Cross-Attention between models to compose their representations and enable new capabilitiesarxiv.orghttps://arxiv.org/pdf/2401.02412.pdf