Reproduce “SUBQRAG: Sub-Question Driven Dynamic Graph RAG” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

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

Northeastern University's SubQRAG framework enhances Retrieval-Augmented Generation by enabling sub-question-driven multi-step reasoning and dynamic knowledge graph updates for complex queries. This approach consistently outperforms existing methods on multi-hop QA benchmarks, notably achieving an Exact Match score of 61.90% on 2WikiMultiHopQA, an 11.3 percentage point improvement over prior state-of-the-art.

Key insights

  • Decomposing complex questions into sequential sub-questions and contextually rewriting them for each step is paramount for effective multi-hop reasoning.
  • Real-time, dynamic updating of the Knowledge Graph by extracting new triples on the fly is essential for adaptability and addressing knowledge gaps in static KGs.
  • Aggregating utilized knowledge triples into a structured "graph memory" significantly improves evidence traceability, reduces error accumulation, and optimizes LLM context usage.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.