UrbanDS: A Graph-Guided LLM Multi-Agent System for Data-Intensive Urban Tasks UrbanDS：一种面向数据密集型城市任务的图引导LLM多智能体系统

TL;DR

• UrbanDS introduces a graph-guided multi-agent system designed to handle data-intensive urban tasks by leveraging large-scale, heterogeneous datasets.
• The system constructs a unified dataset graph using Data Profiling and Relation Agents to organize reusable skills and inter-dataset relationships.
• A Planner Agent retrieves relevant datasets and generates execution plans, while Execution Agents process data collaboratively via shared memory, culminating in Report Agent synthesis of results.
• UrbanDS-Bench, a new benchmark for urban data science tasks, demonstrates UrbanDS’s superior performance over existing agents on both general and domain-specific challenges.
• Real-world deployment in Wuhan’s Dongxihu District validates the system’s practical utility in urban operations.

Why It Matters

This work addresses a critical gap in LLM-based automation: the inability of current systems to effectively navigate and integrate complex, large-scale, multi-source data typical in urban environments. By introducing a structured, graph-guided multi-agent architecture, UrbanDS enables scalable, collaborative data analysis that is essential for smart city applications, offering a blueprint for future AI systems operating in data-rich, real-world domains.

Technical Details

• Unified Dataset Graph: A central knowledge structure that encodes dataset metadata, processing skills (via Data Profiling Agent), and semantic/spatial/temporal relationships (via Relation Agent), enabling efficient retrieval and reuse.
• Multi-Agent Architecture: Five specialized agents—Data Profiling, Relation, Planner, Execution, and Report—each with distinct roles, coordinated through a shared memory for state and intermediate result exchange.
• Execution Workflow: Planner identifies relevant datasets from the graph, generates a task-specific plan, delegates sub-tasks to Execution Agents, which collaborate iteratively before Report Agent compiles final outputs.
• UrbanDS-Bench: A curated benchmark comprising representative urban data analysis and modeling tasks (e.g., traffic prediction, resource allocation), used to evaluate agent performance under realistic constraints.
• Real-World Deployment: Integrated into Wuhan’s urban operations platform, demonstrating operational viability and adaptability to dynamic, real-time urban data streams.

Industry Insight

The success of UrbanDS underscores the necessity of moving beyond monolithic or isolated LLM agents toward modular, graph-augmented multi-agent systems capable of managing complexity in large-scale data environments. For AI practitioners, this suggests investing in structured data representation (like dataset graphs) and agent coordination mechanisms as foundational components for deploying robust, scalable solutions in sectors like urban planning, public health, and logistics. Furthermore, the integration of human feedback loops in report refinement highlights the importance of iterative, user-centered design in real-world AI deployments.

TL;DR

• 提出UrbanDS，一种基于图引导的LLM多智能体系统，用于处理数据密集型城市任务。
• 构建统一数据集图谱，组织可复用的数据集技能及数据集间关系，提升数据发现与利用效率。
• 引入Data Profiling Agent、Relation Agent、Planner Agent、Execution Agents和Report Agent等分工协作的智能体架构。
• 构建UrbanDS-Bench基准测试集，涵盖典型城市数据分析与建模任务，验证系统在复杂场景下的优越性。
• 已在武汉东西湖区城市运营平台部署，证明其在真实城市治理中的实用价值。

为什么值得看

该工作针对城市数据规模大、来源异构、关系复杂的特点，创新性地将图谱结构与多智能体协同机制结合，为AI从业者提供了可扩展、可解释的数据科学自动化新范式，对推动城市智能决策系统落地具有重要意义。

技术解析

• 核心架构采用“图谱+多智能体”设计：通过Data Profiling Agent为每个数据集生成技能描述，Relation Agent识别并建模数据集间的空间、时间与语义关联，形成统一的数据集图谱作为全局知识索引。
• Planner Agent根据任务需求从图谱中检索相关数据集并生成执行计划，多个Execution Agents并行处理数据，共享中间结果至公共内存，实现动态协作与状态同步。
• Report Agent自动整合实验日志生成结构化报告，支持用户反馈迭代优化输出内容，形成闭环分析流程。
• UrbanDS-Bench benchmark包含交通预测、污染溯源、人口分布建模等12个典型城市任务，覆盖数据清洗、特征工程、模型训练与可视化全流程，用于系统评估泛化能力与鲁棒性。
• 实验表明，在UrbanDS-Bench及通用数据科学基准上，UrbanDS相比传统单智能体或静态数据供给方案，在任务完成率和准确率上平均提升23.7%，尤其在跨源数据融合场景中优势显著。

行业启示

• 城市级数据智能系统需突破单一数据源限制，构建具备自发现、自组织能力的图谱驱动架构，以应对日益复杂的异构数据环境。
• LLM多智能体系统应强调角色分工与信息共享机制的设计，通过模块化组件（如规划、执行、报告）提升系统的可维护性与扩展性。
• 技术落地需结合具体场景进行定制化部署，如本案例中在城市运营平台的成功应用，表明“算法+业务+基础设施”三位一体推进是实现规模化价值的关键路径。