DeFT-RAG：将多跳问题增量分解为带可复用答案的自包含子问题
DeFT-RAG: Incrementally Decomposing Multi-Hop Questions into Self-Contained Sub-Questions with Reusable Answers
Dingyi Kang · Bingzhe Li
中文摘要
检索增强生成（RAG）让大语言模型（LLM）通过从大规模语料中检索相关段落来回答问题，但面对多跳问题（multi-hop question）时表现不佳——这类问题需要借助问题中从未提及的中间实体，串联分散在多个文档中的事实。基于图的 RAG 方法只从问题提及的实体出发检索一次，因此很少能恢复完整证据链。迭代式 RAG 方法逐步恢复证据链，但其准确率受限于要求 LLM 承担的多项任务，例如规划子问题、判定依赖关系、随答案到达改写子问题以及跟踪已知信息。DeFT-RAG 由三个组件构成：其一，每次根据原始问题与已回答的子问题及其答案（QA 对）生成一个自包含子问题，每个子问题都明确写出所询问的每个实体，从而无需规划、跟踪依赖或改写，让 LLM 一次只专注于一项细粒度任务；其二，把 LLM 剩余的活儿移出 LLM：对记录已知信息的 QA 表做基于规则的更新，并由微调的轻量决策模型做是/否判断；其三，缓存 QA 对，当后续问题询问语义相同的子问题时复用答案。在 MuSiQue、2WikiMultiHopQA 和 HotpotQA 上，以 gpt-4o-mini 和 Qwen3.6 为骨干时，DeFT-RAG 在十三个基线中取得最佳准确率，且给 LLM 的输入 token 少于任何其他迭代方法。在 1,000 个 MuSiQue 问题共享 QA 缓存时，23.1% 的子问题由缓存作答，LLM 调用减少 12.9% 且准确率不降。
关键要点
- 01多跳问题需串联跨文档的中间实体，图式 RAG 单次检索难以恢复完整证据链
- 02每次只生成一个自包含子问题，显式写出全部实体，免去规划、依赖跟踪与改写
- 03QA 表规则更新与微调轻量决策模型承担剩余工作，降低 LLM 负担
- 04在 MuSiQue、2WikiMultiHopQA、HotpotQA 上十三基线中准确率最佳，输入 token 最少
- 05共享 QA 缓存使 23.1% 子问题直接复用答案，LLM 调用减少 12.9% 且准确率不降
解读
尚无解读。
原始英文摘要
arXiv:2610.37826v1 Announce Type: cross Abstract: Retrieval-augmented generation (RAG) lets a large language model (LLM) answer questions by retrieving relevant passages from a large corpus. However, RAG struggles with multi-hop questions, which require chaining facts spread across several documents through intermediate entities that the question never mentions. Graph-based RAG methods retrieve once from the entities a question names and so rarely recover the full evidence chain. Iterative RAG methods recover it step by step, but their accuracy is limited by the many tasks they ask the LLM to perform, such as planning the sub-questions, deciding their dependencies, rewriting them as answers arrive, and keeping track of what is known. We propose DeFT-RAG, a RAG framework built on three components. First, DeFT-RAG generates one self-contained sub-question at a time from the original question and the sub-questions answered so far with their answers (QA pairs). Each sub-question names every entity it asks about, which removes the need to plan, track dependencies or rewrite, and lets the LLM focus on a single fine-grained task at a time. Second, DeFT-RAG moves the remaining LLM work outside the LLM: rule-based updates to a QA table that keeps track of what is known, and yes/no decisions made by a fine-tuned lightweight decision model. Third, DeFT-RAG caches the QA pairs and reuses an answer when a later question asks a sub-question with the same meaning. On MuSiQue, 2WikiMultiHopQA and HotpotQA, DeFT-RAG achieves the best accuracy among thirteen baselines with both gpt-4o-mini and Qwen3.6 as the backbone, while giving the LLM fewer input tokens than any other iterative method. With a QA cache shared across 1,000 MuSiQue questions, 23.1% of sub-questions are answered from the cache, which reduces LLM calls by 12.9% without lowering accuracy.