Large language models (LLMs) can produce fluent but incorrect or fabricated statements, called hallucinations, which limits their use in education, healthcare and other high-stakes areas. Most existing safeguards act only after generation, and few correct errors automatically. This paper presents a structured narrative review of [N] verified sources published between [2022] and 2026, organized into seven themes: taxonomies, sampling-consistency detection, semantic-clustering methods, retrieval and fact-verification, internal-state detection, correction and integrated architectures, and evaluation practices. Black-box methods can detect hallucinations without external knowledge or model access, but only 8.8 % of reviewed systems combine detection and mitigation. Evidence relies heavily on question-answering benchmarks, small datasets, automatically generated labels and non-comparable metrics, and correction layers are rarely validated. The paper contributes a thematic synthesis, a comparison of [N] studies, a conceptual seven-layer self-correcting architecture and a prioritized research agenda.