Retrieval-augmented generation (RAG) is meant to supply up-to-date knowledge, but this paper shows a critical flaw: when the retrieved document is outdated, it can override the model's own correct knowledge. The authors built a of 317 verified knowledge reversals across medicine, law, software, and platform policy, and tested 12 models. They found that outdated retrieval flipped 30% of Llama and 37% of Qwen answers even without instruction to trust the document, and up to 66–75% when instructed. In contrast, matched up-to-date evidence was followed 97–100% of the time.
This matters because RAG is widely used to address stale parametric knowledge, yet the retrieval itself can be stale and actively cause harm. The failure is not general inability to use retrieval, but a lack of selective trust: models often cannot tell whether retrieved evidence still applies. The paper also shows that explicitly providing validity boundaries helps larger models switch answers almost perfectly, and that a recency-aware re-ranker can reduce poisoning by 4.6–10 percentage points when dates are reliable.