Reproduce “Multi-hop Question Answering via Reasoning Chains” 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

Jifan Chen, Shih-ting Lin, Greg Durrett

Researchers at The University of Texas at Austin developed a two-stage model for multi-hop question answering that first explicitly extracts discrete reasoning chains before predicting answers. This approach achieved a new state-of-the-art accuracy of 76.5% on WikiHop and a strong 74.11% F1 on HotpotQA, demonstrating improved interpretability for complex question answering.

Key insights

  • Encoding sentences within their paragraph context significantly improves chain extraction by capturing cross-sentence relations and reducing distractor selection.
  • Training with heuristically derived 'Question Overlap' reasoning chains is more effective for supporting fact retrieval than 'Shortest Path' chains, leading to more comprehensive evidence.
  • Explicitly modeling sequential reasoning with a pointer network for chain extraction yields consistently higher QA performance compared to unordered sentence selection, especially for complex questions.

Summary by alphaXiv.

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