# Survey on Factuality in Large Language Models: Knowledge, Retrieval and Domain-Specificity

**Authors:** Cunxiang Wang, Xiaoze Liu, Yuanhao Yue, Xiangru Tang, Tianhang Zhang, Cheng Jiayang, Yunzhi Yao, Wenyang Gao, Xuming Hu, Zehan Qi, Yidong Wang, Linyi Yang, Jindong Wang, Xing Xie, Zheng Zhang, Yue Zhang  
**arXiv:** [2310.07521](https://arxiv.org/abs/2310.07521)  
**Date:** 2023-10-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.07521v1/

## Key points

- Survey focuses on factuality in Large Language Models (LLMs)
- Implications and challenges of factual inaccuracies in LLM outputs
- Analysis of mechanisms for storing and processing facts in LLMs
- Methodologies for evaluating LLM factuality, including metrics, benchmarks, and studies
- Strategies for enhancing LLM factuality, tailored for specific domains
- Discussion of standalone LLMs and Retrieval-Augmented LLMs configurations
- Introduction of a comprehensive framework to reduce factual inaccuracies in LLM outputs
- Entity extraction and keyword distillation techniques to ascertain pivotal concepts within contextual sentences
- Use of confidence estimates as surrogates to determine if additional information is needed from external sources
- Exploration of retrieval adaptation strategies: prompt-based methods, SFT-based methods, RLHF-based methods
- Example system "LLM-Augmenter" that combines a fixed LLM with a retrieval module to improve performance and address factual errors
- Acknowledgment of unique challenges in summarization tasks but not heavily focused on in the survey

## Summary

This survey focuses on the issue of factuality in Large Language Models (LLMs) and aims to provide researchers with a structured guide to enhance the factual reliability of LLMs. The authors first discuss the implications of factual inaccuracies in LLM outputs and highlight the potential consequences and challenges posed by these errors. They then analyze the mechanisms through which LLMs store and process facts, identifying the primary causes of factual errors. The survey delves into methodologies for evaluating LLM factuality, emphasizing key metrics, benchmarks, and studies. It also explores strategies for enhancing LLM factuality, including approaches tailored for specific domains. Two primary LLM configurations are discussed: standalone LLMs and Retrieval-Augmented LLMs that utilize external data. The unique challenges and potential enhancements for each configuration are detailed. The authors introduce a comprehensive framework aimed at reducing factual inaccuracies in LLM outputs. This framework utilizes models to recognize entities and generate questions, acting as tools for the LLM-based agent. Pivotal concepts such as names, geographical locales, and temporal references are ascertained within contextual sentences using entity extraction or keyword distillation techniques. Confidence estimates in the form of logit output values are used as surrogates to determine if additional information is needed from external sources. The survey also discusses retrieval adaptation strategies that enable LLMs to better adapt to retrieved data and produce more accurate content. Three methodological approaches are explored: prompt-based methods, SFT-based methods, and RLHF-based methods. Prompt-based methods leverage prompts to navigate the retrieval process and extract pertinent factual data. Additionally, the survey presents an example system called "LLM-Augmenter" that combines a fixed LLM with a plug-and-play retrieval module to improve performance in tasks sensitive to factual errors. This system allows the LLM to interact with external knowledge modules and uses automated feedback generated by utility functions to modify candidate response options. The authors acknowledge that while summarization tasks have seen research on factuality, they chose not to heavily focus on this domain in the survey due its unique challenges such as coherence, conciseness, and relevance which deviate from addressing factual errors in LLMs directly . Overall, this survey provides a comprehensive overview of methodologies for evaluating and enhancing factuality in Large Language Models (LLMs), retrieval adaptation strategies tailored for specific domains as well as an example system "LLM Augmenter" designed reduce factual inaccuracies in its outputs..