让大模型在推理时智能决定何时检索，提升准确率同时减少检索次数。
When to Retrieve During Reasoning: Adaptive Retrieval for Large Reasoning Models
- 按推理步骤检测知识缺口，精准判断何时需要外部证据。
- 相比传统方法，平均提升10.1%答案准确率，检索调用减少47%。
- 适合需要多步推理的复杂问答任务，尤其擅长少次检索高效求解。
大型推理模型如 DeepSeek-R1 和 OpenAI o1 会生成长达数千个 token 的思维链，但其与检索增强生成（RAG）的集成仍存在根本性错配。现有 RAG 系统优化的是推理前提供上下文，而推理模型需要在多步推理过程中动态注入证据。本文提出 ReaLM-Retrieve，一种面向推理的检索框架，包含三项创新：（1）步骤级不确定性检测器，能在推理步骤层面识别知识缺口，而非仅限于词或句子层级；（2）检索干预策略，学习在何时引入外部证据能最大化推理收益；（3）效率优化的集成机制，使每次检索开销降低 3.2 倍。在 MuSiQue、HotpotQA 与 2WikiMultiHopQA 上的实验表明，ReaLM-Retrieve 在三个基准上平均比标准 RAG 提升 10.1% 的答案 F1（范围为 9.0–11.8%），同时相比固定间隔方法如 IRCoT 减少 47% 的检索调用（所有提升均在 p<0.01 时显著，采用配对自助法检验）。在需 2–4 步推理的 MuSiQue 难题上，达到 71.2% F1，平均每题仅需 1.8 次检索。分析显示，该方法还提升了检索质量，在支持证据上的 Recall@5 达 81.3%，且精度和 MRR 均优于固定间隔基线，确立了推理密集型检索任务的新性能-效率标杆。
原文摘要 · Abstract (English)
Large reasoning models such as DeepSeek-R1 and OpenAI o1 generate extended chains of thought spanning thousands of tokens, yet their integration with retrieval-augmented generation (RAG) remains fundamentally misaligned. Current RAG systems optimize for providing context before reasoning begins, while reasoning models require evidence injection during multi-step inference chains. We introduce ReaLM-Retrieve, a reasoning-aware retrieval framework that addresses this mismatch through three key innovations: (1) a step-level uncertainty detector that identifies knowledge gaps at reasoning-step granularity rather than token or sentence level; (2) a retrieval intervention policy that learns when external evidence maximally benefits ongoing reasoning; and (3) an efficiency-optimized integration mechanism that reduces per-retrieval overhead by 3.2x compared to naive integration. Experiments on MuSiQue, HotpotQA, and 2WikiMultiHopQA demonstrate that ReaLM-Retrieve achieves on average 10.1% absolute improvement in answer F1 over standard RAG (range: 9.0-11.8% across the three benchmarks) while reducing retrieval calls by 47% compared to fixed-interval approaches like IRCoT (all improvements significant at p<0.01, paired bootstrap). On the challenging MuSiQue benchmark requiring 2-4 hop reasoning, our method achieves 71.2% F1 with an average of only 1.8 retrieval calls per question. Analysis shows that ReaLM-Retrieve also improves retrieval quality itself, achieving 81.3% Recall@5 with consistently higher precision and MRR than fixed-interval baselines on supporting evidence, establishing new state-of-the-art efficiency-accuracy trade-offs for reasoning-intensive retrieval tasks.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。