Unveiling and Consulting Core Experts in Retrieval-Augmented MoE-based LLMs
This work addresses the underexplored internal mechanisms in LLMs for RAG systems, offering incremental improvements for knowledge-intensive AI tasks.
The paper tackled the problem of understanding internal mechanisms in Mixture-of-Expert-based LLMs that affect Retrieval-Augmented Generation (RAG) performance, and demonstrated that identifying and consulting core expert groups can improve RAG efficiency and effectiveness across various datasets and models.
Retrieval-Augmented Generation (RAG) significantly improved the ability of Large Language Models (LLMs) to solve knowledge-intensive tasks. While existing research seeks to enhance RAG performance by retrieving higher-quality documents or designing RAG-specific LLMs, the internal mechanisms within LLMs that contribute to the effectiveness of RAG systems remain underexplored. In this paper, we aim to investigate these internal mechanisms within the popular Mixture-of-Expert (MoE)-based LLMs and demonstrate how to improve RAG by examining expert activations in these LLMs. Our controlled experiments reveal that several core groups of experts are primarily responsible for RAG-related behaviors. The activation of these core experts can signify the model's inclination towards external/internal knowledge and adjust its behavior. For instance, we identify core experts that can (1) indicate the sufficiency of the model's internal knowledge, (2) assess the quality of retrieved documents, and (3) enhance the model's ability to utilize context. Based on these findings, we propose several strategies to enhance RAG's efficiency and effectiveness through expert activation. Experimental results across various datasets and MoE-based LLMs show the effectiveness of our method.