Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by integrating external knowledge for question-answering (QA) tasks. However, effectively utilizing multiple retrieved contexts remains a challenge. Traditional sequential RAG methods process contexts incrementally, leading to incorrect answer revisions, as our experiments show. On the other hand, fusion-in-decoder approaches process all contexts simultaneously. Inspired by this method, we propose fusion-in-LLM, a parallel approach in which LLMs generate independent answers for each retrieved context, selecting the final answer via majority voting or LLM-based selection. This method allows simultaneous processing, mitigating context interference, while improving efficiency. We evaluated fusion-in-LLM on quiz-based QA tasks using the AI-O dataset. Experimental results show that fusion-in-LLM outperforms sequential RAG, particularly when utilizing the top 5 to 7 retrieved contexts. Additionally, fusion-in-LLM reduces computational overhead compared to sequential methods by using parallel processing. These findings demonstrate that fusion-in-LLM effectively improves retrieval-augmented QA in terms of accuracy and efficiency, making it a promising approach to handle multiple contexts in LLM-driven QA tasks.