Reproduce “Revisiting RaBitQ and TurboQuant: A Symmetric Comparison of Methods, Theory, and Experiments” 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

Jianyang Gao, Yutong Gou, Yuexuan Xu, Jifan Shi, Yongyi Yang, Shuolin Li, Raymond Chi-Wing Wong, Cheng Long

This study offers a symmetric, reproducible comparison of RaBitQ and TurboQuant, two high-dimensional vector quantization methods, clarifying their theoretical properties and empirical performance across tasks like nearest neighbor search and KV cache management. The evaluation demonstrates RaBitQ's competitive or superior performance, particularly at lower bit widths, while addressing inconsistencies in prior claims regarding TurboQuant.

Key insights

  • RaBitQ's theoretical guarantees for inner-product estimation align with optimal scaling (log log(1/δ)), while TurboQuant's yield an exponentially worse rate (log(1/δ)).
  • Prior reported performance claims for TurboQuant were often based on asymmetric comparisons (e.g., CPU vs. GPU) or were not reproducible under stated hardware and software configurations.
  • Both methods are unbiased inner-product estimators, but RaBitQ generally offers more stable and accurate estimates, along with superior efficiency on identical hardware.

Summary by alphaXiv.

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