Transformer^2: Self-Adaptive LLMs