This research introduces RAGuard, a framework to evaluate how Large Language Models handle misleading information when using Retrieval-Augmented Generation (RAG).

• RAG systems can perform worse than zero-shot when faced with misleading retrievals
• LLMs often adopt misinformation from retrieved content rather than maintaining accurate reasoning
• Political topics are particularly vulnerable due to polarized framing and selective evidence
• The study reveals the need for robust verification mechanisms in RAG systems deployed in real-world scenarios

Security Implications: As RAG becomes more prevalent in production systems, its vulnerability to misinformation presents significant security risks, especially in domains where factual accuracy is critical.

Worse than Zero-shot? A Fact-Checking Dataset for Evaluating the Robustness of RAG Against Misleading Retrievals