The paper addresses the preference gap between retrievers and Large Language Models (LLMs) in Retrieval-Augmented Generation (RAG) systems. This gap arises from the different preferences of retrievers, which are designed to be human-friendly, and LLMs, which require specific formatting for optimal performance. The authors propose a novel framework called BGM (Bridging the Gap between retrievers and LLMs) to bridge this gap. BGM is a sequence-to-sequence model that chains supervised learning (SL) and reinforcement learning (RL) to train a bridge model that optimizes the connection between the retriever and the LLM. The bridge model adapts retrieved information to be LLM-friendly, addressing issues such as ranking and selection. Empirical results demonstrate that BGM enhances the performance of various downstream tasks, including question answering and personalized generation, across different datasets. The paper also discusses the limitations of BGM, such as its current inability to generalize across different datasets and LLMs, and suggests future directions for improvement.The paper addresses the preference gap between retrievers and Large Language Models (LLMs) in Retrieval-Augmented Generation (RAG) systems. This gap arises from the different preferences of retrievers, which are designed to be human-friendly, and LLMs, which require specific formatting for optimal performance. The authors propose a novel framework called BGM (Bridging the Gap between retrievers and LLMs) to bridge this gap. BGM is a sequence-to-sequence model that chains supervised learning (SL) and reinforcement learning (RL) to train a bridge model that optimizes the connection between the retriever and the LLM. The bridge model adapts retrieved information to be LLM-friendly, addressing issues such as ranking and selection. Empirical results demonstrate that BGM enhances the performance of various downstream tasks, including question answering and personalized generation, across different datasets. The paper also discusses the limitations of BGM, such as its current inability to generalize across different datasets and LLMs, and suggests future directions for improvement.