The Coordinates Were Always There: Detecting and Quantifying Pretraining Contamination from Geospatial Data in LLMs

Overview

This paper studies whether large language models have been contaminated by geospatial data during pretraining, and proposes methods to detect and quantify that contamination.

Venue

ACM SIGSPATIAL 2026, Riverside, CA

Recommended Citation

Awasthi, N., Abrar, S., Bardhoshi, D., Buddi, B., Kurella, E., & Frias-Martinez, V. (2026). The Coordinates Were Always There: Detecting and Quantifying Pretraining Contamination from Geospatial Data in LLMs. In *ACM SIGSPATIAL 2026. (Accepted)

* Equal contribution.

Accepted papers

Recommended citation: Awasthi, N.*, Abrar, S.*, Bardhoshi, D.*, Buddi, B.*, Kurella, E.*, & Frias-Martinez, V. (2026). The Coordinates Were Always There: Detecting and Quantifying Pretraining Contamination from Geospatial Data in LLMs. In ACM SIGSPATIAL 2026. (Accepted) *Equal contribution.
Download Paper