TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation

Researchers introduce TrajGenAgent, an LLM-based framework that generates realistic synthetic human mobility trajectories without model fine-tuning by combining hierarchical agent design with deterministic workflows. The approach addresses privacy and cost constraints in trajectory data collection while maintaining semantic coherence and behavioral realism.

TrajGenAgent represents a meaningful advance in synthetic data generation for mobility research by solving a fundamental tradeoff between reasoning capability and computational efficiency. Traditional approaches force developers to choose between prompt engineering that preserves general LLM reasoning but lacks spatial-temporal precision, or fine-tuning that improves accuracy at the cost of significant compute overhead and potential loss of generalization. This framework circumvents that dilemma through intelligent orchestration rather than parameter updates.

The two-stage design separates high-level semantic reasoning from low-level grounding. An LLM synthesizes activity chains using in-context learning—leveraging the model's existing knowledge without retraining—while a deterministic workflow handles the mechanistic details: POI retrieval, distance-aware location selection, kinematics-aware travel time, and duration estimation. This separation allows each component to excel at its specific task. The introduction of an anomaly-detection-based evaluation framework signals growing recognition that trajectory quality extends beyond aggregate statistics to individual behavioral plausibility, a critical metric for transportation simulations and urban planning applications.

For practitioners in transportation, urban planning, and epidemiology, this work reduces barriers to accessing high-quality synthetic training data while preserving privacy. Organizations can now generate realistic mobility patterns without costly collection campaigns or privacy risks. The computational efficiency gains matter significantly for institutions with limited ML infrastructure. The framework's ability to preserve semantic coherence—ensuring generated trajectories remain logically consistent—makes outputs suitable for downstream analysis rather than mere statistical benchmarking. Future relevance depends on whether the approach scales to diverse geographic regions and mobility contexts beyond the tested datasets.

• →TrajGenAgent generates synthetic mobility trajectories using LLM hierarchical agents without fine-tuning, reducing computational costs while preserving reasoning capability.
• →Two-stage orchestrator-worker design separates semantic activity planning from deterministic spatiotemporal grounding for improved efficiency.
• →Novel anomaly-detection evaluation framework assesses behavioral and semantic plausibility beyond traditional aggregate statistics.
• →Approach addresses privacy constraints and cost barriers in large-scale trajectory data collection for transportation and urban planning applications.
• →Framework avoids parameter updates while improving spatiotemporal fidelity and individual-specific behavioral realism over existing neural and LLM baselines.