Unsupervised Information Refinement Training of Large Language Models for Retrieval-Augmented Generation

LLMs can refine information for better text generation by integrating knowledge from retrieved texts and model parameters.