AgentMove: A Large Language Model based Agentic Framework for Zero-shot Next Location Prediction

TL;DR

AgentMove tackles zero-shot next-location prediction by integrating a modular, agentic design with LLMs to model mobility through memory, world knowledge, and collective pattern extraction. It decomposes the task into personal pattern mining, collective pattern discovery, and urban-structure modeling, implemented via a spatial-temporal memory, a world knowledge generator, and a collective knowledge extractor, followed by a final reasoning step. Across mobility data from 12 cities, AgentMove outperforms strong baselines on most metrics, demonstrating robustness to different LLMs and reduced geographical bias. The framework highlights how explicit memory, structured world knowledge, and graph-based reasoning can leverage LLMs for scalable and generalizable spatial-temporal prediction tasks.

Abstract

Next location prediction plays a crucial role in various real-world applications. Recently, due to the limitation of existing deep learning methods, attempts have been made to apply large language models (LLMs) to zero-shot next location prediction task. However, they directly generate the final output using LLMs without systematic design, which limits the potential of LLMs to uncover complex mobility patterns and underestimates their extensive reserve of global geospatial knowledge. In this paper, we introduce AgentMove, a systematic agentic prediction framework to achieve generalized next location prediction. In AgentMove, we first decompose the mobility prediction task and design specific modules to complete them, including spatial-temporal memory for individual mobility pattern mining, world knowledge generator for modeling the effects of urban structure and collective knowledge extractor for capturing the shared patterns among population. Finally, we combine the results of three modules and conduct a reasoning step to generate the final predictions. Extensive experiments utilizing mobility data from two distinct sources reveal that AgentMove surpasses the leading baseline by 3.33% to 8.57% across 8 out of 12 metrics and it shows robust predictions with various LLMs as base and also less geographical bias across cities. Our codes are available via https://github.com/tsinghua-fib-lab/AgentMove.

Figures (5)

• Figure 1: The framework of AgentMove, including three key components: spatial temporal memory unit for capturing individual mobility pattern, world knowledge generator for multi-level urban structure, and collective knowledge extractor for extracting shared mobility patterns among users.
• Figure 2: Deep learning and LLM-based mobility predictors work in different ways. Deep learning models need to learn from training data for specific regions, while LLMs predict directly using zero-shot reasoning with its world knowledge.
• Figure 3: Illustration of spatial-temporal memory.
• Figure 4: Geospatial bias analysis of various methods in mobility prediction across 12 cities, where AgentMove outperforms most methods and exhibits lower geospatial bias.
• Figure 5: The effects of LLM with varying sizes and sources on the prediction performance of three LLM based methods.

Theorems & Definitions (4)

• Definition 1: Location
• Definition 2: User Trajectory
• Definition 3: Contextual Stays
• Definition 4: Historical Stays