Research map / Large language models / Retrieval and knowledge

From Complex to Atomic: Enhancing Augmented Generation via Knowledge-Aware Dual Rewriting and Reasoning

ICML 2025 · Large language models › Retrieval and knowledge · approach: RAG robustness and evaluation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes KAR3-RAG, a training-free iterative RAG framework that atomizes chunks into question-style tags and dynamically decomposes multihop questions via atomic query proposals, atomic retrieval, and selection.

Problem

Standard RAG and fixed decomposition pipelines struggle with domain-specific multihop questions because retrieval is misaligned with corpus phrasing and sub-questions are generated without regard to available knowledge.

Approach

Atomize each chunk into question-form atomic tags indexed in a knowledge base, have the LLM propose multiple atomic queries per iteration conditioned on accumulated context, retrieve (tag, chunk) pairs, and let an LLM atomic selector pick the next chunk to add, iterating until no more information is needed.

Key results

• With GPT-4, KAR3 reaches 88.00 Acc on HotpotQA, 82.20 on 2Wiki, and 62.60 on MuSiQue, the best among the compared methods in Table 1.
• With Llama 3.1-70B, KAR3 improves MuSiQue accuracy to 59.70 versus 49.60 for IRCoT, a gain of +10.1 (20.4%) over the second-best method.
• Ablations show removing atomic-question tags, multi-query proposals, pair retrieval, or atomic selection reduces accuracy by up to 15.1%, 16.6%, 15.3%, and 16.2% respectively.

Measured effects reported in the paper

Surprising findings

• With GPT-3.5, KAR3 does not clearly outperform IRCoT and Self-Ask with retrieval and sometimes falls behind, suggesting gains depend on the LLM's reasoning ability.
• GraphRAG local achieves high accuracy but very low EM and F1 on HotpotQA, because its outputs echo the question and include meta-information.

Limitations

• Performance depends heavily on the reasoning capability of the underlying LLM; gains shrink with GPT-3.5.
• Higher token consumption than some baselines, with completion tokens about one quarter of total, raising API cost.
• Evaluation uses sampled 500-question subsets and a single shared knowledge base per dataset, and the learnable query proposer is left as future work.

Experimental setup

Datasets: HotpotQA 2WikiMultiHopQA MuSiQue LawBench Open Australian Legal QA

Benchmarks: HotpotQA 2WikiMultiHopQA MuSiQue LawBench Open Australian Legal QA

Metrics: F1 Accuracy (GPT-4 judged) EM Precision Recall

Baselines: Zero-Shot CoT Naive RAG Self-Ask w/ Retrieval IRCoT Iter-RetGen ProbTree SearChain GraphRAG Local GraphRAG Global

Builds on: Self-Ask IRCoT Iter-RetGen ProbTree GraphRAG

Closest papers by meaning

• StructRAG: Boosting Knowledge Intensive Reasoning of LLMs via Inference-time Hybrid Information Structurization (ICLR 2025 · similarity 0.89)
• Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.89)
• Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.88)
• Think-on-Graph 2.0: Deep and Faithful Large Language Model Reasoning with Knowledge-guided Retrieval Augmented Generation (ICLR 2025 · similarity 0.88)
• Retriever-Free Retrieval-Augmented Reasoning via Corpus-Traversing MCTS (NeurIPS 2026 · similarity 0.88)
• Can Knowledge-Graph-based Retrieval Augmented Generation Really Retrieve What You Need? (NeurIPS 2025 · similarity 0.88)