Enhancing Multi-Step Mathematical Reasoning in Large Language Models with Step-by-Step Similarity Prompts and Answer Voting

Complex reasoning problems, especially multi-step mathematical reasoning problems, are a difficult class of NLP tasks to solve. Existing methods such as Manual-CoT improve the accuracy of reasoning tasks by manually designing prompts to allow large models to output reasoning paths. However, the quality of the inference steps generated by this method is not high, resulting in many calculation and planning errors. To tackle the problems, we propose a method that combines enhanced similar step-by-step prompts with an answer voting mechanism. Specifically, we first design a comprehensive prompt template that integrates task prompts, CoT prompts, and format prompts, and then use two similar templates to guide the Large Language Model in generating better inference paths. Furthermore, we use ChatGLM for efficient information retrieval and determine the most accurate answer through a majority voting system. We evaluate our method in five mathematical datasets and one symbolic dataset. The experimental results over GPT-3 show that our proposed method outperforms Zero-shot-CoT and Zero-shot-Program-of-Thought Prompting across all datasets by a large margin of 7.3\% and 4.4\% respectively, and exceeds Plan-and-Solve in five of six datasets. Particularly, on symbolic datasets our method completely outperforms all comparable methods by a large margin of an average of 13\%. Our code and data are publicly available at https://anonymous.4open.science/r/ESPDE-2740.