Reproduce “ProofOptimizer: Training Language Models to Simplify Proofs without Human Demonstrations” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Alex Gu, Bartosz Piotrowski, Fabian Gloeckle, Kaiyu Yang, Aram H. Markosyan

ProofOptimizer introduces a system for training language models to simplify formal Lean proofs through self-supervision, eliminating the need for human demonstrations. It achieves substantial length reductions, such as 87% on miniF2F and 49% on complex Seed-Prover IMO proofs, which in turn improves the quality of training data for future provers and significantly reduces verification times.

Key insights

  • Leveraging Lean's mechanical verification capabilities as an automated feedback mechanism effectively overcomes the data scarcity problem for training proof simplification models.
  • Simplified proofs serve as higher-quality training data, leading to a 2% improvement in pass@32 accuracy for downstream proof generation models.
  • The iterative proof shortening workflow is highly effective, enabling progressive reductions in proof length and demonstrating steady improvements across multiple iterations.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.