Title: Enhancing Mobility Prediction Efficiency and Reliability via an LLM-Powered Agent

Abstract: Predicting individual movement patterns is a critical component for urban simulation, transportation logistics, and policy assessment. While supervised sequence models deliver high accuracy, they are constrained by the need for task-specific training and provide limited insight into decision-making processes. Although recent approaches leveraging Large Language Models (LLMs) have enhanced interpretability, they typically depend on static prompts and single-pass inference, which hinders their capacity to gather further evidence when mobility cues are faint or contradictory. To address these limitations, we introduce \method{}, a training-free framework driven by LLM agents that frames next-location prediction as a decision-making process governed by adaptive evidence. This system handles routine scenarios through a rapid pathway based on historical consistency, while complex or ambiguous situations initiate iterative tool usage that analyzes recent trajectories, past behavior, stay-move probabilities, and geographical data. Evaluated across three mobility datasets, AgentMob demonstrates superior performance among training-free LLM-based methods. Specifically, GPT-5.4 achieved an Acc@1 of 71.42% on the BW dataset, 33.14% on YJMob100K, and 33.50% on Shanghai ISP. In non-fast-path instances on the BW dataset, the LLM controller raised Acc@1 from 30.65% to 48.62% compared to a statistical baseline using identical tools, highlighting that the primary advantage of this approach is the resolution of ambiguous predictions via adaptive evidence collection. Our source code is accessible at https://github.com/Unknown-zoo/AgentMob.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC