Latex-to-Lean4 automatic formalization and verification of mathematical proofs.
Given a .tex file, the system extracts theorem/proof blocks, generates Lean 4 code via an LLM backend (API or self-hosted), sends it to the Lean 4 kernel through the Axiom API, and returns whether it type-checks.
Correctness is determined solely by the trusted kernel.
This project was started during the 2026 Mistral AI Hackathon (NYC) and achieved 2nd place :)
Python 3.10+
Unix system recommanded
Python environment:
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
API Keys:
export AXLE_API_KEY=axle_key
export MISTRAL_API_KEY=mistral_key
Axiom Lean Engine (replaces local Lean execution since it is faster, released on March 5, 2026)
To use another model, adapt api_client.py for other API / local models.
Run the app with
streamlit run app.py --server.address 127.0.0.1
The frontend was largely vibecoded during a hackathon and may not be super safe. I tried to add some safety checks, but I recommended to run it locally and avoid exposing it to the public network.
I may try to turn it into a VSCode extension or something similar.
Note that this is not tactic-level proof interaction (no info on the intermediate goals).
It is also dependent on LLM quality (Mistral works well here).