Arithmetic-Aware Question-Answering on Tabular Data Using a Large Language Model-Based Code Generation Agent

Large language models excel in language understanding but struggle with precise numerical reasoning, making accurate tabular data question-answering challenging, especially with arithmetic operations. This study proposes an arithmetic-aware question-answering agent that enhances the accuracy of tabular question-answering by integrating code generation and knowledge injection techniques. We employ structured prompt engineering, table preprocessing, and dynamic function synthesis to improve numerical computations without applying additional model fine-tuning. Our results show that a native large language model-based approach achieves only 29% exact match accuracy, while our code generation agent reaches 58%. If we further exclude scale mismatches, the accuracy improves to 80%, which also shows an increase in arithmetic precision.