This paper explores the use of generative models for open-domain question answering, focusing on the benefits of retrieving text passages to enhance performance. The authors investigate how the number of retrieved passages affects the model's performance, achieving state-of-the-art results on the Natural Questions and TriviaQA benchmarks. They observe that increasing the number of retrieved passages significantly improves performance, indicating that sequence-to-sequence models are effective at aggregating evidence from multiple passages. The method, called Fusion-in-Decoder, involves retrieving support passages using either sparse or dense representations and then processing them with a sequence-to-sequence model. This approach scales well with the number of retrieved passages and demonstrates superior performance compared to extractive models. The paper also discusses related work and provides a detailed experimental setup, including datasets and evaluation metrics.This paper explores the use of generative models for open-domain question answering, focusing on the benefits of retrieving text passages to enhance performance. The authors investigate how the number of retrieved passages affects the model's performance, achieving state-of-the-art results on the Natural Questions and TriviaQA benchmarks. They observe that increasing the number of retrieved passages significantly improves performance, indicating that sequence-to-sequence models are effective at aggregating evidence from multiple passages. The method, called Fusion-in-Decoder, involves retrieving support passages using either sparse or dense representations and then processing them with a sequence-to-sequence model. This approach scales well with the number of retrieved passages and demonstrates superior performance compared to extractive models. The paper also discusses related work and provides a detailed experimental setup, including datasets and evaluation metrics.