Retrieval-Augmented Selective QA (Title Withheld for Anonymous Review)

This manuscript is being prepared for ARR submission; the title and details are withheld to preserve anonymous review. The work studies retrieval-augmented selective QA: for each question, compare the model’s direct answer with the answer produced after retrieving evidence and revising, then measure whether the revision helped, hurt, or made no difference.

The operating question is practical. Confidence estimates whether the draft answer is already correct; a separate signal is needed to estimate whether revising with retrieved evidence will improve it. The current manuscript reports results on NQ-Open, TriviaQA, and PopQA. No acceptance or public paper link is claimed here.