LLM-CDSim: A Simulation-Based Framework for Language-Model-Assisted Customs Declaration

Customs declaration automation is a pressing need in the digital transformation of international trade. Existing systems lack adaptability to diverse document types, languages, and regulatory changes. This paper introduces LLM-CDSim, A Simulation-Based Framework for Language-Model-Assisted Customs Declaration, a role-driven agent framework that simulates customs processes using large language models. We design structured role prompts and integrate retrieval-augmented generation to model declarants, inspectors, and regulators. Scenario-based experiments on food trade show LLM-CDSim generates compliant declarations and adapts to complex rules. Our work provides a scalable simulation tool for testing policies, training agents, and advancing intelligent customs systems.