Vibeleaderboard
Index / agent
Visit github.com
Category
AI Agents
Rank
No. 3857Tools index

Previous survey · No. 3665 ·

Pricing
Open Source
Type
AGENT
Use case
Research & Education
GitHub
2 stars
Date

About

ProofForge is an AI-agent pipeline that decomposes math problems into pieces, proves them, and formalizes the proofs in Lean 4/Mathlib, so every step is rechecked by the Lean kernel rather than trusted from the model's output. It has produced six pull requests merged into Google DeepMind's formal-conjectures repository, including a proved upper bound for Erdős problem #1084 and the computed 24-digit 5th unitary perfect number for Erdős problem #1052.

At a glance

ProofForge documents an agent workflow for decomposing mathematical problems and checking formal proofs with Lean 4. It is worth reading if you want examples of moving from a proof idea to a machine-checked statement.

Good fit

Lean users reviewing how to split a problem into lemmas and audit the formal result.

Consider something else when

Use a turnkey proving tool if you need a hosted interface with a documented execution workflow.

Start here

Choose one included proof. Compare its theorem statement with the intended result, then check the proof with Lean and inspect its axioms.

Cost and requirements

Apache-2.0 repository with Lean proof sources and references to upstream contributions. Verification needs an appropriate Lean and Mathlib environment.

Limits to know

The kernel verifies the formal statement under its assumptions. It cannot decide whether that statement faithfully represents the original problem.

Why it made the leaderboard

An agent pipeline that only counts a proof as done when Lean's kernel accepts it, with several proofs already merged into Google DeepMind's formal-conjectures repository as concrete, checkable validation.

Tags

lean4formal-verificationai-agentsmathlibtheorem-provingdeepmind

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.