The data generation and evaluation methods driven by large language models have become the core paradigm of modern artificial intelligence model development. Through streamlined data annotation and performance evaluation, it has significantly improved the efficiency and effectiveness of model training. However, the problem of preference leakage caused by the implicit correlation between the text generation model and the evaluation model, undermines the fairness and reliability of evaluations and threatens the generalization of text generation models. This problem necessitates urgent in-depth research and systematic solutions. To address these challenges, we propose a multi-perspective cross-lingual alignment framework (MPCLA). The proposed framework alleviates the preference leakage problem by facilitating cross-lingual alignment prompts across different language representations, thereby effectively reducing the potential bias between models. Additionally, we introduce a novel preference leakage score calculation method to quantify preference leakage. The extensive experimental results demonstrate that our method significantly mitigates the impact of preference leakage, and providing an extensible solution for the LLM-as-a-Judge paradigm. According to our investigation, our novel methodology is the first solution to the preference leakage in LLMs.