TrajGenAgent generates realistic mobility trajectories without fine-tuning

Human mobility data is critical for transportation, urban planning, and epidemic control, but collecting real trajectories is costly and privacy-invasive. Existing LLM-based generators either rely on prompt engineering (lacking fine-grained spatiotemporal grounding) or trajectory-level fine-tuning (computationally expensive and risking loss of general reasoning). To address this, researchers have developed TrajGenAgent, a semantic-aware hierarchical LLM-agent framework that generates synthetic trajectories without any parameter updates. The system employs a two-stage architecture: first, an orchestrator LLM synthesizes an individual- and weekday-conditioned activity chain from historical data using in-context learning; second, a deterministic worker workflow converts each activity into a complete visit using personalized POI retrieval, distance-aware location selection, kinematics-aware travel-time propagation, and LLM-based duration estimation.

To evaluate realism beyond aggregate statistics, the team introduced a novel anomaly-detection-based evaluation framework with two complementary detectors that assess both behavioral and semantic plausibility. Experiments on benchmark datasets and large-scale simulations show that TrajGenAgent outperforms representative neural and LLM-based baselines across spatiotemporal fidelity, semantic coherence, and individual-specific behavioral realism—all while avoiding expensive fine-tuning. Accepted at MDM 2026, this work offers a practical, privacy-preserving solution for generating high-quality mobility data, with potential applications in urban infrastructure modeling, epidemic spread simulation, and transportation demand forecasting.