Reproduce “Question Decomposition for Retrieval-Augmented Generation” 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

Paul J. L. Ammann, Jonas Golde, Alan Akbik

Researchers at Humboldt-Universität zu Berlin introduced an enhancement for Retrieval-Augmented Generation (RAG) that decomposes complex queries into sub-queries using a zero-shot LLM and then refines retrieved passages with a cross-encoder reranker. This combined strategy improved retrieval recall and answer accuracy, achieving an 16.5% higher Hits@10 on MultiHop-RAG and an F1 score of 35.0 on HotpotQA for answer generation.

Key insights

  • Question decomposition effectively broadens retrieval coverage for multi-hop queries, while reranking precisely filters this expanded pool, demonstrating a synergistic relationship between the two components.
  • The use of off-the-shelf, pre-trained LLMs and cross-encoders enables a practical, zero-shot 'drop-in' enhancement for existing RAG pipelines, avoiding specialized training or architectural changes.
  • Achieving substantial performance gains for complex queries through decomposition incurs significant computational overhead, primarily from LLM inference, highlighting a trade-off between quality and latency.

Summary by alphaXiv.

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