Bayesian optimization relies on posterior uncertainty to guide exploration, but that uncertainty carries an implicit spatial preference. A new paper on arXiv (2610.14696) points out that on bounded domains, Gaussian-process posterior variance is typically larger near faces, edges, and corners than in the interior.
The effect means that standard acquisition strategies may systematically favor boundary regions, not because they are more promising, but because the model's uncertainty is inflated there. The authors describe this as a hidden spatial preference that can misdirect the search process.
Because the abstract only introduces the phenomenon, the paper's full implications—such as possible corrections or alternative uncertainty measures—are not yet detailed in the available text. Still, the observation is a useful caution for practitioners who assume exploration is spatially neutral.