SignalTech/AI
DeFT-RAG: Decomposing Multi-Hop Questions into Self-Contained Sub-Questions with Reusable Answers
High-priority alert based on source significance and keyword relevance.
Context
What happened: DeFT-RAG is a retrieval-augmented generation framework that generates one self-contained sub-question at a time, moving rule-based QA table updates and yes/no decisions to a fine-tuned lightweight decision model, and caching answers for reuse.
Key numbers: Best accuracy among 13 baselines on MuSiQue, 2WikiMultiHopQA and HotpotQA with gpt-4o-mini and Qwen3.6; +5.2 points on MuSiQue over Hi-Q;
Why it matters: The method gives the LLM fewer input tokens than any other iterative RAG approach while improving accuracy, and cache reuse reduces LLM calls without a drop in accuracy.
What changed
DeFT-RAG is a retrieval-augmented generation framework that generates one self-contained sub-question at a time, moving rule-based QA table updates and yes/no decisions to a fine-tuned lightweight decision model, and caching answers for reuse.
Why it matters
The method gives the LLM fewer input tokens than any other iterative RAG approach while improving accuracy, and cache reuse reduces LLM calls without a drop in accuracy.