This paper addresses the issue of hallucinations, or factual errors, in large language models (LMs). The authors propose a comprehensive taxonomy of hallucinations and introduce a novel task of automatic fine-grained hallucination detection. They construct the FAVA BENCH benchmark, which includes about 1,000 fine-grained human judgments on three LM outputs across various domains. The analysis reveals that ChatGPT and Llama2-Chat (70B, 7B) exhibit diverse types of hallucinations in most of their outputs, highlighting the need for fine-grained systems. To address this, the authors train FAVA, a retrieval-augmented LM, using synthetic data to detect and correct fine-grained hallucinations. FAVA significantly outperforms retrieval-augmented ChatGPT and GPT-4 on fine-grained hallucination detection and editing tasks, improving the factuality score of Alapaca 7, 13B, and ChatGPT by 4.4, 9.3, and 3.3%, respectively. The paper also discusses the challenges and future directions in fine-grained hallucination detection and editing.This paper addresses the issue of hallucinations, or factual errors, in large language models (LMs). The authors propose a comprehensive taxonomy of hallucinations and introduce a novel task of automatic fine-grained hallucination detection. They construct the FAVA BENCH benchmark, which includes about 1,000 fine-grained human judgments on three LM outputs across various domains. The analysis reveals that ChatGPT and Llama2-Chat (70B, 7B) exhibit diverse types of hallucinations in most of their outputs, highlighting the need for fine-grained systems. To address this, the authors train FAVA, a retrieval-augmented LM, using synthetic data to detect and correct fine-grained hallucinations. FAVA significantly outperforms retrieval-augmented ChatGPT and GPT-4 on fine-grained hallucination detection and editing tasks, improving the factuality score of Alapaca 7, 13B, and ChatGPT by 4.4, 9.3, and 3.3%, respectively. The paper also discusses the challenges and future directions in fine-grained hallucination detection and editing.