This paper proposes a system to automate the tax filing process, which requires complex reasoning and numerical calculations. Based on the fact that, according to the Internal Revenue Service, the average American spends $270 and 13 hours filing their taxes, this paper proposes a system to automate the tax filing process. Because existing large-scale language models (LLMs) have limitations in terms of accuracy and auditability, this paper presents an approach that integrates LLMs with symbolic solvers. Using the SARA dataset, we evaluate several variants of the system and propose a novel method for estimating the system's implementation costs based on penalties for real-world tax errors. Furthermore, we demonstrate how to improve performance and reduce costs by transforming plaintext rules into formal logic programs and intelligently searching for examples to formalize case representations. Ultimately, we demonstrate the potential and economic feasibility of leveraging a neural symbolic architecture to increase equitable access to reliable tax assistance.
