ADORE: Iterative Query Expansion with Retrieval-Grounded Relevance Feedback

ADORE (ADapt, Observe, Relevance Evaluate) is an iterative framework for query expansion that utilizes retrieval-grounded feedback to enhance the relevance of generated queries. By combining LLM-generated pseudo-passages, corpus responses, and relevance assessments, ADORE mitigates issues like retrieval drift and misleading vocabulary. It demonstrates significant performance improvements, achieving a 24.5% increase in average nDCG@10 over BM25 on the TREC Deep Learning benchmark and a 122.9% improvement on the BRIGHT dataset, indicating its effectiveness for practitioners focused on optimizing retrieval systems.