Standard conformal prediction is usually applied to scalar or vector outputs, but this framework targets a different setting: regression where each response is a point on the Grassmann manifold, meaning a linear subspace of a vector space. Instead of adjusting intervals in Euclidean space, the method forms prediction regions as metric balls directly on the manifold.
The construction is basis-invariant, so the resulting balls do not depend on a particular choice of basis for the subspace. That is an important property because Grassmannian points are often represented by many equivalent matrices, and a reliable prediction region should treat those representations consistently.
The source is an arXiv abstract, so the description focuses on the methodological framework rather than experimental results. Full details on the behavior of these prediction regions would need to be checked in the complete paper.