Reproduce “Dynamic Superblock Pruning for Fast Learned Sparse Retrieval” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Parker Carlson, Wentai Xie, Shanxiu He, Tao Yang

The University of California, Santa Barbara, developed Dynamic Superblock Pruning (SP), a hierarchical scheme for learned sparse retrieval that significantly accelerates query processing on CPU-only servers. Incorporating CPU cache locality optimizations, this method offers up to 2.9x faster retrieval than prior block-based approaches while preserving high relevance, demonstrating a 32% speedup for rank-safe k=10 on SPLADE.

Key insights

  • Introducing a hierarchical pruning layer with superblocks enables more aggressive and earlier skipping of irrelevant index segments, effectively reducing computational overhead.
  • Explicitly optimizing for CPU cache locality during score bound computations, such as with a "Superblock-at-a-time" processing order, yields substantial real-world performance gains.
  • Significant retrieval speedups for learned sparse models are achievable while maintaining high relevance, offering controllable trade-offs through probabilistic safeness parameters.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.