Reproduce “DeepRAG: Integrating Hierarchical Reasoning and Process Supervision for Biomedical Multi-Hop QA” 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
Yuelyu Ji, Hang Zhang, Shiven Verma, Hui Ji, Chun Li, Yushui Han, Yanshan Wang
DeepRAG, a framework from the University of Pittsburgh, integrates hierarchical reasoning with process supervision to enhance biomedical multi-hop question answering. It achieved an Exact Match score of 62.4% and a Concept Accuracy of 71.8% on the MedHopQA dataset, demonstrating improved performance over baseline RAG-Gym and standalone DeepSeek models by leveraging specific biomedical semantic matching.
Key insights
- The combination of hierarchical question decomposition and reward-guided fine-tuning is effective for improving performance in complex biomedical multi-hop QA tasks.
- Explicitly tracking nested dependencies through hierarchical indicators significantly enhances the clarity of reasoning pathways and optimizes the precision of information retrieval.
- Incorporating concept-level rewards, based on semantic matching with specialized ontologies like UMLS, is critical for achieving high semantic precision and biomedical alignment in generated answers and retrieved information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.