Samuel Korn

The paper evaluates four RAG architectures under knowledge base poisoning, demonstrating that advanced architectures significantly improve robustness against adversarial contradictions, localizing the vulnerability to the content-reasoning stage.