Mobility Anomaly Generation using LLM-Driven Behavior with Kinematic Constraints

Researchers have developed an LLM-driven framework to generate synthetic human trajectory anomalies with kinematic constraints, addressing the critical shortage of ground-truth anomaly datasets in spatial data mining. The system combines large language models with map-constrained routing and context-aware noise modeling to create realistic, annotated mobility anomalies at scale while respecting physical constraints.

The scarcity of labeled anomaly datasets has long constrained progress in trajectory analysis and spatial data mining. This work tackles a fundamental challenge in machine learning: the practical impossibility of collecting large-scale real-world anomaly examples due to statistical rarity, privacy regulations, and prohibitive acquisition costs. Researchers have engineered a hybrid synthetic approach that bridges simulation and reality, moving beyond purely artificial data that lacks physical plausibility.

The framework's innovation lies in its multi-stage architecture. LLM agents inject semantically meaningful behavioral anomalies—such as atypical check-ins or skipped routine visits—that reflect how humans actually deviate from normal patterns. Rather than accepting these modifications as-is, the system reconstructs valid spatial paths using map-constrained routing, ensuring generated trajectories remain physically feasible. This constraint-aware approach addresses a critical gap where purely synthetic anomalies often violate real-world geography or movement physics.

The context-aware spatial noise model represents another advancement, accounting for heterogeneous GPS sensor degradation across different environments. This heterogeneity matters substantially: GPS accuracy varies dramatically between urban canyons, rural areas, and indoor spaces. By parameterizing noise to match real-world conditions, the framework narrows the simulation-to-reality gap that typically degrades the performance of models trained on synthetic data when applied to real trajectories.

For researchers in anomaly detection, autonomous systems, and location-based services, this work enables systematic training and evaluation of detection algorithms against validated ground truth. The approach could accelerate development of surveillance systems, fraud detection in location-based services, and autonomous navigation safety validation.

• →LLM agents generate semantically meaningful trajectory anomalies while preserving physical feasibility through map-constrained routing reconstruction
• →Context-aware spatial noise modeling accounts for environment-specific GPS degradation, bridging the simulation-to-reality gap
• →The framework solves the ground-truth dataset scarcity problem that has constrained spatial data mining research
• →Generated datasets enable systematic training and evaluation of anomaly detection algorithms without privacy or cost barriers
• →Multi-stage architecture combines behavioral realism from LLMs with physical constraints and sensor-level accuracy emulation