AI & Maths watch GitHub

Retrospective research digest

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

At a glance3 entries · 8 linked sources Partial coverage

Three substantive entries established for October 1, 2026: an account of AI-assisted operator-system research, a benchmark separating mathematical idea discovery from execution, and arXiv's response to AI-related submission growth. Coverage of the supplied social accounts remains incomplete.

Research updatesSelect an entry to read more

Van Suijlekom describes AI-assisted computations of operator-system K-theoretic invariants Walter van Suijlekom Published: 3 sources

In a newly dated research account, van Suijlekom explains how iterative exchanges with generative AI helped develop computations for Toeplitz, Fourier-truncation and graph operator systems. He reports checking proposed arguments and references by hand before incorporating them into papers. His publication approach pairs an extensive technical treatment with a six-page, self-contained account of selected three-dimensional examples.

Research relevance:

A concrete workflow for using AI to explore unfamiliar techniques while retaining mathematical responsibility. The associated papers compute invariants that sometimes distinguish an operator system from its C*-envelope.

Claim status:
Author's firsthand account of AI assistance and personal handwritten checking; proof manuscripts are available. This is not independent scrutiny or a claimed machine-checked formalization.
Limitations:
The October 1 news is the workflow account; both underlying preprints were submitted September 30. Their abstracts and submission histories were consulted, but their proofs were not independently verified. The reported acceleration from weeks or months to days is the author's estimate.

Sources & further reading

Link to this entry
The Missing Primitive tests whether models discover the organizing idea of a mathematical solution Shuo Xing, Zilin Dai, Chengyuan Qian, Fangzhou Lin, Wenjing Chen, Ping He, Pan Lu, Alvaro Velasquez, Mohit Bansal and Zhengzhong Tu Published: 3 sources

The preprint introduces mathematical primitives—compact structural ideas such as invariants or reductions—and evaluates discovery, answer generation, recognition in worked solutions and execution with guidance. The authors report that supplying the primitive exposes substantial latent execution ability, while independent discovery is the main bottleneck in their evaluations. They also propose primitive-guided self-distillation and release a public implementation.

Research relevance:

Provides a useful distinction for research assistance: producing a plausible solution and finding the decisive idea are different capabilities. The benchmark and code offer material for studying that distinction.

Claim status:
Available experimental preprint and public repository; performance and diagnostic conclusions are the authors' claims. No new research theorem or checked formalization is established by these experiments.
Limitations:
The findings depend on the benchmark's selection and scoring of mathematical primitives and the evaluated models. They do not establish general competence at open-problem research. Experiments were not reproduced. The original submission was October 1 at 17:59:32 UTC, within the Paris window; repository availability on that precise day was not independently archived.

Sources & further reading

Link to this entry
arXiv introduces monthly submission limits in response to AI-related moderation pressure Kat Boboris / arXiv Published: 2 sources

arXiv announces an October 1 policy limiting each submitter to two submissions per calendar month across categories, while retaining a maximum of three active submissions. It attributes part of the growing moderation burden to AI-assisted production of narrow, fragmented and dense manuscripts. The FAQ specifies that rejected submissions count, whereas submissions deleted before announcement do not; the limits apply to the submitting person rather than every coauthor.

Research relevance:

Directly affects dissemination workflows for mathematicians producing AI-assisted results and teams coordinating multiple manuscripts. It also reinforces the practical value of substantial, self-contained papers.

Claim status:
Official dated policy announcement, also highlighted that day by Terence Tao. The explanation of submission growth is arXiv's institutional assessment, not a controlled measurement of AI's causal contribution.
Limitations:
This concerns submission and moderation, not mathematical correctness or peer review. The digest reports the October 1 announcement rather than assuming that subsequently accessible policy pages preserve every historical detail.

Sources & further reading

Link to this entry
Coverage and limitations
  • Reconstructed the Paris civil day October 1, 2026 using public web searches and original dated sources. Search-index and crawl dates were not treated as publication dates.
  • Read Walter van Suijlekom's October 1 blog entry and the original arXiv abstracts and submission histories of its two September 30 companion papers.
  • Read the original Missing Primitive arXiv record, consulted its HTML paper and public GitHub repository, and read the dated MathsAI report that supplied the lead. No experiments or source code were executed.
  • Read Terence Tao's October 1 post, the linked original arXiv announcement and its FAQ. Consulted Tao's UCLA page and October blog archive.
  • Confirmed the Xena blog address through Kevin Buzzard's Imperial College webpage and read its October 1 essay. Consulted Gowers's blog and the official Lean, Axiom, Epoch AI and Harmonic homepages. Project Numina's homepage returned no readable content.
  • Identity checks were incomplete: the supplied X handles were not all corroborated through official websites. Indexed X and mirror snippets were leads only; no complete October 1 timelines were available. The other supplied individual accounts, Tao's Mathstodon timeline and Buzzard's Bluesky timeline were not comprehensively consulted.
  • Excluded Cogentic as a new October 1 preprint: its original submission was September 30 at 17:55:22 UTC, despite October 1 dates on aggregators. Its October 1 revision fell within the Paris window, but the substantive changes were not established, so no update item is included.
  • Excluded repeat coverage of the September Thomson-problem and zeta-zero claims. Consulted Quanta's Transformation page and indexed news leads, but did not establish a sufficiently supported new mathematical announcement from them. Buzzard's October 1 opinion essay was omitted to prioritize concrete research and workflows.
  • This is a selective archival reconstruction, not evidence that no other relevant announcements occurred. No proof was independently verified and no Lean checker was run.

Requested coverage window: 2026-10-01T00:00:00+02:00 — 2026-10-02T00:00:00+02:00.

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

Archives 32
Older editions (2)