LLM-driven multi-agent framework for enhancing human-digital twin interaction in built infrastructure management

Digital twins (DTs) are becoming a transformative technology for built infrastructure management. However, their potential remains constrained by the limitations of existing human-DT interaction modes, namely graphical user interfaces (GUIs), which impose steep learning curves, rigid workflows, and high cognitive loads. This paper introduces a large language model (LLM)-driven multi-agent system (MAS) framework as a complementary human-DT interaction paradigm. Within this framework, a Leader Agent, supported by a Reasoning Agent, coordinates specialized Member Agents to decompose tasks, orchestrate workflow, and orchestrate heterogeneous DT functions. A proof-of-concept prototype, termed HighwayMAS, was developed for highway infrastructure management and evaluated through a mixed-design experiment involving 18 highway experts. Results showed that HighwayMAS significantly outperformed the traditional GUI in reducing cognitive workload, improving task performance, and improving usability and user experience. Importantly, participants without prior DT experience completed tasks successfully, signifying MAS as a promising direction for more intuitive, adaptive, and human-centered DT interaction. • Digital twins (DTs) have emerged as a promising solution for intelligent built infrastructure management. • The existing GUI-based interaction modes impose learning barriers, rigid workflows, and high cognitive loads. • An LLM-driven multi-agent system (MAS) framework is proposed as an alternative human-DT interaction mode. • A proof-of-concept prototype, HighwayMAS, was implemented for highway infrastructure management. • User study showed HighwayMAS reduced workload and improved task performance, usability, and user experience.