Google Research’s TurboQuant faces scrutiny after ETH Zurich postdoc alleges misleading comparisons to RaBitQ

Allegations from ETH Zurich researcher

Google Research on March 25 unveiled TurboQuant, an extreme compression algorithm the company says can cut AI memory use to one-sixth and boost inference speeds eightfold by addressing large-model KV‑cache bottlenecks. But the paper is now under scrutiny after ETH Zurich postdoctoral researcher Gao Jianyang (高健扬), first author of the RaBitQ algorithm, publicly accused the TurboQuant authors of “serious problems” in how RaBitQ is described — including incorrect technical statements and misleading theoretical and experimental comparisons. Gao says these concerns were raised with the Google team before submission and were allegedly acknowledged but not corrected.

The contested experiments

Among the specific complaints: the experimental setup appears mismatched. Gao has said that RaBitQ results in the paper were produced on a single‑core CPU while TurboQuant was benchmarked on an NVIDIA A100 GPU — an apples‑to‑oranges comparison if true. The TurboQuant paper has been accepted to ICLR 2026 and, it has been reported, received heavy promotion by Google with view counts in the tens of millions, meaning any uncorrected claims reach a wide audience and risk becoming treated as consensus.

Why this matters beyond academia

This is more than a lab feud. Efficiency gains for large models are strategically important as countries and companies vie for AI leadership and as export controls on high‑end chips change what hardware is accessible in different markets. A method that materially reduces memory and compute needs could reshape where and how large models are deployed. At the same time, reproducibility and fair benchmarking are essential to maintain trust in papers that inform product and policy decisions. Who polices claims when a paper goes viral?

Next steps and community reaction

It has been reported that Google has not publicly replied to Gao’s detailed critique at the time of writing. The ICLR acceptance remains in place, and the research community will likely call for clarifications, replication of experiments, and, if warranted, corrections or an erratum. For now, the dispute underscores how quickly high‑impact claims can spread — and how fragile the gatekeeping around experimental fairness can be when research crosses into mass publicity.