This paper addresses the preference gap between retrievers and large language models (LLMs) in retrieval-augmented generation (RAG). The key challenge is that retrievers, designed to be human-friendly, often fail to align with LLMs' preferences, leading to suboptimal performance in RAG systems. To bridge this gap, the authors propose BGM (Bridging the Gap between retrievers and LLMs), a novel framework that trains a bridge model to adapt retrieved information to be more LLM-friendly. BGM operates by chaining supervised learning (SL) and reinforcement learning (RL). The bridge model is a sequence-to-sequence (seq2seq) model that transforms retrieved passages into a format that aligns with LLM preferences. SL is used to provide an initial model that can rank and select passages, while RL refines this model to optimize performance on downstream tasks. This approach allows the bridge model to dynamically select, re-rank, and even repeat passages, which is crucial for LLMs. The authors evaluate BGM on various datasets, including open-domain QA (Natural Questions, HotpotQA) and personalized generation (Avocado Email, Amazon Book). Results show that BGM outperforms existing baselines in all tasks, demonstrating its effectiveness in bridging the preference gap. BGM is particularly effective in scenarios where retrieval is not essential, such as in the Book dataset, where it can generate accurate responses without relying on retrieved passages. The study also explores the impact of different bridge model sizes and LLM sizes on performance, showing that BGM performs well even with smaller models. Additionally, the bridge model's ability to generalize across different datasets and LLMs is investigated, though current results suggest limitations in this area. Overall, BGM provides a practical and effective solution to the preference gap problem in RAG, enhancing the performance of LLMs by aligning retrievers with their preferences through a bridge model that combines SL and RL. This approach has the potential to significantly improve the effectiveness of RAG systems in various NLP tasks.This paper addresses the preference gap between retrievers and large language models (LLMs) in retrieval-augmented generation (RAG). The key challenge is that retrievers, designed to be human-friendly, often fail to align with LLMs' preferences, leading to suboptimal performance in RAG systems. To bridge this gap, the authors propose BGM (Bridging the Gap between retrievers and LLMs), a novel framework that trains a bridge model to adapt retrieved information to be more LLM-friendly. BGM operates by chaining supervised learning (SL) and reinforcement learning (RL). The bridge model is a sequence-to-sequence (seq2seq) model that transforms retrieved passages into a format that aligns with LLM preferences. SL is used to provide an initial model that can rank and select passages, while RL refines this model to optimize performance on downstream tasks. This approach allows the bridge model to dynamically select, re-rank, and even repeat passages, which is crucial for LLMs. The authors evaluate BGM on various datasets, including open-domain QA (Natural Questions, HotpotQA) and personalized generation (Avocado Email, Amazon Book). Results show that BGM outperforms existing baselines in all tasks, demonstrating its effectiveness in bridging the preference gap. BGM is particularly effective in scenarios where retrieval is not essential, such as in the Book dataset, where it can generate accurate responses without relying on retrieved passages. The study also explores the impact of different bridge model sizes and LLM sizes on performance, showing that BGM performs well even with smaller models. Additionally, the bridge model's ability to generalize across different datasets and LLMs is investigated, though current results suggest limitations in this area. Overall, BGM provides a practical and effective solution to the preference gap problem in RAG, enhancing the performance of LLMs by aligning retrievers with their preferences through a bridge model that combines SL and RL. This approach has the potential to significantly improve the effectiveness of RAG systems in various NLP tasks.