MechGeo: Autoformalizing and Proving Euclidean Geometry in Lean 4
AI Summary
MechGeo is a Mathlib-native agentic framework for faithful formalization and certified proving of Euclidean geometry in Lean 4. GeoFormalizer maps informal problems into GeoIR and Lean, then repairs candidate statements using structural diagnostics and semantic evaluation. GeoProver builds proof plans, derives intermediate lemmas, and algebraizes selected subgoals with Lean-verified libraries. Across 43 historical IMO geometry problems, it proved 29 formalized statements, produced Lean-checked counterexamples for 14, and proved all repaired versions after expert correction. On the 14-problem Lean-IMO-Bench geometry subset, it proved 12 and formally refuted two.
Why it's worth reading
The work is timely because it treats incorrect formalizations as first-class outputs: Lean checks proofs, counterexamples, and repaired statements, addressing a central reliability problem in LLM-assisted theorem proving.
Deep Read
What happened
Original facts: The paper introduces MechGeo, a Mathlib-native Lean 4 agentic framework for autoformalizing and proving Euclidean geometry. It combines GeoFormalizer and GeoProver and reports results on historical IMO geometry problems, IMO 2026 Problem 2, and the geometry subset of LEAP's Lean-IMO-Bench.
Core technology
Original facts: GeoFormalizer represents informal statements in GeoIR, deterministically translates them into Lean 4, and iteratively repairs candidates using structural diagnostics and semantic evaluation. GeoProver constructs geometric proof plans, derives intermediate lemmas, and algebraizes suitable subgoals. Singular or SymPy may produce algebraic certificates, but Lean's kernel checks every resulting proof and counterexample.
Key evidence & numbers
Original facts: The experiments use seven LLM backbones. Among 43 historical IMO geometry problems, GeoProver proves 29 formalized statements. For the other 14, the system constructs Lean-verified counterexamples; after expert correction, it proves all repaired statements. On 14 geometry problems from LEAP's Lean-IMO-Bench, MechGeo proves 12 for the first time, formally refutes two, and proves both repaired statements.
Why it matters
Analysis: Formal theorem proving can fail because the translated proposition is semantically wrong, not merely because proof search is incomplete. By treating counterexamples as diagnostic outputs, MechGeo separates provable statements, false statements, and corrected statements. This offers a more informative evaluation protocol for LLM-based mathematical reasoning than proof success alone.
Practical impact
Analysis: GeoIR, automated geometric lemma generation, and selective algebraization could reduce manual effort when building Lean geometry libraries. For benchmark design, publishing counterexamples and repaired proofs can expose errors in either problem statements or formal translations. The requirement that external computer-algebra outputs be checked by Lean makes the workflow more suitable for auditable proof pipelines.
Limitations & uncertainty
Original facts and uncertainty: The abstract does not identify the seven backbones, provide per-stage success rates, report runtime or cost, give proof-size statistics, or present a complete baseline comparison. It also does not explain how experts selected the 14 repaired statements. The claim of the “largest reported collection” is scoped by the authors and requires checking the paper and related work. The abstract provides no ablations for GeoIR errors, representation choices, or problem complexity, so generalization to unseen geometry styles cannot be inferred from these results alone.
Original sources
- arXiv abstract page
- Paper: MechGeo: Autoformalizing and Proving Euclidean Geometry in Lean 4
- arXiv ID: 2608.02295