Tag
This paper audits eight automatic attribution metrics across three evaluation constructs for RAG systems, finding that no single metric transfers across datasets within the same construct, challenging the common practice of treating them as interchangeable.