Archived edition

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

2 entries · 5 linked sources Partial coverage
Axiom releases a claimed Lean formalization of the four Weil conjectures Axiom Math; lead contributor Tony Yue Yu Published: 3 sources

Axiom published a public Lean repository claiming formal proofs of the rationality, functional equation, Riemann-hypothesis, and Betti-number conclusions of the Weil conjectures. The release is a formalization of known mathematics, not a new solution: it says the decisive weight estimate is supplied by Axiom's earlier Stacks development, via a later Fourier-transform route, rather than reproducing Deligne's 1974 proof line by line. This is potentially useful as reusable algebraic-geometry infrastructure; the stated Riemann-hypothesis and point-counting scope is for smooth proper varieties, while rationality is stated more broadly.

Research relevance:

This is the first in-window public artifact that attempts to turn Axiom's recently announced Stacks foundation into a major downstream theorem. The repository is available for future independent builds and audits, and the claimed conclusions could be reusable if the dependency stack is validated.

Claim status:
Announcement and public code are available. Axiom states that the repository is sorry-free and uses no added axioms, but also states that its AI-agent-written code has not yet been reviewed by humans. Independent commentary found no public CI run, external build, comparator result, or expert review as of 2026-11-06. This digest did not run Lean or verify the proof.
Limitations:
The formalization depends on Axiom's unreviewed StacksFrontier layer for the key weight bound. It is neither a new mathematical result nor a line-by-line formalization of Deligne's original proof; the Ramanujan tau bound is explicitly outside its scope. The independent assessment is itself not peer review.

Sources & further reading

LeanContext releases a context-controlled benchmark and code for Lean whole-proof generation Hang Zhang, Qihao Ye, Litong Liu, Yunbei Zhang, Yugyung Lee, Yue Zhou, Bowen Liu, Emily First, Cheng Han, and Qiao Zhuang Published: 2 sources

This arXiv preprint introduces LeanContext: 129,826 whole-proof targets from 4,469 Mathlib files, each with aligned views separating selected preceding mathematical declarations from local environment commands. The authors report that prioritizing omitted target dependencies improved Pass@32 by 1.40 percentage points for LLaDA and 1.70 points for Llama on a 2,000-target experiment. The linked MIT-licensed repository provides the dataset, extraction and audit tooling, training configurations, and Lean-verification workflow.

Research relevance:

For researchers training or evaluating Lean agents, it directly targets a practical confound: whether apparent gains from long context come from relevant mathematical declarations rather than surrounding syntax. The aligned views and dependency audit can support more controlled premise-selection and context-ablation experiments.

Claim status:
A public preprint and companion code/data release are available. The numerical results are author-reported experimental claims, not a mathematical proof and not independently replicated in the sources consulted.
Limitations:
The work evaluates specified LLaDA and Llama-3.1 configurations on a Mathlib snapshot with fixed prompt and proof budgets; the reported gains need not transfer to other models, libraries, retrieval systems, or research-level developments. The repository documents a Lean-checking procedure, but this digest did not execute it or reproduce the experiments.

Sources & further reading

Coverage and limitations
  • Primary sources read: Axiom Math's 2026-11-06 release and the public AxiomMath/Weil repository; arXiv's full record for LeanContext and its linked public GitHub repository. An independent critical analysis, AI Proved What?, was also read for the Weil item. No code was executed and no Lean development was built.
  • Related blogs consulted: Terence Tao's blog archive showed no new substantive post in this narrow window beyond material already covered; Kevin Buzzard's official Xena blog was identified as https://xenaproject.wordpress.com/ but no relevant new post was found. Simon Frieder's blog did not yield a substantive in-window item in the searches performed.
  • Direct Bluesky collection: xenaproject.bsky.social status ok, but no posts returned; replies and reposts were excluded and pagination was capped at 300. Direct Mathstodon collection: mathstodon.xyz unavailable due to HTTPError; replies and reposts were excluded and the intended cap was 300.
  • Direct X collection: __alpoge__, scottnarmstrong, wtgowers, sebastienbubeck, thomasfbloom, littmath, alexkontorovich, f_charton, leonard41111588, boazbaraktcs, leanprover, projectnumina, harmonicmath, axiommathai, and epochairesearch all returned status ok under bounded API collection; replies and reposts were excluded, cached posts may be present, and API resolution does not independently authenticate identities. Only supplied API posts were treated as direct X access; no complete X timeline was inferred. Axiom's supplied post was used only as a lead and the release was checked on Axiom's official website and repository.
  • The supplied Armstrong, de Moura, Lean, and other social leads were announcements, event notices, reposts, or unrelated material rather than independently supported mathematical research results in the window. The SnapJudge site linked by Armstrong could not be read because its public page returned 403, so it is not included.
  • Search is not exhaustive: X pagination is incomplete, Mathstodon was inaccessible, and neither formal artifact received a local build, formal proof audit, peer review, or independent replication in the consulted sources.

Requested coverage window: 2026-11-05T21:15:01.404343+01:00 — 2026-11-07T09:15:01.342393+01:00.

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

Archives 79
Older editions (49)