Exploring the Efficacy of Generative Pretrained Transformers for Arithmetic Problem Solving: A Comparative Analysis (KEY IJP************343)

• Mohit Santosh Pawaskar

Abstract

In this project, we investigate the effectiveness of using generative pretrained transformers (GPT) in solving math problems without resorting to calculators. Through a comparative analysis, we investigate the performance of different GPT variants, including COHERE, GEMMA, ZEPHYR, Meta-Llama, and ChatGPT, as well as DeepSeekMath. We perform arithmetic computations across different domains and complexity levels and evaluate the accuracy and efficiency of these models. Our results shed light on the capabilities of GPT-based approaches in mathematical problem solving tasks and provide insights into their potential applications in education, computing, and practice. We have developed the API for arithmetic computations and perform arithmetic operations using the DeepSeekMath-7B-instruct LLM model.