AI-assisted results, 20 Nov
Partly confirmed
The takeaway
The sweep listed 361 arXiv papers with AI statements (207 new on 20 Nov), 63 conjecture papers without one (9 new), 69 Hexagon works, 10 watched repos and 7 result-repo commits. Every new line was read and its AI statement checked in the PDF or the Hexagon provenance record. Six results got their own entries (two promoted from the 19 Nov listing), two existing entries were extended, and 97 arXiv papers and 62 Hexagon works where AI supplied proofs, constructions, code or Lean are listed here; 106 arXiv papers and 6 Hexagon works were dropped as editorial. All are unrefereed.
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 2 of 5
- Last verified
- 20 November 2026
Your AI and this story
- GPT-6.2 Astra173 days after its cutoff
- Claude Opus 5.5143 days after its cutoff
- Gemini 4 Hydrogen234 days after its cutoff
- Grok 4.7173 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 143 days before it.
Key facts
- Method: all 207 lines of ‘arXiv papers that disclose or mention AI use’ dated 20 Nov were read and each PDF’s AI statement extracted (about 70 sweep lines quoted only a heading or unrelated text); the 154 lines dated 12–19 Nov are re-listings read in earlier runs. 3 own entries and 1 extension from this list; 97 listed below; 106 dropped as editorial: about 55 disclose only editing, proofreading, literature search, figures, translation, routine code or proof checks, about 24 state that no AI was used or that the results predate it, and about 27 mention AI only in references, funding, the topic or not at all
- Conjecture papers without a sweep disclosure: the 9 new PDFs (20 Nov) were scanned in full; one has an AI line (Mahler measures and a cubic Dedekind zeta value in three dilogarithms, ‘AI-based tools assisted with parts of the computations’; 2611.26076); the others (strong chromatic index for Ore-degree ≤ 9, the random triadic process, Dimca–Sernesi curves, fixed-point-free automorphisms, Auslander bounds, V-variable fractals, minimal models, Chern–Simons couplings) credit no AI. The other 54 lines are re-listings
- Algebra, number theory and geometry: algebraically elliptic nonrational quartic threefolds, ‘the authors directed GPT-6 Astra to search for further examples … This search led to the class considered in this paper’ (2611.25122); F-signature bounds, GPT-6 Astra’s ‘main mathematical contribution was the development of Theorem 4.1’ (2611.25974); uniform positivity of F-signature and pure descent of klt singularities, theorems 2.9, 2.11, 3.7 and 4.2 reached with Claude Opus 5.5 starting from [XZ26, Problem 46] (Ajit, Apsley; 2611.26163); unirationality of supersingular K3 surfaces, the idea of Theorem A ‘first emerged from occasional conversations with ChatGPT 6 Astra’ (2611.26115); Harish-Chandra isomorphism over arbitrary base rings, ‘GPT-6 Astra generated an initial version of the proofs of the main theorems’ (2611.26401); exceptional zeros for GSp(4)×GL(2), Astra 6 and Opus 5.5 developed the coherent period formula and test-vector arguments (2611.25278); Stein degree examples suggested by GPT-6 Astra (2611.25501); homogeneous nilpotent-Jacobian maps in dimension four, GPT-6 Astra for ‘mathematical development’ (2611.25630); Hessenberg cohomology generators found with Claude Opus 5.5 and GPT Astra 6 (2611.26246); v-adic Carlitz polylogarithms, an AI audit of a cited proof fed Proposition 4.11 (2611.25267); D(n)-pairs, two cases eliminated with ChatGPT-5.2 (2611.25270); every rational spherical 4-design on the circle has size divisible by 2304, ‘the AI produced a proof’ (Misawa, Špiranec; 2611.25516); algebraic matroids with non-algebraic duals in every positive characteristic, Theorem 3.3 found with Claude Opus (2611.26429); Schubert calculus, ‘Gemini completed the case classification for Proposition 4.2’ (2611.25956); root-of-unity limits of qKZ connections, ‘developed through a series of conversations with GPT-5.6 Sol’ (2611.26375); a noncommutative-geometry lemma ‘obtained with the help of Claude AI (Opus 5.5)’ (2611.25392); also self-adjoint skew quartics with GPT-6 and Codex finite-field programs (2611.25159), class-number sums (2611.24948), twisted Verlinde formulas (2611.26067), subfactor symmetries of real Hadamard matrices (2611.25616), principal-minor detectors with GPT-6.1 Sol and DeepSeek V4.1 (2611.25715), critical groups with the spectrum of S10 (2611.25766), closed G2-structures (2611.26292), monotone Lagrangians (2611.25699), semi-Fibonacci partitions (2611.25331), an ε-geodesic suggested by ChatGPT (2611.25768), level-2 Eisenstein integrals (2611.26597)
- Analysis, PDE, probability and physics: homogeneous convex non-scattering domains that are neither disks nor ellipses, ‘The main idea of the paper arose from discussions with ChatGPT 6 and Claude Fable 5.2’ (Cakoni, Levitin; 2611.259997); rational maps with at most three critical values have no Herman rings, generalizing Shishikura and Milnor, ‘Part of the proof idea was suggested by ChatGPT’ (2611.25996); Lotka–Volterra cycle exclusion, ‘The key arguments in these parts were proposed by AI’ (2611.26473); Sine_β pair correlations, ChatGPT ‘used extensively to develop the random walk coupling, the pair correlation identity, and the analytic continuation argument’ (2611.26344); lopsided Sylvester solver, ‘Propositions A.1 and 4.1 were suggested, along with their proofs, entirely by Claude’ (2611.25245); KdV analyticity radius, a multiplier bound found in discussions with ChatGPT (2611.25439); disk pseudo-rotations below C², GPT-6 worked out the technical details (2611.25362); MPPI control, a reduction ‘found by this AI model’ (Claude Opus 5.5) and finite-sample results with ChatGPT-6 (2611.26406); also non-Hermitian spectral measures with GPT-6 Astra (2611.25547), a Langevin example by Claude (2611.25961), tanh-network sampling (2611.25867), matrix-point interpolation, proof of Theorem 5.1 filled in by GPT-5.6 Sol (2611.26240), memory-induced blow-up with Claude (2611.25600), compressible Navier–Stokes for γ > 4/3 (2611.25843), Neumann eigenfunctions on corner domains (2611.26403), dark solitons (2611.25343), spectral isometries (2611.26462), wave-envelope estimates (2611.26581), anytime-valid inference (2611.26536), Ritz projection (2611.26099), PINN examples (2611.25515), core–periphery graphs with Claude Sonnet 5.5 (2611.25144), Gaussian-mixture clustering proofs written with Claude Code (2611.25163), a quasi-Newton method with a section-by-section AI ledger (2611.25133), certified quantum phase diagrams (2611.25064), fully connected entanglement (2611.26430), fermion ground energies (2611.26382), bosonic Lindbladian fixed points (2611.25519), the Clifford hierarchy with Fable 5.2’s initial exploration (2611.25025)
- Combinatorics, algorithms and complexity: an improved Moore bound for girths 7 and 11, after ‘more than sixty years’ without a substantial improvement beyond an additive constant, Opus 5.5 ‘used to assist with the research’ on the bounds and the auxiliary-polynomial search (Garbe; 2611.260375); quasipolynomial pyjamas, exp((log 1/ε)^50000) rotations against a triple exponential, ChatGPT 5.6 Pro ‘proposed the reduction to continuous characters to shave an exponential’ (Sida Li; 2611.266226); a 2.8027-approximation for moldable precedence scheduling, GPT-6 Astra ‘derived an initial 3.004-approximation guarantee on September 25, 2026’ (Jansen, Ohnesorge; 2611.25953); twelve running-time lower bounds under FPT ≠ W[1], ‘ChatGPT was used to develop all proofs in this manuscript’ (Molter; 2611.260808); deterministic W[1]-hardness of approximating SVP and MDP, constructions ‘discovered by GPT-5.5’, also citing Guruswami–Ren’s unverified GPT-6 reduction (Hirahara, Lin, Liu, Ogitsuka; 2611.26273); subexponential exact k-means, ‘The core idea of the algorithm was generated by the AI tool’ (GPT-6; 2611.25305); combinatorial Dvoretzky theory, ChatGPT’s ‘contribution was indispensable’ (2611.26253); matroid cups (2611.26602) and distance-unbiased sorting (2611.26135) from interactions with GPT-6 Astra; sorting by block-reversals, AI ‘used interactively to derive and check essential parts of the result’ (2611.25842); also k-rollercoasters with ChatGPT 6 Astra and Claude Fable 5.1 (2611.26297), nonadaptive sparse Fourier transform sample bound optimized with ChatGPT 6 and Claude 5.5 (2611.25334), PSD random submatrices, approach suggested by ChatGPT Astra (2611.25662), edge cuts of bounded degree (2611.26556), Turán densities of daisies (2611.25244), induced bipartite subgraphs (2611.26277), critical graphs (2611.26330), k-cross-free families (2611.26539), a complete intersection theorem (2611.25346), Max-Cut on token graphs (2611.25153), Szemerédi’s distinct-distance bound refined ‘in collaboration with ChatGPT’ (2611.25148), compact packings with ChatGPT-written interval arithmetic (2611.26189), six hypergraph Ramsey lower bounds (2611.25143), thirteen constant-weight codes (2611.24958), certified pandigital approximations (2611.25358), patrol widths with Lean tactics (2611.26140), ID-algorithm examples found with Claude (2611.25877), metacomplexity (only the Theorem 1.8 prime ideal from recent AI discussions; 2611.25046)
- Smaller roles (exploration, checks of computations, code for results): 2611.25182, 2611.25982, 2611.25421, 2611.25698, 2611.25242, 2611.25660, 2611.26354, 2611.26575, 2611.25499, 2611.26582, 2611.25129, 2611.26226, 2611.25940, 2611.26076
- Hexagon, 69 works read through the API (provenance fields): 1 extension (Jaswal’s obtuse lattice triangles, answering Hooper, added to the non-split Cartan entry); 6 dropped as editorial (human-written with no AI or AI only for figures or review: 2611.00573, 00575, 00648, 00611, 00559, 00587); 62 listed below
- Hexagon, named problems: Bianchi’s covariogram problem and Kobayashi’s Fourier zero-set problem claimed for C²₊ convex bodies in every dimension, ‘the AI supplied the critical techniques and detailed calculations’ with GPT-6.2 Sol, author verified key derivations (Sihan Chen; 2611.006799); the Lotoreichik–Lu conjecture (AIM list) that the equilateral triangle is extremal for the lowest Robin eigenvalue for every boundary parameter, by Claude Opus 5.5 and Fable 5.2, ‘Checking so far has been done only by AI systems and by computer’ (Jason Davies; 2611.0069110); the strong AJ conjecture fails for the hyperbolic prime knot 8_11, AI-assisted with partial Lean, human-checked (RJ Acuna; 2611.0068611); Erdős Problem #708, g(4) = 5 and g(5) = 6, Lean via Palomar, author ‘does not claim personal understanding’ (John Tripodo; 2611.0069412); the unfair 0–1 polynomial conjecture for trinomial factors, GPT-6.2 Astra proof and Lean (Baravelli; 2611.00674); t^{log t} transcendental for almost all integers t and odd q-zeta irrationality at real quadratic units, Claude Opus 5.5, no human check (Shawn Krisman; 2611.00695, 2611.00693); plane Jacobian conjecture, residue obstructions excluding the family F₂ of Guccione–Guccione–Valqui (Niedbala Giraudin with Claude, ‘No professional mathematician has reviewed the note yet’; 2611.00685, 2611.00675); Conjecture 1.2(3) of Bandini–Valentino on Drinfeld cusp forms (2611.00607); a candidate proof that c-vectors are roots of the Cartan companion, ‘None of this contribution has been checked’ (2611.00606); convex time-decreasing functions need not preserve semimartingales, ‘Core oscillation construction found by GPT 6 Astra’ (2611.00684); root separation over ℤ and F_q[T], ChatGPT 6.2 Astra constructions and proofs (Tomislav Pejković; 2611.00687, 2611.00692); regular pentagon and hexagon maximize the first Neumann eigenvalue among simple polygons (Dong, Yang; 2611.00690); iso-Delaunay domains with Claude Opus 5.5 (Dutour Sikirić; 2611.00678); linear-size sparsification from self-concordant barriers, GPT-6 Astra Pro (Yilin Liu; 2611.00696)
Show 5 more
- Hexagon, series and smaller deposits: twinheterodox with Codex GPT-6.2-Astra max, eight more irrationality deposits on Dirichlet beta values, Lambert sums and Lucas sums (2611.00600, 00601, 00605, 00615, 00623, 00624, 00638, 00639); Samuil Petkov, eight GPT-5.6 Sol and Claude Opus 5.5 deposits, Claude credited with whole theorems in several (2611.00656–00663); Roy van Rijn, five no-human-check deposits (2611.00593, 00595, 00596, 00672, 00673); William Howe’s group with Codex, nine puzzle and enumeration results incl. all 97 Chaos Tiles duos for Ed Pegg Jr. (2611.00670, 00680–00683, 00688, 00689); Banica and Palcoux with ChatGPT, Grok, Codex and Claude Fable (2611.00603, 00616, 00632); Niedbala Giraudin (2611.00617, 00618); Couret, Lean via Palomar (2611.00676); Ramos, Hulak, de Queiroz, Lean (2611.00677); Arhancet (2611.00592); Babanskyy (2611.00590); Diakin (2611.00604); Fradj (2611.00627); MacGregor, Albert-algebra measurements (2611.00620); Vilela Jato (2611.00655); Hackbarth (2611.00647); Zeng Chunping (2611.00572, 00608, 00621)
- GitHub, watched accounts (10): mahdi-jfri/GraphIsomorphismLog2 got its own entry; the 8 new hajiaghayi repos are zip archives of Lean files for the group’s Astra-assisted papers (Bertrand networks 2611.22216, random-order matching 2611.21294, submodular secretary 2611.17292, dynamic diameter 2611.17050, excluded grid, online survivable network design 2611.16521, prophet secretary 2611.06891, spanning-tree embeddings 2611.05189); google-deepmind/efficient_evals is tooling for an evaluation paper (arXiv 2603.02029), no result. Also found: Hajiaghayi and Yun’s (R*+ε)-approximation, R* ≈ 12.26, for asymmetric TSP and path TSP, down from 14.7 and 15 (no AI statement; 2611.1849813)
- Result-repo commits (7): all in google-deepmind/formal-conjectures (Erdős 625, 790, 792, Kourovka 20.51 statements; fixes to Erdős 494, 596 and Hardy–Littlewood); statements only, no new solutions
- Outside the sweep: the Busy Beaver Challenge’s November report lists a 192-state Turing machine (later 185) that halts iff New Foundations is inconsistent, found with ChatGPT 6 Astra, with a Lean proof that NF has the interpretability strength of NFU+Infinity, ‘novel as of the time of this writing’, the idea ‘invented by GPT Astra 6’ and formalized by Claude Fable 5.2; also a 209-state ZFC machine found with ChatGPT 6 Astra, BB(2,5) holdouts closed in Lean with GPT-6 and Aristotle, and BB(17)–BB(19) bounds found with Claude and Gemini (found via a Mathstodon post by David Roberts)
- Reactions (task: Shallit, McKay–Praeger, Graph Minors, Hayden, Davis): none found on HN, Mathstodon, Shallit’s blog or the web for Shallit, McKay–Praeger, Hayden or Davis by 20 Nov; Ross Kang (Mathstodon, 19 Nov) on the arXiv stream: ‘It feels like every day or two there is some new disappointment’
What happened
The 20 Nov sweep listed 207 new arXiv papers with AI statements and 69 Hexagon works. Each new arXiv line was read and the AI statement found in the PDF; each Hexagon record’s provenance was read through the API. Four results from this sweep and two from the 19 Nov listing got their own entries; the strongest, a deterministic n^O(log n) graph isomorphism algorithm with a full Lean proof, surfaced only through our watch of the authors’ GitHub accounts. The substantive rest is grouped above by field. Statements are quoted from the papers or from Hexagon’s provenance fields. Hexagon does not judge correctness, and many deposits declare that no human has checked them.
Why it matters
Several established researchers (Pak; Jovinelly, Lehmann and Riedl; Carioni and Patrini; Garbe; Cakoni and Levitin) now credit named models with specific ideas or steps in results on long-open questions, while others (Chen, Mao and Zhang; Molter) say the model did essentially all of it. Lean proofs are increasingly attached, including one that rests on the AI-built formalization of the classification of finite simple groups.
Sources
16 sources from 4 sites. Numbers match the chips in the text.
16 sources: 13 primary, 3 reactions
Primary
- arXiv 2611.01399: Deterministic n^O(log n) Graph Isomorphism (Hajiaghayi, JafariRaviz)arxiv.org, paper
- arXiv 2611.25164: Shortest paths on 4-polytopes are hard to compute (Pak)arxiv.org, paper
- arXiv 2611.26245: Mumford-Shah convex lifting gap (Carioni, Patrini)arxiv.org, paper
- arXiv 2611.25068: Globally F-regular varieties are rationally chain connected (Jovinelly, Lehmann, Riedl)arxiv.org, paper
- arXiv 2611.26037: An improved Moore bound for girths seven and eleven (Garbe)arxiv.org, paper
- arXiv 2611.26622: Quasipolynomial pyjamas (Sida Li)arxiv.org, paper
- arXiv 2611.25999: Homogeneous convex non-scattering domains (Cakoni, Levitin)arxiv.org, paper
- arXiv 2611.26080: Running-Time Lower Bounds under FPT ≠ W[1] (Molter)arxiv.org, paper
- Hexagon 2611.00679: Bianchi’s covariogram problem for C²₊ bodies (Sihan Chen)hexagonmath.org, paper
- Hexagon 2611.00691: Equilateral triangle extremal for the lowest Robin eigenvalue (Jason Davies)hexagonmath.org, paper
- Hexagon 2611.00686: The strong AJ conjecture for hyperbolic prime knots is false (Acuna)hexagonmath.org, paper
- Hexagon 2611.00694: Erdős–Surányi g(4) = 5 and g(5) = 6 (Tripodo)hexagonmath.org, paper
Show all 13 sources
Reactions
- Busy Beaver Challenge: TMBR November 2026 (Cryptids)wiki.bbchallenge.org, discussion
- David Roberts on Mathstodon quoting the NF machine (19 Nov)mathstodon.xyz, discussion
- Ross Kang on Mathstodon (19 Nov)mathstodon.xyz, discussion
Changes
- Filed (science agent; 216 arXiv PDFs, 69 Hexagon provenance records, 10 repos, 7 commits; bbchallenge report found via Mathstodon)