Learning to Solve Geometry Problems via Simulating Human Dual-Reasoning Process

Tong Xiao, Jiayu Liu, Zhenya Huang, Jinze Wu, Jing Sha, Shijin Wang, Enhong Chen, Tianming Yang, Wang-Zhou Dai

Introduction

Geometry Problem Solving (GPS) has long been considered a "grand challenge" for artificial intelligence. Unlike standard text-based tasks, GPS requires a sophisticated integration of visual perception, linguistic understanding, and logical deduction. A successful solver must parse a textual description, interpret a complex geometric diagram, recall relevant mathematical theorems, and apply them in a precise sequence to reach a solution.

Traditional approaches to GPS typically fall into two categories: symbolic solvers and neural solvers. Symbolic solvers rely on handcrafted rules, making them interpretable but brittle and difficult to generalize. Conversely, neural solvers treat GPS as an end-to-end translation task, mapping inputs directly to mathematical programs. While flexible, these models often lack the "common sense" geometry knowledge and structured reasoning strategies necessary for multi-step problems. To bridge this gap, we introduce DualGeoSolver, a neuro-symbolic framework that mimics the human "dual-process" of reasoning—combining intuitive knowledge retrieval with deliberate, step-by-step calculation.

Human-like dual-reasoning process for solving a geometry problem, decomposing the task into solvable sub-goals.

The DualGeoSolver Architecture

DualGeoSolver is designed to move away from monolithic end-to-end models toward a more modular, step-by-step abstraction process. The framework consists of two interacting systems: the Knowledge System (System 1) and the Inference System (System 2). This architecture allows the model to set intermediate goals, focus on relevant visual features, and apply explicit theorems iteratively until the problem is solved.

The DualGeoSolver architecture, featuring the interaction between the Knowledge System (System 1) and the Inference System (System 2).

System 1: The Knowledge System (Intuition and Memory)

Acting as the model's "intuition," the Knowledge System handles the retrieval of relevant information based on a specific reasoning goal (e.g., "find the measure of $\angle ABC$"). It comprises two primary modules:

  1. Visual Spotlight Module: Instead of processing the entire diagram statically, this module uses an attention mechanism to focus on specific geometric primitives—such as lines, points, and angles—that are relevant to the current sub-goal. This mimics how humans focus their gaze on specific parts of a figure while thinking.
  2. Knowledge Selection Module: This module retrieves relevant theorems from a manually curated Geometry Knowledge Base ($K_G$). For instance, if the current goal involves parallel lines, the system might retrieve the rule: "If two lines are parallel, then consecutive interior angles are supplementary."

System 2: The Inference System (Logical Calculation)

The Inference System acts as the "logical calculator," executing the steps suggested by the Knowledge System. It operates through two main components:

  • Token Prediction: Utilizing the guiding information (visual features and retrieved theorems) from System 1, it generates a sequence of program tokens (e.g., Minus, Half, C_3) to solve the current sub-goal.
  • Goal Generation: Once a reasoning step is completed, this module identifies the next solvable sub-goal based on the updated state of the problem. This iterative process continues until the final answer is reached.

Experimental Evaluation

DualGeoSolver was evaluated on two major benchmarks: GeoQA (5,000+ problems) and GeoQA+, an expanded dataset featuring more complex reasoning steps and a wider variety of problem types, including area calculations. To support the model, the authors developed a comprehensive knowledge base ($K_G$) sourced from Wikipedia and Baidu-Baike and verified by human experts.

Key Results

  • State-of-the-Art Performance: DualGeoSolver achieved 65.2% accuracy on GeoQA and 65.1% on GeoQA+, significantly outperforming previous specialized solvers like SCA-GPS.
  • Comparison with LLMs: The model dramatically outperformed massive multi-modal LLMs. While CogVLM and VisualGLM struggled (scoring 8.09% and 11.2% respectively), even GPT-4V only reached 29.7% accuracy on GeoQA. Interestingly, when GPT-4V was provided with the same explicit theorems used by DualGeoSolver, its performance improved, highlighting the importance of knowledge injection.
  • Ablation Studies: Removing the Knowledge Selection Module led to a nearly 2-point drop in accuracy, confirming that explicit theorem retrieval is critical for bridging the gap between visual patterns and mathematical operations.

Conclusion

DualGeoSolver demonstrates that solving complex, multi-modal reasoning tasks like geometry requires more than just large-scale data; it requires a structured approach to abstraction and knowledge application. By mimicking the human dual-reasoning process, the framework provides a robust method for decomposing complex problems into manageable steps. Future work will focus on automating the expansion of the knowledge base and exploring how this dual-reasoning framework can be used to guide LLMs in other specialized scientific domains.

Summary updated

BibTeX
@inproceedings{xiao2024learning,
  author = {Tong Xiao and Jiayu Liu and Zhenya Huang and Jinze Wu and Jing Sha and Shijin Wang and Enhong Chen and Tianming Yang and Wang-Zhou Dai},
  title = {{Learning to Solve Geometry Problems via Simulating Human Dual-Reasoning Process}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=NBmVr6VJ7j},
}
Figure