author’, not to be submitted (Olivier Hénard; 2611.21862); KPZ fixed point edge behavior (2611.22770); Universal 8-vertex origami torus triangulation, ChatGPT 5.6 code and a formula simplification (Schwartz et al.; 2611.21909); Unique continuation for Pauli–Fierz, GPT-6 Astra, Opus 5.5, Fable 5.2 and Lean (2611.23062, also Hexagon 2611.00569)
- Combinatorics and information theory: strong chromatic index χ′_s(G) ≤ 1.642Δ², improving 1.73Δ², GPT-6.1 Sol did calculations and generated the flag-algebra certificates (Runze Wang; 2611.220312); binary non-overlapping codes, Levenshtein’s 1970 bound improved by 5/6, ‘The algebraic combinatorics proof in Section 3 was generated by OpenAI GPT-6 Astra’ (Simon Blackburn; 2611.219133); relatively Sidorenko oriented graphs, ChatGPT 5.5 answered the draft’s question and proved the converse (2611.23222); independent transversals, main ideas of Theorems 15 and 19 from ChatGPT-5.5 (2611.21966); boxes in infinite words, ChatGPT ‘proceeded to give the statement of Theorem 1.1’ (2611.21772); Hales–Jewett and van der Waerden lower bounds by Claude and Kimi searches (2611.22205); Golomb rulers with real marks agree with integer optima up to ten marks, Claude programs (2611.22656); heap-type permutation generation, Claude proposed the induction proofs (2611.22780); cubic graphs with circumference ratio 21/22 (2611.22037); weighted bisections (2611.22760); V-toothpick sequence (2611.23172); zigzag Eulerian polynomials, human core, AI-designed verification code (2611.22110); Hegselmann–Krause cycles (2611.21656)
- Algorithms, complexity and statistics: EF1 allocations with O(n⁵ log² m) value queries, ‘The proofs in this paper were found with the assistance of ChatGPT Astra’ (Dupré la Tour, Igarashi; 2611.22376); #P-hardness dichotomy for exact quantum circuit amplitudes with GPT-6 Astra (2611.22899); sigmoid network gate and wire bounds, Fable 5 found the upper bound that disproved the authors’ own lower bound (2611.21910); Colombo, a Peking University autoresearch system on GPT-5.6 Sol with Lean certification, claims optimal or near-optimal rates in ‘more than fifty settings’ of oracle-based optimization (2611.230805); competitive traveling repairperson below 4, idea from ChatGPT (2611.21943); Metropolis EPTAS on minor-free graphs (2611.22023); COSI inference, ChatGPT 5.6 Sol found the Wilcoxon confidence result (2611.22096); graded Gaussian mixtures (2611.21863); anytime-valid hazard inference (2611.22477); online conformal coverage (2611.22646)
- Smaller or less specific AI roles (one lemma, exploration, derivations): 2611.21844, 2611.21888, 2611.21944, 2611.21724, 2611.21795, 2611.22014, 2611.22468, 2611.22568, 2611.22585, 2611.22603, 2611.22692, 2611.22745, 2611.22799, 2611.22796, 2611.22830, 2611.22870, 2611.22915, 2611.22928, 2611.22932, 2611.22976, 2611.23004, 2611.23117, 2611.23145, 2611.23164, 2611.22463, 2611.22860, 2611.22018; relative computability for nonregular oracles, ‘The idea for the proof of Theorem 1.1 was suggested by GPT-6 Astra’ (2611.22871); equilateral small polygons (2611.23066); matrix-valued log-Sobolev on spheres (2611.21704)
- Hexagon, 54 works: 10 by Dave Goel join his entry; Aronow–Lopatto (AKLT) and Xi Chen (Bourgain) have own entries; Jit Wu Yap’s Mordell candidate joins the Chabauty–Kim entry. Listed: Sebastien Palcoux, the least α with τ_n ≤ αⁿ for all kissing numbers is √6, ‘GPT-6 Astra generated the original mathematical arguments, manuscript, and Lean 4/Mathlib formalization’, revision with GPT-6.2 Astra, Palomar certificate (2611.005866); Fang, Hu, Li and Quanyu Tang, Gundersen’s three-IM-one-CM problem answered in full, ChatGPT proof, Lean with Aristotle and Codex on Palomar (2611.005687); Liam Kruer’s group (Howe, Huffman, Kohlmeyer), five deposits no human has checked: Erdős problems 1171 (2611.005848) and 1156 (2611.00582), the Deshouillers–Erdős exponents 2+2√2 and (5+√17)/2 (2611.00581), the sharp Erdős maximum-term constant (2611.00583) and consistency of ¬Q from a Woodin cardinal (2611.00585); Sihan Chen, convex bodies in ℝ^d (d ≥ 4) with no equipartition of volume by d hyperplanes, ‘the key improvements were contributed by the AI’ (2611.00588); Ryuya Hora, Back–Caragiu’s GPF-Tribonacci question answered by gpt-6-astra and gpt-6.1-sol in Codex at a ZEN University workshop on 14 Nov, Claude Fable 5.2 independently (2611.00579); Jennifer Voitle, self-contained proof of OpenAI’s Subset Sum exponent 0.4875754036 with Claude and GPT-6 Astra cross-derivations (2611.00558); Samuil Petkov, five deposits with GPT-5.6 Sol, GPT-6 Pro, Codex, Claude Opus 5.5 and Aristotle: χ − ζ ≥ 0.063 n/(ln n)³ for G(n,1/2) along all n (Erdős–Gimbel, Heckel; 2611.00562), obligatory triple systems for Erdős problem 593 (2611.00563), WOW-284 refuted (2611.00564), Clark’s repeated-pair SNIEP conjecture in dimension five (2611.00565), Wampler–Barato’s three-state envelope (2611.00566); Mihaly Csirik, Pauli–Fierz unique continuation and a Hohenberg–Kohn theorem at positive coupling (2611.00569); Roy van Rijn, Mathieu–Zhao property classification extended with ChatGPT and Codex (2611.00556); Oleksiy Babanskyy, Zhi-Hong Sun’s Conjecture 3.7 supercongruences (2611.00551, 2611.00550), M(4,3) ≤ 68 for three-sunflower-free families (2611.00552), decycling sets and digit-reversal thresholds (2611.00554, 2611.00555), no human check declared; David Niedbala Giraudin with Claude, upper density of p + 2^k below 0.4893562 and Prouhet–Tarry–Escott 2-adic valuations (2611.00570, 2611.00571); Tiancheng Xu and Zhengtai Xiong with GPT-6 Astra, Codex and Lean on SPR numerators (2611.00553, 2611.00557); Jason Turner with GPT, Svinin’s Conjectures 5.2, 5.3, 3.1 and cluster monomials (2611.00567, 2611.00589); Zhenyu Sun with Codex (2611.00574); Samir Hamad, Daugavet renormings, GPT-6 Astra (2611.00561); Ivan Besevic, Claude-written polynomial-system solver with Lean (2611.00560). Dropped as editorial: 10 (eight human-written with no or figure-only AI use, two with Codex for review or SageMath only)
Show 4 more
- Zenodo (lines in the sweep’s Stories list, not edited there): Steven Minh Phat Duong’s exact divisor densities for Erdős problem 692, n = 4–20, AI-assisted, ‘No general solution … is claimed’ (24338898); about 45 ‘Smithian Fold Theory’ deposits by Maria Smith, including a rebuttal titled ‘OpenAI’s Ten Mathematical Advances Fail the Reality Test’ (24337824), are crank material and dropped; Castro Marín’s Kung–Traub barrier preprint (24338406) has no AI statement and concerns a restricted model of the conjecture
- GitHub: google-deepmind/efficient_evals is evaluation tooling (no result). Result repos: of 18 google-deepmind/formal-conjectures commits, 14 dated 16 Nov were read on 17 Nov; the 4 new ones (#7983, #7926, #7927, #7984) fix statements of Erdős 785, 494, 596 and the Hardy–Littlewood conjecture, no new results. No openai/math or anthropics/formal-math commits
- GPT-6.2 Astra (released 17 Nov): first research uses seen in the sweep are Xi Chen’s NLS claim and Palcoux’s revision; OpenAI’s own maths figure is Epoch’s FrontierMath Erdős score (7 of 68, see the release entry). No separate OpenAI maths results announced with it were found
- Reactions (searched 18 Nov: Tao, Kalai, Shtetl-Optimized, Not Even Wrong, Xena, Mathstodon tag timelines, HN Algolia, MathHelm search, web): none found yet for Reinhardt, Green’s conjecture counterexamples, the Ricci soliton, quartic hypersurfaces, k-Clique or ℤ in ℚ; MathHelm’s threads so far cover OpenAI’s October release papers
What happened
The 18 Nov sweep listed 208 new arXiv papers with AI statements and 39 new Hexagon works. Each new arXiv line was read and the AI statement found in the PDF. Five results got their own entries; the rest of the substantive ones are grouped above by field. Statements are quoted from the papers or from Hexagon’s provenance fields. Hexagon does not judge correctness, and several deposits declare that no human has checked them.
Why it matters
Two of the day’s strongest claims come from senior researchers who say a model did most of the detailed work: Patrick Hayden (Gemini 4 Argon) and Alekseev–Thom (unnamed LLMs). Two others are machine-written or partly checked deposits on long-open problems (AKLT, Bourgain’s NLS question). A cryptographic reduction now builds on OpenAI’s unrefereed 7/8 zero-free region, so the standing of downstream results depends on that release.
Sources
8 sources from 2 sites. Numbers match the chips in the text.
8 sources: 8 primary
Primary
- arXiv 2611.22465: Power loss for the local Mizohata–Takeuchi estimate on smooth hypersurfacesarxiv.org, paper
- arXiv 2611.22031: An improved bound on the strong chromatic index via local flag algebrasarxiv.org, paper
- arXiv 2611.21913: An upper bound on non-overlapping codesarxiv.org, paper
- arXiv 2611.23094: Regularity of solutions to complex Monge–Ampère equations in big classesarxiv.org, paper
- arXiv 2611.23080: Colombo autoresearch systemarxiv.org, paper
- Hexagon 2611.00586: A universal square-root-of-six bound for kissing numbershexagonmath.org, paper
- Hexagon 2611.00568: Meromorphic functions sharing three values IM and one CMhexagonmath.org, paper
- Hexagon 2611.00584: A finite-color partition relation for ω₁² (Erdős 1171)hexagonmath.org, paper
Changes
- Filed (science agent; 208 + 16 arXiv PDFs, 54 Hexagon provenance records, Zenodo records, 18 commits, 1 repo)