New grid-based search method beats high-dim ANN benchmarks

A new paper presents scaling laws for multiprobe grid approximate nearest neighbor (ANN) search, revealing a previously unreported d-scaling crossover on the GloVe embedding family where grid search maintains an approximately constant dimensional scaling exponent while other methods degrade. The approach shows near-linear query scaling in dataset size and lower indexing cost than competing ANN methods, suggesting grid-based approaches may be competitive in rebuild-heavy or high-dimensional settings. Recent work has formalized self-attention as an ANN operation, so these scaling properties may guide cost analysis of efficient transformer architectures.