Wings: Learning Multimodal LLMs without Text-only Forgetting | Read Paper on Bytez