Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation

In 30 Sekunden

Was passiert ist

Researchers found that retrieval-augmented generation systems cite sources for claims they actually derive from internal model memory, not retrieved context.

Warum es zählt

Engineers building or evaluating RAG systems need to understand that citations may be misleading and standard verification methods fail to catch this problem.

Achtung

The detection method developed measures what models know rather than what they actually used, leaving the core problem of separating knowledge from usage unsolved.