The recent advancements in the field of large language models (LLMs) and multimodal large language models (MLLMs) have shown a significant shift towards enhancing adaptability and efficiency in knowledge-based tasks. Researchers are increasingly focusing on developing methods that allow these models to self-learn and adapt to specific knowledge bases without relying heavily on external annotations or human intervention. This trend is evident in the development of frameworks that enable iterative training with self-annotated data, such as Q&A pairs and revision suggestions, which significantly boost model performance in downstream tasks. Additionally, there is a growing interest in integrating external knowledge sources into MLLMs to improve their adaptability and accuracy in tasks like visual question answering. These innovations aim to reduce the model's reliance on pre-trained knowledge and enhance its ability to manage and utilize external knowledge dynamically. Furthermore, the field is witnessing advancements in multi-modal emotion recognition, where LLMs are being prompted with attention-weighted inputs to improve their understanding and prediction capabilities. Overall, the current direction in this research area is towards creating more autonomous, adaptable, and efficient models that can handle complex, knowledge-intensive tasks with minimal external support.