A Survey on Hallucination in Large Language Models: Definitions, Detection, and Mitigation

Despite the strong capabilities of Large Language Models (LLMs) across a wide range of natural language processing tasks, their outputs can still be unreliable. Hallucination is widely used as an umbrella label for several related failures, but its operational meaning varies across tasks and studies. This work provides a comprehensive review of recent advances in understanding, detecting, and mitigating hallucinations in LLMs. We first distinguish hallucination from the separate evaluation dimensions of factuality, faithfulness, and internal consistency. We then introduce an author-proposed organisational taxonomy covering manifestations commonly discussed under hallucination, including factual contradictions, context conflicts, fabrication, self-contradiction, and emerging failures in reasoning traces and agentic trajectories. We review detection methods based on level of access to the model, ranging from white-box approaches to grey-box uncertainty estimation and black-box consistency and verification methods. We further organize these approaches according to the signals they exploit, including uncertainty, semantic consistency, external evidence, and internal representations, and discuss their strengths and limitations across different deployment settings. Finally, we survey mitigation strategies across the LLM lifecycle, including (1) data-centric approaches such as data curation and retrieval-augmented pre-training, (2) model-centric approaches such as supervised fine-tuning, preference optimization, abstention-aware reinforcement learning, and knowledge editing, and (3) inference-time approaches such as retrieval-augmented generation, self-verification, and decoding or activation-level interventions. Rather than viewing these strategies as isolated solutions, we argue that reliable hallucination mitigation requires complementary controls across multiple stages of the model lifecycle. Key challenges include scalable and provenance-aware data curation, the alignment–capability trade-off, reliable composition of mitigation techniques, and the need to evaluate and modify reasoning and agentic processes rather than only isolated factual associations.