Manuscript Note: Knowing When to Answer, Refine, or Abstain

This work studies retrieval-augmented selective QA. For each question, it compares the model’s direct answer with the answer produced after retrieving evidence and refining, then records whether the outcome was preserved, repaired, harmed, or unrecovered.

The analysis separates two questions: whether the draft is already correct and whether an available refinement process is likely to repair or harm it. The evaluation then compares answer, refine, or abstain policies for one recorded model-retriever-corpus stack. It spans 25,870 held-out questions from NQ-Open, TriviaQA, and PopQA, with DPR and BM25 retrieval over a shared Wikipedia source.

The manuscript is under review at ACL Rolling Review. Its submission title, author list, and preprint remain withheld during double-blind review.