ADORE iteratively expands queries to boost retrieval by up to 122% over BM25
Fixes retrieval drift by feeding real corpus responses back into the LLM for smarter query expansion.
Traditional LLM-based query expansion suffers from retrieval drift—generating plausible but irrelevant terms that degrade results. ADORE (ADapt, Observe, Relevance Evaluate) introduces a closed-loop system: at each round, an LLM creates pseudo-passages, a retriever fetches actual corpus documents, and a relevance assessor scores them against the original query. This feedback pinpoints which expansions to reinforce, which to suppress, and which gaps remain. The process iterates until convergence.
Tested across TREC Deep Learning, BEIR, and BRIGHT benchmarks, ADORE consistently outperforms both BM25 and strong prior query expansion methods. On BEIR, it achieves a 24.5% average nDCG@10 improvement over BM25 and 3.6% over the best previous method. On BRIGHT, gains jump to 122.9% over BM25 and 9.2% over the best baseline. The code and data are publicly available, making ADORE immediately applicable for production retrieval pipelines.
- ADORE closes the expansion loop: LLM generates pseudo-passages → retriever fetches real docs → assessor scores relevance → feedback refines next expansion.
- On BEIR, ADORE improves nDCG@10 by 24.5% over BM25 and 3.6% over the strongest prior query expansion method.
- On BRIGHT, improvements reach 122.9% over BM25 and 9.2% over the best baseline, showing dramatic gains in domain-specific retrieval.
Why It Matters
Makes RAG and search systems far more accurate by grounding LLM-generated expansions in actual corpus feedback.