GenesisGeo Model Overview

ZJUVAI/GenesisGeo is a specialized 0.8 billion parameter language model, based on the Qwen3-0.6B-Base architecture, designed for automated geometric theorem proving. Its primary function is to propose auxiliary constructions, which are crucial steps in solving complex geometry problems. This model is a core component of the GenesisGeo neuro-symbolic system, which aims to reproduce the AlphaGeometry framework.

Key Capabilities & Training

• Specialized Geometric Reasoning: Focuses exclusively on generating auxiliary points and constructions for geometric proofs.
• Neuro-Symbolic Integration: Designed to operate within a larger neuro-symbolic reasoning loop, enhancing its problem-solving capacity.
• Large-Scale Synthetic Training: Developed using a massive GenesisGeo Dataset comprising 21.8 million synthetically generated geometric theorems.

Performance

When integrated into the neuro-symbolic system, GenesisGeo demonstrates robust performance:

• IMO-AG-30 Benchmark: Solves 24 out of 30 problems on the challenging IMO-AG-30 benchmark, closely matching the original AlphaGeometry's performance (25/30).

Use Cases

This model is ideal for research and applications requiring advanced geometric reasoning, particularly in automated proof generation and educational tools for geometry.