OT on the Map: Quantifying Domain Shifts in Geographic Space

The core contribution isn't a new model architecture but a diagnostic tool: GeoSpOT gives practitioners a way to estimate deployment risk before a model ever runs in a new region, which is a different problem than improving accuracy. Most geospatial ML work focuses on training better models; this focuses on knowing when not to trust the one you already have.

This connects loosely to the LLM generalization paper from arXiv covered around the same time ('Generalization in LLM Problem Solving: The Case of the Shortest Path'), which found that models transfer well across similar problem structures but break down when the underlying distribution shifts in scale. GeoSpOT is essentially attacking the same failure mode from a different angle: rather than discovering the breakdown after the fact, it tries to predict it upfront using optimal transport distances. The broader archive here skews toward application and infrastructure stories, so this paper sits in a quieter corner of the coverage, closer to ML reliability research than to product launches.

The meaningful test is whether GeoSpOT's predicted domain distances actually correlate with observed accuracy drops across a diverse benchmark of real-world geospatial deployments, not just the datasets used in the paper. If an independent replication on held-out regions confirms the correlation holds, this becomes a practical pre-deployment checklist item for geospatial teams.