interpretabilityfactbullishConcept Activation Vectors extracted via Sparse Autoencoders can explain image similarity in a model- and metric-agnostic wayComputer Vision02 Aug 2026http://arxiv.org/abs/2607.28386v1