# Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering

**Authors:** Gautier Izacard, Edouard Grave  
**arXiv:** [2007.01282](https://arxiv.org/abs/2007.01282)  
**Date:** 2020-07-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2007.01282v2/

## Key points

- Authors explore the effectiveness of generative models in open domain question answering
- Generative models have competitive performance without relying on external knowledge
- Investigating benefits of retrieving text passages containing relevant evidence
- Incorporating retrieved passages into generative models leads to state-of-the-art results
- Performance improves with increasing number of retrieved passages
- Generative models excel at aggregating and combining evidence from multiple sources
- Passage retrieval techniques enhance generative models for open domain question answering
- Models achieve superior performance without relying on external knowledge.

## Summary

In their paper titled "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering," authors Gautier Izacard and Edouard Grave explore the effectiveness of generative models in open domain question answering. These models have shown competitive performance without relying on external knowledge. To address the limitation of large-scale models with billions of parameters which are both expensive to train and query, the authors investigate the potential benefits of retrieving text passages that may contain relevant evidence. By incorporating retrieved passages into their generative models, they achieve state-of-the-art results on the Natural Questions and TriviaQA open benchmarks. Interestingly, a significant improvement in performance is observed when increasing the number of retrieved passages. This finding suggests that generative models excel at aggregating and combining evidence from multiple sources. Overall, this research highlights the importance of leveraging passage retrieval techniques to enhance generative models for open domain question answering. By effectively integrating evidence from multiple passages, these models can achieve superior performance without relying on external knowledge.