Gold-Medal-Level Olympiad Geometry Solving with Efficient Heuristic Auxiliary Constructions

摘要

Automated theorem proving in Euclidean geometry, particularly for International Mathematical Olympiad (IMO) level problems, remains a major challenge and an important research focus in Artificial Intelligence. In this paper, we present a highly efficient method for geometry theorem proving that runs entirely on CPUs without relying on neural network–based inference. Our initial study shows that a simple random strategy for adding auxiliary points can achieve ”silver-medal” level human performance on IMO. Building on this, we propose HAGeo, a Heuristic-based method for adding Auxiliary points in Geometric deduction that solves 28 of 30 problems on the IMO-30 benchmark, achieving “gold-medal” level performance and surpassing AlphaGeometry, a competitive neural network–based approach, by a notable margin. To evaluate our method and existing approaches more comprehensively, we further construct HAGeo, a benchmark consisting of 409 geometry problems with human-assessed difficulty levels. Compared with the widely used IMO-30, our benchmark poses greater challenges and provides a more precise evaluation, setting a higher bar for geometry theorem proving.