Reproduce “PAR$^2$-RAG: Planned Active Retrieval and Reasoning for Multi-Hop 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
Xingyu Li, Rongguang Wang, Yuying Wang, Mengqing Guo, Chenyang Li, Tao Sheng, Sujith Ravi, Dan Roth
The PAR²-RAG framework improves multi-hop question answering by decoupling broad evidence collection from iterative reasoning refinement, using a two-stage agentic approach. It achieves superior answer quality, with up to a 23.5% gain over IRCoT on MuSiQue, and enhanced retrieval performance across several benchmarks compared to existing RAG methods.
Key insights
- Decoupling broad evidence collection (coverage) from specific reasoning path commitment is crucial for robust multi-hop question answering.
- An adaptive depth control mechanism, rather than a fixed large hop budget, allows for more efficient and stable performance.
- The initial broad evidence frontier (Coverage Anchor) provides a more stable foundation for reasoning, especially as generator capabilities increase.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.