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 scores 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
Qwen-7B show that CrAM improves the RAG performance of LLMs against
misinformation pollution by over 20%, even surpassing supervised fine-tuning
methods.

通过引用外部文档，检索增强生成（RAG）可以缓解大型语言模型（LLMs）的错觉。然而，外部文档中的错误信息可能会误导 LLMs 的生成。为了解决这个问题，我们探索了 “可信度感知 RAG” 的任务，其中 LLMs 根据文档的可信度自动调整检索文档的影响力，以抵消错误信息。为此，我们引入了一种名为 CrAM 的即插即用方法。CrAM 通过识别 LLMs 中具有影响力的注意力头部，并根据文档的可信度调整它们的注意力分数，从而减少低可信度文档的影响。在使用 Llama2-13B、Llama3-8B 和 Qwen-7B 的自然问题和 TrivaQA 进行的实验证明，CrAM 改善了 LLMs 对错误信息污染的 RAG 性能，提高了超过 20%，甚至超过了监督微调方法。