CrAM: Credibility-Aware Attention Modification in LLMs for Combating Misinformation in RAG

TL;DR

This paper introduces CrAM, a method that enhances retrieval-augmented generation in LLMs by adjusting attention based on document credibility, effectively reducing misinformation influence and improving factual accuracy.

Contribution

CrAM is a novel plug-and-play approach that identifies influential attention heads and modulates their weights according to document credibility, improving LLM reliability without fine-tuning.

Findings

CrAM improves RAG performance by over 20% against misinformation.

CrAM surpasses supervised fine-tuning methods in accuracy.

Effective across multiple LLM architectures and datasets.

Abstract

Retrieval-Augmented Generation (RAG) can alleviate hallucinations of Large Language Models (LLMs) by referencing external documents. However, the misinformation in external documents may mislead LLMs' generation. To address this issue, we explore the task of "credibility-aware RAG", in which LLMs automatically adjust the influence of retrieved documents based on their credibility scores to counteract misinformation. To this end, we introduce a plug-and-play method named $\textbf{Cr}$edibility-aware $\textbf{A}$ttention $\textbf{M}$odification (CrAM). CrAM identifies influential attention heads in LLMs and adjusts their attention weights based on the credibility of the documents, thereby reducing the impact of low-credibility documents. Experiments on Natual Questions and TriviaQA using Llama2-13B, Llama3-8B, and Qwen1.5-7B show that CrAM improves the RAG performance of LLMs against misinformation pollution by over 20%, even surpassing supervised fine-tuning methods.