Rethinking Multiple-Query Optimization for Approximate Nearest Neighbor Search

ICLR 2026 Conference Submission19834 Authors

Keywords: approximate nearest neighbor search, multiple-query optimization

TL;DR: We present a multiple-query optimization framework tailored to approximate nearest neighbor search.

Abstract: Approximate nearest neighbor search (ANNS) over vector databases is a fundamental operation for many modern applications, where rapid processing of queries is of critical importance. In traditional database systems, which face the same requirement, multiple-query optimization (MQO) has been extensively studied to address this challenge. Although MQO is a general technique that exploits shared computation to process a set of queries more efficiently than evaluating each query in isolation, no analogous algorithmic strategy has yet been proposed for ANNS. To this end, we present a novel algorithmic MQO framework tailored to ANNS. The framework is universally applicable to graph-based ANNS methods, delivering significant speedups while keeping both the underlying index and the search algorithm intact. Specifically, we construct a minimum spanning tree over the query vectors and initialize each query’s search entry using the nearest neighbor returned by its parent in the tree, thereby revealing and exploiting opportunities for computation reuse. We empirically validate our framework across multiple ANNS methods and datasets, demonstrating its feasibility and effectiveness.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 19834