DeFT-RAG: Decomposing Multi-Hop Questions into Self-Contained Sub-Questions with Reusable Answers

DeFT-RAG. Decomposing Multi-Hop Questions into Self-Contained Sub-Questions with Reusable Answers. By Dingyi Kang and one other, from the University of Texas at Dallas. Published on the thirtieth of October.

Retrieval-augmented generation lets a large language model answer a question by first retrieving relevant passages from a large corpus. It struggles with multi-hop questions. These need facts spread across several documents to be chained together, through intermediate entities the question itself never mentions. To answer, the system has to find something it was never told to look for.

The two main families of approach each fall short. Graph-based methods retrieve once, starting from the entities the question names, and so they rarely recover the full chain of evidence. Iterative methods do recover it, step by step, but their accuracy is limited by how much they ask of the language model. It has to plan the sub-questions, decide how they depend on one another, rewrite them as answers come in, and keep track of what is already known. That is many jobs at once for a single model.

DeFT-RAG is built to take most of those jobs away. Its central move is to generate one self-contained sub-question at a time. Each new sub-question is produced from the original question together with the sub-questions answered so far and their answers, the question and answer pairs. And each sub-question names every entity it asks about. Because nothing in it points back to an earlier step, there is no plan to maintain, no dependency to track and no rewriting to do. The model focuses on one fine-grained task at a time.

Two further components complete the framework. The second moves the remaining work out of the language model altogether. What is known is kept in a table of question and answer pairs, updated by rules. Decisions that reduce to yes or no are made by a lightweight decision model that has been fine-tuned for the purpose.

The third component is a cache. DeFT-RAG stores the question and answer pairs, and when a later question asks a sub-question with the same meaning as one already answered, the stored answer is reused. Since every sub-question is self-contained, an answer written for one question can be picked up by another.

The evaluation uses MuSiQue, 2WikiMultiHopQA and HotpotQA. With either gpt-4o-mini or Qwen3.6 as the backbone, DeFT-RAG achieves the best accuracy among thirteen baselines. It also gives the language model fewer input tokens than any other iterative method.

The cache was tested with a single shared store across one thousand MuSiQue questions. In that setting, twenty-three point one percent of sub-questions were answered from the cache. That cut the number of language model calls by twelve point nine percent, without lowering accuracy.

One design choice carries the whole system. Making every sub-question self-contained is what removes planning, dependency tracking and rewriting from the model's workload, what lets bookkeeping move into rules and a small decision model, and what makes answers reusable across questions. The reported results line up with that: higher accuracy, less input for the model, and fewer calls once a cache has been built.