DESA cuts retrieval depth 37% while boosting nDCG by 3.8%

Most hybrid retrieval systems fuse dense and sparse rankings with a fixed top-L cutoff, which conflates how much each channel contributes with how much of each ranking is accessed. Chunran Zhang's new paper, "Dense Expands, Sparse Anchors: Channel-Asymmetric Query Expansion for Hybrid Retrieval," separates those effects by evaluating under complete-list fusion and reporting per-channel stopping depths. The proposed method, DESA, exploits LLM-generated reference passages in a channel-specific way: orthogonal residual expansion adds new semantic directions to the dense query representation, while score-product anchoring injects lexical cues into sparse retrieval without broadening the original query's lexical support.

On seven BEIR datasets, DESA improves nDCG@10 by 3.82% and Recall@20 by 2.38% over unexpanded queries, while reducing dense and sparse access depths by 36.90% and 36.56%. Notably, 63.31% of queries become shallower in both channels, meaning the system retrieves better results while scanning less data. However, the paper notes a caveat: depths increase with Contriever on the Touché-2020 dataset, suggesting the approach's effectiveness depends on the underlying retriever. The findings advocate for channel-specific integration of generated passages and for evaluating retrieval quality alongside access depth—a practical consideration for production search systems aiming to balance latency and accuracy.