This paper highlights that existing prompt tuning approaches treat prompts at each network layer independently, overlooking the complex interactions between layers. To address this, we propose a novel framework that explicitly models higher-order interactions among multilayer prompts. This framework treats prompts at different layers as a system of interconnected entities rather than individual elements. Through innovative interaction modules, it captures the complex nonlinear correlations between prompts, enhancing the expressiveness and semantic richness of the model. Experimental results on eight benchmark datasets demonstrate that the proposed method outperforms state-of-the-art prompt tuning-based models, particularly in small-shot scenarios.
