Formal verification depends on precise specifications, yet the natural-language task descriptions given to LLMs are often vague. The paper argues that such descriptions may omit verifier-relevant edge cases, so the resulting Lean specifications are frequently imperfect. This frames the problem as one of underspecification and ambiguity rather than raw model capability.
The proposed direction is interactive clarification: asking users questions before writing the specification. The central question is when this extra interaction improves the output and when it does not. The abstract does not report a numerical result, so the contribution here is the setup itself—treating specification elicitation as a distinct phase from specification generation.
This digest is based on a single arXiv abstract, so there are no independent findings to compare. The paper is available as arXiv:2610.22382v1.