This paper introduces a novel decoding strategy for large language models (LLMs) to enhance contextual understanding by integrating contrastive decoding with adversarial irrelevant passages as negative samples. The method operates during inference without requiring further training and aims to balance parametric and non-parametric knowledge sources. The approach involves incorporating both relevant and irrelevant contexts, where irrelevant contexts are adversarially crafted to guide the model away from incorrect responses. The method dynamically adjusts the influence of parametric knowledge based on the model's confidence in the relevant non-parametric knowledge. Experiments on diverse datasets such as TriviaQA, Natural Questions, and PopQA demonstrate that the proposed method outperforms existing decoding approaches in handling knowledge conflicts and integrating contextual information. The method is effective across various model sizes, particularly with larger models, and shows consistent performance improvements in answering questions with varying levels of knowledge popularity. The study also explores the impact of different retrieval sources on the decoding strategy, emphasizing the importance of refining retrieval mechanisms for better performance. The results indicate that the proposed decoding method significantly improves the accuracy of LLMs in open-domain question answering by effectively reconciling conflicting knowledge sources. The method is applicable to various generative tasks beyond question answering and has potential for future research in summarization and other areas. The study acknowledges limitations, including the use of a single prompt template and the computational complexity of the method, but highlights its effectiveness in enhancing contextual understanding and reducing factual inconsistencies in generated text.This paper introduces a novel decoding strategy for large language models (LLMs) to enhance contextual understanding by integrating contrastive decoding with adversarial irrelevant passages as negative samples. The method operates during inference without requiring further training and aims to balance parametric and non-parametric knowledge sources. The approach involves incorporating both relevant and irrelevant contexts, where irrelevant contexts are adversarially crafted to guide the model away from incorrect responses. The method dynamically adjusts the influence of parametric knowledge based on the model's confidence in the relevant non-parametric knowledge. Experiments on diverse datasets such as TriviaQA, Natural Questions, and PopQA demonstrate that the proposed method outperforms existing decoding approaches in handling knowledge conflicts and integrating contextual information. The method is effective across various model sizes, particularly with larger models, and shows consistent performance improvements in answering questions with varying levels of knowledge popularity. The study also explores the impact of different retrieval sources on the decoding strategy, emphasizing the importance of refining retrieval mechanisms for better performance. The results indicate that the proposed decoding method significantly improves the accuracy of LLMs in open-domain question answering by effectively reconciling conflicting knowledge sources. The method is applicable to various generative tasks beyond question answering and has potential for future research in summarization and other areas. The study acknowledges limitations, including the use of a single prompt template and the computational complexity of the method, but highlights its effectiveness in enhancing contextual understanding and reducing factual inconsistencies in generated text.