Context-aware Prompt Tuning: Advancing In-Context Learning with Adversarial Methods