interpretabilitycritiquebearishExisting gradient-based explainability methods struggle to provide global insights into what specifically drives similarity in regions of an embedding spaceComputer Vision02 Aug 2026http://arxiv.org/abs/2607.28386v1