TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

TrajGenAgent is a novel hierarchical LLM-agent framework designed for generating realistic human mobility trajectories without model fine-tuning. It employs a two-stage orchestrator–worker design, decoupling macro-level activity chain planning from micro-level spatiotemporal grounding. The orchestrator, powered by Qwen2.5-32B-Instruct, synthesizes activity chains via in-context learning, while a deterministic LangGraph worker workflow instantiates each activity with distance-aware location retrieval and kinematics-aware temporal generation. The framework introduces an anomaly-detection-based evaluation using ICAD and BeSTAD to assess behavioral fidelity beyond aggregate spatiotemporal statistics. Experiments on NumoSim and MobilitySyn datasets demonstrate that TrajGenAgent outperforms existing baselines, including Geo-Llama and Geo-CETRA, in both statistical alignment and semantic coherence, while significantly reducing computational overhead by avoiding parameter updates.

Key takeaway

For Machine Learning Engineers developing synthetic mobility data generators, TrajGenAgent offers a compelling, cost-effective approach. Its zero-shot hierarchical agent design, which separates high-level planning from low-level grounding, achieves superior realism and semantic coherence compared to fine-tuned LLMs, without incurring substantial computational overhead. You should explore adopting similar workflow-managed agent architectures and integrate anomaly-detection frameworks like ICAD and BeSTAD to ensure generated data exhibits high behavioral fidelity.

Key insights

TrajGenAgent's hierarchical LLM-agent framework generates realistic human mobility trajectories by decoupling semantic planning from spatiotemporal grounding without fine-tuning.

Principles

• Decouple macro-level activity planning from micro-level spatiotemporal realization.
• Leverage in-context learning for fine-grained personalization without model fine-tuning.
• Utilize deterministic workflows for stable, long-horizon tool execution in agents.

Method

An LLM orchestrator synthesizes an activity chain via in-context learning, followed by a LangGraph worker workflow that deterministically grounds each activity using rule-based POI retrieval and kinematics-aware temporal generation.

In practice

• Apply hierarchical orchestrator–worker agent designs for complex generative tasks.
• Augment traditional metrics with anomaly detectors like ICAD and BeSTAD for behavior-aware evaluation.

Topics

• Human Mobility Generation
• LLM Agents
• Hierarchical AI
• Spatiotemporal Data
• Anomaly Detection
• LangGraph
• Zero-shot Learning

Code references

• Emory-AIMS/TrajGenAgent

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation — TrajGenAgent generates realistic human mobility trajectories using a hierarchical LLM-agent framework without fine-tuning.
• Winning a Kaggle Competition with Generative AI–Assisted Coding — LLM agents, paired with GPU acceleration, dramatically speed up ML experimentation by automating code generation and execution.
• GPS-Enhanced Tourist Mobility Modeling with Seasonal Spatial Priors and LLM-Based Activity Chain Generation — The framework combines GPS data, demographics, and LLMs to simulate realistic, demographically-aligned tourist mobility for urban planning.
• Harnessing Generalist Agents for Contextualized Time Series — TimeClaw's agentic harness framework enables LLMs to perform robust, contextualized time series reasoning by providing native temporal runtime support.
• WorkflowGen:an adaptive workflow generation mechanism driven by trajectory experience — WorkflowGen optimizes LLM agent workflow generation via adaptive experience reuse and minimal rewriting, reducing token costs and improving robustness.
• MolmoWeb: Generating Synthetic Data — Generate synthetic web browsing data at scale using LLM or vision-based agents for training.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.