DeFT-RAG decomposes multi-hop questions for RAG
Original title: DeFT-RAG: Decomposing Multi-Hop Questions into Self-Contained Sub-Questions with Reusable Answers
DeFT-RAG reduces LLM calls by 12.9% through QA cache reuse while achieving top accuracy on multi-hop benchmarks.
DeFT-RAG generates one self-contained sub-question at a time from the original question and prior QA pairs. It uses a rule-based QA table for updates and a lightweight decision model for yes/no decisions. Sub-questions are answered from cache when identical. On MuSiQue 2WikiMultiHopQA and HotpotQA it outperforms thirteen baselines with gpt-4o-mini and Qwen3.6 backbones. With a 1000-question cache 23.1% of sub-questions are answered from cache.
- Generates self-contained sub-questions without planning or rewriting
- Moves most LLM work outside using rule-based QA table
- 23.1% sub-questions answered from cache reducing calls 12.9%
- Best accuracy among thirteen baselines with both backbones
- AI models
- gpt-4o-miniQwen3.6
From the original article
1 min read at huggingface.co
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.
By Dingyi Kang, Bingzhe Li · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.