MathEye turns flawed LLM proof drafts into verified Lean proofs with fewer human steps
The arXiv paper behind MathEye describes angelic proof checking plus a MaxSAT-based selection step that pinpoints the few obligations a mathematician must resolve.
In plain words
Language models can draft mathematical proofs in Lean, a system that checks proofs line by line, but small errors often make the whole draft fail. MathEye keeps the draft's structure, marks exactly what the model could not prove, and asks a human only the few questions needed to finish. The authors report it completed most test proofs while cutting the number of human interactions.
What was announced
Large language models can write Lean proof scripts that capture the high-level structure of many mathematical arguments, yet a single missing side condition or misapplied lemma can make the prover abort and discard an otherwise viable outline, the paper's abstract states. MathEye addresses this with angelic proof checking: an optimistic execution model that replaces failing fragments with placeholders while preserving the argument's structure, producing an angelic proof that isolates the exact obligations the model fails to discharge.
Because the unresolved obligations are often logically interdependent, asking a mathematician to answer every placeholder would be inefficient, the authors write. MathEye instead runs a residual analysis over a proof dependence graph that tracks symbol-level data flow, and formulates obligation selection as a MaxSAT optimization problem to pick a minimal-cost set of essential obligations whose resolution completes the proof. It presents these obligations in ordinary mathematical language and uses the answers to reconstruct the full Lean proof automatically. On 30 real-world benchmarks in analysis and algebra, the authors report MathEye solved 93.3% of problems while reducing required user interactions by 55%.
What it does not show
- Evidence is limited to the paper's abstract; the full article's methods, baselines and evaluation details are not available for verification.
- The 93.3% solve rate and 55% interaction reduction are the authors' own reported results on 30 benchmarks and have not been independently confirmed.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.