Hallucination in Large Language Models: Causes, Detection, and Mitigation Strategies

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

One of the common problems with LLMs (Large Language Models) is that they sometimes produce fluent text that deviates from the given source or from world knowledge, a phenomenon known as hallucination.In search engines, healthcare, legal and enterprise applications, the need for factually reliable LLM is now placed in the first place of the agenda of both the researchers and practitioners.This paper reviews the causes of hallucination, categorizes hallucination detection into reference-based and black-box families and examines different categories of approaches for mitigating hallucination from the training-time, retrieval-augmented, decoding-time and post-hoc verification.In addition to this survey, we conduct a comparative study of six mitigation setups and five detection techniques for short-form question answering and long-form generation.The rate of hallucination decreases by about 66.7% from direct prompting for long-form generation by retrieval-augmented generation and post-hoc verification, and by smaller and complementary amounts by contrastive decoding and chain-of-verification.Retrieval-grounded factual-precision scoring outperforms consistency-based black-box detection approaches, which rely on no knowledge of the model's inner workings, when it comes to sentence-level detection accuracy among the detection methods.We find that none of the single-behaviour systems are found to be capable of eliminating hallucination and that the most reliable systems are those that use a training-time or decoding-time intervention followed by an external, retrievalgrounded verification step.