UrbanDS: Graph-Guided Multi-Agent LLM System for Urban Data Tasks

Large language model (LLM) agents are increasingly used to automate data science tasks, but they often struggle with data-intensive scenarios requiring the discovery and integration of information from vast, heterogeneous data repositories. Urban tasks exemplify this challenge, as urban data is large-scale, multi-sourced, and characterized by complex spatial, temporal, and semantic relationships. To overcome these hurdles, researchers introduce UrbanDS, a graph-guided LLM multi-agent system specifically tailored for data-intensive urban tasks. UrbanDS begins by constructing a unified dataset graph that organizes reusable dataset skills and their interrelationships. A Data Profiling Agent creates a skill for each dataset, while a Relation Agent identifies and integrates relationships into the graph. At runtime, a Planner Agent retrieves relevant datasets from the graph and generates execution plans. Multiple Execution Agents then perform data processing and analysis, sharing progress and intermediate results via a common memory. Finally, a Report Agent synthesizes experimental logs into a report, which can be refined with user feedback. Evaluated against both general and urban benchmarks, UrbanDS consistently outperforms existing data science agents on data-intensive tasks and has been successfully deployed on an urban operations platform in Wuhan, demonstrating its real-world effectiveness.