TrajGenAgent generates realistic mobility trajectories without fine-tuning
TrajGenAgent is a hierarchical LLM-agent framework that generates synthetic human mobility trajectories using a two-stage orchestrator-worker design, outperforming neural and LLM-based baselines without any model fine-tuning.
Score breakdown
TrajGenAgent demonstrates that a fine-tuning-free, hierarchical agent design can match or exceed the trajectory realism of computationally expensive fine-tuned models, lowering the barrier to generating privacy-safe synthetic mobility data for transportation, urban planning, and epidemic control applications.
- 01Proposed by Siyu Li, Toan Tran, and Lingyi Zhao; published June 10, 2026 on ArXiv.
- 02TrajGenAgent requires no model fine-tuning, relying entirely on in-context learning and deterministic workflows.
- 03Uses a two-stage orchestrator-worker design: an LLM generates an activity chain, then a deterministic workflow grounds it into full visit records.
Siyu Li, Toan Tran, and Lingyi Zhao introduce TrajGenAgent to address a core tension in LLM-based trajectory generation: prompt-engineering approaches preserve zero-shot reasoning but lack fine-grained spatiotemporal grounding, while trajectory-level fine-tuning improves statistical precision at the cost of substantial compute and potentially weakened general reasoning. TrajGenAgent sidesteps both drawbacks by operating entirely through in-context learning and deterministic post-processing, requiring no parameter updates.
In the first stage, an LLM synthesizes an activity chain conditioned on individual profiles and the day of the week, drawing on historical evidence via in-context learning.
The framework follows a two-stage orchestrator-worker architecture. In the first stage, an LLM synthesizes an activity chain conditioned on individual profiles and the day of the week, drawing on historical evidence via in-context learning. In the second stage, a deterministic workflow converts each activity into a complete visit record by applying personalized POI retrieval, distance-aware location selection, kinematics-aware travel-time propagation, and LLM-based duration estimation. The paper also contributes a new evaluation methodology: an anomaly-detection-based framework using two complementary detectors to measure behavioral and semantic plausibility, going beyond the aggregate spatiotemporal statistics that prior work typically relies on. Experiments on both benchmark and large-scale simulation datasets show TrajGenAgent improves spatiotemporal fidelity, semantic coherence, and individual-specific behavioral realism over representative neural and LLM-based baselines.
Key facts
- 01Proposed by Siyu Li, Toan Tran, and Lingyi Zhao; published June 10, 2026 on ArXiv.
- 02TrajGenAgent requires no model fine-tuning, relying entirely on in-context learning and deterministic workflows.
- 03Uses a two-stage orchestrator-worker design: an LLM generates an activity chain, then a deterministic workflow grounds it into full visit records.
- 04The deterministic grounding stage applies personalized POI retrieval, distance-aware location selection, kinematics-aware travel-time propagation, and LLM-based duration estimation.
- 05Activity chains are conditioned on individual profiles and day of the week using historical evidence.
- 06The paper introduces an anomaly-detection-based evaluation framework with two complementary detectors for behavioral and semantic plausibility.
- 07Outperforms representative neural and LLM-based baselines on spatiotemporal fidelity, semantic coherence, and individual-specific behavioral realism.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 12, 2026 · 10:05 UTC. How this works →