Return or Revise? Learning When Revision Helps Retrieval-Augmented QA

2026-09-24

AI Summary

This study addresses the absence of effective decision mechanisms for determining when to adopt an initial draft or perform revision in retrieval-augmented generation (RAG) question answering. To this end, it introduces the concept of “recoverability” to quantify revision benefits and trains a scorer on paired data. By integrating offline evaluation with reinforcement learning to optimize policies, the approach enables dynamic decision-making within Llama and OLMo architectures coupled with dense retrieval. The proposed method significantly improves the accuracy-revision rate trade-off curve, narrowing the gap toward oracle strategies and outperforming conventional confidence estimation techniques. Furthermore, the analysis reveals that directly selecting standard RAG answers remains advantageous in most scenarios.

Abstract

We consider the decision of whether to return an existing draft answer or revise it using retrieved evidence, as in answer-revision systems. Draft confidence estimates whether the current answer is correct, but the decision requires estimating the effect of a specified revision. For offline training and evaluation, we grade both the returned draft and its candidate revision under the same correctness judge, which makes repair, harm, and the gap to an oracle observable. We call this paired effect its recoverability, and we train policies to predict it before revision. On 25,870 held-out open-domain questions across three revision setups, a scorer trained on the paired outcome has greater area under the accuracy--revision-rate curve than a matched draft-correctness scorer in all nine Llama setup--seed fits, and gains 0.23--0.68 accuracy points on average at development-selected thresholds, a difference significant across training runs only for dense retrieval. The resulting policy improves on always revising and on average closes more than a third of the oracle gap, although it still applies 38--46% of the harmful revisions. When a draft-free standard-RAG answer is also available, however, choosing between the draft and that answer is stronger by about two points for Llama and four for OLMo, and adding candidate revision as a third option yields no significant gain. Recoverability describes one revision; its value as an available action also depends on the alternatives.

Problem

• Retrieval-Augmented QA
• Answer Revision
• Draft Confidence
• Recoverability
• Decision Policy

Innovation

• Recoverability
• Retrieval-Augmented QA
• Answer Revision
• Paired Outcome Scoring
• Revision Policy