Birds of a Feather Flock Together: Background-Invariant Representations via Linear Structure in VLMs

Linear additivity in VLM embeddings (CLIP, SigLIP 2) enables decomposition of scene representations into foreground and background components to mitigate spurious background-object correlations.