Truman: A Large Language Model-based Multi-agent Simulator for Synthetic Money Laundering Data Generation

Money laundering (ML) facilitates the cross-border movement of illicit funds, enabling organized crime by disguising the origins of illegal money. Financial institutions face significant challenges in combating it, primarily due to barriers in adopting advanced technologies such as machine learning, caused by restricted access to sensitive transaction data. Existing synthetic datasets often lack critical customer information and realism, reducing their utility for ML detection. This study presents Truman, an innovative data generator that leverages Large Language Model (LLM) based agents to create realistic financial transaction data, incorporating simulation of ML patterns. Expert validation confirms the dataset's quality and applicability for anti-money laundering research.