Understanding Language Prior of LVLMs by Contrasting Chain-of-Embedding