Reproducibility artifact for the CIKM '26 paper: one query-agnostic preprocessing step takes HNSW from 21.1% to 95.8% recall@10 on attention-derived inner-product workloads, with no index changes.
reproducibility approximate-nearest-neighbor-search nsg hnsw vector-search out-of-distribution maximum-inner-product-search cikm2026 flatnav roargraph
-
Updated
Aug 19, 2026 - Python