LLM-agent-based urban trajectory prediction for behavior-aware and sustainable mobility modeling under geographical constraints

Understanding and modeling urban mobility behaviors are essential for sustainable transportation planning and intelligent mobility management. However, existing trajectory prediction approaches largely rely on historical mobility patterns and task-specific training, limiting their ability to capture the contextual reasoning, heterogeneous behaviors, and adaptive decision-making processes underlying traveler movements. To address this gap, this study proposes TransAgent, an interpretable LLM-agent framework that leverages the semantic reasoning and contextual understanding capabilities of Large Language Models (LLMs), integrating geo-semantic context and spatiotemporal mobility knowledge to predict behavior-aware urban trajectories under geographical constraints. The proposed framework integrates spatiotemporal mobility context, road network semantics, and historical travel behaviors to model adaptive traveler decision-making and heterogeneous mobility patterns. TransAgent adopts a synergistic three-module architecture. First, a long-term memory module employs Retrieval-Augmented Generation (RAG) to retrieve candidate destinations and mobility patterns from historical origin–destination (OD) trajectories. Second, a short-term memory module integrates real-time trajectory observations with road network constraints to infer context-consistent travel paths. Third, a reflection module dynamically refines predictions through multi-perspective reasoning and iterative self-correction, improving behavioral plausibility and interpretability. The framework is empirically evaluated using large-scale taxi trajectory data from Chengdu, China, capturing diverse urban travel behaviors and spatiotemporal mobility dynamics. Experimental results demonstrate that TransAgent achieves state-of-the-art zero-shot prediction performance, reaching 71.32% Accuracy@5 and 68.02% NDCG, outperforming the strongest LLM-based baseline (LLM-Move) by 2.77% and 2.02% on Acc@5 and NDCG, respectively, and achieving the highest Acc@5 among all evaluated methods, including supervised models. Further analyses indicate that integrating geo-semantic context and mobility reasoning substantially improves the framework’s ability to capture context-dependent travel behaviors and complex urban mobility dynamics. Overall, this study contributes a novel LLM-agent framework for behavior-oriented travel behavior modeling and urban mobility simulation, providing new methodological insights into the application of LLMs for intelligent transportation and sustainable urban mobility research.