Abstract
Geometry Problem Solving (GPS) represents a significant challenge for artificial intelligence, requiring the integration of multimodal perception and logical reasoning. This paper provides a comprehensive survey of the field, synthesizing over 310 studies to map the transition from symbolic solvers to modern Multimodal Large Language Models (MLLMs). We present a unified taxonomy of tasks, datasets, and architectures, while identifying the persistent "perception bottleneck" and "reasoning gap" that hinder current neural approaches.
Introduction
Unlike standard text-based word problems, geometry requires a system to simultaneously parse natural language instructions and interpret complex spatial relationships within diagrams. Historically, the field relied on symbolic solvers that were logically rigorous but brittle. Deep learning has introduced flexibility, yet researchers face significant hurdles in bridging the gap between visual perception and deductive logic. This survey organizes the current landscape of GPS into a structured framework of tasks, methods, and evaluations.
Methodology
The survey categorizes deep learning GPS methods into three distinct stages:
- Architectures: Modern systems typically employ Encoder-Decoder frameworks, utilizing specialized text and diagram encoders (e.g., Transformers, ViT, CNNs) fused via co-attention. Recently, Decoder-Only MLLMs like GPT-4V have emerged, treating visual information as sequential tokens.
- Training Strategies: Beyond supervised fine-tuning, Reinforcement Learning (RL) techniques such as PPO and GRPO are used to optimize reasoning paths. Data augmentation remains crucial for generating synthetic problems to mitigate data scarcity.
- Inference Scaling: To address complex "Olympiad-level" problems, researchers utilize test-time scaling techniques, including Chain-of-Thought (CoT), Monte Carlo Tree Search (MCTS), and Program-of-Thought (PoT).
Task and Dataset Taxonomy
The landscape of GPS is divided into fundamental, core, and composite tasks, each supported by specific benchmark datasets.
- Fundamental Tasks: Focus on geometric perception, such as identifying points and lines (e.g., PGDP5K, GeoC50).
- Core Tasks: Address numerical calculations (e.g., Geometry3K, GeoQA) and the more rigorous challenge of theorem proving (e.g., UniGeo, Proving2H).
- Composite Benchmarks: Test general-purpose MLLMs on high-level competition problems (e.g., MathVista, OlympiadBench).
Key Insights and Results
The field has experienced exponential growth, with GPS-related publications increasing from a single paper in 2018 to over 110 in 2024. Key findings include:
- The Perception Bottleneck: Visual encoders often struggle with precise spatial coordinates, sometimes leading to lower accuracy when diagrams are included compared to text-only inputs.
- Neural-Symbolic Hybrids: Systems like AlphaGeometry demonstrate that combining neural models for theorem prediction with symbolic engines for logical deduction yields superior results.
- The Reality Gap: Models trained on synthetic data often fail to generalize to the hand-drawn or scanned diagrams found in real-world educational contexts.
Conclusion and Future Directions
While AI has become proficient at standard numerical calculations, theorem proving remains a "Grand Challenge." Future research must address data scarcity in 3D and analytic geometry, implement more rigorous process-based evaluations, and develop models capable of human-like creative strategies, such as the construction of auxiliary lines.
Summary updated