This report provides a comprehensive review of the current literature on hallucination detection and mitigation in large language models (LLMs). Hallucination refers to the generation of seemingly correct but factually incorrect responses by LLMs, which can have catastrophic consequences in critical applications. The report aims to serve as a reference for engineers and researchers interested in applying LLMs to real-world tasks. **1. Introduction** - LLMs are trained to generate symbols based on a sequence of previous symbols (prompts). - Concerns about hallucinations, where LLMs produce misleading and factually incorrect outputs, limit their widespread use. - Detecting and mitigating hallucinations is crucial for real-world applications. **2. Introduction to Common Metrics** - **Classification Metrics**: Accuracy, Precision, Recall, F1-Score, AUC, BSS, G-Means. - **NLG Metrics**: N-grams, BLEU, SacreBLEU, ROUGE, METEOR, BERTScore, BARTScore, BLEURT, FEQA, QuestEval, DAE. **3. Hallucination Detection** - **Token-Level Detection**: - **HADES**: A token-level reference-free hallucination detection benchmark for free-form text generation. - **NPH**: Neural Path Hunter for reducing hallucinations in dialogue systems via path grounding. - **HALLUCINATED BUT FACTUAL**: Inspecting the factual nature of hallucinations in abstractive summarization. - **Detecting Hallucinated Content in Conditional Neural Sequence Generation**: A pipeline for token-level hallucination detection in conditional sequence generation tasks. - **Sentence-Level Detection**: - **Self CheckGPT**: Zero-resource black-box hallucination detection for generative LLMs. - **ALIGNSCORE**: A unified alignment function for evaluating factual consistency across various tasks. - **"Why is this misleading?"**: Detecting news headline hallucinations with explanations. - **HaRiM+**: Evaluating summary quality with hallucination risk. The report reviews existing approaches to hallucination detection and mitigation, providing a structured overview of the key characteristics and methodologies. It aims to guide researchers and engineers in developing more robust and reliable LLMs by addressing the issue of hallucinations.This report provides a comprehensive review of the current literature on hallucination detection and mitigation in large language models (LLMs). Hallucination refers to the generation of seemingly correct but factually incorrect responses by LLMs, which can have catastrophic consequences in critical applications. The report aims to serve as a reference for engineers and researchers interested in applying LLMs to real-world tasks. **1. Introduction** - LLMs are trained to generate symbols based on a sequence of previous symbols (prompts). - Concerns about hallucinations, where LLMs produce misleading and factually incorrect outputs, limit their widespread use. - Detecting and mitigating hallucinations is crucial for real-world applications. **2. Introduction to Common Metrics** - **Classification Metrics**: Accuracy, Precision, Recall, F1-Score, AUC, BSS, G-Means. - **NLG Metrics**: N-grams, BLEU, SacreBLEU, ROUGE, METEOR, BERTScore, BARTScore, BLEURT, FEQA, QuestEval, DAE. **3. Hallucination Detection** - **Token-Level Detection**: - **HADES**: A token-level reference-free hallucination detection benchmark for free-form text generation. - **NPH**: Neural Path Hunter for reducing hallucinations in dialogue systems via path grounding. - **HALLUCINATED BUT FACTUAL**: Inspecting the factual nature of hallucinations in abstractive summarization. - **Detecting Hallucinated Content in Conditional Neural Sequence Generation**: A pipeline for token-level hallucination detection in conditional sequence generation tasks. - **Sentence-Level Detection**: - **Self CheckGPT**: Zero-resource black-box hallucination detection for generative LLMs. - **ALIGNSCORE**: A unified alignment function for evaluating factual consistency across various tasks. - **"Why is this misleading?"**: Detecting news headline hallucinations with explanations. - **HaRiM+**: Evaluating summary quality with hallucination risk. The report reviews existing approaches to hallucination detection and mitigation, providing a structured overview of the key characteristics and methodologies. It aims to guide researchers and engineers in developing more robust and reliable LLMs by addressing the issue of hallucinations.