An arXiv paper (2610.12990) studies how offline data can help online learning in stochastic linear bandits. The setting is partially covered offline data: only k of the d feature directions are poorly covered, rather than the entire feature space being unavailable.
The paper introduces Gaussian-width bounds for offline-to-online linear bandits. These bounds are relevant to algorithms that build on offline data, where existing analyses have limitations in the partially covered case.
The abstract does not report specific numerical results, but it frames the contribution as a theoretical step toward understanding when limited offline coverage can still improve online decision-making.