[Paper Review] A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models

S. M Towhidul Islam Tonmoy, S M Mehedi Zaman+5 more|arXiv (Cornell University)|Jan 2, 2024

COVID-19 diagnosis using AIMedicine109 citations

TL;DR

This survey catalogs over thirty-two techniques for mitigating hallucinations in LLMs, presenting a taxonomy and insights to guide future research.

ABSTRACT

As Large Language Models (LLMs) continue to advance in their ability to write human-like text, a key challenge remains around their tendency to hallucinate generating content that appears factual but is ungrounded. This issue of hallucination is arguably the biggest hindrance to safely deploying these powerful LLMs into real-world production systems that impact people's lives. The journey toward widespread adoption of LLMs in practical settings heavily relies on addressing and mitigating hallucinations. Unlike traditional AI systems focused on limited tasks, LLMs have been exposed to vast amounts of online text data during training. While this allows them to display impressive language fluency, it also means they are capable of extrapolating information from the biases in training data, misinterpreting ambiguous prompts, or modifying the information to align superficially with the input. This becomes hugely alarming when we rely on language generation capabilities for sensitive applications, such as summarizing medical records, financial analysis reports, etc. This paper presents a comprehensive survey of over 32 techniques developed to mitigate hallucination in LLMs. Notable among these are Retrieval Augmented Generation (Lewis et al, 2021), Knowledge Retrieval (Varshney et al,2023), CoNLI (Lei et al, 2023), and CoVe (Dhuliawala et al, 2023). Furthermore, we introduce a detailed taxonomy categorizing these methods based on various parameters, such as dataset utilization, common tasks, feedback mechanisms, and retriever types. This classification helps distinguish the diverse approaches specifically designed to tackle hallucination issues in LLMs. Additionally, we analyze the challenges and limitations inherent in these techniques, providing a solid foundation for future research in addressing hallucinations and related phenomena within the realm of LLMs.

Motivation & Objective

• Characterize the spectrum of hallucination mitigation methods for LLMs across models and tasks.
• Provide a structured taxonomy to organize retrieval, prompting, model development, and evaluation approaches.
• Analyze challenges, limitations, and potential directions to guide future research and deployment.
• Highlight practical considerations for faithfulness, reliability, and safety in LLM outputs.

Proposed method

• Construct a comprehensive taxonomy of hallucination mitigation techniques covering prompting, model development, and evaluation.
• Synthesise techniques including retrieval-augmented generation, self-refinement, prompt tuning, knowledge graphs, faithfulness-based losses, and supervised fine-tuning.
• Discuss end-to-end RAG systems and decoding strategies that influence factuality and grounding.
• Summarize notable systems and frameworks (e.g., RAG, D&Q, EVER, CoVe, CoNLI) and their reported roles in reducing hallucinations.
• Provide cross-cutting analysis of dataset usage, feedback mechanisms, and retriever types to guide future research.

Experimental results

Research questions

• RQ1 What are the major categories and techniques used to mitigate hallucinations in LLMs?
• RQ2 How do retrieval, prompting, model development, and evaluation strategies compare in effectiveness and applicability?
• RQ3 What are the key limitations and challenges in current hallucination mitigation approaches, and what directions are promising for future work?
• RQ4 How can faithfulness and grounding be quantified and improved across different modalities and tasks?

Key findings

• A wide taxonomy of hallucination mitigation techniques is presented, spanning prompt engineering, end-to-end retrieval, self-refinement, and model architecture changes.
• Retrieval-Augmented Generation (RAG) and knowledge-grounding approaches are repeatedly highlighted as effective mechanisms for grounding outputs.
• Several self-checking and feedback-based strategies (e.g., EVER, CoVe, CoNLI, SC reasoning) show reductions in hallucinations across tasks.
• Knowledge graphs and faithfulness-based losses contribute to improving factual alignment, with several architecture-level methods proposed.
• The survey identifies limitations such as data quality, evaluation challenges, and the need for robust, task- and domain-specific solutions.