Large language models reason about space through text: locations, directions, containment, distance, maps, and object relations. Yet it remains unclear whether these abilities rely on shallow linguistic associations or on internal structures that behave like spatial representations.
This project investigates how spatial information is represented in LLMs. The goal is to understand whether models form coherent internal geometries for spatial relations, how those representations vary across layers and prompts, and whether they can support reliable spatial reasoning.
More broadly, the work asks how abstract structure emerges from language-only training, and what this reveals about representation learning in modern AI systems.