Although Large Language Models (LLMs) have demonstrated significant transformative potential in multiple domains, the prevalence of hallucination has severely limited the reliability and accuracy of model outputs, and has become a core challenge in practical deployment. Although the current research on LLM hallucination has shown rapid growth, there are two main limitations: the existing literature focuses on listing the phenomena, and lacks a systematic categorization framework for detection methods. Therefore, this review innovatively constructs an analytical framework: based on the model interpretability dimension, mainstream detection techniques are classified into two categories: white-box detection and black-box detection, which provides a fine-grained method selection guide for different application scenarios, and at the same time, systematically summarizes the factors that contribute to the hallucination of LLM. This analytical framework not only facilitates the deepening of domain cognition, but also provides an actionable improvement path for building a credible LLM system.