Additional Details

Urban renewal rarely involves planning cities from scratch. Instead, real-world urban development proceeds through incremental and interdependent modifications over existing geospatial layouts. However, current computational approaches predominantly treat urban planning as a one-shot generation problem, lacking support for dependency-aware, iterative editing of structured spatial plans. In this talk, we introduce City Geospatial Editing as a new computational paradigm that reframes urban renewal as an executable, constraint-aware modification process over structured GeoJSON representations. We propose CEAE (City Editing via Agentic Execution), a hierarchical agentic framework that decomposes natural-language editing instructions into geometry-level intents and performs coarse-to-fine execution across polygon, line, and point abstractions. To ensure reliability, CEAE integrates explicit dependency propagation and self-reflective validation, enabling intermediate state verification, rollback, and controlled re-execution. Extensive experiments across multi-level urban editing tasks demonstrate significant improvements in execution correctness, structural validity, and robustness compared to single-pass baselines. This work highlights a paradigm shift from generative urban planning toward structured, validation-aware urban editing, and opens new research directions at the intersection of agentic AI, geospatial reasoning, and execution-centric foundation models for structured environments.