From Answers to Relationships: Dual Digital Twins for Well-Being-Oriented Face-to-Face Service

Abstract

Advances in artificial intelligence, particularly large lan-guage models, have accelerated the development of an-swer-oriented systems optimized for correctness, effi-ciency, and task completion. However, many face-to-face human services—such as caregiving, rehabilitation, and education—operate within relationship-oriented con-texts in which value emerges through engagement, emo-tional alignment, and motivational support rather than informational accuracy alone. This paper introduces the Dual Digital Twin (DDT) framework as a computational architecture for modeling and analyzing such relational interactions. The framework simulates both participants in a service encounter through a Resident Twin and Staff Twin, while an Observer Twin evaluates relational dy-namics including engagement persistence, emotional res-onance, and motivational impact. Implemented as a mul-ti-LLM agent system on the Dify platform, the frame-work enables controlled scenario simulation and reflec-tive training. We conduct a rehabilitation motivation study centered on daily walking goals monitored via wearable devices, comparing task-oriented, relationship-oriented, and hybrid dialogue strategies. Simulation re-sults indicate that while task-oriented feedback ensures performance clarity, relational continuity is essential for sustaining engagement following negative outcomes. Hybrid strategies integrating performance transparency with motivational scaffolding yield the most balanced results. These findings position AI not as a replacement for relational labor, but as a reflective partner that illu-minates the hidden mechanisms sustaining meaningful human interaction and well-being.