Reproduce “IterCQR: Iterative Conversational Query Reformulation with Retrieval Guidance” 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
Yunah Jang, Kang-il Lee, Hyunkyung Bae, Hwanhee Lee, Kyomin Jung
IterCQR introduces an iterative, self-improving framework for conversational query reformulation, directly optimizing queries with information retrieval signals. It achieves state-of-the-art retrieval performance on TopiOCQA and QReCC datasets for both dense and sparse retrievers, notably without relying on costly human-annotated rewrites, and demonstrates strong generalization in low-resource settings.
Key insights
- Direct optimization with information retrieval signals, rather than human judgment or separate models, effectively generates retriever-friendly queries.
- An iterative, self-improving loop can overcome the limitations of initial LLM-generated data, progressively refining queries for better retrieval performance.
- The combination of MBR training (exploration) and top-1 candidate selection (exploitation) is crucial for both exploring diverse query candidates and stably optimizing for high-quality rewrites.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.