[D] TurboQuant author replies on OpenReview

A technical controversy has erupted in the AI quantization community, centered on the paper 'TurboQuant.' The authors have posted a detailed response on OpenReview to address claims that their core method was derived from the earlier RaBitQ technique. They firmly state that random rotation is a standard, ubiquitous technique in quantization literature, citing several pre-existing works, and that TurboQuant's novelty lies specifically in their derivation of the exact statistical distribution followed by the coordinates of rotated vectors. This derivation, they argue, enables optimal coordinate-wise quantization and constitutes their primary contribution.

The response tackles several points of contention. First, they address the characterization of RaBitQ as 'suboptimal,' explaining that while the main theorem's formal statement lacked an explicit optimality guarantee, a deeper investigation of its appendix revealed a strict bound. Consequently, they are updating the TurboQuant manuscript to accurately credit RaBitQ's bounds. Second, they defend their experimental focus, arguing that runtime benchmarks are 'immaterial' to their findings, which are centered on the compression-quality tradeoff at extreme compression levels. Finally, they note the paper has been on arXiv since April 2025 and that prior communication with RaBitQ authors occurred, questioning the timing of the raised concerns.