Concerns Raised About Accuracy of Google's TurboQuant Paper

The author, a researcher at ETH Zurich, claims that Google's TurboQuant paper accepted at ICLR 2026 has serious problems in its description and comparison of the prior RaBitQ vector quantization method, which the author developed during their PhD.

Details

The author, Jianyang Gao, is a postdoctoral researcher at ETH Zurich and the first author of the RaBitQ line of work, which was published in 2024. RaBitQ proposed a high-dimensional vector quantization method and proved it achieves asymptotically optimal error bounds. One of the key ideas in RaBitQ is to apply a random rotation to the input vector before quantization, similar to a Johnson-Lindenstrauss transform. The author claims that TurboQuant, a paper accepted to ICLR 2026 from Google Research, has serious problems in how it describes and compares the prior RaBitQ method. Specifically, TurboQuant systematically avoids acknowledging the methodological similarity between the two approaches, even though the TurboQuant authors had detailed understanding of RaBitQ's technical details beforehand. The author argues that this omission is problematic, as RaBitQ was the first work to combine random rotations with vector quantization and prove optimal theoretical guarantees under the same problem setting. Despite reviewer feedback and the author's direct requests, the TurboQuant authors refused to add a fuller discussion and comparison of RaBitQ in the final paper.