“UrbanTrace: LLM-Assisted Discovery and Semantics-Aware Integration of Spatial Data,” by Sonia Castelo, Eden Wu, Joao Rulff, Harish Doraiswamy, Juliana Freire and Claudio Silva, was accepted to IEEE VIS 2026 and will appear in IEEE Transactions on Visualization and Computer Graphics.

UrbanTrace turns manual spatial data wrangling into a transparent, node-based workflow driven by context-aware AI agents. An offline profiler extracts semantic and geometric metadata that grounds the agents in real data distributions, so they can find datasets from an analyst’s high-level goals and enforce valid spatial aggregations. On 28 urban scenarios spanning 112 datasets, its profiling reached 100% semantic validity and 87% geometric validity in spatial mapping, well above baseline LLMs, and its provenance graph records how every transformation was derived.

Preprint: arXiv 2607.25124. Code: VIDA-NYU/urbanTrace.