Taming the mirage: a taxonomy of detection methods for LLM hallucinations

Large language models often produce hallucinated content involving factual errors, input contradictions, or internal inconsistencies. Existing detection methods typically address only one type, leading to incomplete coverage. This paper presents a hierarchical detection framework that sequentially evaluates fact, input, and context conflicts using early-stopping logic. Once a hallucination is identified, processing halts with a confidence-graded classification reflecting severity. The framework achieves computational efficiency, comprehensive coverage, and interpretable confidence levels that prioritize factual accuracy over subjective consistency.