[arXiv]score: 0.42

Stale-Document Poisoning Flips 30-37% of LLM Answers in RAG

September 28, 2026

Retrieving outdated documents can cause models to override correct internal knowledge, a phenomenon termed stale-document poisoning. In testing, outdated retrieval flipped 30% of Llama and 37% of Qwen answers, with rates rising to 75% when models were explicitly told to follow the document.

HOW THIS AFFECTS YOU

●

builderYou must implement temporal filtering in your RAG pipelines to prevent outdated data from degrading performance.

●

researcherThis identifies a critical temporal alignment failure mode in current RAG evaluation frameworks.

●

policyThis highlights reliability risks in deploying LLMs for legal or medical information retrieval.

read original ↗arxiv.org

DAILY DIGEST

you don't check 9 sources — we do. one email every morning, read in 2 min. free. unsubscribe anytime. privacy

← back to feed