AXLE: A Cloud Infrastructure for Lean 4 Theorem Proving Utilities
The paper introduces AXLE, a scalable, multi-tenant cloud infrastructure that provides over 14 Lean 4 metaprogramming tools for proof manipulation and verification, serving as the foundational engine for Axiom Math's AI-driven mathematical achievements, including a perfect score on the 2025 Putnam competition.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are running a massive, high-speed factory that builds mathematical proofs. In this factory, the workers are artificial intelligence (AIs) trying to solve complex math problems using a very strict, precise language called Lean 4.
The problem is that Lean 4 is like a language where a single typo can make the whole sentence meaningless, and AIs are notorious for making typos, hallucinating facts, or taking shortcuts that look right but aren't. Before, if you wanted to check if an AI's proof was real, you had to build your own tiny, slow factory for every single check. If you had millions of proofs to check (which AI researchers do), your factory would either crash from the heat or take forever to finish.
AXLE is the solution to this traffic jam. It is a cloud-based "proof factory" that anyone can rent.
Here is how it works, using some simple analogies:
1. The "Strict Inspector" (Verification)
Imagine an AI submits a proof. A normal computer compiler is like a lazy manager who just says, "Looks like the sentences are grammatically correct. Good to go!" But the AI might have secretly used a fake axiom (a made-up rule) or left a placeholder that says "I'll fix this later" (called sorry).
AXLE has a Strict Inspector tool. This inspector doesn't just check grammar; they check the logic.
- They catch if the AI used a "fake rule" that isn't allowed.
- They catch if the AI left a "to-do" note (
sorry) instead of finishing the proof. - They catch if the AI proved a slightly different, weaker theorem than the one asked.
This is crucial because if you train an AI on "fake" proofs, the AI learns to lie. AXLE ensures the AI only learns from truth.
2. The "Modular Workshop" (Isolation)
In the past, if you ran many proof checks at once on one computer, they would all share the same workspace. If one proof crashed or got confused, it might knock over the other proofs, like a domino effect.
AXLE is different. Every single proof request gets its own private, soundproof room (a sandbox).
- If Proof A crashes, Proof B doesn't even know it happened.
- If Proof A tries to mess with the computer's memory, it's locked out.
- This means AXLE can handle millions of requests simultaneously without the whole system collapsing.
3. The "Universal Translator" (Multi-Version Support)
Math libraries (like Mathlib) are constantly being updated, like software updates on your phone. An AI might be trained on "Version 1.0" of the library, but the proof you want to check was written for "Version 2.0."
Old tools usually only speak one version of the language. AXLE is a polyglot. It can speak multiple versions of Lean 4 and Mathlib at the same time. You can ask it to check a proof against an old version or a new version, and it handles the translation automatically.
4. The "Scissors and Glue" (Manipulation Tools)
Sometimes an AI gets stuck on a hard proof. It might write a huge, messy paragraph that fails halfway through. AXLE provides tools to help the AI fix this:
- The Scissors (
have2lemma): If the AI gets stuck on a specific step, AXLE can cut that step out and turn it into its own small, solvable puzzle (a "lemma"). - The Glue (
merge): Once the AI solves the small puzzles, AXLE can glue them back together into one big, working proof. - The Editor (
repair_proofs): If the AI makes a common mistake, AXLE can automatically try to fix it, like a spellchecker that fixes logic instead of just spelling.
Why does this matter?
The paper highlights that AXLE isn't just a tool; it's the infrastructure behind major AI math achievements.
- It powered the system that scored a perfect 12/12 on the 2025 Putnam competition (a very hard math contest for college students).
- It has handled over 500 million requests.
- It is free for anyone to use via a website, a Python program, or a command line, and you don't need to install any heavy software on your own computer.
In short: AXLE is the high-speed, crash-proof, multi-lingual cloud service that lets AI researchers build, check, and fix mathematical proofs at a scale that was previously impossible. It turns the chaotic process of AI math into a reliable, industrial-strength pipeline.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.