Prompt Engineering for Virtual Asset Transaction Analysis using Large Language Models: A Pilot Study

Since their emergence, virtual assets have rapidly become a primary tool in criminal activities. Their advantages of speed and low cost, combined with the anonymity afforded by decentralization, necessitate that law enforcement agencies analyze virtual asset flows by cross-referencing vast numbers of transactions to identify suspicious behavioral patterns. The specialized skills and experience required for virtual asset flow analysis, coupled with the extensive time needed for case-by-case analysis, overwhelm law enforcement personnel in the face of a rapidly increasing number of cases. The rapid iterative development of large language models (LLMs) has shown impressive performance across diverse tasks in various fields. With their understanding of human semantics and needs, and through a reasonable logical thinking process, LLMs are well-equipped to serve as crucial assistants for law enforcement agencies in analyzing virtual asset flows. This study serves as a pilot investigation presenting the effectiveness of different prompting techniques on the analysis of temporal numerical data structures by large language models. It aims to stimulate further research in relevant academic fields and provide a basis for inference for law enforcement personnel during the initial stages of evidence investigation.