A Chain-of-Thought Subspace Meta-Learning for Few-shot Image Captioning with Large Vision and Language Models