Google DeepMind Open-Sources AlphaProof, Bringing Formal Math Verification to the LLM Era
The system that solved International Mathematical Olympiad problems in 2024 has its source code and weights released, letting researchers couple language models to formal provers like Lean
Google DeepMind has released the source code and weights of AlphaProof, the system that in 2024 solved International Mathematical Olympiad (IMO) problems by coupling a language model to a formal prover. The release puts one of the most influential architectures for verifiable automated reasoning into the hands of the research community.
The decision turns what was a closed laboratory experiment into open infrastructure for universities, startups and independent groups working with machine-assisted mathematical proofs.
How the system works in practice
AlphaProof does not generate loose text: it produces proofs written in the Lean formal language, in which every logical step is checked by an independent kernel. If a step does not hold up, the proof is rejected — which eliminates much of the hallucination typical of language models.
The architecture combines a language model that proposes proof strategies with a search mechanism that tests and refines each attempt against the verifier. Learning happens through reinforcement, with the system accumulating valid proofs over the course of training.
Why formal verification matters
Verification via Lean has become one of the most contested fields in AI research. Unlike a chatbot that confidently answers a wrong equation, a system that writes formal proofs must convince a mathematical verifier — not a human reader.
This property attracts interest from laboratories seeking reliable models for software engineering, cryptography and scientific discovery, areas where a silent error is costly.
What changes with open source
With access to the code and weights, researchers can reproduce the training pipeline, adapt the system to other formal provers and measure its performance on their own benchmarks. The release also allows auditing DeepMind's design choices.
The move follows the laboratory's strategy of publishing research artifacts in formal mathematics, a field in which machine verification has become a standard of scientific quality.