ADORE: Iterative Query Expansion with Retrieval-Grounded Relevance Feedback

A new iterative framework called ADORE improves information retrieval by grounding query expansion in actual retrieval outcomes rather than relying solely on language-model generation, according to a paper submitted June 11, 2026 [1]. The framework, whose name stands for ADapt, Observe, Relevance Evaluate, operates in three steps per round: an LLM generates pseudo-passages conditioned on the original query and prior feedback, a retriever exposes how the target corpus responds, and a relevance assessor evaluates the retrieved documents against the original query [3]. Those judgments are converted into structured feedback that identifies what to reinforce, what remains undercovered, and what to suppress [3]. By anchoring all relevance assessments to the original query, ADORE aims to prevent the retrieval drift that can occur when expansions are generated without checking corpus behavior [3]. Most existing LLM-based query expansion methods remain generation-driven, producing plausible pseudo-documents or expansions without verifying how the target corpus responds [1]. That approach can amplify misleading vocabulary or miss terms that distinguish relevant from non-relevant documents [2]. Earlier work relied on pseudo-relevance feedback, which extracts terms from initially retrieved documents, but those documents can be noisy and undermine ranking quality [6]. A separate recent study proposed ProQE, a progressive query expansion algorithm that combines classic pseudo-relevance feedback with LLMs and uses an LLM as a relevance judge for each returned result [6]. ADORE differs by separating generation from evaluation explicitly, turning retrieval outcomes into structured feedback rather than feeding raw passages back to the generator and expecting it to infer what matters [4]. On the BEIR benchmark, ADORE improved average nDCG@10 by 24.5% over BM25 and by 3.6% over the strongest prior query expansion method [1]. On the reasoning-intensive BRIGHT benchmark, the gains were larger: 122.9% over BM25 and 9.2% over the best query expansion baseline [5]. The framework was also evaluated on TREC Deep Learning and consistently outperformed strong baselines across nearly all settings [1]. The authors have released the code and data publicly [2].