Magenta: Closing the Loop Between Mathematical Reasoning and Lean Verification

TL;DR - Magenta is a training-free agentic pipeline that converts natural-language math solutions into Lean 4 statements and machine-checked proofs. It closes the informal–formal reasoning loop with judges that detect faulty formalizations and route errors for targeted correction.

• Generates an answer, formalizes it in Lean 4, and constructs a verified proof from only a natural-language problem.
• Uses a statement judge to check semantic fidelity and an error-attribution judge to choose between mathematical re-derivation and local Lean repair.
• Reports 100% accuracy on evaluated olympiad benchmarks, including AIME 2025/2026 and HMMT February 2026.
• Paired with the open-weight K2-Horizon-7B reasoner, it reportedly solves all six IMO 2026 problems.