Comparison of hallucination detection techniques in LLMs

Hallucination detection aims to detect LLM-generated outputs that are factually incorrect, logically incoherent, or ungrounded in the provided context. In contrast to traditional fact verification which relies on external evidence sources hallucination detection involves a broader evaluation of semantic consistency, reasoning validity, and contextual faithfulness. In this comparison total 28 contributions are compared for exploring the various hallucination detection techniques in LLMs. In this comparison hallucination detection methods are categorizede into uncertainty, retrieval, learning, embedding, and self-consistency-based techniques.