TurboQuant paper sparks RaBitQ dispute

This looks less like ordinary conference chatter and more like a real attribution-and-reproducibility fight. TurboQuant may still be a strong compression result, but the public conversation has shifted to whether the paper's framing gives a fair picture of novelty and speedups.

• –Google's official write-up positions TurboQuant as training-free KV-cache compression plus vector-search acceleration, with no fine-tuning required and strong benchmark results.
• –RaBitQ authors say the paper describes their method too narrowly, downplays the Johnson-Lindenstrauss/random-rotation connection, and moves key context into the appendix.
• –They also allege the RaBitQ baseline was run under weaker conditions than TurboQuant, which makes the performance comparison hard to read at face value.
• –If those complaints hold up, the damage is bigger than one paper: it chips away at trust in the benchmark narrative around fast LLM compression.
• –For developers, the practical move is to wait for code, exact setup details, and the final conference version before treating the headline gains as settled.