Show HN: ProofForge, AI agents whose proofs have to compile in Lean (github.com)

A new AI-agent pipeline called ProofForge has been announced that produces machine-verified proofs in Lean 4, a programming language and theorem prover widely used in formal mathematics. This system enhances the integrity of automated proof generation by ensuring that each proof generated by the AI must compile correctly in Lean, meaning that any erroneous proofs will not be accepted. This eliminates blind trust in the AI's output, as the Lean kernel rigorously verifies each step. Since the proofs are successfully merged into the google-deepmind/formal-conjectures repository, the initiative has already demonstrated its capability by solving significant problems related to Erdős conjectures, contributing six formal proofs that had previously remained unverified. The significance of ProofForge for the AI and ML community lies in its advancement toward reliable and rigorous machine-generated proofs, pushing the boundaries of formal verification. This initiative not only aids mathematicians by providing reliable tools for formal proofs but also opens avenues for further research and development in AI's ability to assist in complex problem-solving scenarios. With detailed processes like decomposition, adversarial verification, and formalization playing key roles, ProofForge stands as a pivotal integration of AI into the future of mathematics and machine learning, promoting greater accuracy and trustworthiness in computational proofs.