Updates on artificial intelligence for mathematical research, proof discovery and formal verification, with links to papers, code and tools.

Research updatesSelect an entry to read more

4 entries · 6 linked sources News found
Axiom announces a Lean formalization of the Minimal Model Program through dimension four Axiom Math Published: 2 sources

Axiom says it has released a reusable Lean library of roughly 2.1 million lines formalizing the Minimal Model Program in dimensions 2, 3, and 4, building on its Stacks Project work and including results related to OpenAI’s recent four-dimensional work. The announcement describes a large formalization effort in birational geometry and may be useful to researchers inspecting or extending its stated formal infrastructure.

Research relevance:

This is a substantial formalization claim about central results in birational geometry, with a public repository that could support inspection and reuse. The announcement highlights how formalized dependencies might enable later theorem development.

Claim status:
Axiom’s announcement claims a complete formalization and machine-checked code. The repository is public. These facts establish a release and a claim, not independent confirmation of the mathematical scope or correspondence between the formal statements and the advertised informal results.
Limitations:
The announcement is from the project itself. I found no independent review or reproduction in the consulted sources and did not build or inspect the Lean development. The post’s broad completeness language should be assessed against the repository’s precise theorem statements and dependencies.

Sources & further reading

Lean 4.36.0 improves proof automation and Mathlib build performance Lean developers Published: 1 source

The release notes describe improved `grind` support for bit vectors, characters, and `Fin`, `bv_decide` support for limited congruence reasoning with uninterpreted functions, and exception postconditions in function contracts. They report that Mathlib builds roughly 12% faster, alongside fixes to slowdowns that grew quadratically with file size. The release also bundles Lean Beam, a language-server tool and MCP server for working with Lean projects.

Research relevance:

The automation and performance changes may affect day-to-day formalization, larger proof projects, and agent-assisted Lean workflows. The release notes give concrete feature and limitation details for evaluating upgrade costs.

Claim status:
Official Lean release notes document Lean 4.36.0, dated November 16, and describe the changes. This is a software release, not evidence that any particular mathematical theorem has been formalized correctly.
Limitations:
The roughly 12% Mathlib build improvement is a release-note report; performance can vary by setup and project. The uninterpreted-function support in `bv_decide` is limited, including restrictions on function return types and arguments. This release is infrastructure news rather than a new mathematical result.

Sources & further reading

Epoch AI reports seven Lean solutions on its FrontierMath Erdős benchmark for GPT-6.2 Astra Epoch AI; model results also summarized in OpenAI’s system card Published: 2 sources

Epoch AI reports that GPT-6.2 Astra produced solutions to 7 of 68 problems in its FrontierMath Erdős set, each formally verified in Lean, at a reported budget of about $240 per question; its post compares this with a prior record of 3. OpenAI’s system card also reports seven solutions but gives a $300 per-question budget. Both sources say one solution had no prior known solution. Epoch AI separately reports that Astra remained below its internal human standard on its open-ended work tasks.

Research relevance:

The result is relevant to research-level problem solving and to evaluating formal proof search beyond fixed benchmark exercises. The Lean verification claim is stronger evidence for correctness of the submitted formal statements than an informal model answer alone, while the budget discrepancy merits attention.

Claim status:
Epoch AI’s benchmark and OpenAI’s system card report the result and Lean verification. The two sources disagree on the per-question budget ($240 versus $300), so that detail is unresolved here. A checked Lean solution supports the encoded proposition; it does not by itself establish that the formal statement fully captures a broader informal problem or that the result has been independently reviewed.
Limitations:
This is one curated benchmark with 68 problems, and the set was open as of August 2026. Epoch AI’s reported ECI and task evaluations are organization-specific and may change as its benchmarks are updated. The sources do not provide independent rechecking by this digest, and the budget discrepancy should not be collapsed into a single figure.

Sources & further reading

Older material found late: Lean-verified cop-number results for flower snarks and generalized Petersen graphs Aarav Ghai Published: 1 source

The paper proves that the flower snark J₃ has cop number 2 and Jₙ has cop number 3 for n ≥ 5, also determining its lazy cop number for n ≥ 5. It proves a conjecture classifying generalized Petersen graphs with cop number 2. The author says the flower-snark theorem and new lower bounds were formalized in Lean; the paper describes a rolling-wall strategy and explicit robber strategies. It acknowledges Claude assistance with strategy exploration, drafting, computation, and formalization.

Research relevance:

This is a concrete combinatorics result with both explicit mathematical arguments and a stated formal verification, including a reusable strategy for graphs with cyclic dominated decompositions. It was not included in the supplied previous-items list.

Claim status:
The paper is directly available on arXiv and states the results and formalization claim. It reports two independent solvers for its small-parameter computations. The Lean claims are the author’s; I did not inspect or build the code. The paper itself notes that the formalized lower bounds do not formalize matching upper bounds for all listed generalized Petersen graphs.
Limitations:
The original arXiv submission date is November 12, outside this digest’s window; this is older material found late, not a new in-window publication. The results are presented in a preprint, and the paper reports no independent mathematical review. Some matching upper bounds rely on earlier literature rather than the formal development.

Sources & further reading

Coverage and limitations
  • Window used: 2026-11-16 18:15 to 2026-11-18 06:15 Europe/Paris. Read the official Axiom MMP announcement and repository, official Lean 4.36.0 release notes, Epoch AI’s model page, OpenAI’s system card, and the arXiv abstract and paper text for the graph-theory result. Search results were used as leads; relevant primary pages were opened directly. No proof was independently reviewed or Lean code executed.
  • Direct Bluesky collection: xenaproject.bsky.social returned no posts; up to 300 posts checked, with replies and reposts excluded. mathstodon.xyz collection was unavailable (HTTPError); up to 300 posts intended, replies and reposts excluded. Empty results do not show that no relevant posts existed. Official websites were not independently linked to these supplied social handles during this search.
  • Direct X API collection was supplied for __alpoge__, scottnarmstrong, wtgowers, sebastienbubeck, thomasfbloom, littmath, alexkontorovich, f_charton, leonard41111588, boazbaraktcs, leanprover, projectnumina, harmonicmath, axiommathai, and epochairesearch. It was bounded, excludes replies and reposts, includes cached posts, and does not verify account identity; Thomas Bloom’s feed was partial and reached its pagination or post cap. No supplied X post was used as the sole support for an item. X timelines for other named accounts were not directly collected, so this is not an exhaustive survey.
  • The result pages provide official project, organization, and author affiliations where available, but those pages do not independently authenticate supplied social accounts. No independent scrutiny or external reproduction of Axiom’s MMP formalization was found. The arXiv graph-theory paper is directly readable, but the linked Lean development was not built or checked in this research.

Requested coverage window: 2026-11-16T18:15:01.598483+01:00 — 2026-11-18T06:15:02.106652+01:00.

Public web research; coverage is not exhaustive. A source link is not a certification of a claim.

Archives 79
Older editions (49)