DeFT-RAG decomposes multi-hop questions into self-contained sub-questions with reusable answers
Dingyi Kang · hf · 2026-11-03
UT Dallas proposes DeFT-RAG, improving retrieval-augmented generation for multi-hop QA: graph-based RAG's one-shot retrieval rarely recovers full evidence chains through bridge entities (only 6-10% of 3-hop MuSiQue questions), while iterative RAG is limited by heavy LLM-driven planning and rewriting.
DeFT-RAG's three components:
- Generates one self-contained sub-question at a time (each naming every entity it asks about), eliminating planning, dependency tracking and rewriting
- Offloads reasoning to rule-based updates and a lightweight decision model
- Caches sub-question answers for reuse, reducing LLM usage
Improves accuracy on knowledge-intensive tasks without retraining the model.
More from Research
- VideoRebirth Report: Agent Harnesses for Agentic Video Generation, World Models and Recursive Self-Improvement — VideoRebirthAI · 2026-11-03
- Huawei's TinyAstral Distills Self-Evolving Harnesses into 4B Robot Agents, Wins Across Six Benchmarks — Kehong Gong · 2026-11-03
- KAIST's ReTrace Demystifies Computer-Use Agent Recovery, Finds Reflection and Rollback Fragile — kaist-ai · 2026-11-03
- Luxembourg study: empathetic framing pushes Gemini to human-level anxiety scores on GAD-7 — xiaohu · 2026-11-03
- Three wild predictions after MIT's weight-space programming work — Yuchenj_UW · 2026-11-03
- Mistral Large 4 tech report rejected by arXiv, team releases it on alphaXiv — _AndrewZhao · 2026-11-03