The emergence of large foundation models, particularly Large Language Models (LLMs) and Vision-Language Models (VLMs), has marked a transformative era in artificial intelligence, characterized by the manifestation of emergent capabilities that are not explicitly programmed during training. These emergent properties represent qualitatively new abilities that arise from the complex interactions within large-scale neural architectures, enabling models to perform tasks and exhibit behaviors beyond their initial training objectives [6, 29, 47].