Better Retrievers, Worse Labels: How Incomplete Relevance Judgments Distort Abstention Evaluation in Conversational RAG
A convenient shortcut reads sufficiency off the relevance judgments and treats a context without gold documents as unanswerable, and is released, and recommended scoring refusals against content labels, not the presence of gold.
Ivan Kochergin, Vasily Konovalov
· 0 citations