<p>Multi-task prompt tuning has recently gained significant attention due to its modular design and potential for parameter-efficient transfer learning across diverse domains, including natural language and vision tasks. This paper investigates strategies to enhance multi-task performance by enabling effective knowledge transfer between task-specific prompts. We propose a novel framework that constructs each target task’s prompt as a combination of shared source prompts and a task-specific private prompt. Several methods for integrating these components are presented and compared, with a detailed analysis of the respective roles and contributions of source and private prompts. Based on this analysis, we introduce flexible configurations that adaptively balance shared and task-specific knowledge to optimize performance. Empirical results demonstrate consistent improvements in both accuracy and robustness over standard prompt tuning and related baselines. Our approach particularly excels in few-shot scenarios, achieving superior performance on the GLUE benchmark and other tasks, while requiring substantially less training data. These findings highlight the effectiveness of our method for data-efficient, multi-task learning.</p>