AdaWidth Adapts Embedding Dimensions for Efficient Dense Retrieval

_reachsumit · x · 2026-08-26

The paper proposes AdaWidth, a query-adaptive embedding width method for dense retrieval. Addressing the variance in dimensions needed for different queries, AdaWidth uses an orthogonal prefix adapter to concentrate discriminative signals in leading coordinates, while a lightweight router decides whether to evaluate more dimensions based on current rankings. Analysis shows required dimensions grow logarithmically with corpus size. Across six tasks, AdaWidth matches state-of-the-art NDCG@10 using 55% to 84% fewer dimensions.

Original post →

More from Research

Research channel →