# Siren's Song in the AI Ocean: A Survey on Hallucination in Large Language Models

**Authors:** Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, Longyue Wang, Anh Tuan Luu, Wei Bi, Freda Shi, Shuming Shi  
**arXiv:** [2309.01219](https://arxiv.org/abs/2309.01219)  
**Date:** 2023-09-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.01219v1/

## Key points

- Large language models (LLMs) pose challenges in terms of hallucinations, generating content that deviates from user input or established knowledge.
- Recent efforts have focused on detecting, explaining, and mitigating hallucinations in LLMs, with a particular emphasis on the unique challenges they present.
- Taxonomies of LLM hallucination phenomena and evaluation benchmarks are presented, along with an analysis of existing approaches to mitigate them.
- Curating pre-training corpora is important to reduce hallucinations during training; strategies like up-sampling data from factual sources and adding topic prefixes to sentences can improve LLM performance and reduce hallucinations.
- Mitigation of hallucinations during supervised fine-tuning (SFT) is crucial, emphasizing the need for well-designed SFT strategies to prevent inaccurate responses from LLMs.
- Effective selection and filtering strategies for data curation could help mitigate hallucinations in LLMs; future research may explore new approaches to improve the reliability of LLMs in real-world scenarios.

## Summary

In this paper, we focus on the challenges posed by large language models (LLMs) in terms of hallucinations. LLMs occasionally generate content that deviates from user input or established knowledge. The paper surveys recent efforts in detecting, explaining, and mitigating hallucinations in LLMs with a particular emphasis on the unique challenges they present. Taxonomies of LLM hallucination phenomena and evaluation benchmarks are presented along with an analysis of existing approaches to mitigate them. We also discuss the importance of curating pre-training corpora to reduce hallucinations during training. Strategies such as up-sampling data from factual sources like Wikipedia and adding topic prefixes to sentences have been proposed to improve LLM performance and reduce hallucinations. Additionally, we explore the mitigation of hallucinations during supervised fine-tuning (SFT), highlighting the need for well-designed SFT strategies to prevent inaccurate responses from LLMs. Overall, our paper suggests that more effective selection and filtering strategies for data curation could help mitigate hallucinations in LLMs. Future research directions may include exploring new approaches to address this challenge and improve the reliability of LLMs in real-world scenarios.