Understanding and Mitigating Hallucinations in Large Language Models: A Systematic Survey of Mechanisms, Detection Paradigms, and Architectural Defenses

Large Language Models (LLMs) based on the autoregressive Transformer architecture have achieved remarkable empirical capabilities across open-domain dialogue, reasoning, code synthesis, and analytical tasks. Despite these milestones, their operational deployment in high-stakes settings—such as clinical informatics, jurisprudence, and autonomous systems—remains bottlenecked by hallucination: the generation of fluent, syntactically coherent text that is factually erroneous, unfaithful to provided source material, or entirely fabricated. This survey provides a comprehensive, rigorous synthesis of the LLM hallucination phenomenon. We formalize the taxonomy of hallucinations, distinguishing between factuality violations and context-faithfulness discrepancies. We investigate the underlying computational and statistical etiology of hallucinations across data pipelines, objective functions, representational bottlenecks, and decoding dynamics. Furthermore, we systematically categorize contemporary detection paradigms—ranging from internal activation probing and semantic entropy estimation to external verification networks—and examine defense strategies spanning Retrieval-Augmented Generation (RAG), internal representation steering, contrastive decoding, and alignment algorithms. Finally, we highlight unresolved challenges across multimodal models and propose actionable frontiers for achieving epistemically grounded artificial intelligence.