Unveiling and Consulting Core Experts in Retrieval-Augmented MoE-based LLMs

TL;DR

This paper investigates internal expert mechanisms in MoE-based LLMs to improve Retrieval-Augmented Generation by identifying core experts that influence knowledge utilization and document assessment, leading to enhanced RAG performance.

Contribution

It uncovers core expert groups responsible for RAG behaviors in MoE-LLMs and proposes strategies to leverage expert activations for better retrieval-augmented generation.

Findings

Core experts indicate model's knowledge sufficiency

Expert activations assess retrieved document quality

Strategies improve RAG efficiency and effectiveness

Abstract

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.