Reproduce “Navier-Stokes lost in translation: Why Lean verification of AI autoformalisation does not guarantee correct natural language proofs” 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
Alexander Bastounis, Fabian Circelli, Anders C. Hansen
The paper examines whether Lean formalizations preserve natural-language proofs. Lean checks proofs against encoded statements, but compilation does not verify that translation preserves the source. Comparing OpenAI’s Navier–Stokes paper with its Lean formalization, the authors find an estimate requiring five rather than four extra derivatives, concluding the formalization is not faithful, not that the natural-language proof is wrong.
Key insights
- Formal verification certifies a proof relative to its formal premises; it does not, by itself, certify that those premises capture the intended natural-language mathematics.
- The examined Lean results differ from the source in both quantitative assumptions and reasoning, while general decisions about whether definitions are well-posed can encode problems beyond computable resolution.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.