New AI Can Guess Your Next Stop — Using Your Neighborhood

Researchers propose a multi-agent LLM framework that predicts a user's next point of interest by breaking the task into three stages: a Pattern Extraction Agent that captures temporal and categorical mobility patterns from trajectory history, a Spatial Reasoning Agent that combines behavioral preferences with real-world spatial constraints like geographic distance, road network distance, and neighborhood affiliation, and a Decision Synthesis Agent that integrates both for the final prediction. The motivation: LLM-based approaches have focused on semantic reasoning from previous mobility records while neglecting real-world spatial context, and LLMs often struggle with spatial reasoning tasks such as distance estimation and geographically biased prediction. Experiments on the NYC benchmark dataset with two LLM backbones showed improvements over baseline methods, with up to 493% Hit@1 improvement and 37% relative improvement in Hit@5. Ablations showed that combining neighborhood affiliation with distance-based features generally outperforms distance-only settings, and that the Spatial Reasoning Agent plays a crucial role, especially for smaller models.