Hallucination detection in large language models based on multi-granularity consistency

Large Language Models (LLMs) have achieved notable advancements in Natural Language Processing. However, these models are prone to hallucinations, generating factually incorrect content, and often exhibit over-confidence in their responses, which can lead to misguided decisions. This phenomenon is particularly evident in Q&A systems and dialog generation, which may lead to the propagation of false or misleading information. To address this problem, this paper proposes an innovative method to detect hallucinations through multi-granularity consistency verification and cosine similarity analysis. The method leverages dual generation pathways—Forward Language Modeling (FLLM) and Backward Language Modeling (BLLM)—to produce multiple candidate answers, the reliability of which is determined by a consistency estimation algorithm. Experimental results on the Global Country Information Dataset 2023 demonstrate that the proposed framework achieves superior detection stability and precision compared to SelfCheckGPT-style NLI baselines. By leveraging structured multi-granularity consistency patterns, our method effectively identifies hallucinations through a self-contained, reference-free process without requiring external knowledge. This provides a robust pathway for quantifying uncertainty and enhancing the reliability of LLMs in factual diagnostic tasks.