# Active Retrieval Augmented Generation

**Authors:** Zhengbao Jiang, Frank F. Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, Graham Neubig  
**arXiv:** [2305.06983](https://arxiv.org/abs/2305.06983)  
**Date:** 2023-05-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.06983v1/

## Key points

- The paper proposes a retrieval-augmented generation method to address the issue of large language models generating factually inaccurate output.
- Existing retrieval-augmented LMs retrieve information only once based on the input, limiting their ability to generate long texts.
- The authors propose a generalized view of active retrieval augmented generation which actively decides when and what to retrieve across the course of the generation process.
- This is implemented in Forward-Looking Active Retrieval augmented generation (FLARE), a generic retrieval-augmented generation method that iteratively uses a prediction of the upcoming sentence to anticipate future content and retrieve relevant documents to regenerate low-confidence tokens.
- FLARE is evaluated along with baselines over four knowledge-intensive datasets and demonstrates superior or competitive performance on all tasks.
- Code and datasets are available at https://github.com/jzbjyb/FLARE.
- The paper is authored by Zhengbao Jiang, Frank F. Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan and Graham Neubig.

## Summary

The paper titled "Active Retrieval Augmented Generation" addresses the issue of large language models (LMs) generating factually inaccurate output by proposing a retrieval-augmented generation method. Existing retrieval-augmented LMs typically retrieve information only once based on the input, limiting their ability to generate long texts. To address this limitation, the authors propose a generalized view of active retrieval augmented generation which actively decides when and what to retrieve across the course of the generation process. This is implemented in Forward-Looking Active Retrieval augmented generation (FLARE), a generic retrieval-augmented generation method that iteratively uses a prediction of the upcoming sentence to anticipate future content and retrieve relevant documents to regenerate low-confidence tokens. The authors evaluate FLARE along with baselines over four knowledge-intensive datasets and demonstrate its superior or competitive performance on all tasks. Code and datasets are available at https://github.com/jzbjyb/FLARE. The paper is authored by Zhengbao Jiang, Frank F. Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan and Graham Neubig.