A learned prover exposed ambiguous translations of unique-answer math problems
Source check: Source checked. Not independent scientific confirmation.
Finding: A learned prover found two distinct Lean-accepted answers for 5.4% of the tested formal statements pooled across three autoformalizers. The original problems had unique answers, so these examples expose a mismatch between the formal statements and their intended problems.
Technology: Gobble-Prover 1.7B and 4B; Lean 4
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- Gobble-Prover models generated verifiable proof terms for numerical candidates under independently autoformalized competition problems.
- Test and result
- The audit combined evidence from two learned provers across three autoformalizers and five contest problem datasets. Lean checked acceptance of distinct answers.
- Limitation / still unknown
- The detected rate is a lower bound for these tested versions. A translation can be wrong in ways this multiple-answer test does not flag.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.34490v1 · https://arxiv.org/html/2610.34490v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: v1 Methods, unique-answer argument, Appendix D Table 5.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.