Research map / Large language models / Retrieval and knowledge
Cooperative Retrieval-Augmented Generation for Question Answering: Mutual Information Exchange and Ranking by Contrasting Layers
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 CoopRAG, a RAG framework where an LLM unrolls questions into sub-questions and a masked reasoning chain, a fine-tuned MPNet retriever reranks documents by contrasting layers (RaLa), and the LLM fills masks before answering.
Problem
Single-step and multi-hop QA RAG systems suffer from incorrect retrievals and LLM hallucinations, partly because short questions lack information and retrievers rely on final-layer, shallow similarity.
Approach
Cooperative retriever-LLM loop: the LLM unrolls the question into sub-questions and a reasoning chain with low-confidence entities masked; the retriever reranks by the score gap between last and intermediate encoder layers (RaLa), and the LLM fills masks from retrieved documents before answering.
Key results
- CoopRAG (GPT-4o-mini) reaches 88.8 R@2 on HotpotQA and 80.8 R@2 on NQ, improving NQ R@2 by 35.2% over HippoRAG2.
- CoopRAG (GPT-4o-mini) gets 65.6 EM on HotpotQA and 72.0 EM on NQ in single-step QA.
- On SimpleQA, CoopRAG reaches 58.3 EM vs 50.2 for HopRAG (best baseline).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoopRAG retrieval instead of HippoRAG2 | Single-step retrieval · NQ | R@2: 46.2 -> 80.8 | Helped |
| Use CoopRAG instead of HopRAG | Question answering · SimpleQA | EM: 50.2 -> 58.3 | Helped |
| Remove uncertainty masks and reasoning chain completion (ablation) | CoopRAG · MuSiQue | EM: 52.3 -> 47.2 | Hurt |
| Remove gap-aware weight ω_U,D (ablation) | CoopRAG retrieval · MuSiQue | R@2: 59.6 -> 57.0 | Hurt |
| Remove difficulty-aware weight α_U (ablation) | CoopRAG retrieval · MuSiQue | R@2: 59.6 -> 58.2 | Hurt |
| Remove both α_U and ω_U,D (ablation) | CoopRAG retrieval · 2Wiki | R@2: 81.4 -> 76.1 | Hurt |
| Add sub-questions and a masked reasoning chain (ablation) | CoopRAG retrieval · MuSiQue | R@2: 43.7 -> 59.6 | Helped |
| Add sub-questions without a reasoning chain (ablation) | CoopRAG retrieval · HotpotQA | R@2: 71.6 -> 80.3 | Helped |
| Add a masked reasoning chain without sub-questions (ablation) | CoopRAG retrieval · HotpotQA | R@2: 71.6 -> 86.4 | Helped |
Surprising findings
- Gemma2-9B-based CoopRAG outperforms HippoRAG2 with Llama3.3-70B on single-step retrieval, despite being much smaller.
Limitations
- Pretrained LMs cannot embed long documents due to sequence length limits.
- MaxSim operations incur rising computational cost as token count grows, as in ColBERT.
- Evaluated only on passage-based retrieval; KBQA and domain-specific datasets left to future work.
Experimental setup
Datasets: HotpotQA MuSiQue 2WikiMultihopQA NaturalQuestions
Benchmarks: HotpotQA MuSiQue 2WikiMultihopQA NaturalQuestions SimpleQA FreshQA
Metrics: Recall@2 Recall@5 EM F1 Correct/Incorrect/Not Attempted
Baselines: BM25 Contriever ColBERTv2 GTR ReSCORE HippoRAG2 SiReRAG HopRAG RAPTOR IRCoT
Builds on: ColBERT DoLa HippoRAG IRCoT
Closest papers by meaning
- RankRAG: Unifying Context Ranking with Retrieval-Augmented Generation in LLMs (NeurIPS 2024 · similarity 0.90)
- Autofocus Retrieval: An Effective Pipeline for Multi-Hop Question Answering With Semi-Structured Knowledge (NeurIPS 2026 · similarity 0.89)
- Chain-of-Retrieval Augmented Generation (NeurIPS 2025 · similarity 0.89)
- RAG-IGBench: Innovative Evaluation for RAG-based Interleaved Generation in Open-domain Question Answering (NeurIPS 2025 · similarity 0.89)
- UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-World Document Analysis (NeurIPS 2024 · similarity 0.89)
- ChatQA: Surpassing GPT-4 on Conversational QA and RAG (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.