Abstract
Geometry Problem Solving (GPS) requires a solver to comprehensively understand both text and diagram, master essential geometry knowledge, and appropriately apply it in reasoning. Inspired by dual-process theory, we propose DualGeoSolver to simulate the human dual-reasoning process. It consists of a Knowledge System (implicit reasoning) and an Inference System (explicit reasoning) that work iteratively to solve problems step-by-step. Experiments on GeoQA and GeoQA+ datasets demonstrate its superiority in accuracy and robustness from explicitly modeling human reasoning and knowledge application.
Introduction
Geometry Problem Solving (GPS) is a notoriously difficult task for AI because it requires "multi-modal" reasoning. To solve a single problem, a system must simultaneously parse textual descriptions, interpret geometric diagrams, and recall specific mathematical theorems (like the properties of parallel lines or angle bisectors).
Most existing neural solvers treat GPS as a "translation" task—mapping the input text and image directly to a solution program. This "black box" approach ignores how humans actually solve these problems: we identify a sub-goal (e.g., "I need to find this angle first"), recall a relevant theorem, look at the specific part of the diagram that theorem applies to, and then perform the calculation. DualGeoSolver addresses this by mimicking this step-wise, goal-oriented process.
The DualGeoSolver Framework
DualGeoSolver is built on the psychological "dual-process theory," consisting of two interacting systems that mimic human intuition and analytical reasoning:
- The Knowledge System (System 1 - Implicit): This acts as the "intuition" and "memory." Given a specific reasoning goal, it uses a Visual Spotlight Module to focus on relevant geometric primitives in the diagram and a Knowledge Selection Module to retrieve the appropriate theorem from a curated external knowledge base.
- The Inference System (System 2 - Explicit): This is the "analytical" side. It takes the "guiding information" from System 1 to generate program tokens (like
MinusorHalf) to solve the current sub-goal. Crucially, it includes a Goal Generation Module that identifies what the next logical step should be based on what has already been solved.
The two systems work iteratively. System 2 sets a goal, System 1 provides the "how-to" (theorems and visual focus), System 2 executes the step, and then they repeat the process until the final answer is reached.
Experimental Evaluation
The model was evaluated on two primary benchmarks: GeoQA (over 5,000 problems) and GeoQA+ (an expanded version with more complex problems and higher reasoning steps). DualGeoSolver set new performance benchmarks across the board:
- Accuracy: It achieved 65.2% on GeoQA and 65.1% on GeoQA+, surpassing previous state-of-the-art models.
- Comparison with LLMs: It significantly outperformed general-purpose multi-modal models. For context, GPT-4V only managed 29.7% on GeoQA. Even when GPT-4V was given the "ground truth" theorems, it only reached 34.4%, highlighting the advantage of DualGeoSolver's specialized reasoning structure.
Key Insights and Ideas
- The "Goal" is the Glue: Moving from "solving the problem" to "solving the next goal" allows the model to maintain a logical thread that standard Sequence-to-Sequence models often lose.
- Theorems are Bottlenecks: Ablation studies found that the Knowledge Selection Module was the most critical component. When the model couldn't retrieve the specific theorem, its accuracy plummeted.
- Interpretability: Because the model generates explicit goals and selects specific theorems, researchers can "peek under the hood" to see exactly where a reasoning chain breaks down.
Conclusion
DualGeoSolver demonstrates that explicitly modeling the human reasoning process—combining implicit knowledge retrieval with explicit logical inference—is highly effective for complex multi-modal tasks. While currently dependent on a manually curated knowledge base, future work aims to integrate these dual-reasoning capabilities into Large Language Models to combine broad linguistic power with precise, rule-based logic.
Summary updated