From Compression to Execution: What Helps Large Language Models Digest Urban Graphs for Spatial QA?

Urban graphs are central to smart-city analytics, but they remain difficult for language-guided models, specifically LLMs, to use effectively in spatial question answering. This paper studies how urban graph structure can be exposed to such models through different graph-language interaction designs. We construct a controlled multi-task urban graph QA benchmark over three cities that covers adjacency count, adjacency binary, reachability, shortest path, and centrality, and compare seven mechanisms spanning compression, memory, retrieval, execution, and structural reasoning, alongside text-only and GNN-only baselines. Within this benchmark, methods using more explicit structural evidence generally outperform compression and memory-based approaches: Path Tokens reaches 78.1% average accuracy, Graph Tool 87.4%, and Counterfactual 71.7%, while centrality remains the most difficult task. These results should be interpreted as a comparison of graph-language interaction designs rather than a fully input-matched ablation, but they suggest that exposing task-relevant structural information is an important factor for urban spatial QA.