$\text{Transformer}^2$: Self-adaptive LLMs