We propose SpatialLLM, an integrated framework advancing spatial intelligence tasks in complex urban scenes. Unlike previous methods requiring specialized geographic analysis tools or domain expertise, SpatialLLM leverages the inherent reasoning capabilities of pre-trained Large Language Models (LLMs) to address various spatial intelligence tasks. The core of SpatialLLM lies in constructing detailed and structured scene descriptions from raw spatial data to prompt LLMs for scene-based analysis. Extensive experiments demonstrate that, with our designs, general-purpose LLMs can accurately perceive spatial distribution information and execute advanced spatial intelligence tasks, including urban planning, ecological analysis, traffic management. We further investigate key factors influencing LLM performance in urban analysis, such as multi-field knowledge, context length, and reasoning ability. We hope that SpatialLLM offers a viable perspective for intelligent urban analysis and management. The code and dataset are available at https://github.com/WHU-USI3DV/SpatialLLM . • Propose SpatialLLM, a framework for urban intelligence without fine-tuning. • Propose MDJD Module fusing urban spatial data into detailed scene description. • Reveal critical factors governing LLM spatial intelligence performance. • A novel dataset with spatial QA cases for evaluating LLM spatial perception.
Chen et al. (Wed,) studied this question.