[Paper Review] A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions

Lei Huang, Weijiang Yu+9 more|arXiv (Cornell University)|Nov 9, 2023

Text Readability and Simplification207 citations

TL;DR

This survey provides a refined taxonomy of LLM hallucinations, analyzes data–training–inference causes, reviews detection benchmarks, and outlines mitigation approaches.

ABSTRACT

The emergence of large language models (LLMs) has marked a significant breakthrough in natural language processing (NLP), fueling a paradigm shift in information acquisition. Nevertheless, LLMs are prone to hallucination, generating plausible yet nonfactual content. This phenomenon raises significant concerns over the reliability of LLMs in real-world information retrieval (IR) systems and has attracted intensive research to detect and mitigate such hallucinations. Given the open-ended general-purpose attributes inherent to LLMs, LLM hallucinations present distinct challenges that diverge from prior task-specific models. This divergence highlights the urgency for a nuanced understanding and comprehensive overview of recent advances in LLM hallucinations. In this survey, we begin with an innovative taxonomy of hallucination in the era of LLM and then delve into the factors contributing to hallucinations. Subsequently, we present a thorough overview of hallucination detection methods and benchmarks. Our discussion then transfers to representative methodologies for mitigating LLM hallucinations. Additionally, we delve into the current limitations faced by retrieval-augmented LLMs in combating hallucinations, offering insights for developing more robust IR systems. Finally, we highlight the promising research directions on LLM hallucinations, including hallucination in large vision-language models and understanding of knowledge boundaries in LLM hallucinations.

Motivation & Objective

• Define and justify a refined taxonomy for LLM hallucinations focused on factuality and faithfulness.
• Analyze the root causes of hallucinations across data, training, and inference stages.
• Review detection methods and evaluation benchmarks for LLM hallucinations.
• Present mitigation strategies that address underlying causes and practical deployment considerations.
• Highlight open challenges and future research directions in trustworthy LLMs.

Proposed method

• Propose a layered, granular taxonomy distinguishing factuality hallucination (inconsistency, fabrication) from faithfulness hallucination (instruction, context, logical) with subcategories.
• Link hallucination causes to data quality, training dynamics, and inference/decoding processes, with illustrative examples.
• Survey existing detection techniques and benchmarks for factuality and faithfulness hallucinations.
• Outline comprehensive mitigation strategies, including data enhancement, debiasing, knowledge boundary management, model editing, retrieval augmentation, and decoding improvements.
• Compare the proposed taxonomy with prior surveys to emphasize cohesive, cause-driven mitigation.

Experimental results

Research questions

• RQ1 What are the main categories of LLM hallucinations and how do they relate to factuality and faithfulness?
• RQ2 What data, training, and inference factors cause hallucinations in LLMs?
• RQ3 How can hallucinations be detected and measured, and what benchmarks exist?
• RQ4 Which mitigation strategies best address the identified root causes of hallucinations?

Key findings

• The authors define a two-pronged taxonomy (factuality vs faithfulness) with clear subtypes to capture LLM-specific hallucination phenomena.
• Data-related causes include flawed sources, knowledge boundaries, and inferior data utilization leading to factuality and bias issues.
• Training-related causes cover pre-training, alignment, and objective-related factors that can induce or fail to mitigate hallucinations.
• Inference-related causes focus on decoding randomness, representations, and context attention that affect output fidelity.
• The survey documents a range of detection methods and benchmarks for both factuality and faithfulness, and discusses mitigation strategies mapped to root causes.