Reasoning Over Space: Enabling Geographic Reasoning for LLM-Based Generative Next POI Recommendation

TL;DR

This work tackles the challenge of integrating geographic signals into generative POI recommender systems that use large language models. It introduces Reasoning Over Space (ROS), which builds a Hierarchical Spatial SID to discretize locality and semantics and trains LLMs with a three-stage Mobility Chain-of-Thought (CoT) that reasons over personality, intent, and locality. ROS further aligns geographic behavior with real-world geography via spatial-guided reinforcement learning that combines a SID-based correctness reward and a distance-based feasibility reward. Empirical results on three LBSN datasets show that ROS achieves substantial improvements in HR@1 over strong LLM baselines while using a smaller backbone, and demonstrates strong cross-city generalization and improved geographic plausibility. The approach offers a principled framework for geography-aware reasoning in open-vocabulary POI recommendation with practical implications for location-based services.

Abstract

Generative recommendation with large language models (LLMs) reframes prediction as sequence generation, yet existing LLM-based recommenders remain limited in leveraging geographic signals that are crucial in mobility and local-services scenarios. Here, we present Reasoning Over Space (ROS), a framework that utilizes geography as a vital decision variable within the reasoning process. ROS introduces a Hierarchical Spatial Semantic ID (SID) that discretizes coarse-to-fine locality and POI semantics into compositional tokens, and endows LLM with a three-stage Mobility Chain-of-Thought (CoT) paradigm that models user personality, constructs an intent-aligned candidate space, and performs locality informed pruning. We further align the model with real world geography via spatial-guided Reinforcement Learning (RL). Experiments on three widely used location-based social network (LBSN) datasets show that ROS achieves over 10% relative gains in hit rate over strongest LLM-based baselines and improves cross-city transfer, despite using a smaller backbone model.

Figures (9)

• Figure 1: Comparison of different paradigms for incorporating geographic signals into LLM recommendation.
• Figure 2: Overall Framework of our method.
• Figure 3: Impact of reward weighting in spatial-guided RL.$\alpha$ and $\beta$ indicate the weights of the hierarchical correctness reward and the distance-based grounding reward, respectively. Results are reported in HR@1 on NYC.
• Figure 4: Cumulative Distribution of three datasets. Larger area under curve indicating better performance.
• Figure 5: NYC distance error percentiles versus sampling temperature $\tau$.