ARMove: Learning to Predict Human Mobility through Agentic Reasoning

A research team led by Chuyue Wang has introduced ARMove, a groundbreaking framework that revolutionizes human mobility prediction through agentic reasoning. Unlike traditional black-box approaches, ARMove employs a sophisticated architecture featuring four standardized feature pools for foundational knowledge, user profiles for segmentation, and an automated generation mechanism that integrates LLM knowledge. The system's core innovation lies in its agentic decision-making process, which dynamically adjusts feature weights to maximize prediction accuracy while maintaining full transparency in decision paths. This addresses long-standing challenges in the field, including limited interpretability and poor transferability across different populations and regions.

ARMove achieves robust generalization through a unique large-small model synergy approach, where strategies learned from massive 72B parameter LLMs are distilled into more efficient 7B models. This not only reduces computational costs but also enhances performance ceilings. Extensive testing across four global datasets demonstrated significant improvements, with ARMove outperforming state-of-the-art baselines on 6 out of 12 key metrics, achieving gains ranging from 0.78% to 10.47%. The framework's transferability was rigorously validated, showing consistent performance across diverse regions, user groups, and model scales while maintaining interpretability through transparent decision-making processes.