# Check Your Facts and Try Again: Improving Large Language Models with External Knowledge and Automated Feedback

**Authors:** Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, Jianfeng Gao  
**arXiv:** [2302.12813](https://arxiv.org/abs/2302.12813)  
**Date:** 2023-02-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.12813v1/

## Key points

- Challenges of applying large language models (LLMs) to real-world applications
- Issues with LLMs such as hallucinations and lack of external knowledge integration
- Introduction of LLM-Augmenter system to enhance LLMs
- LLM-Augmenter incorporates plug-and-play modules for external knowledge integration
- Iterative revision of LLM prompts using feedback generated by utility functions
- Empirical validation in task-oriented dialog and open-domain question answering scenarios
- Reduction of hallucinations without sacrificing response fluency and informativeness compared to ChatGPT alone
- Availability of source code and models for public use
- Evaluation results on Wiki QA showing top-5 answer recall of consolidated evidence (CORE) at 50.83%
- Conclusion: LLM-Augmenter effectively improves large language models by incorporating external knowledge and automated feedback mechanisms

## Summary

The paper titled "Check Your Facts and Try Again: Improving Large Language Models with External Knowledge and Automated Feedback" addresses the challenges of applying large language models (LLMs) like ChatGPT to real-world, mission-critical applications. While LLMs are capable of generating human-like responses for various tasks, they often suffer from issues such as hallucinations and a lack of external knowledge integration. To overcome these limitations, the authors propose a system called LLM-Augmenter. This system enhances a black-box LLM by incorporating plug-and-play modules. These modules enable the LLM to generate responses grounded in consolidated external knowledge stored in task-specific databases. Additionally, the system iteratively revises LLM prompts using feedback generated by utility functions, such as factuality scores of LLM-generated responses. The effectiveness of LLM-Augmenter is empirically validated in two mission-critical scenarios: task-oriented dialog and open-domain question answering. The results demonstrate that LLM-Augmenter significantly reduces hallucinations without sacrificing response fluency and informativeness compared to ChatGPT alone. The paper also highlights the availability of the source code and models for public use. Furthermore, it provides evaluation results on Wiki QA showing that the top-5 answer recall of consolidated evidence (CORE) is 50.83%. In conclusion, this paper presents an approach that improves large language models by augmenting them with external knowledge and automated feedback mechanisms. The proposed LLM Augmenter system effectively addresses issues related to hallucinations while maintaining response quality.