Discovering and Detecting Tax Avoidance Using Natural Language Processing and Coevolutionary Algorithms

Abstract

The Internal Revenue Service estimates that unpaid taxes cost the United States up to $1 trillion each year, much of it through tax avoidance schemes such as the Installment Bogus Optional Basis strategy (iBOB). This thesis focuses on discovering and detecting iBOB schemes within a tax network by using a coevolutionary framework powered by two large language model (LLM) agents: a tax planner, which generates transactions to reduce tax and suspicion, and an auditor, which generates patterns to detect tax avoidance. This thesis shows that as the tax planner and auditor evolve against each other, the planner consistently uncovers iBOB transactions and even discovers more effective variants than those previously known. In turn, the auditor learns to generate increasingly targeted detection patterns that adapt to the planner’s changing strategies. Their interaction leads to oscillating behavior between avoidance and enforcement, providing a new LLM-based framework for studying the dynamics of tax compliance.