Reproduce “Decomposition-Driven Multi-Table Retrieval and Reasoning for Numerical 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

Feng Luo, Hai Lan, Hui Luo, Zhifeng Bao, Xiaoli Wang, J.Shane Culpepper, Shazia Sadiq

A framework named DMRAL improves numerical Multi-Table Question Answering (MTQA) over large-scale, heterogeneous table collections by explicitly modeling diverse inter-table relationships and employing a decomposition-driven approach for retrieval and reasoning. The method achieved an average of 24% improvement in table retrieval and 55% improvement in answer accuracy compared to state-of-the-art baselines on new large-scale benchmarks.

Key insights

  • Explicitly modeling both joinability and unionability via a Table Relationship Graph is crucial for effective MTQA over diverse, large-scale table collections.
  • Decomposition-driven, table-aligned sub-question generation significantly reduces redundancy and improves the accuracy of downstream table retrieval and programmatic reasoning.
  • Incremental program generation via Chain-of-Thought prompting combined with execution-guided refinement substantially enhances the robustness and accuracy of numerical answer generation.

Summary by alphaXiv.

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