GPT-6 Astra produced explicit mutually orthogonal Latin-square constructions that improve known lower bounds for N(n) at 33 values of n from 22 through 494. The standout construction gives four MOLS of order 22, completing the last open case needed to show N(n)≥4 for every n>10. The same paper reports resulting bounds at other orders and derives a strongly regular graph with parameters (484,126,40,30) from the order-22 construction; these connected consequences are counted once as one paper-level result.
Technology: OpenAI GPT-6 Astra, used through ChatGPT Pro and the OpenAI Codex harness
Family GPT-6: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: GPT-6 Astra produced explicit mutually orthogonal Latin-square constructions that improve known lower bounds for N(n) at 33 values of n from 22 through 494.
- Still unknown: This is an arXiv preprint. The constructions are finite combinatorial objects, checked by code written by the author and separately reconstructed/checked by Claude Opus 5.5 as reported in the paper. Counted once despite 33 parameter values and multiple corollaries.
- AI role: The author states GPT-6 Astra found all listed MOLS constructions and wrote their descriptions/code. The author wrote independent verification code for each construction; the paper also reports an independent Claude Opus 5.5 recreation and verification from the descriptions without the supplied code.
Evidence sources: https://arxiv.org/html/2611.16299
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The paper proves that every unital surjective complex-linear spectral isometry between unital C*-algebras is a Jordan isomorphism; without unitality, such a map is a central-unitary multiple of a Jordan isomorphism. The proof establishes that entire one-parameter groups of unital spectral isometries are algebra automorphisms, then reduces the general case to the commutative case. The authors report that the result was formally verified in Lean.
Technology: OpenAI GPT-6 Astra
Family GPT-6: profile · models page
Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: prospective or independent check
- Tested: The paper proves that every unital surjective complex-linear spectral isometry between unital C*-algebras is a Jordan isomorphism; without unitality, such a map is a central-unitary multiple of a Jordan isomorphism.
- Still unknown: ArXiv preprint. The authors explicitly describe the article as a human exposition/digestion of an AI-found proof. Their manuscript says formal verification is in Lean; The reported formalization was not independently audited in the source review.
- AI role: Authors state GPT-6 Astra found the proof after a sequence of simple prompts. They then discussed it with ChatGPT 6 to understand the mechanisms, wrote the paper themselves, and used ChatGPT 6 only for minor language editing; the Lean formalization is evidence for proof checking.
Evidence sources: https://arxiv.org/abs/2610.40470
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The paper proves that a uniformly random permutation of length 32n² contains every permutation of length n as a pattern with probability at least 1−exp(−c n/log n), for an absolute c>0 and sufficiently large n. This removes the prior log-log factor from the sufficient-length bound, resolves Alon’s open 1000n² problem, and establishes the correct Θ(n²) order.
Technology: OpenAI GPT-6 Pro
Family GPT-6: profile · models page
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The paper proves that a uniformly random permutation of length 32n² contains every permutation of length n as a pattern with probability at least 1−exp(−c n/log n), for an absolute c>0 and sufficiently large n.
- Still unknown: ArXiv preprint. Authors report checking and digesting the proof; no separate independent validation is claimed here. The result is an asymptotic sufficient bound with constant 32 (the paper explains the constant can be reduced to any value greater than 1), not a sharp threshold constant.
- AI role: The authors state that GPT-6 Pro found the proof; they do not claim credit for its proof ideas, and say they verified and digested it and take responsibility for the manuscript.
Evidence sources: https://arxiv.org/abs/2610.13530
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The source presents a proof discovered autonomously by GPT-6 Pro establishing an O_d(r^{-1/2}) zero-probability bound for degree-d multilinear polynomials with r disjoint degree-d monomials, improving the prior bound by removing a polylogarithmic factor and resolving the Nguyen–Vu conjecture. The key argument also proves a total-influence estimate for bounded-degree rational functions, resolving a recent conjecture of Kothari, Kovacs-Deak, Wang and Yang; these are one connected paper-level result, not separate entries.
Technology: OpenAI GPT-6 Pro
Family GPT-6: profile · models page
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The source presents a proof discovered autonomously by GPT-6 Pro establishing an O_d(r^{-1/2}) zero-probability bound for degree-d multilinear polynomials with r disjoint degree-d monomials, improving the prior bound by removing a polylogarithmic factor and resolving the Nguyen–Vu conjecture.
- Still unknown: This is an arXiv preprint. The author reports personally verifying and rewriting the proof; this review has not independently formalized it. Bound and related rational-function influence result are counted once because the paper derives them as a connected argument.
- AI role: The author states that GPT-6 Pro autonomously discovered the proof and that the author claims no credit for the proof ideas; the author subsequently verified, digested and rewrote the argument.
Evidence sources: https://arxiv.org/abs/2610.08708
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Zhao Song and Xinzhi Zhang prove a universal, dimension-independent bound for the Kannan–Lovász–Simonovits constant of isotropic log-concave probability measures, resolving the KLS conjecture. Their proof repeatedly refines bounds on polynomial variance and curvature; the authors report that AI tools helped explore more than 100 approaches, obtain the earlier log-star bound, and develop and verify the final proof. Independent researchers also posted proofs, and the paper’s streamlined proof is being formalized in Lean. The 1 October preprint gave a slowly dimension-dependent bound; the 4 October version established the dimension-independent theorem described here.
Technology: GPT-6 Astra; GPT-5.6 Sol; Claude Fable 5; Fable (as named in the paper’s AI disclosure)
Family Claude: profile · models page
Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: prospective or independent check
- Tested: Zhao Song and Xinzhi Zhang prove a universal, dimension-independent bound for the Kannan–Lovász–Simonovits constant of isotropic log-concave probability measures, resolving the KLS conjecture.
- Still unknown: The proof is a recent arXiv preprint, not yet journal peer-reviewed in the source reviewed. Independent proofs were posted, and the Microsoft Research author account reports later Lean formalization of a streamlined proof; the paper version and posted revisions should be checked as the work develops. This entry counts the theorem result once, not the intermediate log-star bound separately.
- AI role: The authors state that they explored candidate approaches collaboratively with AI tools; the tools helped obtain the earlier log-star bound and develop detailed proof steps, while the researchers selected the direction and completed the repeated refinement argument. The authors state they checked the proofs with AI tools and take responsibility for correctness.
Evidence sources: https://arxiv.org/pdf/2610.01447v2 · https://arxiv.org/abs/2610.01447 · https://www.microsoft.com/en-us/research/blog/solving-a-31-year-old-grand-challenge-in-mathematics-the-kls-conjecture/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The authors present a proof of the Linear Hadwiger conjecture: there is a constant C such that every graph with no K_t minor is colorable with at most Ct colors. GPT-6 Astra developed the proof after the authors directed it to the dense case and then asked it to bridge existing bounds; the resulting proof combines several known frameworks with a new bootstrap argument. The result is an asymptotic linear bound, not the stronger Hadwiger conjecture χ(G)≤h(G).
Technology: OpenAI GPT-6 Astra (described in the paper as ChatGPT model Astra 6)
Family GPT-6: profile · models page
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The authors present a proof of the Linear Hadwiger conjecture: there is a constant C such that every graph with no K_t minor is colorable with at most Ct colors.
- Still unknown: The result is an arXiv preprint and is not the full Hadwiger conjecture. It proves existence of an absolute linear-coloring constant for K_t-minor-free graphs. The authors take responsibility for the proof; Independent formal verification was not assessed in the source review.
- AI role: Authors disclose that the proof was developed by the model following their research directions and feedback; the model supplied a new bootstrap argument after being prompted first on a dense case and then to bridge existing bounds. Authors independently wrote the exposition, reviewed the result and take responsibility for correctness.
Evidence sources: https://arxiv.org/abs/2610.05291
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The authors prove that, assuming Zieve-admissible branch data are realizable for three branch points, all such data with four or more branch points are realizable. This establishes that no additional obstruction arises solely from adding branch points; the underlying three-branch-point conjecture remains open. The paper discloses GPT-5.6 Sol and GPT-6 Astra assistance in exploring reduction strategies and in proof development and checking.
Technology: OpenAI GPT-5.6 Sol and GPT-6 Astra
Family GPT-5: profile · models page
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The authors prove that, assuming Zieve-admissible branch data are realizable for three branch points, all such data with four or more branch points are realizable.
- Still unknown: This is a new conditional mathematical reduction theorem, not a resolution of Zieve’s full conjecture: the three-branch-point case remains open. The source is an arXiv preprint; human authors report reviewing AI-assisted work and accept responsibility, but no independent verification of the proof is claimed here.
- AI role: Authors explicitly report that the models assisted throughout the research, including exploration of reduction strategies and proof development/checking, as well as notation consistency and manuscript work. Humans reviewed all AI-assisted content and take responsibility for the proof.
Evidence sources: https://arxiv.org/abs/2609.23663
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The paper gives a negative answer, already for q=2, to Talagrand’s question whether an operator’s Rademacher cotype is universally bounded by its Gaussian cotype and its (q,1)-summing norm. For every sufficiently large power-of-two dimension it constructs a Banach space and operator violating such a universal bound. The same construction also yields a negative answer to Kwapień’s decomposition problem.
Technology: ChatGPT GPT-5.6
Family GPT-5: profile · models page
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The paper gives a negative answer, already for q=2, to Talagrand’s question whether an operator’s Rademacher cotype is universally bounded by its Gaussian cotype and its (q,1)-summing norm.
- Still unknown: This is a single-author arXiv preprint, not a peer-reviewed publication. GPT-5.6 is credited with discovering the counterexample during discussions; the author says they independently verified all arguments. No Lean or other proof-assistant formalization is reported. The mathematical scope is the stated q=2 counterexample and its stated corollary, not a complete characterization of operator cotype.
- AI role: The author states that the counterexample was discovered during discussions with GPT-5.6. The paper presents an explicit Walsh–Hadamard-based construction and proof; the author independently verified the arguments.
Evidence sources: https://arxiv.org/abs/2609.19731
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A counting argument proves that every finite graph with average degree greater than t−2 contains every tree on t vertices, resolving the Erdős–Sós conjecture. The proof counts adjacency-marked vertex orderings and uses an induction on the tree to bound those not supporting a rooted copy. The exposition states the sharp classical theorem and records its proof was discovered by GPT-6 Astra; a Lean repository checks the benchmark formalization of the result. Related expository papers of the same proof are aliases, not separate discoveries.
Technology: Pre-release OpenAI GPT-6 Astra in the FrontierMath Erdős benchmark
Family GPT-6: profile · models page
Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: prospective or independent check
- Tested: A counting argument proves that every finite graph with average degree greater than t−2 contains every tree on t vertices, resolving the Erdős–Sós conjecture.
- Still unknown: The arXiv paper is an exposition of the proof, not the original benchmark output. The linked Lean proof is machine-checked against the benchmark formalization; the registry notes a minor parity-case difference in that compared statement, while the paper’s counting lemma derives the sharp classical average-degree threshold. This remains a preprint and the formalization is not independently audited here.
- AI role: GPT-6 Astra autonomously found the counting proof while solving Erdős problem 548 in the FrontierMath Erdős benchmark; the exposition author reports the discovery and lays out the proof. A Lean 4 repository states the proof was machine-checked by a Comparator against the benchmark’s formal statement.
Evidence sources: https://arxiv.org/abs/2609.17877
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 4/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The paper proves that for every finite family of Euclidean-norm-at-most-one vectors, signs can be chosen so the signed sum has infinity norm below 3√(2π), independent of vector count and dimension. It derives a corresponding square-root-in-sets bound for Beck–Fiala discrepancy. The authors explicitly credit the Odin Automatic AI Research Agent with discovering the proof; a separate elementary proof was posted independently. The algorithmic construction of such signings is a separate follow-up result and is not claimed here.
Technology: Odin Automatic AI Research Agent
Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: prospective or independent check
- Tested: The paper proves that for every finite family of Euclidean-norm-at-most-one vectors, signs can be chosen so the signed sum has infinity norm below 3√(2π), independent of vector count and dimension.
- Still unknown: The result is reported in an arXiv preprint, not a peer-reviewed journal source in this review. A separate elementary proof of the conjecture appeared later, providing an independent proof route; the paper’s AI attribution is author-reported. The result establishes existence, not a polynomial-time signing algorithm.
- AI role: The authors credit Odin with discovering the proof of their vector-balancing theorem; Duke’s official account says the system discovered the proof and human authors scrutinized and organized the argument into the paper.
Evidence sources: https://arxiv.org/abs/2609.11189
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
OpenAI announces a proof that three-dimensional solutions to the Navier–Stokes equations can develop a singularity in finite time and publishes a text formalized in Lean. This is the source's claim: this record does not establish community consensus, correctness of the artifacts, or recognition as a Millennium Prize problem.
Technology: OpenAI internal model
Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: prospective or independent check
- Tested: OpenAI shared an analytical proof and a Lean formalization that an initially smooth fluid at rest can develop a finite-time singularity under a smooth force with finite energy.
- Still unknown: Not stated in the reviewed material
- AI role: An internal OpenAI model, described as more capable than GPT-6 Astra, produced the proof and the Lean formalization.
Evidence sources: https://openai.com/index/navier-stokes-solution/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 5/5 · Validation 4/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The authors prove the classical complex Grothendieck constant exceeds 1.35584631827168, closing more than one quarter of the gap between Davie’s lower bound and Haagerup’s upper bound reported in the paper. The numerical search used Hermite projections and a radial weight; interval arithmetic on exact rational inputs rigorously verifies the stated bound. The authors credit the Odin Automatic AI Research Agent with discovering both the lower bound and its proof.
Technology: Odin Automatic AI Research Agent
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: The authors prove the classical complex Grothendieck constant exceeds 1.35584631827168, closing more than one quarter of the gap between Davie’s lower bound and Haagerup’s upper bound reported in the paper.
- Still unknown: The source is an arXiv preprint. Its stated lower bound is computer-assisted and the proof includes interval arithmetic with exact inputs; the paper does not determine the exact complex Grothendieck constant. The authors’ AI attribution is not independent evidence about the system’s internal process.
- AI role: The authors explicitly state that Odin discovered the improved lower bound and its proof; the preprint describes numerical optimization that found a candidate, followed by exact rational inputs and interval-arithmetic verification by the authors’ supplied code.
Evidence sources: https://arxiv.org/abs/2609.07000
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A proof of Fermat's Last Theorem fully machine-verified in Lean: Claude produced most of it independently over 11 days, with about 13 million lines and 29,500 intermediate theorems, and a comparator confirms that its statement matches Mathlib's. It follows a simplified version of Wiles's proof, uses Mathlib and parts of an Imperial College project, and is probably longer than necessary.
Technology: Claude, Lean
Family Claude: profile · models page
Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: prospective or independent check
- Tested: The proof was checked by Lean using only standard axioms; a comparator tool confirmed the theorem statement matches Mathlib's, and mathematician Kevin Buzzard reviewed it.
- Still unknown: The proof is about 13 million lines and likely much longer than needed; it follows a simplified version of Wiles's proof and is presented as verification, not new mathematics.
- AI role: Dozens of agents worked for 11 days on the Prove2Me platform, producing about 30,300 computer-verified theorems; humans gave occasional high-level instructions.
Evidence sources: https://www.anthropic.com/research/formalizing-fermats-last-theorem
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
OpenAI announces ten results on codes, non-sofic groups, arithmetic circuits, quantum games, lattices, and combinatorics, accompanied, according to the source, by Lean certificates. This record groups one announcement; it does not mean this collection independently verified ten results.
Technology: OpenAI internal model
Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: prospective or independent check
- Tested: Each of the ten arguments was prepared into a manuscript by humans with the model and then formalized by the model in a Lean certificate.
- Still unknown: Not stated in the reviewed material
- AI role: An internal version of Astra found the results; humans prepared the manuscripts with the same model.
Evidence sources: https://openai.com/index/ten-advances-in-mathematics/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 5/5 · Validation 4/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A research project used ChatGPT 5.5 Pro and Codex agents to develop proof ideas and proofs for five Banach-space problems. The resulting paper gives affirmative solutions to questions on toroidal Elton–Odell separation, non-Calkin unital Banach algebras, strict cosingularity and adjoints, weakly compact basis factorization, and primariness of Lp(L1). The authors report that models generated the core mathematics; human mathematicians checked and refined the proofs. The results are presented in an unrefereed preprint, and the authors take responsibility for correctness.
Technology: ChatGPT 5.5 Pro and Codex agents using an automated open-problem search pipeline
Family Codex: profile · models page
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: A research project used ChatGPT 5.5 Pro and Codex agents to develop proof ideas and proofs for five Banach-space problems.
- Still unknown: Unrefereed arXiv preprint. Authors state that humans verified and refined the proofs and are responsible for correctness; the proofs are not formally machine-checked. The five problems were selected by mathematicians, while a separate pipeline component automatically found other candidate problems. The five results remain bundled as one paper-level report.
- AI role: In the pipeline component, agents searched arXiv literature for open problems and attempted solutions. For the five selected Banach-space problems, the authors say model attempts generated the proof ideas and essentially complete arguments before human verification, editing and integration; models were not limited to exposition.
Evidence sources: https://arxiv.org/abs/2607.17388
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Giorgio Parisi and Francesco Zamponi provide an analytic proof that the critical exponents a and b in the full replica-symmetry-breaking scaling solution for infinite-dimensional dense hard spheres obey a+b=1. The identity links gap, contact-force and overlap scaling relations and had previously been known numerically but not analytically. The authors report that Claude Opus 4.7 derived the core proof idea with minimal supervision, with Sonnet 4.6 helping refine it; the authors found and corrected errors in an early draft, then verified and edited the final proof.
Technology: Anthropic Claude Opus 4.7 and Claude Sonnet 4.6
Family Claude: profile · models page
Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: direct test
- Tested: Giorgio Parisi and Francesco Zamponi provide an analytic proof that the critical exponents a and b in the full replica-symmetry-breaking scaling solution for infinite-dimensional dense hard spheres obey a+b=1.
- Still unknown: The theorem is scoped to the fullRSB scaling equations for the infinite-dimensional dense hard-sphere jamming solution; it is not a general proof about every finite-dimensional jamming system. The result is peer-reviewed and published, but proof validation and correction are attributed to the human authors.
- AI role: After numerical investigation, Opus generated the analytic proof strategy and proof draft for the previously unproven scaling identity; Sonnet refined minor steps. The authors report 40 prompts, corrected early inconsistencies and verified/edited the published proof.
Evidence sources: https://doi.org/10.1088/1742-5468/ae7bd7
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
An agent that alternates LLM proof generation with Lean verification autonomously resolved 9 of 353 open Erdős problems and proved 44 of 492 OEIS conjectures, at a cost of a few hundred dollars per problem according to the abstract. Proofs are Lean-verified, so the checking is mechanical; the abstract does not name the system and the results are limited to the problem sets tested.
Technology: LLM-based agent with Lean-based proof verification (system unnamed in abstract)
Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: prospective or independent check
- Tested: Proofs generated by the agent were checked in Lean; the agent resolved 9 of 353 open Erdos problems and proved 44 of 492 OEIS conjectures, at a stated cost of a few hundred dollars per problem.
- Still unknown: The abstract does not describe how the problem sets were selected, and notes the simpler agent cost more on the hardest problems.
- AI role: An agent alternates LLM proof generation with Lean verification to prove open problems autonomously.
Evidence sources: https://arxiv.org/abs/2605.22763
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
OpenAI presents a construction of planar point configurations with more unit-distance pairs than predicted by the Erdős conjecture. The announcement says mathematicians outside the team checked it and provides a complementary explanation; this record checks the source, not an independent reconstruction of the proof.
Technology: OpenAI internal model
Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: prospective or independent check
- Tested: The proof was checked by a group of external mathematicians, who also wrote a companion paper explaining the argument.
- Still unknown: Not stated in the reviewed material
- AI role: A general-purpose internal OpenAI reasoning model produced the proof while being evaluated on Erdős problems.
Evidence sources: https://openai.com/index/model-disproves-discrete-geometry-conjecture/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 5/5 · Validation 4/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Math Inc's autoformalization model Gauss completed the final stages of the Lean verification of Viazovska's proof that the E8 lattice is the optimal sphere packing in dimension 8, working with the human formalization team. The paper describes the human-AI collaboration and the verification in the Lean kernel, with the milestone dated February 2026 in the abstract. Formalization of an existing 2016 result, not a new theorem, and the paper is a progress report.
Technology: Gauss (Math Inc autoformalization model) with Lean
Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: prospective or independent check
- Tested: The completed Lean formalization of Viazovska's dimension-8 sphere packing proof was verified in Lean, with the milestone dated February 2026 in the abstract.
- Still unknown: The paper discusses project objectives that remain, which the abstract page does not list; no other limitations are stated.
- AI role: Math Inc.'s autoformalization model Gauss performed the final stages of the Lean verification, working with the human formalization team.
Evidence sources: https://arxiv.org/abs/2604.23468
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 4/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
OpenAI reports that an internal version of GPT-5.5, with a purpose-built scaffold, found a proof of a known asymptotic result on off-diagonal Ramsey numbers, then verified in Lean. The announcement appeared with the model launch on April 23, 2026; the full proof is in a separate PDF now linked. Page rechecked November 12, 2026: the text matches.
Technology: GPT-5.5 (internal version)
Family GPT-5: profile · models page
Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: prospective or independent check
- Tested: GPT-5.5 found a proof of an asymptotic fact about off-diagonal Ramsey numbers, later verified in Lean.
- Still unknown: Not stated in the reviewed material
- AI role: An internal version of GPT-5.5 with a custom harness found the proof.
Evidence sources: https://openai.com/index/introducing-gpt-5-5/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
An informal reasoning agent (Rethlas) and a formal verification agent (Archon) resolve an open problem in commutative algebra posed by D. D. Anderson in 2014, as a counterexample. The Lean 4 formalization has about 19,450 lines and was produced in roughly 80 agent hours, with the code public on GitHub. Caveat: arXiv preprint; the Lean check covers the formal statement as written.
Technology: Rethlas (informal agent) with Archon (formal Lean agent)
Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: prospective or independent check
- Tested: The informal proof was formalized and machine-checked in Lean 4 for the formal statement as written.
- Still unknown: The page describes the case studies as illustrations of capabilities rather than broad evaluations.
- AI role: Rethlas, an informal reasoning agent, built candidate proofs; Archon, a formal verification agent, translated arguments into Lean 4 projects.
Evidence sources: https://arxiv.org/abs/2604.03789
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The authors prove f(m) = min(m, ceil(2 sqrt(m))) for all m, resolving Erdős Problem #650 on Bloom's site, where f(m) counts disjoint divisibility pairs in an interval for an m-element set. The detailed argument was formally verified in Lean by Harmonic's Aristotle, following a strategy proposed by ChatGPT. Caveat: arXiv preprint; the AI contribution is the strategy and formalization, not an independently checked autonomous solution.
Technology: ChatGPT (strategy) and Aristotle (Lean formalization)
Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: prospective or independent check
- Tested: The detailed argument for f(m) = min(m, ceil(2 sqrt(m))) was formally verified in Lean by Harmonic's Aristotle.
- Still unknown: The abstract lists no explicit limitations; the authors state the exposition and final proofs are entirely human-written.
- AI role: ChatGPT proposed the proof strategy and Aristotle produced the Lean formalization.
Evidence sources: https://arxiv.org/abs/2603.28636
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 2/5 · Validation 4/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
AlphaEvolve-generated search algorithms improved bounds such as R(3,13) from 60 to 61, R(4,16) from 170 to 174 and R(4,19) from 213 to 219. The paper does not describe independent verification; search code is public. arXiv preprint, latest version read.
Technology: AlphaEvolve
Family AlphaEvolve: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: Search algorithms evolved by AlphaEvolve improved nine Ramsey lower bounds; the authors report recovering bounds for all known exact Ramsey numbers and matching best known bounds in many other cases. No independent verification is described.
- Still unknown: Not reported
- AI role: LLM-based code mutation agent that produced the search algorithms behind all results.
Evidence sources: https://arxiv.org/abs/2603.09172
The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A collaborative LLM-plus-symbolic-computation study proved that every n×n Latin square with n ≡ 1 (mod 3) has imbalance at least 4n(n−1)/9. It introduced near-perfect permutations that attain the bound and computationally found independently verified matching examples for every such order from 4 through 52. The universal lower-bound theorem was formally checked in Lean 4.
Technology: Claude Opus 4.5 LLM agent with SageMath, Rust exhaustive solver, simulated annealing and Lean 4 verification tools
Family Claude: profile · models page
Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: prospective or independent check
- Tested: A collaborative LLM-plus-symbolic-computation study proved that every n×n Latin square with n ≡ 1 (mod 3) has imbalance at least 4n(n−1)/9.
- Still unknown: This is an arXiv preprint (v2 dated 2026-08-13), not yet peer-reviewed in the sources checked. The universal proven theorem is a lower bound for n ≡ 1 mod 3. Matching constructions, and hence demonstrated tightness, are computationally established only for n=4,7,…,52; the paper does not prove that the bound is attained for every larger order. Lean 4 formalizes the lower-bound inequality, not the finite construction search or universal tightness.
- AI role: The agent identified an all-even row-distance parity pattern in numerical search data, formulated and helped prove the lower bound, and used model review to catch proof errors. SageMath and a Rust solver supported symbolic analysis and enumeration; simulated annealing found matching constructions through n=52. Human strategic direction supplied the decisive pivot to the minimum-positive-imbalance question. Lean 4 formally verifies the lower-bound theorem itself.
Evidence sources: https://arxiv.org/html/2603.08322v2
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The Aletheia mathematical agent scored up to 90% on IMO-ProofBench Advanced and independently solved four open Erdős questions; work with experts covered 18 research problems, including a disproof of a submodular optimization conjecture. Its score on doctoral-level problems falls to about 38–46%; no result is claimed as a breakthrough, and the papers were submitted to journals.
Technology: Gemini Deep Think, Aletheia agent
Family Gemini: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: Aletheia was evaluated on IMO-ProofBench Advanced (about 90%, human experts graded), an internal PhD-level benchmark (about 46%), a semi-autonomous run over 700 open Erdős problems (four autonomous solutions), and expert collaborations; Level 2 papers were submitted to journals.
- Still unknown: The article states that no Level 3 or Level 4 results are claimed; the abstract states no limitations, and the PhD-level score falls to about 38-46%.
- AI role: Aletheia, a math research agent built on Gemini Deep Think, generated, verified and revised proofs and used search, with varying autonomy across projects.
Evidence sources: https://deepmind.google/blog/accelerating-mathematical-and-scientific-discovery-with-gemini-deep-think/ · https://arxiv.org/abs/2602.03837
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Axiom Prover generated a proof of Fel's conjecture on syzygies of numerical semigroups, verified in Lean 4 with Mathlib, without hints, lemmas, or manual corrections after the run began. This is a case study and an unreviewed preprint, with no data on attempts or time taken.
Technology: Axiom Prover, Lean
Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: prospective or independent check
- Tested: The proof of Fel's conjecture was generated by AxiomProver with no hints, lemmas or manual corrections after kickoff and checked by Lean 4 with Mathlib; the formalization is publicly available and the mathematical results appear in a human-written arXiv preprint.
- Still unknown: The authors state the limiting factor is often the quality of the shared library (Mathlib), and the post describes a single case study with no claim about automating research.
- AI role: AxiomProver translated the statement into Lean, generated intermediate lemmas and assembled the complete Lean proof.
Evidence sources: https://axiommath.ai/research/proof-of-concept · https://arxiv.org/abs/2602.03716
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
GPT-5.2 Pro produced the argument and Harmonic's Aristotle a Lean proof of infinitely many triples with a logarithmic gap in a factorial divisibility condition. The arXiv writeup calls it resolved autonomously, which is the author's claim; no independent review is stated.
Technology: GPT-5.2 Pro and Aristotle (Lean)
Family GPT-5: profile · models page
Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: prospective or independent check
- Tested: The argument produced by GPT-5.2 Pro was formalized as a Lean proof by Harmonic's Aristotle; the proof is machine-checked, and the abstract describes no independent human review.
- Still unknown: The abstract states no limitations and describes only the formal Lean proof as verification.
- AI role: GPT-5.2 Pro produced the argument; Aristotle produced the Lean proof, which the author translated into informal mathematics.
Evidence sources: https://arxiv.org/abs/2601.07421
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
GPT-5.2 Pro solved an open problem in statistical learning theory (learning curves are monotonic in the Gaussian case with known mean and unknown standard deviation), then extended it to higher dimensions. The authors checked the proof, including with outside experts. Reported by OpenAI.
Technology: GPT-5.2 Pro
Family GPT-5: profile · models page
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The authors verified and transcribed the model's proofs of monotonicity of the maximum likelihood estimator's KL divergence for Gaussian and Gamma families, with no outside expert review described.
- Still unknown: The guarantees cover only the listed well-specified parametric settings, and the reverse-KL result relies on a folklore trick; the page shows only the abstract.
- AI role: All results were derived by variants of GPT-5.2 Pro, with humans prompting for further results and verifying the proofs.
Evidence sources: https://arxiv.org/abs/2512.10220
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Ernest Ryu (UCLA) proved why Nesterov's accelerated gradient method, introduced in 1983, remains stable despite the extra momentum that makes it faster. GPT-5's turning point was a rewrite of the equations: not correct as given, but with a structure Ryu then developed and proved himself, over about twelve hours across three days and after a dozen approaches. OpenAI's account says the model created no new mathematical tools and often produced plausible arguments that failed verification.
Technology: GPT-5
Family GPT-5: profile · models page
Publication: Preprint · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: not reported
- Tested: Not reported
- Still unknown: OpenAI states GPT-5 often produced plausible but flawed arguments; its restructuring suggestion was not correct as written. The proof was developed and written by Ryu. Whether peer review will confirm it is not reported.
- AI role: Per OpenAI, GPT-5 suggested approaches and restructured equations; Ryu checked each idea and wrote the final proof. About twelve hours over three days and nearly a dozen approaches. GPT-5 is not a co-author.
Evidence sources: https://openai.com/index/gpt-5-mathematical-discovery/
The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Tested on 67 problems in analysis, combinatorics, geometry, and number theory, AlphaEvolve matched the best known solutions in most cases and improved them in several; in some cases it generalized finite instances into formulas valid for every input. Preprint not yet peer reviewed; source checked was only the team's article.
Technology: AlphaEvolve
Family AlphaEvolve: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: The system was run on 67 problems in analysis, combinatorics, geometry and number theory and compared with the best known solutions. It rediscovered the best known solutions in most cases and improved several.
- Still unknown: The abstract and page state no explicit limitations, and no verification details or failure cases are given.
- AI role: An LLM-guided evolutionary coding agent proposed, tested and refined algorithmic constructions, sometimes combined with Deep Think and AlphaProof.
Evidence sources: https://arxiv.org/abs/2511.02864
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Physics-informed neural networks trained to near machine precision revealed a new family of unstable singularities (blow-ups) in three fluid-dynamics equations, with a clear pattern between blow-up speed and instability order for the incompressible porous-medium and Boussinesq equations. Work with Brown, NYU, and Stanford: a step toward computer-assisted proofs, not a proof; the Navier–Stokes problem remains open.
Technology: Physics-informed neural networks
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: Physics-informed neural networks trained to near machine precision found unstable singularities, and the accuracy was assessed numerically against the governing equations.
- Still unknown: The source states the initial conditions must be tuned with extreme precision, the scaling formula is empirical, Navier-Stokes remains open, and supplementary information is not yet available.
- AI role: Neural networks trained on the equation residuals, with a high-precision optimizer, located new families of unstable singularities, with humans supplying the mathematical insight.
Evidence sources: https://deepmind.google/blog/discovering-new-solutions-to-century-old-problems-in-fluid-dynamics/ · https://arxiv.org/abs/2509.14185
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Math Inc reports that Gauss produced most statements and proofs for a Lean formalization of the strong Prime Number Theorem, with about 25,000 lines and 1,100 theorems/definitions in three weeks. Humans supplied mathematical scaffolding and reviewed strategies, and some files adapt prior work. The date follows the original project blueprint dated 11 September 2025, rather than an inferred announcement timestamp. This verifies a formalization of an existing theorem, not a new theorem.
Technology: Gauss (Lean autoformalization agent)
Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: prospective or independent check
- Tested: A published Lean formalization of the encoded theorem; this review did not compile it.
- Still unknown: The original dated blueprint and published project/verification context were checked; the Lean project was not rebuilt in this review.
- AI role: Gauss generated most statements and proofs with human scaffolding and reuse of prior formalization.
Evidence sources: https://math-inc.github.io/strongpnt/ · https://github.com/math-inc/strongpnt · https://www.math.inc/gauss · https://math-inc.github.io/strongpnt/blueprint.pdf
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: other. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Interpretable decision trees applied to Dedekind-zeta coefficients helped the authors identify patterns in Galois extensions of degrees 4, 6, 8, 9 and 10. They formulated and proved mathematical criteria, including a nonic result: for the stated class, a₁₀₀₀(K) ≤ 4.5 implies the extension is cyclic. This theorem is a specific result; dataset classification accuracy is not a universal guarantee, and the paper reports a counterexample to an apparently perfect classifier branch. The arXiv v1 of 2025-08-08 and the 2026 journal version are one study.
Technology: Interpretable decision trees and logistic regression
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The authors report conventional proofs of criteria motivated by model interpretation.
- Still unknown: The criteria apply to stated classes; model accuracy is data-dependent and a branch with apparent perfect performance has a stated out-of-sample counterexample.
- AI role: Decision-tree patterns in number-field coefficients motivated mathematical criteria that the authors then proved using conventional reasoning.
Evidence sources: https://arxiv.org/html/2508.06670v1 · https://link.springer.com/article/10.1007/s40687-026-00620-3
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 16 Nov 2026. Review scope: Primary arXiv v1 and publisher full text, including result, proof scope, dates and affiliations..
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
An advanced version of Gemini Deep Think solved 5 of 6 problems at the 2025 IMO, scoring 35 out of 42 and working in natural language within the 4.5-hour limit. IMO coordinators assessed and certified the score, while clarifying that they do not validate the system.
Technology: Gemini Deep Think
Family Gemini: profile · models page
Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: direct test
- Tested: An advanced Gemini Deep Think version solved the official 2025 IMO problems within the time limit, and IMO coordinators graded and certified the 35 of 42 score with the same criteria as for students.
- Still unknown: The IMO did not validate the system or model, and this version is not yet public and was first given to trusted testers.
- AI role: Gemini Deep Think worked end-to-end in natural language, exploring multiple solution paths in parallel.
Evidence sources: https://deepmind.google/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
An evolutionary agent based on Gemini and automated evaluators. It recovered an average of 0.7% of Google's computing resources, sped up a Gemini training kernel by 23%, and found an algorithm using 48 multiplications for 4×4 complex matrices. Across more than 50 open math problems, it matched the state of the art in about 75% and improved it in about 20%.
Technology: AlphaEvolve (Gemini Flash and Pro)
Family Gemini: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: Results were checked by automated evaluators, and the agent was applied to over 50 open math problems and deployed tasks, matching state of the art in about 75% and improving it in about 20%.
- Still unknown: The source states the approach applies only to problems whose solutions can be checked automatically, and broader access is not yet available.
- AI role: An ensemble of Gemini models proposed code in an evolutionary loop with automated evaluators.
Evidence sources: https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/ · https://arxiv.org/abs/2506.13131
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
PatternBoost alternates a local search that generates candidate constructions with a transformer trained on the best ones. The authors report the best known solutions to several extremal combinatorics problems, including a counterexample to a conjecture that had been open for 30 years. Candidates are scored mechanically, so each construction is checked by computer; the abstract does not name the conjecture and the paper was not peer-reviewed at the time of posting.
Technology: PatternBoost (transformer trained on search outputs, combined with local search)
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: The abstract reports best known solutions to several extremal combinatorics problems, including a counterexample to a 30-year-old conjecture, but describes no verification procedure.
- Still unknown: The abstract states that PatternBoost's performance varies across problems.
- AI role: A local search generated candidate constructions, and a transformer trained on the best candidates seeded the next search round.
Evidence sources: https://arxiv.org/abs/2411.00566
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The two systems solved 4 of 6 problems at IMO 2024, scoring 28 of 42 points, one point below the gold threshold. Problems were manually translated into formal language, and some required up to three days of computation instead of the official 4.5 hours.
Technology: AlphaProof (Lean) and AlphaGeometry 2
Family AlphaGeometry: profile · models page
Publication: Peer-reviewed paper · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: prospective or independent check
- Tested: The systems solved official IMO 2024 problems, graded under IMO rules by Prof. Timothy Gowers and Dr Joseph Myers, reaching 28 of 42 points; problems were manually translated into formal language first.
- Still unknown: Two combinatorics problems remained unsolved, translation was manual, and some solutions took up to three days.
- AI role: AlphaProof solved three problems with formal proofs, and AlphaGeometry 2 proved the geometry problem within 19 seconds of formalization.
Evidence sources: https://deepmind.google/blog/ai-solves-imo-problems-at-silver-medal-level/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review date: 16 Nov 2026. Review scope: Original institutional announcement checked for the specific results, AI contribution and limitations cited in the antibiotic dossier or mathematical follow-up pilot. Full-paper reanalysis, external replication and an exhaustive follow-up search were not performed..
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 146 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
- publication · 12 Nov 2025: The updated original announcement dates publication of the AlphaProof methodology in Nature to 12 November 2025. This journal transition is not an independent replication and does not apply to every component of the combined system. Source. Recorded 16 Nov 2026.
Follow-up checked 16 Nov 2026 within scope: Bounded check of the original announcements cited by the events recorded here, including explicit links to later program work. This is not an exhaustive literature search or ongoing monitoring commitment..
Directed relationships from this entry; they do not merge catalog counts.
- 2024-01-alphageometry · extends: AlphaGeometry 2 is explicitly described as an improved AlphaGeometry version; this link concerns the geometry component, not attribution of all four IMO solutions. Source
A custom machine-learning approach suggested constructions that authors formalized as two six-colorings. The primary source reports an off-diagonal distance range of 0.354 to 0.657, extending the previously known range. This concerns a specified off-diagonal variant and does not solve the original Hadwiger–Nelson problem.
Technology: Custom machine-learning approach; neural networks; differentiable optimization
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The source reports the specific mathematical or algorithmic result summarized here.
- Still unknown: The result is attributed to the paper; no independent proof review beyond the authors’ source text is claimed.
- AI role: Neural optimization suggests geometric configurations; authors formalize valid colorings and prove the construction.
Evidence sources: https://arxiv.org/abs/2404.05509 · https://proceedings.mlr.press/v267/mundinger25a.html
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review date: 16 Nov 2026. Review scope: Primary preprint full text or abstract reviewed; evidence scope is specified in source_checks..
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A language model proposing constructions and a symbolic engine deriving proofs solved 25 of 30 problems in the IMO-AG-30 benchmark (the previous method solved 10; the average gold medalist score was 25.9), trained on 100 million synthetic examples. It covers geometry only and a 30-problem benchmark.
Technology: AlphaGeometry (language model and symbolic engine)
Family AlphaGeometry: profile · models page
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: On the IMO-AG-30 benchmark of 30 Olympiad geometry problems, the system solved 25 versus 10 for the previous method, with each solution checked by computer.
- Still unknown: It covers geometry only, roughly a third of Olympiad problems, and approaches but does not match gold-medalist performance.
- AI role: A language model proposes auxiliary constructions and a symbolic engine derives proofs, looping until a proof is found.
Evidence sources: https://deepmind.google/blog/alphageometry-an-olympiad-level-ai-system-for-geometry/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review date: 16 Nov 2026. Review scope: Original institutional announcement checked for the specific results, AI contribution and limitations cited in the antibiotic dossier or mathematical follow-up pilot. Full-paper reanalysis, external replication and an exhaustive follow-up search were not performed..
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 494 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
- same-team extension · 25 Jul 2024: DeepMind describes AlphaGeometry 2 as an improved version used for the IMO 2024 geometry problem. The combined result also used AlphaProof; it does not belong to AlphaGeometry alone. Source. Recorded 16 Nov 2026.
Follow-up checked 16 Nov 2026 within scope: Bounded check of the original announcements cited by the events recorded here, including explicit links to later program work. This is not an exhaustive literature search or ongoing monitoring commitment..
No source-supported relationships recorded.
A language model (PaLM 2) paired with an automatic evaluator found the largest cap sets on record in some dimensions, the biggest increase in 20 years, and bin-packing heuristics using fewer bins than established methods. The problem must first be expressed as code with an evaluation procedure.
Technology: FunSearch (PaLM 2 and automatic evaluator)
Family FunSearch: profile · models page
Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: direct test
- Tested: Programs were scored automatically by an evaluator; on the cap set problem they found larger cap sets than the best known, and on online bin packing their heuristics beat established baselines.
- Still unknown: The abstract does not explicitly state limitations; the blog says the method needs an automatic evaluation procedure and that the bin-packing heuristics were tailored to specific data.
- AI role: A pretrained large language model (PaLM 2) generated candidate programs in an evolutionary search guided by an automatic evaluator.
Evidence sources: https://deepmind.google/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06924-6&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 990 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A neural network-guided AlphaZero search with a curriculum generated triangle- and 4-cycle-free graphs that improve known lower bounds for the extremal function ex(n,{C3,C4}) over finite sizes; the proceedings report AlphaZero improvements for n=64–136 and release the graph constructions. The paper also finds improvements using a separate tabu-search method, which often matches or outperforms AlphaZero. This does not solve the asymptotic Erdős conjecture. The result was present in the arXiv v1 posted 2023-11-06; the 2024 IJCAI proceedings paper is a later version.
Technology: AlphaZero reinforcement learning with neural Pairformer representation and curriculum
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The paper reports released graphs with improved lower bounds for ex(n,{C3,C4}) for node counts 64–134. Incremental AlphaZero (neural network-guided Monte Carlo tree search) is explicitly reported to improve bounds for sizes 64–136 in the proceedings version; incremental tabu search also improves bounds and the authors report it can match or exceed AlphaZero. This is a finite-graph construction result, not a proof of the asymptotic Erdős conjecture.
- Still unknown: The result consists of explicit finite constructions/lower bounds; the asymptotic conjecture remains open. The work does not establish that learned search outperforms strong local search.
- AI role: AlphaZero with a neural Pairformer representation and a curriculum generated extremal graphs improving the established lower bounds for n=64–136; tabu search is a separate, non-neural method and its results are not attributed to AI.
Evidence sources: https://arxiv.org/html/2311.03583v1 · https://ijcai.org/proceedings/2024/0772 · https://ijcai.org/proceedings/2024/0772.pdf
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 17 Nov 2026. Review scope: Primary arXiv v1 full text and metadata plus IJCAI 2024 proceedings page/PDF for AlphaZero versus tabu-search graph construction results, theorem limits and author affiliations..
Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A neural-network classifier trained on eight-dimensional, Picard-rank-two toric varieties achieved 95% accuracy on a five-million-example test set and enabled rapid generation of a much larger predicted landscape of Q-Fano varieties. The authors report that the model’s classification pattern led them to formulate a new combinatorial criterion for terminal singularities, which they prove mathematically. The learned model suggested the route to the theorem; it did not supply the proof.
Technology: Feed-forward neural-network classifier trained on toric-variety weight matrices
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The authors report a 95% classification accuracy on a balanced test split of five million eight-dimensional toric-variety examples and use the model to explore a larger Q-Fano landscape. They formulate and prove a new combinatorial criterion for terminal singularities, motivated by the ML analysis.
- Still unknown: The 95% is internal held-out classification performance, not independent external validation; the authors state the neural network does not itself perform the theorem’s divisibility-based test or proof.
- AI role: The trained neural-network classifier revealed a tractable pattern in the classification landscape and directly motivated the authors’ new combinatorial criterion; the authors separately establish its correctness by proof.
Evidence sources: https://arxiv.org/abs/2310.20458 · https://arxiv.org/html/2310.20458v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 17 Nov 2026. Review scope: Primary arXiv v1 metadata and full text for posting date, author affiliations, learned-classifier results and connection to the stated mathematical criterion; linked NeurIPS proceedings identity checked..
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A feed-forward neural network classified the dimension of Fano varieties from quantum-period coefficients with 98% accuracy on the authors’ dataset. The authors report that machine learning helped formulate rigorous asymptotic results, which prove dimension recovery for specified low-Picard-rank toric Fano classes. The broader conjecture that the quantum period uniquely determines every Fano variety remains open; the empirical classification accuracy is not itself a proof. The journal article appeared 2023-09-08, before the same work’s 2023-09-11 arXiv posting.
Technology: Feed-forward neural network
Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: direct test
- Tested: The article combines a dataset classification result with rigorous asymptotic proofs for specified classes.
- Still unknown: 98% is empirical accuracy on the paper’s dataset; the broad uniqueness conjecture remains open. Journal publication preceded arXiv submission.
- AI role: A neural network analyzed quantum-period coefficients and helped identify dimension-related patterns used to formulate asymptotic results, which the authors prove for restricted classes.
Evidence sources: https://kasprzyk.work/research/pdf/ml_dimension.pdf · https://www.nature.com/articles/s41467-023-41157-1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 16 Nov 2026. Review scope: Primary author-hosted full manuscript plus primary publisher article metadata; checked authors, method, theorem scope and chronology..
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The authors used computational data and machine-learning pattern exploration to study affine Deligne–Lusztig varieties. They report rediscovering the previously known virtual-dimension formula, then using the exploration to formulate and prove a new lower bound on actual dimension. The mathematical theorem, not model accuracy or rediscovery of the known formula, is the specific result. The 2024 journal article and 2023 arXiv v1 are versions of one study.
Technology: Linear models, support-vector machines and neural networks
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The paper reports mathematical proof of a lower bound prompted by machine-learning exploration.
- Still unknown: The virtual-dimension formula was previously known and is not the new theorem; the work is not a formal automated proof.
- AI role: Models explored generated mathematical data and highlighted patterns that led mathematicians to formulate and prove a new lower bound.
Evidence sources: https://arxiv.org/html/2308.11355v1 · https://link.springer.com/article/10.1007/s42543-024-00086-8
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 16 Nov 2026. Review scope: Primary arXiv v1 full text and primary journal indexed full text, including methods, theorem, affiliations and version dates..
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A one-hidden-layer multilayer perceptron classified examples of three-strand braids and suggested algebraic criteria. Authors converted the observed patterns into proofs. The paper proves a complete criterion for triviality of flat braids on three strands; its result does not establish a complete invariant for arbitrary braids or the neural network itself as a proof system.
Technology: Supervised multilayer perceptron
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The source reports the specific mathematical or algorithmic result summarized here.
- Still unknown: The result is attributed to the paper; no independent proof review beyond the authors’ source text is claimed.
- AI role: The MLP classified examples and exposed patterns used by human authors to formulate and prove exact invariants.
Evidence sources: https://arxiv.org/html/2307.12185
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review date: 16 Nov 2026. Review scope: Primary preprint full text or abstract reviewed; evidence scope is specified in source_checks..
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Bayesian optimization and reinforcement learning search for sequences of knot moves that certify a knot is ribbon. The primary v1 reports that both knots in 843 specified pairs were shown ribbon, eliminating those pairs as potential counterexamples in the studied strategy. It also recovers a separately independently established set of 1,705 ribbon knots; these are not claimed as novel. This does not resolve the smooth four-dimensional Poincaré conjecture.
Technology: Bayesian optimization; reinforcement learning
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The source reports the specific mathematical or algorithmic result summarized here.
- Still unknown: The result is attributed to the paper; no independent proof review beyond the authors’ source text is claimed.
- AI role: The algorithms search move sequences; a successful path is a verifiable certificate that authors use to prove ribbon status.
Evidence sources: https://arxiv.org/html/2304.09304v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review date: 16 Nov 2026. Review scope: Primary preprint full text or abstract reviewed; evidence scope is specified in source_checks..
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A reinforcement-learning agent found matrix-multiplication algorithms better than known ones for many sizes (for example, 76 multiplications instead of 80 for 4×5 by 5×5). Algorithms optimized for hardware were 10–20% faster on V100 GPUs and TPU v2. The improvement over Strassen applies in a finite field.
Technology: AlphaTensor (based on AlphaZero)
Family AlphaTensor: profile · models page
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Algorithms are provably correct by construction and their multiplication counts are compared with known methods; hardware-targeted algorithms were benchmarked on V100 GPU and TPU v2.
- Still unknown: The page lists no formal limitations; it notes the search space is very large and the work is specific to matrix multiplication.
- AI role: Reinforcement-learning agent searched for tensor decompositions giving new matrix-multiplication algorithms.
Evidence sources: https://deepmind.google/blog/discovering-novel-algorithms-with-alphatensor/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1038/s41586-022-05172-4&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 613 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Classification experiments and analysis of elliptic-curve data show oscillations in the mean of Frobenius traces that vary with rank. A numerical observation, not proof of a general law; the 2022 preprint was later published in Experimental Mathematics in 2024.
Technology: Logistic regression and PCA
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: Classification experiments and analysis of elliptic-curve data show oscillations in the average Frobenius trace that vary with rank.
- Still unknown: Not reported in the arXiv abstract read.
- AI role: Machine-learning classification experiments were used to classify elliptic curves by rank.
Evidence sources: https://arxiv.org/abs/2204.10140
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
With mathematicians Williamson (Sydney) and Lackenby and Juhász (Oxford), supervised-learning techniques revealed patterns leading to an unexpected link between a knot’s algebraic signature and its geometry (the ‘natural slope’), and to a new algorithm for the combinatorial invariance conjecture. The conjectures and proofs are the mathematicians’; the algorithm was checked on over 3 million examples, not proved in general.
Technology: Supervised learning and attribution techniques
Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: direct test
- Tested: Conjectures and proofs were established by mathematicians, and the new algorithm was computationally verified on over 3 million examples.
- Still unknown: Certain patterns still elude ML, and the representation-theory structure appears only at 100s or 1000s of vertices.
- AI role: Supervised learning revealed patterns that led to the knot and representation-theory results and a new algorithm.
Evidence sources: https://deepmind.google/discover/blog/exploring-the-beauty-of-pure-mathematics-in-novel-ways/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1038/s41586-021-04086-x&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 551 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The deep cross-entropy method (reinforcement learning) finds constructions and counterexamples to open conjectures in extremal combinatorics and graph theory, including a Brualdi–Cao question about permanents. The counterexamples are verifiable; source checked was only the arXiv paper, with no institutional announcement.
Technology: Deep cross-entropy (reinforcement learning)
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: The deep cross-entropy method found explicit constructions and counterexamples, verifiable by checking, including a refutation of a Brualdi and Cao question.
- Still unknown: Not reported in the arXiv abstract read.
- AI role: Reinforcement learning (deep cross-entropy method) searched for constructions and counterexamples.
Evidence sources: https://arxiv.org/abs/2104.14516
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.