Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SUBQRAG: Sub-Question Driven Dynamic Graph RAG

Jiaoyang Li, Junhao Ruan, Shengwei Tang, Saihan Chen, Kaiyan Chang, Yuan Ge, Tong Xiao, Jingbo Zhu

TL;DR

SubQRAG tackles the limitations of static Graph RAG in multi-hop QA by decomposing a complex question into an ordered chain of verifiable sub-questions and dynamically updating the knowledge graph with new triples from source documents when needed. Each sub-question retrieves relevant graph triples, and a collective graph memory is formed to ground the final answer, mitigating error accumulation from undirected graph traversal. The approach yields consistent EM improvements across HotpotQA, MuSiQue, and 2WikiMultiHopQA, demonstrating enhanced multi-hop reasoning and traceable evidence paths. The framework balances reasoning depth with knowledge-base breadth, enabling more robust, up-to-date question answering with interpretable reasoning traces.

Abstract

Graph Retrieval-Augmented Generation (Graph RAG) effectively builds a knowledge graph (KG) to connect disparate facts across a large document corpus. However, this broad-view approach often lacks the deep structured reasoning needed for complex multi-hop question answering (QA), leading to incomplete evidence and error accumulation. To address these limitations, we propose SubQRAG, a sub-question-driven framework that enhances reasoning depth. SubQRAG decomposes a complex question into an ordered chain of verifiable sub-questions. For each sub-question, it retrieves relevant triples from the graph. When the existing graph is insufficient, the system dynamically expands it by extracting new triples from source documents in real time. All triples used in the reasoning process are aggregated into a "graph memory," forming a structured and traceable evidence path for final answer generation. Experiments on three multi-hop QA benchmarks demonstrate that SubQRAG achieves consistent and significant improvements, especially in Exact Match scores.

SUBQRAG: Sub-Question Driven Dynamic Graph RAG

TL;DR

SubQRAG tackles the limitations of static Graph RAG in multi-hop QA by decomposing a complex question into an ordered chain of verifiable sub-questions and dynamically updating the knowledge graph with new triples from source documents when needed. Each sub-question retrieves relevant graph triples, and a collective graph memory is formed to ground the final answer, mitigating error accumulation from undirected graph traversal. The approach yields consistent EM improvements across HotpotQA, MuSiQue, and 2WikiMultiHopQA, demonstrating enhanced multi-hop reasoning and traceable evidence paths. The framework balances reasoning depth with knowledge-base breadth, enabling more robust, up-to-date question answering with interpretable reasoning traces.

Abstract

Graph Retrieval-Augmented Generation (Graph RAG) effectively builds a knowledge graph (KG) to connect disparate facts across a large document corpus. However, this broad-view approach often lacks the deep structured reasoning needed for complex multi-hop question answering (QA), leading to incomplete evidence and error accumulation. To address these limitations, we propose SubQRAG, a sub-question-driven framework that enhances reasoning depth. SubQRAG decomposes a complex question into an ordered chain of verifiable sub-questions. For each sub-question, it retrieves relevant triples from the graph. When the existing graph is insufficient, the system dynamically expands it by extracting new triples from source documents in real time. All triples used in the reasoning process are aggregated into a "graph memory," forming a structured and traceable evidence path for final answer generation. Experiments on three multi-hop QA benchmarks demonstrate that SubQRAG achieves consistent and significant improvements, especially in Exact Match scores.

Paper Structure

This paper contains 16 sections, 5 equations, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Offline Indexing
  4. Question Decomposition and Rewriting
  5. Question Decomposition
  6. Question Rewriting
  7. Retrieval and Dynamic Graph Updating
  8. Answer Generation
  9. EXPERIMENTS
  10. Datasets
  11. Baselines
  12. Metrics
  13. Implementation Details
  14. Results
  15. Ablation Study
  16. ...and 1 more sections

Figures (1)

  • Figure 1: The SubQRAG framework consists of four stages: (I) Offline Indexing: the corpus is processed into triples to construct the initial KG. (II) Question Decomposition + Rewriting: the original question is decomposed into verifiable sub-questions. (III) Retrieval and Dynamic Graph Updating: each sub-question retrieves supporting triples from the graph. If unanswered, documents are consulted to extract new triples and update the graph. (IV) Answer Generation: supporting triples are assembled into graph memory to generate the final answer based on the original question.