The Grothendieckian Programme is a set of seven works on machine reasoning about mathematics (34 volumes, 26,021 pages), released openly in October 2026 as first drafts that have not been peer-reviewed. This paper draws from them twelve theses on how mathematics and machine reasoning should be combined. Each thesis is stated in one sentence, argued, tied to the passages of the seven works on which it rests, related to existing work, and paired with an observation that would refute it. The theses concern the unit of mathematical knowledge for a reasoning system, the place of epistemic status, the latent space of mathematical reasoning, attention, failure, authority, and the method of discovery. The paper adds no mathematical result to the seven works; none of them claims a proof of the Riemann Hypothesis, the Birch and Swinnerton-Dyer conjecture, the Hodge conjecture, the abc conjecture or the Langlands correspondence.
Author comments: 26 pages, 5 figures. Version 1.0 of 28 October 2026; the same text is deposited at Zenodo, DOI 10.5281/zenodo.23734424 (all versions: 10.5281/zenodo.23734423). This copy was typeset with pdfLaTeX for Hexagon: its text is identical to version 1.0 at Zenodo, whose sealed PDF was typeset with LuaLaTeX, except that a broken cross-reference on page 3 ("igref{works}") is corrected to "Figure 1"; a few line and page breaks differ. The twelve theses are also set out at https://github.com/AlienMuse/grothendieckian-problems/blob/main/THESES.md; the seven works on which they rest, each with its DOI, are listed at https://github.com/AlienMuse/grothendieckian-programme. Corrections recorded since version 1.0 to the seven works and to the papers: https://github.com/AlienMuse/grothendieckian-problems/blob/main/ERRATA.md. Companion papers: The Grothendieckian Programme: An Overview (DOI 10.5281/zenodo.23734421) and The 57 Grothendieckian Problems (DOI 10.5281/zenodo.23734426).
A mix of human-written and AI-generated textHuman understanding: all partsGrothendieckian Programmedescentepistemic statusformal verificationmachine reasoningrecursive mathematical discovery
The Grothendieckian Programme is a series of seven works on machine reasoning about mathematics (34 volumes, 26,021 pages), released openly on Zenodo in October 2026 with an overview volume. Two of the works specify a reasoning architecture — a core that holds mathematical worlds as typed, versioned and verifiable objects, and a kernel that reasons inside such worlds by sites, sheaves and descent. Five apply one method, recursive mathematical discovery, to the Riemann Hypothesis, the conjecture of Birch and Swinnerton-Dyer, the Hodge conjecture, the abc conjecture and the Langlands programme, and record what it proves, what it reduces to stated hypotheses and what it leaves open. This paper describes the architecture, the five research subjects, the method, the published corpus and the open process by which the works are to be reviewed and continued. None of the works claims a proof of any of these conjectures; they are first drafts and have not been peer-reviewed.
Author comments: 10 pages, 3 figures. Version 1.0 of 28 October 2026; the same text is deposited at Zenodo, DOI 10.5281/zenodo.23734422 (all versions: 10.5281/zenodo.23734421). This copy was typeset with pdfLaTeX for Hexagon: its text, and its line and page breaks, are the same as in version 1.0 at Zenodo, whose sealed PDF was typeset with LuaLaTeX. The seven works it describes, each with its DOI, are listed at https://github.com/AlienMuse/grothendieckian-programme. Corrections recorded since version 1.0 to the seven works and to the papers: https://github.com/AlienMuse/grothendieckian-problems/blob/main/ERRATA.md. Companion papers: The Rising-Sea Theses (DOI 10.5281/zenodo.23734423) and The 57 Grothendieckian Problems (DOI 10.5281/zenodo.23734426).
A mix of human-written and AI-generated textHuman understanding: all partsGrothendieckian Programmedescentmachine reasoningmathematical worldsopen sciencerecursive mathematical discovery
We state fifty-seven open problems drawn from the seven works of the Grothendieckian Programme (34 volumes, 26,021 pages, released openly in October 2026 as first drafts that have not been peer-reviewed). Thirty-five come from the two works on the architecture of a reasoning system and twenty-two from the five works that each take a named conjecture or programme as research subject; no problem is one of those conjectures. Each problem is stated in the words of its source, located by work, chapter and page, given the status the source assigns it, and paired with a criterion that decides when it is solved, a note on existing work and the date of a literature check. The problems are of three kinds: twenty-five mathematical problems, nineteen machine problems with success criteria fixed in advance, and thirteen formalization problems for Lean 4 and Mathlib. Every quotation from the sources has been checked against the published PDFs, and each problem has been checked against the literature as of 28 October 2026; a machine-readable version accompanies the paper.
Author comments: 54 pages, 5 figures. Version 1.0 of 28 October 2026; the same text is deposited at Zenodo, DOI 10.5281/zenodo.23734427 (all versions: 10.5281/zenodo.23734426). This copy was typeset with pdfLaTeX for Hexagon: its text is identical to version 1.0 at Zenodo, whose sealed PDF was typeset with LuaLaTeX, and a few line and page breaks differ. Each problem has a public issue at https://github.com/AlienMuse/grothendieckian-problems, which also holds the machine-readable data, the literature-check records and the Lean statement of Problem 49. Corrections recorded since version 1.0, including literature missed by the check for Problem 18: https://github.com/AlienMuse/grothendieckian-problems/blob/main/ERRATA.md. Companion papers: The Grothendieckian Programme: An Overview (DOI 10.5281/zenodo.23734421) and The Rising-Sea Theses (DOI 10.5281/zenodo.23734423).
A mix of human-written and AI-generated textHuman understanding: all partsGrothendieckian ProgrammeLeandescentformal verificationmachine reasoningopen problems
In-context learning is often explained as implicit Bayesian inference: given labeled examples of a task drawn from a prior, the learner predicts like the posterior. We determine the computational price of this explanation for task families whose labels depend on few bits of the task description, such as quantized models on sparse data, and we compare it with the price of merely learning. The comparison uses the recent refutations of the Exponential Time Hypothesis by OpenAI and Anthropic.
Bayesian inference in context is counting. For κ-local task families with N-bit descriptions under the uniform prior, the posterior predictive and posterior sampling reduce to #κ-SAT. Without label noise they run in time poly(m)⋅2o(N) on contexts of every length m when κ=o((logN)1/4/loglogN), and label noise is handled by error profiles. A two-variable gadget gives contexts that are always consistent and whose posterior predictive is 2#F/(2#F+1) for a CNF F. So approximating the posterior predictive within 1/4 needs time 2(1−γ)N at locality O(logN) under CNF-SETH, and time 2N/polylogN at locality 3 under a length form of ETH. At constant locality, Bayesian inference in context is SAT parameterized by formula length.
Learning in context is cheap. The Bayes risk of any class with N-bit descriptions is at most N/(m+1) after m examples. For binarized perceptrons and literal disjunctions on sparse data, the very families of the lower bounds, a linear program is a polynomial-time in-context learner with essentially the same sample efficiency. Every κ-local family is linear over ∑j≤κ(jN) local Fourier features, so regression in context learns it from NO(κ) examples, in time polynomial in Nκ and the inverse accuracy.
The average case. For Goldreich's random local functions, the Bayes risk is at most N/(m+1), while every in-context learner yields a distinguisher for Goldreich's generator, with the same running time up to a linear term, whose advantage is 1/2 minus the risk. Under a local pseudorandom generator hypothesis with polynomial stretch, polynomial-time in-context learners need polynomially more examples than the Bayes predictor. Under its length form at linear stretch, which the new algorithms refute below rate 1/2, the Bayes predictor of the refutations is optimal up to polylogarithmic factors in the exponent.
For transformers, these bounds limit the length of a chain of thought that computes the posterior predictive, and they separate Bayes-optimal in-context learning from the regression-type in-context learning that transformers are known to implement.
Primarily AI-generated textHuman understanding: some partsBayesian inferenceCNF-SETHExponential Time HypothesisGoldreich's generatorfine-grained complexityin-context learninglocal pseudorandom generatorsmodel countingposterior predictivetransformers
Alman and Song showed that attention with head dimension O(logn) can be approximated in almost linear time when the entries are o(logn), and needs quadratic time at Θ(logn) under the Strong Exponential Time Hypothesis. In October 2026 the Exponential Time Hypothesis was refuted, and with it SETH. The quadratic lower bound now rests on the hardness of Orthogonal Vectors in dimension clogn (LDOV), which no open hypothesis about satisfiability is known to imply. We measure an instance by its spread W, the largest gap between two logits of a row. At every nontrivial constant precision and in every dimension between ω(logn) and no(1), we show that under LDOV the time exponent of attention is exactly 1+min{W/lnn,1}: uniform sampling of keys is optimal, the cost is min{neW,n2} up to factors no(1), and the transition to quadratic time sits at W=lnn, the spread at which one key can outweigh all others. At polynomial precision and in dimension O(logn) the threshold becomes Θ(logn), with a constant that tends to zero as the precision grows, and the dimension moves it: at precision n−1−O(1/loglogn), spread O(logn/loglogn) is hard in dimension (logn)1+Ω(1) but easy in dimension O(logn). From satisfiability we obtain hardness at spread (1+θ)lnn in dimension no(1), provided the hard CNF formulas have Nc0 clauses for some c0<2, and we show that the construction behind this bound, which builds one instance for each proof of a Merlin–Arthur protocol for disjointness, needs c0<2 for every such protocol. For the entry bound, every reduction that embeds Orthogonal Vectors vector by vector, with polynomial blowup and logit gap Ω(logn), yields a Merlin–Arthur protocol for disjointness; by Klauck's lower bound, such reductions cannot derive B=Θ(logn) from Orthogonal Vectors in dimension (logn)1+Ω(1). We also map the conditional lower bounds for attention in the literature to the hypotheses that now support them, and find two families whose hypotheses are false, one of them in part of its parameter range.
Primarily AI-generated textHuman understanding: some partsCNF-SETHMerlin-Arthur protocolsOrthogonal VectorsStrong Exponential Time Hypothesisattentioncommunication complexityfine-grained complexitypolynomial methodsoftmaxtransformers
Deciding which conjectures are worth proving is a crucial part of mathematical discovery, and a system that learns mathematics from its axioms alone must make that decision without human guidance. Self-play provers such as Minimo learn a conjecturer alongside the prover and reward it for producing hard conjectures, but a conjecture that is harder to prove is not necessarily more useful for improving a prover: the learned conjecturer drifts toward long, artificial statements. We study conjecturing in a controlled, single-round setting in the Peano formal environment, across propositional logic, abelian groups, and natural-number multiplication, evaluating a set of conjectures by how much a fixed prover improves on held-out theorems after training on their proofs. We show that a far simpler generator does better: making small structural edits to statements the prover already knows yields conjectures that, at an equal training budget and under best-first search, improve the prover by 6 to 13 percentage points more than Minimo's learned conjectures. The gain holds even when conjectures are drawn at random from the generated pool. In contrast to the common practice of selecting the most difficult statements, we also propose Hard Neighbours (HN), a score that favors conjectures whose proof difficulty grows disproportionately relative to their structural distance from easier reference statements; HN gives the largest selection gains in propositional logic, but no selection method is best in every domain. In the current setting, how conjectures are generated matters more than how they are selected, and structural perturbation is the baseline that future conjecturers should be measured against. Apart from improvements in proving ability, we also study other attributes of generated proofs: the amount of search required and the length of the proofs.
Primarily human-written textHuman understanding: all partsconjecturinglanguage models
Two recent works by OpenAI and Anthropic refuted the Exponential Time Hypothesis (ETH). They count the satisfying assignments of a k-CNF on n variables in time 2n/(logn)1/2−o(1). Many fine-grained lower bounds in learning theory rest on ETH and its relatives. We determine what the refutation changes for learning.
The answer is a locality principle: the new algorithms speed up learning when each labeled example constrains few bits of the hypothesis, and under the CNF form of the Strong Exponential Time Hypothesis, which survives, the speed-up ends at Θ(logN) bits. Three transitions make the principle precise.
Locality. Call a class with N-bit descriptions κ-local if each label depends on at most κ description bits, as for quantized models on sparse data. Then proper PAC learning, exact version-space counting and sampling, and the halving algorithm run in time poly(m)⋅2o(N) on samples of size m for κ=o((logN)1/4/loglogN), and proper PAC learning at constant accuracy does so for κ=o(logN). At inverse-polynomial accuracy, proper learning of κ-local classes is equivalent to κ-SAT, up to a factor 2 in the locality. Under CNF-SETH, consistency for binarized perceptrons on O(logN)-sparse data needs time 2(1−γ)N for every γ>0.
Accuracy. Exact empirical risk minimization, proper agnostic learning and the exponential mechanism reduce to tensor contractions over a truncated polynomial ring. As a consequence, the Sparse Max-SAT Hypotheses of Alman and Vassilevska Williams are false. At constant accuracy, proper learning is SAT parameterized by formula length. With quasi-linear PCPs this gives a 2N/polylogN lower bound under a length form of ETH, which replaces Gap-ETH for reductions of quasi-linear size. At accuracy N−3/8, the Max-3-SAT Hypothesis forces time 2(1−o(1))N.
Treewidth. Exact inference in graphical models of bounded degree and arity takes time 2O(n/logn). On expanders this beats every elimination order and every structured DNNF. Under #CNF-SETH, time 2(1−γ)tw is again necessary once a model may have twO(1) variables.
We also take stock of the conditional lower bounds of learning theory: which ones fall, and which ones survive on which hypothesis. The bound for training one ReLU loses its proof, because its hypothesis about Densest k-Subgraph on bounded-degree graphs is false. The next obstacle is a permutation problem. On difference vectors ei−ej, empirical risk minimization for real-valued halfspaces is weighted Feedback Arc Set, which the new algorithms do not reach.
Primarily AI-generated textHuman understanding: some partsExponential Time HypothesisPAC learningStrong Exponential Time Hypothesiscomputational learning theoryempirical risk minimizationfine-grained complexitymodel countingprobabilistic graphical modelstensor networkstreewidth
Contributed by Isaac Rudich, Louis-Martin Rousseau
The current best upper bound for deterministic factoring is N1/5+o(1). We provide a Lean certificate that there exists a deterministic Turing machine that computes the smallest prime factor of every N in No(1) steps: for every k≥1 it runs within Ck2n/k steps on n-bit inputs. The algorithm reduces factoring to counting divisors with prescribed leading bits, and counts them by leveraging the counting engine used to refute the Exponential Time Hypothesis (ETH). The Lean statement that a reader must check to verify the proof is only 13 lines and only uses Mathlib. The full proof is about 35,000 lines. This note was written less than 24 hours after the ETH refutations were posted. No human-readable explanation of these results exists yet, and their Lean proofs are long: Anthropic's alone runs to about 152,000 lines. Our proof re-proves only the part of the counting engine it needs and imports nothing from those developments. We hope this application, with its short statement, serves as a stepping stone for those working to understand the engine; a follow-up paper will explain this proof in detail.
A mix of human-written and AI-generated textHuman understanding: some partsExponential Time HypothesisLeancounting algorithmsdeterministic factoringformal verificationinteger factorization
Self-improving agent harnesses such as the Darwin Gödel Machine (DGM) let a coding agent edit its own scaffolding, and they keep the edits that score better and report large gains on coding benchmarks. In three pre-registered rounds, we compare what such loops find with standard engineering changes ("textbook moves"), one-shot redesigns by a language model, and placebo loops that make no edits; we also re-run released agents of the DGM and of ADAS with newer models. On hard LiveCodeBench problems with Claude Haiku 5.5, the starting agent often stops on an empty turn because its output-token limit is too small for the model's reasoning. A single textbook move, which raises this limit to 16,000 tokens and makes three other changes to the model loop, recovers about 90% of the gain of the best loop agent and costs less than half as much per task. Four of five one-shot redesigns by Claude Opus 5.5 also raise the limit, to 8,192 tokens, without evaluation feedback. The best loop agent still leads the textbook moves by 1–5 percentage points in pass rate. On held-out SWE-bench Verified tasks, run in the DGM's harness, the DGM's released best agent does not outperform its initial agent; three textbook moves on the initial agent outperform the best agent at a quarter of its cost, whether or not the patches of runs that ended early, most of them the best agent's, are kept. Claims of self-improvement should be measured against textbook moves, one-shot redesigns, placebo loops, held-out tasks and newer models.
A mix of human-written and AI-generated textHuman understanding: all parts
Contributed by Deep Bhattacharjee, Ushashi Bhattacharya
A system that improves itself keeps every change that passes a check, so a check that makes errors lets mistakes pile up. We compute the least checking that keeps such a system sound. It is an entropy of the costs of the changes, and an endless chain of improvements stays sound at finite cost exactly when that entropy is finite. A stack of checked layers costs more than its layers would cost alone. We also compute how many guesses a proposer needs before a right one appears: human authors come close to this least number, and language models need 1.5 to 4.3 times as many. Anaxia, an engine built on these laws, rebuilt the published 2016 counterexample to Steinberg's conjecture from the rules of a colouring problem alone. The engine is a work in progress, and its rules and search are being revised and extended for the open problems it has not yet settled.
A mix of human-written and AI-generated textHuman understanding: some partsautomated discoveryentropyrecursive self-improvementverification
Contributed by Mohammed Abouzaid, Nikhil Srivastava, Rachel Ward, Lauren Williams
To assess the ability of current AI systems to correctly solve research-level mathematics problems, we tested several AI systems on our ``third batch'' of ten problems in a broad range of mathematical fields; these problems arose naturally in the research process of the contributors. This document includes the problems, our methodology, and the results of our testing. We assessed the AI-generated solutions based on mathematical correctness, quality of exposition, and quality
of attribution and citations. We provide links to supplementary documents including the human solutions, the AI-generated solutions, and the referee reports and logs for the AI-generated solutions.
Primarily human-written textHuman understanding: all partsAI systemsbenchmarklarge language models
Chen and Larson study tautological classes on the strata of holomorphic
abelian differentials, where they predict a vanishing result. Using known
tautological relations on the moduli space of curves, they reduce this
prediction to the non-vanishing of certain coefficients of a quotient of
hypergeometric generating series, treating the residue classes
g≡0,2(mod3) and g≡1(mod3) through two separate series; they verify
the non-vanishing by computer for small genus.
We prove it in general: in both cases the relevant coefficient is non-vanishing
for every genus and every stratum, and in the first case it has, more strongly,
a uniform strict sign. Along the way we correct an error in the Chen--Larson
derivation of the g≡1(mod3) series, where a constant of the underlying
relation of Ionel had inadvertently been changed.
The paper falls into two parts. Part~I gives the mathematical proofs; Part~II
documents the Lean~4 and Mathlib formalization --- whose only non-standard trust
assumption is the compiler invoked by \code{native\_decide} for the large finite
computations --- together with the long-horizon, multi-model generative-AI
process that produced the proofs, exposed false intermediate routes, and
uncovered the error in the printed Proposition~5.2 relation noted above.
Primarily AI-generated textHuman understanding: not declaredpower seriestautological classes