"Misleading information introduced through retrieved context can negatively influence language-model decisions."
Hong and colleagues demonstrate that contradictory or incorrect information in retrieved documents can impair retrieval-augmented language models. Their research identifies a concrete mechanism through which additional context can influence a model toward plausible but incorrect answers, making the reliability of contextual inputs an engineering concern.
View Source