Mutli-Step Chain-of-Thought in Geometry Problem Solving

In the realm of natural language processing, large language models (LLMs) have demonstrated superb performance in human-level reasoning and text generation, which has inspired a large number of researchers to conduct extensive research on the automatic solution of mathematical word problems. The latest work has concentrated on the construction of extensive geometry datasets for fine-tuning, leveraging huge datasets to enhance the mathematical reasoning capabilities of models. However, the majority of these work focus on calculation problem solving, and little attention is given to the more challenge proving problems. To address this gap, we introduce a multi-step chain of thought approach. By employing two-step generation approach, our method enables the model to concentrate on distinct task, thereby augmenting the model's comprehension and mathematical reasoning ability. Furthermore, we explore the application of multi-modal large language models in geometry proving problems, proposing an evaluation framework that employ the existing powerful LLMs to evaluate model's performance.