Reproduce “ProofBridge: Auto-Formalization of Natural Language Proofs in Lean via Joint Embeddings” 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

Prithwish Jana, Kaan Kale, Ahmet Ege Tanriverdi, Cruise Song, Sriram Vishwanath, Vijay Ganesh

Researchers from Georgia Institute of Technology and Bogazici University introduce PROOFBRIDGE, a unified framework that automatically translates natural language mathematical theorems and their proofs into Lean 4. This system achieves 62.70% semantic correctness and 95.49% type correctness on Olympiad-level problems by integrating cross-modal retrieval and an iterative repair mechanism.

Key insights

  • Encoding Lean proofs as Directed Acyclic Graphs (DAGs) and aligning them with natural language proofs in a joint embedding space significantly improves cross-modal semantic understanding and retrieval for formalization tasks.
  • Retrieval-augmented fine-tuning, where LLMs are conditioned on semantically relevant retrieved formal examples, is crucial for improving the generation of syntactically and semantically correct Lean proofs.
  • An iterative repair loop, combining feedback from Lean's type checker for syntactic validity and an LLM-based equivalence judge for semantic fidelity, is effective in refining generated formal proofs.

Summary by alphaXiv.

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