# 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/2305.06983v2/

## Key points

- The paper presents the FLARE approach to address factually inaccurate output from large language models due to hallucination.
- FLARE integrates information retrieval from external knowledge resources throughout the text generation process.
- Unlike traditional retrieval augmented LMs, FLARE actively decides when and what to retrieve as the generation progresses.
- It iteratively predicts upcoming sentences and uses this prediction as a query to retrieve relevant documents for regenerating sentences with low-confidence tokens.
- Comprehensive experiments on four long-form knowledge-intensive generation tasks show that FLARE outperforms baseline methods in terms of accuracy.
- The research was presented at EMNLP 2023 and offers insights into enhancing language model outputs through active retrieval augmentation.

## Summary

The paper "Active Retrieval Augmented Generation" presents a solution to the issue of factually inaccurate output generated by large language models (LMs) due to hallucination. The proposed approach, called Forward-Looking Active REtrieval augmented generation (FLARE), integrates information retrieval from external knowledge resources throughout the text generation process. Unlike traditional retrieval augmented LMs, FLARE actively decides when and what to retrieve as the generation progresses. It iteratively predicts upcoming sentences and uses this prediction as a query to retrieve relevant documents for regenerating sentences containing low-confidence tokens. Comprehensive experiments on four long-form knowledge-intensive generation tasks/datasets demonstrate that FLARE outperforms baseline methods in terms of accuracy. This research was presented at EMNLP 2023 and provides valuable insights into enhancing language model outputs through active retrieval augmentation. For more details on FLARE and access to code and datasets used in the study, interested readers can visit https://github.com/jzbjyb/FLARE.