Reproduce “ThinkQE: Query Expansion via an Evolving Thinking Process” on OpenResearch

About this paper

Yibin Lei, Tao Shen, Andrew Yates

ThinkQE, a test-time query expansion framework, integrates an LLM's explicit thinking process with an iterative corpus interaction strategy to enhance query expansion diversity and exploration. The approach outperforms prior zero-shot methods and competes with supervised dense retrievers on diverse web search benchmarks without requiring additional model training.

Key insights

• An explicit LLM-guided thinking process is crucial for generating higher-quality, more exploratory query expansions, outperforming direct generation or passive use of LLMs.
• Iterative query refinement through dynamic corpus feedback, termed "corpus-interaction scaling," is more effective than generating static, isolated expansions.
• Both expansion accumulation and semantic redundancy filtering are essential for promoting diversity in retrieved evidence and preventing query dilution across iterative rounds.
• Strategic use of thinking processes and multiple interaction rounds can yield greater performance gains than simply scaling up LLM model size, offering a more efficient path to improved retrieval effectiveness.

Summary by alphaXiv.