Skip to content
HuggingMathPublic

About

An AI collaborative system that works alongside mathematicians on challenging proofs—handling retrieval, lemma decomposition, and Lean verification.

Topics

Resources

Stars

28 stars

Watchers

1 watching

Forks

Latest commit

 

History

125 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

English · 简体中文

TheoremOS

Bring a mathematical problem to the TheoremOS workbench. Review its Lean 4 statement and guide the proof plan while TheoremOS looks for related results and explores possible lemmas. Completed candidates go through an independent Lean check.

Official website · theoremos.top

Star this repository to find the public archive and verification kits again.

GitHub stars Research preview 0.1.0-preview.1 Lean 4 verification Public materials rights reserved

Proof archive: English · 简体中文

Workbench: Request research preview access · Sign in · Access guide

An illustrated TheoremOS workbench. An anime mathematician compares the problem with its Lean statement. Mathlib results inform the lemma work, and a candidate proof enters an independent Lean 4 check. Below are feedback, a proof revision, the check report, and materials for local verification.

This public repository contains the website (中文), proof records, and local verification materials. The workbench runs online. Runtime source and a public CLI installer are not included.

Try TheoremOS

  1. Browse without an account. Open the proof archive above for statements, Lean source, recorded attempts, and verification reports. The planar Anderson spectrum case (EN) (中文) walks through one example.
  2. Start a research task online. Apply for email allowlist access and sign in with a verification code. Workbench access is free during the preview; you supply a model API key for a supported Responses API provider and pay that provider for inference. Set time and token limits before running. The first-task guide (EN) (中文) shows the workflow.
  3. Check a proof locally. You need no TheoremOS account or model key. Start with the Lean verification guide (EN) (中文), or take the Dimitrov proof kit (EN) (中文) for a frozen benchmark workspace, proof, and comparator instructions.

TheoremOS workbench showing problem input, model settings and proof budgets

Workbench screenshots: English · 简体中文. The preset Lean replay (EN · 中文) shows a candidate being checked, repaired, and accepted on a small example.

How proof work is organized

  • Confirm the statement. Submit natural language, TeX, or a single Lean file. Formalization produces a compilable statement, and a separate review compares it with the original. Compilation checks Lean syntax and types; the researcher still confirms mathematical intent.
  • Check a route before filling in the lemmas. Temporarily assume candidate lemmas to check whether they imply the main goal. Prove the lemmas separately, then combine and check the completed proof without the temporary assumptions.
  • Inspect what was tried. The workbench exposes available proof attempts and compiler feedback. Researchers can steer the plan and reuse verified lemmas, which are checked again in the target environment.
  • Run custom experiments through research collaboration. The CLI is available by arrangement for multi-file projects and comparisons of execution, verification, scheduling, context, and role policies. See capabilities and availability (EN) (中文).

Proof evidence

The catalog lists 10 proofs with official Lean-Eval pass records, 18 OpenAI Math records checked with local Lean and problem comparators, and one in-progress record. The OpenAI Math records have no official OpenAI confirmation. Some records check a helper result or a conditional statement. The evaluation notes (EN) (中文) describe an earlier snapshot of selected runs; the archive does not establish a system-wide success rate or cost.

The source-by-source records below link to the criteria appropriate to each collection.

Lean-Eval: 10 Officially Accepted Proofs

View all 10 proofs officially accepted by Lean-Eval
Proof Package Area Official Result
Unit conjecture: a nontrivial group-ring unit Algebra #1246 · pass
Rademacher type and Enflo type coincide Functional analysis #2115 · pass
Equiangular lines with a fixed angle Discrete geometry #2135 · pass
Fractional expectation thresholds Probabilistic combinatorics #2141 · pass
Dimitrov's lower bound for algebraic integers Algebraic number theory #2143 · pass
Hopf–Rinow theorem Riemannian geometry #2143 · pass
Improved bounds for the sunflower lemma Combinatorics #2144 · pass
Lorentzian polynomials Algebraic combinatorics #2144 · pass
The optimal paper Möbius band Geometry #2144 · pass
Algebraic integers: an upper bound for λ_SSS Algebraic number theory #2144 · pass

Model and provenance metadata are recorded in accepted.json and submitted.json. The final four proofs were accepted together in submission #2144, with source files stored under submitted/.

OpenAI Math: 18 locally verified records

These records link to the saved local Lean and problem-comparator checks. The all-seams entry covers a support result for the stated surface presentation.

View statements, Lean proofs, verification reports, and full research trajectories for all 18 records
Problem Record Area Statement and Proof Verification and Process
Supercritical three-state reconstruction on regular and Poisson trees Probability Statement · Proof Local pass · Process
Face correspondence under triangle expansion Combinatorial optimization Statement · Proof Local pass · Process
Endpoint non-integrability for the Koebe map and its inverse Complex analysis Statement · Proof Local pass · Process
Support results on numerical ranges and finite-dimensional compressions Functional analysis Statement · Proof Local pass · Process
Almost-sure spectrum of the planar Anderson operator Random operators Statement · Proof Local pass · Process
Finiteness of the Ising-perceptron variational value Probability and statistical physics Statement · Proof Local pass · Process
Totient companion: the no-seed zero case Number theory Statement · Proof Local pass · Process
Compactness of normalized Cartier chart retractions Algebraic geometry and topology Statement · Proof Local pass · Process
A colored-graph criterion for finite congruence lattices Universal algebra and lattice theory Statement · Proof Local pass · Process
Existence of the honeycomb free-energy limit Probability and statistical physics Statement · Proof Local pass · Process
A seven-coloring of the Euclidean plane Discrete geometry Statement · Proof Local pass · Process
A legal 35-move winning strategy for Snaky Combinatorial games Statement · Proof Local pass · Process
Snaky certificate verification Combinatorial games Statement · Proof Local pass · Process
Hadamard cube-fiber cancellation Linear algebra Statement · Proof Local pass · Process
A three-dimensional complete normal local complex algebra Commutative algebra Statement · Proof Local pass · Process
A finite counterexample to Truffet's procedure Tropical optimization Statement · Proof Local pass · Process
Supercritical reconstruction for three-state tree channels Probability Statement · Proof Local pass · Process
Marked trace values across all surface seams (support result) Algebraic geometry Statement · Proof Local pass · Process

Repository and reuse

Path What you can inspect
accepted/ and submitted/ Lean-Eval proof packages and official result provenance
data/proofs/ OpenAI Math statements, proof files, local checks and saved process records
in-progress/ Earlier drafts, some with unclosed sorry placeholders
proof-kits/ and examples/ A standalone benchmark kit and a short Lean example
docs/ and CHANGELOG.md Access, verification, evaluation and release information

LICENSE reserves rights to project-authored public materials. Archived proofs and third-party materials retain their own terms; see Public materials, reuse, and citation. Cite the project through CITATION.cff and cite the original sources of any proof you discuss.

For a multi-file Lean project, a custom toolchain, or a controlled CLI experiment, use the collaboration route in the access guide. Open an issue for archive corrections or feedback, and star the repository to keep the verification resources handy.

About

An AI collaborative system that works alongside mathematicians on challenging proofs—handling retrieval, lemma decomposition, and Lean verification.

Topics

Resources

Stars

28 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages