Reproduce “Beyond LLM Judge: Mechanical Verification for Agentic Autoformalization” 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

Anna Mei, Michael R. Douglas

The authors present a Lean theorem verification pipeline that combines language-model comparisons of informal and formal statements with numerical sampling and checks for proofs independent of their goals. In a Taylor-remainder example, sampling exposed a false source claim that semantic comparisons accepted. On Beyond Compilation, the method accepted 3 incorrect-labeled items out of 154 (1.9%), while flagging 38.4% of correct-labeled items; dataset labels are imperfect.

Summary by alphaXiv.

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