Wan Zhang, Jing Zhang
A comprehensive review that systematically organizes the hallucination problem in RAG-based LLMs by causes and mitigation techniques across retrieval and generation phases.
RAG enhances LLM factuality and timeliness by leveraging external knowledge, but hallucinations (confabulations) occur due to inaccuracies in the retriever or misuse by the generator. Existing research is scattered, making it difficult to understand the overall causes and solutions in an integrated manner.
The paper decomposes the RAG pipeline into sub-tasks of retrieval (query understanding, document retrieval, document selection) and generation (document integration, response generation), and classifies hallucination types that can occur at each stage. It then systematically organizes mitigation techniques corresponding to each type (e.g., improving retrieval quality, prompt engineering, model fine-tuning, decoding strategies). Additionally, it separately addresses methods for detecting and correcting hallucinations in generated responses.
As the first comprehensive review on RAG hallucination, it helps researchers systematically understand the problem and choose appropriate mitigation strategies. It emphasizes the interaction between retrieval and generation, and suggests future research directions (e.g., multimodal RAG, dynamic retrieval, evaluation benchmarks).