What the Machine Invented: Mining New Mathematical Instruments from OpenAI's October 2026 Release of AI-Generated Mathematics

Abstract: On 6 October 2026 OpenAI published an initial 722 AI-written mathematical manuscripts, organized into 372 problem families, in the public repository github.com/openai/math. The release includes claimed solutions to many prominent open problems. We approach this collection as a treasure mine of mathematical ideas. Our aim is to extract new lemmas, constructions, inequalities and proof mechanisms introduced in the source arguments, and to distinguish them from established tools and new combinations of those tools. We call these mathematical instruments. Extraction comes first; their validity, novelty and possible reuse then become subjects for further study. The resulting descriptions also help readers understand the selected papers without having to reconstruct every argument at the outset. We examine 38 families in two groups of 17 and 21, spanning number theory, geometry, analysis, combinatorics, complexity, algebra, operator algebras, topology and mathematical physics. Each entry records the claim, extracted instruments, antecedents and caveats. Eleven cross-cutting patterns connect the entries, and a scoreboard records formalization status and dependencies. Novelty assessments are provisional and sometimes mixed. Selected source comparisons correct mathematical statements, hypotheses and counts; they do not independently establish priority or certify the underlying proofs. No Lean development was compiled or kernel-checked for this survey. We also briefly describe a subsequent application of the signed restriction lemma from family 112 to circuit lower bounds at every fixed depth.

• The paper introduces a structured methodology for extracting and validating reusable mathematical instruments, known as 'mathematical devices', from AI-generated proofs, focusing on 38 prominent problem families spanning multiple mathematical fields.
• The study identified several novel instruments, including a shared-leg entropy inequality and a restriction bound, while distinguishing these from previously known elements ensuring the correct assessment of novelty and applicability.
• The paper demonstrates the importance of separating practical reuse from the verification of AI-generated claims, extracting foundational instruments that can be independently validated and used to advance mathematical research.

Scope and objective

“What the Machine Invented: Mining New Mathematical Instruments from OpenAI's October 2026 Release of AI-Generated Mathematics” (2610.17090) is a survey and methodological study of a large corpus of AI-generated mathematical manuscripts. Its central proposal is to read machine-generated proofs not only as attempted solutions to headline problems, but also as repositories of reusable mathematical devices. The paper calls these devices mathematical instruments: lemmas, constructions, inequalities, auxiliary structures, and proof architectures that may retain value independently of the theorem for which they were introduced.

The source corpus was released by OpenAI on 6 October 2026 and initially comprised 722 manuscripts organized into 372 problem families. The survey examines 38 families selected principally for the prominence of their target problems. These span number theory, arithmetic geometry, analysis, geometry, combinatorics, computational complexity, algebra, operator algebras, topology, symplectic geometry, and mathematical physics. The selected claims include alleged resolutions or major advances concerning the Unique Games Conjecture, the matrix multiplication exponent, Hadwiger's conjecture, the chromatic number of the plane, Hilbert's tenth problem over $\mathbb{Q}$, the Haldane gap, Serre positivity, the complete Crouzeix conjecture, and several conjectures of Kaplansky and Thompson.

The paper deliberately separates extraction from validation. Extraction identifies the mathematical mechanism used by a manuscript and traces its role through the argument. Validation concerns correctness, formalization, prior art, and proof completeness. The survey does not claim that its extraction establishes any headline theorem or mathematical priority.

Methodological framework

The initial extraction assigned each family a separate reading based on the available LaTeX source, PDF, Lean files, challenge files, solution modules, and, in selected cases, abridged reasoning traces. For each family, the procedure generally isolated up to three candidate instruments, followed their dependencies through the source argument, and compared them with relevant literature. The resulting novelty judgments use four categories:

• New: no precedent was found for the object, structure, inequality, or proof architecture, even when its components are classical.
• New combination: the ingredients are established, but their combination or application appears new.
• Incremental: the argument refines or varies an established technique.
• Unclear: the available material did not support a reliable classification.

These categories are explicitly provisional. They mean “new relative to the literature examined during the initial extraction,” not a certified priority claim. The second stage of the project performed selected mathematical checks and corrected theorem statements, hypotheses, counts, and formalization descriptions. It did not repeat every literature search or inspect every proof in full.

The survey's principal organizational contribution is its uniform family-card format. Each card records the headline claim, extracted instruments, established ingredients, formalization status, dependencies, suspected antecedents, and caveats. This format allows a reader to identify whether a purported contribution lies in a genuinely new object, an unusual transfer of a theorem, a quantitative refinement, or merely a recombination of known methods.

Scale and formalization status

The numerical scope of the study is substantial but limited relative to the release:

Quantity	Reported value
Initial manuscripts	722
Problem families in the release	372
Families examined in the survey	38
Families in the first group	17
Families in the second group	21
Original headline-formalization labels	19 of 38
Lean projects compiled and kernel-checked for the survey	0

The figure of 19 formalized headline families requires careful interpretation. It includes partial inspection for families 003 and 087, and some entries were classified as companion-only even though additional source files were later located. The presence of hundreds or thousands of Lean files is therefore not treated as evidence that the corresponding theorem has been independently verified. Challenge files may contain placeholders, while separate solution modules may or may not be complete; neither was compiled during the survey.

This distinction is methodologically important because the release combines mathematical prose, generated formalizations, computational certificates, and informal reasoning traces. The paper argues that these artifacts must be tracked separately. A theorem statement may be formally represented without its proof being complete; a source argument may contain a valid lemma even if its headline theorem fails; and a computational result may rely on floating-point assumptions that are not captured by the abstract mathematical statement.

Principal extracted instruments

The survey's family cards contain many technical constructions, but several illustrate the paper's conception of an instrument particularly clearly.

In family 107, the claimed bound $\omega\le 9/4$ is obtained without producing a competitive finite matrix-multiplication algorithm. The extracted device is a shared-leg entropy inequality for points in Strassen's asymptotic spectrum. A Fourier transform over copies preserves tensor blocks sharing a common tensor leg, after which entropy concavity and a tripling argument force polynomial growth of the spectral profile. The numerical claim is substantially stronger than the previously cited bound $2.371177$, but the result is nonconstructive: it constrains asymptotic rank through spectral inequalities rather than exhibiting a practical degeneration or finite matrix size. The survey also identifies an uncited relation to an Alman--Li template, which complicates the novelty assessment.

Family 112 supplies another sharply defined instrument: a restriction bound independent of clause count for restricted width-$k$ CNFs expressed as signed combinations of width-$b$ CNFs with bounded expected coefficient mass. The normalization eliminates dependence on the number of clauses, while a switching estimate yields correlation decay when the restriction parameter satisfies the relevant smallness condition. This supports an explicit depth-three circuit lower bound of

$2^{\omega(\sqrt n)}.$

The paper emphasizes that this is not a fixed improvement to $2^{n^{1/2+\varepsilon}}$; the rate is described as inverse-tower in $\log n$. A subsequent application manuscript reuses this signed restriction lemma to claim, for every fixed depth $D\ge3$,

$$S_D(f_n)\ge 2^{\Omega_D\left((n\log n/\log\log n)^{1/(D-1)}\right)}.$$

The survey presents this as an example of the transition from extraction to reuse, while noting that the application is separate and its proof is not supplied in the survey.

Family 102 introduces a latent alphabet gadget for the Unique Games construction. Its nonlinear shift-equivariant decoder acts on a larger latent space, and a noise distribution is chosen so that the honest decoder is rarely affected while sufficiently high-rank linear families detect the noise. This permits the degree-two KMS inverse theorem to be used without changing its analytic degree. The claimed Unique Games theorem is accompanied by a substantial formal development, including formalizations of a Dinur-style PCP theorem, Håstad's 3LIN theorem, and the Khot--Minzer--Safra Grassmann expansion theorem. The survey classifies the gadget as new relative to the literature examined, while identifying its tower-like alphabet and reduction parameters as a major quantitative cost.

Family 087's treatment of the symmetric Mahler conjecture illustrates an instrument assembled from complex analysis and convex geometry. A conformal lens is used to replace real strips by complex domains with a uniform boundary coordinate. Lelong-number mass is then pushed toward the boundary through a large-power limit, transforming an analytic Jacobian estimate into a sum of feasibility probabilities for signed simplices. The resulting feasible-simplex identity is claimed to yield the volume-product inequality and, through equality conditions, the Hanner classification. The conformal map itself is attributed to Gross, while the survey regards the rescaling from Lelong mass to feasibility probabilities as the potentially novel component. The general-body argument is treated more cautiously because it depends on numerous numerical certificates and uniform scalar inequalities.

The survey also identifies instruments that are primarily proof architectures. In family 159, triangular polynomial cells with descending precision budgets are intended to prevent the loss of precision associated with repeatedly passing to subprogressions. The loss at one weight depends only on higher-weight precisions, producing a quasipolynomial rather than iterated logarithmic recurrence. The claimed consequence is an improved bound for arithmetic progressions in sets with divergent reciprocal sum. However, the survey states plainly that the quantitative upper bound was not verified and that the reasoning trace ends with unresolved circularity concerns.

In family 268, the coupled purity bootstrap relates two purities of a transfer-matrix partition function: one obtained by doubling inverse temperature and another by doubling spatial length. Exact identities couple these operations, allowing a high-purity regime to propagate across scales. The source claims a uniform Haldane gap for the periodic spin-one Heisenberg chain, with

$\omega\le 9/4$0

for $\omega\le 9/4$1, and a stronger bound exceeding $\omega\le 9/4$2 for $\omega\le 9/4$3. These values are far below the numerical gap near $\omega\le 9/4$4, but the conceptual significance lies in the proposed criterion's absence of frustration-freeness assumptions. The survey stresses that the purity inequality alone does not generate convergence from arbitrary initial data; the bootstrap requires explicit high-purity hypotheses and computer-assisted finite-size estimates.

Cross-cutting proof patterns

Rather than treating the 38 families as unrelated case studies, the paper extracts recurring modes of reasoning.

A first pattern is existence without construction. Several arguments use compactness, ultrafilters, spectral duality, random sampling, or terminating oracle procedures to establish existence without producing an explicit object of practical size. This occurs in the claimed matrix multiplication bound, the Hadwiger construction, the Hilbert's tenth reduction, the Kaplansky counterexamples, and the Borsuk construction. The distinction is consequential: an existential theorem may have strong mathematical content while providing no usable algorithm, finite graph, or explicit parameter bound.

A second pattern is the conversion of estimates into extremal or variational problems. The Kakeya arguments introduce a critical-exponent surface and infer structural properties of hypothetical extremizers from differentiability. The Haldane argument turns purity evolution into a coupled dynamical system. The area-law argument constructs optimized positive filters and interprets conditional mutual information through a derivative of a log norm. In each case, the proof attempts to replace a direct global estimate by a rigid extremal configuration that can be ruled out or controlled.

A third pattern is the use of entropy and information quantities as structural proof devices. Mutual information appears in the Kakeya argument to compare dependent events with product projections. Entropy controls container progress in the claimed Hadwiger counterexample and bounds repeated traversals in the zero-divisor construction. In the matrix multiplication argument, entropy supplies the shared-leg inequality. In the area-law argument, conditional mutual information is represented through a transport derivative. These uses differ technically, but the common strategy is to convert combinatorial or analytic dependence into a monotone scalar potential.

A fourth pattern is rigidity as a substitute for regularity. The plane-coloring argument claims that unrestricted proper colorings can be transferred to weak measurable colorings through Haar rigidity of certain character laws. The Hilbert--Smith construction attempts to work with sheaf categories generated by arbitrary continuous maps while preserving a usable duality theory. The Hilbert's tenth construction replaces an existential definition of $\omega\le 9/4$5 with an oracle reduction whose termination follows from compactness and height bounds. In each case, a rigidity principle is used to remove an otherwise unavailable regularity assumption.

A fifth pattern is the transfer of established results into settings where no precedent was found. Examples include applying the Benyamini--Sternfeld fixed-point-free Lipschitz map to Thompson's group $\omega\le 9/4$6, Braverman's bounded-depth circuit theorem to Chowla correlations, Quot-scheme localization to Bloch's conjecture, and a conformal Skorokhod map to Mahler's conjecture. The paper correctly distinguishes novelty in the transfer from novelty in the imported theorem itself.

Finally, the survey documents rediscovery without citation. It identifies close antecedents involving Cheverry--Ibrahim in the Vlasov--Maxwell argument, Masser and Fischler--Nakamaye in the interpolation proof for $\omega\le 9/4$7, Rios--Sawyer in Kakeya theory, Mourrat--Panchenko and Barbier--Panchenko in the Mézard--Parisi argument, and Suslin--Voevodsky in Serre positivity. These cases demonstrate why instrument extraction cannot be separated from careful historical comparison.

Strong numerical and structural claims

Several headline claims are especially strong:

• The claimed matrix multiplication bound is $\omega\le 9/4$8, improving the cited prior value $\omega\le 9/4$9, but without a finite algorithmic construction.
• The claimed Borsuk counterexample occurs in dimension 9, improving the cited previous range beginning at dimension 64. The witness is a continuous compact set isometric to $2.371177$0, not a finite set.
• The claimed Hadwiger construction produces graphs with independence number at most 2, Hadwiger number below approximately $2.371177$1, and fractional chromatic number at least $2.371177$2, but does not provide an explicit graph.
• The claimed plane-coloring result rules out proper 5-colorings of $2.371177$3 in ZFC, while the proof does not exhibit a finite 6-chromatic unit-distance graph.
• The claimed complete Crouzeix theorem establishes the matrix-valued constant 2 for arbitrary bounded operators, including infinite-dimensional Hilbert spaces, rather than merely the scalar inequality.
• The claimed depth-three circuit lower bound is $2.371177$4, explicitly weaker than $2.371177$5 for any fixed $2.371177$6.
• The claimed Vlasov--Maxwell theorem gives global classical solutions for unrestricted smooth compactly supported data under the stated compatibility conditions, with no smallness assumption.
• The claimed nonamenability of Thompson's group $2.371177$7 is obtained through a five-page argument using a recursive Hilbert-ball coloring and a fixed-point-free Lipschitz map.
• The claimed isomorphism of all free group factors would have major consequences for the Connes--Shlyakhtenko $2.371177$8-Betti conjecture, a consequence not discussed in the source manuscript.

The survey does not treat these numerical strengths as evidence of correctness. Instead, the figures help prioritize verification. Large improvements over established bounds, universal quantifiers, nonconstructive existence statements, and dependencies on other unverified manuscripts are all identified as reasons for intensified scrutiny.

Formalization, dependencies, and reuse

The paper's dependency scoreboard is one of its most practically useful components. It records whether a family has an associated Lean artifact, whether the statement was checked for faithfulness, whether it depends on other OpenAI manuscripts, and whether an uncited nearest ancestor was identified.

Some dependencies are mathematically decisive. The Hilbert's tenth argument depends on the claimed Fontaine--Mazur result at $2.371177$9 and a pointwise 2-converse for elliptic curves. The BSD family depends on companion results concerning Selmer converses, the two-primary case, and auxiliary quadratic fields. The 3D Kakeya argument uses another restriction manuscript as an alternative route. The Haldane and area-law entries depend on computer-assisted intermediate data. Such chains mean that even if an individual manuscript is internally coherent, its headline theorem may remain conditional on several other unverified claims.

The paper also distinguishes formalization coverage from formal verification. A repository may contain hundreds of files, but the survey did not compile any Lean development. It therefore refuses to convert file counts into correctness claims. This is particularly important for the claims involving large generated developments, such as the unrestricted plane-coloring result with approximately 24,500 lines, the Unique Games construction with 475 files, and the hyperinvariant-subspace construction with a 1,108-module import closure.

The signed restriction lemma from family 112 provides the clearest example of reuse. The lemma is abstracted from its original depth-three application and combined with adaptive averaging, depth reduction, and quantitative sparsification to obtain a claimed fixed-depth circuit lower bound. The survey does not prove this extension, but its inclusion demonstrates the intended workflow: isolate a device, restate it independently of its original theorem, test its hypotheses in a new setting, and then establish a separate result.

Limitations and open questions

The study's limitations are substantial and are stated directly. It examines only 38 of 372 families, selected for problem prominence, so it is not a representative census of the release. Literature-search coverage is incomplete, especially in the second group, where some novelty comparisons used directly retrieved sources and existing knowledge without a separate fresh search. Consequently, “new” means only that no precedent was found within the search conducted.

The mathematical review checked selected theorem statements and identities but did not verify every proof. No Lean project was compiled or kernel-checked. Several source manuscripts rely on unverified OpenAI manuscripts, large computational certificates, floating-point calculations, nonconstructive compactness, or parameters too large to execute. The MUB claim in dimension six is explicitly conditional on binary64 computations, compiler behavior, and hand-derived rounding margins. The Borsuk claim concerns a continuous set rather than a finite witness. The Hadwiger construction is nonconstructive and does not output an explicit graph. The Hilbert 16 bound is ineffective. These are not merely expositional details; they delimit the mathematical content of the claims.

The paper also leaves unresolved priority questions. It records close antecedents for Naimark's problem, Crouzeix's conjecture, Serre positivity, and other entries, but does not adjudicate priority. In several cases, the survey identifies a potentially new combination while also noting that an uncited source or related construction may substantially reduce the novelty claim. The principal open methodological question is therefore how much of the apparent instrument-level novelty survives exhaustive literature searches and independent proof verification.

Conclusion

“What the Machine Invented” (2610.17090) presents a structured method for analyzing a large corpus of AI-generated mathematics at the level of proof mechanisms rather than headline theorems. Its principal contribution is the distinction between extraction, novelty assessment, formalization status, and mathematical verification. Across 38 selected families, it identifies new candidate gadgets, inequalities, variational principles, transfer mechanisms, and proof architectures, while also documenting extensive dependence on established mathematics and numerous unresolved caveats.

The survey's strongest conclusion is methodological rather than theorem-specific: AI-generated mathematical manuscripts should be mined for reusable intermediate instruments, but those instruments must be detached from their source claims and subjected independently to literature comparison, proof checking, and formal verification. The signed restriction lemma illustrates the potential value of this separation, while the survey's extensive caveats show why extraction alone cannot establish correctness or priority.