Candidate Evidence Reranking and Risk-Aware Answer Selection in Multi-Hop Retrieval-Augmented Generation
Abstract
Abstract Multi-hop retrieval-augmented generation (RAG) can fail to fully exploit useful evidence even after it has entered the candidate pool, because ranking and limited context budgets determine which documents are actually exposed to the reader. Candidate answers produced by multiple retrieval trajectories can likewise be misselected when answer frequency does not align with correction value. Building on the SQL-Retrieval Augmented Generation (SAG) backbone, we develop a two-layer framework for candidate utilization. At the evidence layer, candidate evidence reranking (CAPE) combines trajectory ranks, lexical matching, and event-entity structural features to rerank the full union of existing candidates. At the answer layer, trajectory candidate selection (CTA) and evidence-based candidate recovery (EBC) organize multi-trajectory and single-view candidates; gain-harm estimates are then used for gated selection and joint arbitration (Joint) relative to the current answer. Experiments on 2WikiMultiHopQA, MuSiQue, and HotpotQA show that, under a fixed Top-5 reading budget, CAPE improves Recall@5 by 1.25-3.50 percentage points and answer F1 by 1.38-3.45 percentage points. In the unified end-to-end chained evaluation, CAPE+Joint improves F1 over SAG by 1.63-4.26 percentage points. Comparisons with reciprocal rank fusion (RRF) and three-trajectory majority voting further show that learned reranking and risk-aware answer selection make more effective use of existing candidates.…
Read the paper
Related studies
- Knowledge Management Perspective of Generative Artificial Intelligence
- Navigating Generative Artificial Intelligence Promises and Perils for Knowledge and Creative Work
- Characterizing the evolving research on enterprise content management
- Large language models and multimodal foundation models for precision oncology
- From Generative AI to Extended Reality: Multidisciplinary Perspectives on the Challenges, Opportunities, and Future of Educational Computing
- Generative Artificial Intelligence: A New Engine for Advancing Environmental Science and Engineering
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.