Reproduce “RSLM: Training-Free Vector Quantization for Approximate Nearest Neighbor Search” 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

Rastislav Lenhardt, Teodora Dobos, Thomas Vecchiato, Jiri Isa, Igor Ginzburg

RSLM (Rotated Scaled Lloyd-Max) introduces a family of training-free vector quantization codecs for Approximate Nearest Neighbor (ANN) search, designed to reduce memory and bandwidth while maintaining high recall. These codecs achieve comparable or superior recall to state-of-the-art trained methods, enabling up to 8x DRAM footprint reduction, through efficient block-wise rotations and explicit norm correction for the final reconstructed vector.

Key insights

  • Training-free codecs, when designed with efficient rotations and norm correction, can achieve performance competitive with or superior to complex, data-dependent trained quantization methods.
  • Correcting the L2 norm of the *final reconstructed vector* significantly improves MIPS ranking accuracy, outperforming methods that only correct residual norms.
  • High-dimensional embeddings inherently offer greater resilience to quantization noise, enabling more aggressive compression rates (1-2 bits per dimension) while maintaining high recall.

Summary by alphaXiv.

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