Iterative Generation Method for Factual QA in Large Language Models Based on Semantic Entropy Verification

Retrieval-Augmented Generation (RAG) technology is an effective solution to the hallucination problem in large language models. However, existing techniques require substantial data fine-tuning or rely on validation from ultra-large parameter models to determine the certainty of the generated content, leading to significant resource consumption. To address this issue, this paper proposes a lightweight method based on semantic entropy conflict detection. This method detects the semantic entropy of model responses through an algorithm. On this basis, it dynamically adjusts the length of input information by combining a hierarchical knowledge transfer approach, allowing the model to iteratively optimize the generated content. This enhances the model’s accuracy in answering questions based on knowledge. Experimental results show that the accuracy of this method in judging the consistency of model responses can reach up to 97%. It outperforms traditional RAG methods in improving the ability of large language models to answer questions based on knowledge and demonstrates better performance than the existing iterative generation method, Iter-RetGen, in practical applications.