Exploring Geographic Relative Space in Large Language Models through Activation Patching
For researchers in AI safety and geographic information science, this work provides initial mechanistic insights into LLMs' spatial reasoning, though it is an incremental contribution limited to a specific analysis technique.
This study uses activation patching to investigate how LLMs represent relative geographic space, revealing that spatial information is processed in a distributed manner across multiple layers rather than localized in specific regions.
The increased use of Large Language Models (LLMs) in geography raises substantial questions about the safety of integrating these tools across a wide range of processes and analyses, given our very limited understanding of their inner workings. In this extended abstract, we examine how LLMs process relative geographic space using activation patching, an emerging tool for mechanistic interpretability.