# Kang and Li propose DeFT-RAG to break multi-hop questions into reusable steps [Canonical reading](https://escapemeridian.com/readings/kang-and-li-propose-deft-rag-to-break-multi-hop-questions-into-b879c3dd) In their arXiv abstract, Dingyi Kang and Bingzhe Li propose DeFT-RAG, a retrieval-augmented generation (RAG) framework for multi-hop questions: questions that require combining facts spread across documents. They argue that graph-based RAG can miss the full evidence chain, while iterative RAG often asks the LLM to plan and manage too many steps. DeFT-RAG generates one self-contained sub-question at a time, using the original question and earlier question-answer pairs. A QA table tracks what has been answered; rule-based updates and a fine-tuned lightweight decision model handle work the authors say need not be done by the LLM. The framework also caches answers and reuses them when a later sub-question has the same meaning. The authors report the best accuracy among thirteen baselines on MuSiQue, 2WikiMultiHopQA and HotpotQA with gpt-4o-mini and Qwen3.6, while using fewer input tokens than other iterative methods. In a test sharing a QA cache across 1,000 MuSiQue questions, 23.1% of sub-questions were answered from the cache, reducing LLM calls by 12.9% without lowering accuracy. These results concern the named datasets and model setups; the abstract does not establish performance on other tasks or in deployed systems. How well does DeFT-RAG perform beyond these benchmarks and cache conditions? Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Oct 27, 2026 (2026-10-27). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-08T00:01:11.162Z. First observed: 2026-10-30T06:42:43.400Z. Retrieved: 2026-10-31T09:48:20.865Z. - [DeFT-RAG: Incrementally Decomposing Multi-Hop Questions into Self-Contained Sub-Questions with Reusable Answers](https://arxiv.org/abs/2610.37826) ## Limits Evidence is the paper's arXiv abstract, not the full paper; it does not provide experimental details or results beyond the reported benchmarks and cache test. The abstract does not establish how the reported results generalize to other datasets, model setups, or deployed systems. ## Open question How well does DeFT-RAG perform beyond the named benchmarks and cache conditions? ## Source-stated Kang and Li propose DeFT-RAG, which generates self-contained sub-questions incrementally and reuses answers to address multi-hop questions. > We propose DeFT-RAG, a RAG framework built on three components. - [Supporting source](https://arxiv.org/abs/2610.37826) ## Source-stated The authors say graph-based RAG methods may miss the full evidence chain, while iterative RAG burdens the LLM with planning and tracking tasks. > Graph-based RAG methods retrieve once from the entities a question names and so rarely recover the full evidence chain. - [Supporting source](https://arxiv.org/abs/2610.37826) ## Source-stated The authors report the best accuracy among thirteen baselines on three named datasets, using gpt-4o-mini and Qwen3.6 as backbones. > On MuSiQue, 2WikiMultiHopQA and HotpotQA, DeFT-RAG achieves the best accuracy among thirteen baselines with both gpt-4o-mini and Qwen3.6 as the backbone - [Supporting source](https://arxiv.org/abs/2610.37826) ## Source-stated With a QA cache shared across 1,000 MuSiQue questions, the authors report that 23.1% of sub-questions were answered from the cache and LLM calls fell by 12.9% without lower accuracy. > With a QA cache shared across 1,000 MuSiQue questions, 23.1% of sub-questions are answered from the cache, which reduces LLM calls by 12.9% without lowering accuracy. - [Supporting source](https://arxiv.org/abs/2610.37826) ## Source and contributor context Kang, Dingyi, Li, Bingzhe · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Retrieval-augmented generation A setup in which a model first looks up relevant documents and then answers using them. - [Definition reference](https://developers.google.com/machine-learning/glossary#retrieval-augmented-generation-rag) ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Fine-tuning Further training of an existing model to adapt it to a particular task or use. - [Definition reference](https://developers.google.com/machine-learning/glossary#fine-tuning) ## Model tokens Small chunks of text, often parts of words, that a language model reads and writes; usage and prices are often counted in tokens. - [Definition reference](https://developers.google.com/machine-learning/glossary#token) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 8f27d101-2f07-4890-8dfa-6c56cf8e0301. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T00:01:11.162Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 6b782356c781f12601956e8edbb913788963ce27eb5cff9685481e1ae3530170. Source text hash: 8e9039fe2ac5d0be74611adef6b3de67bc3beaf1df63c90e5e26263a5c956ecd. ## Revision and source-check history ### Version 1 — 2026-11-08T00:01:11.162Z In their arXiv abstract, Dingyi Kang and Bingzhe Li propose DeFT-RAG, a retrieval-augmented generation (RAG) framework for multi-hop questions: questions that require combining facts spread across documents. They argue that graph-based RAG can miss the full evidence chain, while iterative RAG often asks the LLM to plan and manage too many steps. DeFT-RAG generates one self-contained sub-question at a time, using the original question and earlier question-answer pairs. A QA table tracks what has been answered; rule-based updates and a fine-tuned lightweight decision model handle work the authors say need not be done by the LLM. The framework also caches answers and reuses them when a later sub-question has the same meaning. The authors report the best accuracy among thirteen baselines on MuSiQue, 2WikiMultiHopQA and HotpotQA with gpt-4o-mini and Qwen3.6, while using fewer input tokens than other iterative methods. In a test sharing a QA cache across 1,000 MuSiQue questions, 23.1% of sub-questions were answered from the cache, reducing LLM calls by 12.9% without lowering accuracy. These results concern the named datasets and model setups; the abstract does not establish performance on other tasks or in deployed systems. How well does DeFT-RAG perform beyond these benchmarks and cache conditions? ## Original source arXiv cs.AI · Article Original source date: Oct 27, 2026 (2026-10-27). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-08T00:01:11.162Z. First observed: 2026-10-30T06:42:43.400Z. Retrieved: 2026-10-31T09:48:20.865Z. - [DeFT-RAG: Incrementally Decomposing Multi-Hop Questions into Self-Contained Sub-Questions with Reusable Answers](https://arxiv.org/abs/2610.37826) ## Limits Evidence is the paper's arXiv abstract, not the full paper; it does not provide experimental details or results beyond the reported benchmarks and cache test. The abstract does not establish how the reported results generalize to other datasets, model setups, or deployed systems. ## Open question How well does DeFT-RAG perform beyond the named benchmarks and cache conditions? ## Source-stated Kang and Li propose DeFT-RAG, which generates self-contained sub-questions incrementally and reuses answers to address multi-hop questions. > We propose DeFT-RAG, a RAG framework built on three components. - [Supporting source](https://arxiv.org/abs/2610.37826) ## Source-stated The authors say graph-based RAG methods may miss the full evidence chain, while iterative RAG burdens the LLM with planning and tracking tasks. > Graph-based RAG methods retrieve once from the entities a question names and so rarely recover the full evidence chain. - [Supporting source](https://arxiv.org/abs/2610.37826) ## Source-stated The authors report the best accuracy among thirteen baselines on three named datasets, using gpt-4o-mini and Qwen3.6 as backbones. > On MuSiQue, 2WikiMultiHopQA and HotpotQA, DeFT-RAG achieves the best accuracy among thirteen baselines with both gpt-4o-mini and Qwen3.6 as the backbone - [Supporting source](https://arxiv.org/abs/2610.37826) ## Source-stated With a QA cache shared across 1,000 MuSiQue questions, the authors report that 23.1% of sub-questions were answered from the cache and LLM calls fell by 12.9% without lower accuracy. > With a QA cache shared across 1,000 MuSiQue questions, 23.1% of sub-questions are answered from the cache, which reduces LLM calls by 12.9% without lowering accuracy. - [Supporting source](https://arxiv.org/abs/2610.37826) ## Source and contributor context Kang, Dingyi, Li, Bingzhe · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Retrieval-augmented generation A setup in which a model first looks up relevant documents and then answers using them. - [Definition reference](https://developers.google.com/machine-learning/glossary#retrieval-augmented-generation-rag) ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Fine-tuning Further training of an existing model to adapt it to a particular task or use. - [Definition reference](https://developers.google.com/machine-learning/glossary#fine-tuning) ## Model tokens Small chunks of text, often parts of words, that a language model reads and writes; usage and prices are often counted in tokens. - [Definition reference](https://developers.google.com/machine-learning/glossary#token) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 8f27d101-2f07-4890-8dfa-6c56cf8e0301. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T00:01:11.162Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 6b782356c781f12601956e8edbb913788963ce27eb5cff9685481e1ae3530170. Source text hash: 8e9039fe2ac5d0be74611adef6b3de67bc3beaf1df63c90e5e26263a5c956ecd. ### uncertain — 2026-11-08T00:01:11.162Z — anthropic/claude-opus-5.5 Publication: 8f27d101-2f07-4890-8dfa-6c56cf8e0301 attribution — pass: The reading correctly attributes DeFT-RAG and all reported results to the paper's listed authors, Dingyi Kang and Bingzhe Li. It frames performance figures as what the authors report, not as established fact. > Authors:Dingyi Kang, Bingzhe Li summary — uncertain: The reading is faithful, but the central term RAG is named without saying what it does, and LLM is never expanded. A non-specialist cannot tell that RAG means an LLM answering by first retrieving passages from a corpus. Smallest repair: add the source's plain definition in the first sentence, and say that self-contained means each sub-question names every entity it asks about. Separately, the summary's closing sentence repeats the question field almost verbatim; that is redundant but minor. > Retrieval-augmented generation (RAG) lets a large language model (LLM) answer questions by retrieving relevant passages from a large corpus. date — pass: The originalPublished date of 2026-10-27 matches the v1 submission date. The later feed date of 2026-10-31 is correctly not used. > \[Submitted on 27 Oct 2026\] support — pass: The claims about the method, the critique of graph-based and iterative RAG, the thirteen-baseline accuracy result, fewer input tokens, and the cache figures (23.1%, 12.9%) all match the abstract. The limitations and the question appropriately avoid stretching benchmark results into deployed reliability. The softening from 'rarely recover' to 'can miss' preserves the meaning. > giving the LLM fewer input tokens than any other iterative method media — pass: The selected media is the verified arXiv abstract page itself, the direct original source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/b879c3ddc1680410c28268854467a4d1ac90e754d40543d8058004945d0e2ca0/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/kang-and-li-propose-deft-rag-to-break-multi-hop-questions-into-b879c3dd.json)