DeFT-RAG is a retrieval-augmented generation method for multi-hop questions. The LLM writes one self-contained sub-question at a time, and most of the other work moves outside the model.
Approach: Each sub-question names every entity it asks about. A rule-updated QA table tracks what is known, and a fine-tuned lightweight decision model makes the yes/no calls. The author's comment says that model is 4B.
Results: The paper reports the best accuracy among 13 baselines on MuSiQue, 2WikiMultiHopQA and HotpotQA, with gpt-4o-mini and Qwen3.6 as backbones. The author's comment adds +5.2 points on MuSiQue over Hi-Q. It also gives the LLM fewer input tokens than any other iterative method.
Caching: Across 1,000 MuSiQue questions with a shared cache, 23.1% of sub-questions are answered from it, cutting LLM calls by 12.9% without lowering accuracy.
Code: The code is on GitHub, and the decision model is on Hugging Face.
Sign in to comment.
TL;DR of the paper, generated automatically.
DeFT-RAG is a retrieval-augmented generation method for multi-hop questions. The LLM writes one self-contained sub-question at a time, and most of the other work moves outside the model.