# Generate rather than Retrieve: Large Language Models are Strong Context Generators

**Authors:** Wenhao Yu, Dan Iter, Shuohang Wang, Yichong Xu, Mingxuan Ju, Soumya Sanyal, Chenguang Zhu, Michael Zeng, Meng Jiang  
**arXiv:** [2209.10063](https://arxiv.org/abs/2209.10063)  
**Date:** 2022-09-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2209.10063v3/

## Key points

- The paper proposes a new approach to solving knowledge-intensive tasks using large language model generators instead of document retrievers.
- The proposed method is called GenRead and involves prompting a large language model to generate contextual documents based on a given question and then reading the generated documents to produce the final answer.
- The authors also propose a clustering-based prompting method that selects distinct prompts resulting in generated documents that cover different perspectives, leading to better recall over acceptable answers.
- GenRead is demonstrated to be effective through extensive experiments on three different knowledge-intensive tasks, including open-domain QA, fact checking, and dialogue systems.
- GenRead achieves significantly higher exact match scores than state-of-the-art retrieve-then-read pipeline DPR-FiD without retrieving any documents from external knowledge sources.
- However, GenRead has limitations in updating knowledge state and adapting to new domains compared to retrieve-then-read methods' ability to swap in new documents when new information is learned or adding in documents from a new domain for quick adaptation to downstream tasks.
- Future research directions include incorporating new knowledge efficiently into generate then read methods while minimizing hallucination errors in generated documents.

## Summary

This paper presents a novel approach to solving knowledge-intensive tasks, such as open-domain question answering (QA), by replacing document retrievers with large language model generators. The proposed method, called generate-then-read (GenRead), prompts a large language model to generate contextual documents based on a given question and then reads the generated documents to produce the final answer. The authors also propose a clustering-based prompting method that selects distinct prompts resulting in generated documents that cover different perspectives, leading to better recall over acceptable answers. The effectiveness of GenRead is demonstrated through extensive experiments on three different knowledge-intensive tasks, including open-domain QA, fact checking, and dialogue systems. Notably, GenRead achieves 71.6 and 54.4 exact match scores on TriviaQA and WebQ datasets, respectively, significantly outperforming the state-of-the-art retrieve-then-read pipeline DPR-FiD by +4.0 and +3.9 without retrieving any documents from external knowledge sources. Furthermore, the authors conducted a readability study on retrieved documents and generated documents extracted from NQ, TriviaQA, and WebQ datasets when both contain the correct answer. The results show that FiD reader can produce more correct answers when reading generated documents from large language models such as InstructGPT. Despite its strong performance on presented datasets, GenRead has limitations in updating knowledge state and adapting to new domains compared to retrieve-then-read methods' ability to swap in new documents when new information is learned or adding in documents from a new domain for quick adaptation to downstream tasks. In conclusion, this paper offers an innovative perspective for solving knowledge intensive tasks by replacing dense retrieval models with large language model generators using GenRead methodology while acknowledging future research directions for incorporating new knowledge efficiently into generate then read methods while minimizing hallucination errors in generated documents.