Reproduce “Trace Only What You Need: Structure-Aware On-Demand Hypergraph Memory for Long-Document Question Answering” 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

Xiangjun Zai, Xingyu Tan, Chen Chen, Xiaoyang Wang, Wenjie Zhang

DocTrace supports long-document question answering by building a tree once per document and organizing selected evidence into shared working memory on demand. It can save plans for reuse. With GPT-4o-mini, it reports average relative gains over ComoRAG (Best) of 16.91% in F1 and 15.50% in exact match across NarrativeQA and EN.QA; these are benchmark results, not guarantees for every query.

Key insights

  • Dynamic, on-demand knowledge organization significantly reduces computational overhead and token usage compared to static, pre-indexed approaches for long-document QA.
  • Explicitly preserving and leveraging a document's inherent hierarchical and sequential structure (e.g., sections, paragraphs) is crucial for accurate contextual understanding and effective evidence retrieval.
  • Learning from and reusing successful multi-step reasoning trajectories enhances the efficiency and adaptability of RAG systems, enabling them to improve performance on recurring types of complex queries.

Summary by alphaXiv.

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