To be precise, most of the deployed large models are "static" models. They perform well on a series of tasks optimized during pre-training or fine-tuning, but they lack the ability for dynamic ...