[2205.12630] Multimodal Knowledge Alignment with Reinforcement Learning