Emergent Spatio-Semantic Structure in Large Language Model Embedding Spaces

Abstract

This research examines whether high-dimensional embedding spaces generated by Large Language Models (LLMs) encode intrinsic geospatial structures. Analyzing extensive spatial-textual data from urban property descriptions, we demonstrate that latent language representations exhibit emergent spatial-semantic topology capable of reflecting geographical proximity and functional urban clusters without explicit coordinate conditioning.

Yunus Serhat Bıçakçı

Assistant Professor

Assistant Professor specializing in GeoAI, Multimodal Vision-Language Models, and Spatial Data Science.