UrbanDS: A Graph-Guided LLM Multi-Agent System for Data-Intensive Urban Tasks
The paper proposes UrbanDS, a graph-guided LLM multi-agent system that constructs a unified dataset graph to organize skills and relationships for retrieving and processing data, which effectively addresses the challenges of data-intensive urban tasks and outperforms existing methods in both benchmark evaluations and real-world deployment.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where computers are getting really good at talking to us, understanding our questions, and even writing code to solve problems. These smart computers are called Large Language Models (LLMs), and they act like super-powered assistants. But there's a catch: while they are great at chatting, they often struggle when asked to do real data science work, especially when the data is messy, huge, and scattered everywhere. Think of it like asking a brilliant librarian to find a specific fact, but instead of a neat library, you throw them into a giant warehouse where millions of boxes are piled up randomly, and you don't know which box holds the answer. The librarian might know how to read, but they can't find the right book without a map. This is the big problem researchers are trying to solve: how do we teach these AI assistants to not just answer questions, but to actually hunt down the right information from a chaotic ocean of data, figure out how different pieces fit together, and then do the math to give us a real answer?
Enter UrbanDS, a clever new system designed to be the ultimate "data detective" for city-related problems. The researchers behind this project realized that city data is a special kind of mess. It comes from everywhere—traffic sensors, population records, maps, and business logs—and all these pieces are connected in tricky ways (like how a road's location relates to the people living nearby). Existing AI tools usually fail here because they expect the data to be handed to them on a silver platter. UrbanDS changes the game by acting like a team of specialized agents working together. First, they build a giant, living "map" (a graph) of all the data. One team of agents goes through every single file, reads it, and writes a "resume" for it, describing what's inside and how it connects to other files. Another team draws lines between these resumes to show which data pieces are neighbors in time, space, or meaning.
When a human asks a question, like "Where are the traffic jams happening near schools right now?", a "Planner" agent looks at this map. Instead of guessing, it follows the lines to find the exact traffic and school data it needs. Then, a team of "Execution" agents writes and runs the code to crunch the numbers, sharing their progress like a group of friends working on a group project. Finally, a "Reporter" agent writes up the findings. The team tested this system on a massive challenge called UrbanDS-Bench, which included 94 different datasets from 10 major Chinese cities and nearly 500 different tasks. The results were impressive: UrbanDS solved these tough, data-hunting puzzles significantly better than other top AI tools, getting about 70% of the answers right compared to the next best tool's 63%. They even tested it in the real world in Wuhan, where it helped city planners analyze actual municipal data, proving that this "map-guided" approach works not just in a lab, but on the streets of a real city.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.