Self-review of real-analysis proofs
I tested whether asking Qwen to check a proof helped it correct mistakes. On ten handpickedproblems, a neutral request to review its work raised the score from 6/10 to 8/10. Of the four firstattempts graded incorrect, neutral review repaired two, a fresh retry repaired one, and reviewwith an explicit error flag repaired none. A retrospective audit used Lean to check solutions toall ten problems and retained all 28 prose grades. Every follow-up on an initial error reachedthe output limit. The extra repair from neutral review came from one problem
Authors
- Pinglikar Vivaan
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24357992
- Primary Topic
- Artificial Intelligence Applications
- Type
- preprint