Gap between natural language and formal proofs in mathematical papers
Problem
As large language models are applied in mathematics, the gap between natural language proofs and formal systems like Lean becomes increasingly apparent. Users report that formalized Lean proofs often do not correspond to the original natural language proofs, and LLMs may 'get lazy' during translation, satisfying only minimal formal requirements while overlooking stronger claims. Furthermore, mechanizing printed papers often reveals skipped steps and handwaving in the original, causing formal verification results to diverge from the intended meaning.
Pain points
- Formalization processes reveal errors and omissions in original papers, such as type setting errors, mismatched quantifier scopes.
- LLMs 'get lazy' when translating natural language proofs, generating formal code that satisfies the theorem but overlooks stronger claims from the original.
- The gap between Lean proofs and prose (PDF) proofs, while not affecting validity, reduces interpretability, and interpretive issues are often as hard as the proof itself.
- Mechanizing printed papers is highly repetitive labor, often resulting in formal outputs that diverge from the original intent due to skipped steps and handwaving.
Questions to verify
- Would you be willing to pay for or switch to a tool that bridges natural language mathematical proofs and formal systems while preserving original intent?
- How critical is the gap between Lean formalization and the original prose proof for your work's interpretability?
- Have you encountered cases where LLM-generated formalizations unintentionally strengthened or weakened the original claims?
Evidence
In particular, we show that the formalised Lean proof does not correspond to the NL proof of blow-up of solutions to the Navier-Stokes equations.
The paper shows that the Lean proof and the prose (pdf) proof do not match exactly.
I recently spent 3 weeks with claude formalizing a CS paper about a borrow checker in lean
the translator LLM got lazy and wrote the minimal amount of code that satisfied the theorem without the extra stronger claims
Disclaimer: this card is distilled by AI from community discussions as “pain point evidence”. It is a starting point for need analysis and validation, not a “validated market need”. AI has analytical limits; please open the original discussion above to manually verify before key decisions.