Research map / Large language models / Retrieval and knowledge

RVR: Retrieve-Verify-Retrieve for Comprehensive Question Answering

NeurIPS 2026 · Large language models › Retrieval and knowledge · approach: Retrieval-augmented reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Retrieve-Verify-Retrieve (RVR), a multi-round retrieval loop where an LLM verifier filters retrieved documents and the query is augmented with verified documents to retrieve remaining answers, with retrievers fine-tuned for this inference setting.

Problem

Comprehensive retrieval for queries with many valid answers, where standard one-shot retrieval misses long-tail answer-bearing documents.

Approach

Iterate retrieval by conditioning the subsequent retriever on query plus LLM-verified documents from prior rounds, and fine-tune retrievers to retrieve complementary (not yet covered) gold documents, rather than relying on agentic query generation.

Key results

• On QAMPARI, RVR with fine-tuned retrievers reaches MRecall@100 of 33.70 with INF-Retriever (FT(D_i)+FT(D_r)) versus 29.30 for FT(D_i) alone.
• Out-of-domain on QUEST, Base + FT(D_r) raises MRecall@100 from 3.24 to 4.52 for Contriever and from 4.75 to 6.02 for INF.
• With an oracle verifier, MRecall@100 on QAMPARI rises to 36.30 (INF), indicating headroom in verification; the LLM verifier plateaus after two iterations.

Surprising findings

• Agentic baselines (Tongyi, SearchR1) underperform simple base retrievers on QAMPARI MRecall in most settings.
• Fine-tuned initial retriever underperforms the off-the-shelf retriever on QUEST and WebQuestionsSP, so the base retriever is used for initial retrieval there.

Limitations

• Efficiency overhead: RVR is about 2-3x slower than single-pass baseline and needs extra GPU memory.
• LLM verifier tends to select redundant documents in later iterations, limiting multi-turn gains.
• Out-of-domain generalization is mixed; fine-tuned retrievers hurt on WebQuestionsSP due to domain mismatch.

Experimental setup

Datasets: QAMPARI Wikipedia dump 2021-08-01 (25.9M passages)

Benchmarks: QAMPARI QUEST WebQuestionsSP

Metrics: MRecall@100 Recall@100 Precision Recall Accuracy

Baselines: Base retrievers FT (D_i) Tongyi DeepResearch SearchR1 TopK verifier

Builds on: Contriever Search-R1 in-batch negatives

Closest papers by meaning

• Retriever-Free Retrieval-Augmented Reasoning via Corpus-Traversing MCTS (NeurIPS 2026 · similarity 0.86)
• BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval (ICLR 2025 · similarity 0.86)
• Cooperative Retrieval-Augmented Generation for Question Answering: Mutual Information Exchange and Ranking by Contrasting Layers (NeurIPS 2025 · similarity 0.86)
• POQD: Performance-Oriented Query Decomposer for Multi-vector retrieval (ICML 2025 · similarity 0.85)
• SuRe: Summarizing Retrievals using Answer Candidates for Open-domain QA of LLMs (ICLR 2024 · similarity 0.85)
• Autofocus Retrieval: An Effective Pipeline for Multi-Hop Question Answering With Semi-Structured Knowledge (NeurIPS 2026 · similarity 0.85)