Google's TurboQuant KV Cache Paper Accused of Plagiarizing ETH Zurich's RaBitQ Algorithm

Forum topic · ✨步子哥 · 2026-04-27

Summary

A Google research paper called TurboQuant, which claims to compress large language model KV cache memory by at least 6x with up to 8x inference speedup and zero accuracy loss, has been accused of plagiarizing RaBitQ, an algorithm published two years earlier by a postdoctoral team led by Gao Jianyang at ETH Zurich. The accusers published open comments on the ICLR OpenReview platform and filed a formal complaint, citing three allegations: TurboQuant avoided citing RaBitQ despite reviewer warnings, mischaracterized RaBitQ's theoretically optimal error bounds as suboptimal without evidence, and ran unfair benchmarks comparing an unofficial single-core CPU implementation of RaBitQ against TurboQuant on an NVIDIA A100 GPU. The controversy triggered a market shock: memory chip stocks plunged, erasing over $90 billion in market capitalization, since mature KV cache compression would reduce DRAM and HBM demand. Google's first author reportedly promised corrections after the conference but declined to discuss core technical similarities. The episode has sparked broader debate on academic integrity, citation norms, standardized benchmarks, and third-party auditing in AI research.

A Google paper claiming a breakthrough in KV cache compression for large language models has become the center of a rare academic integrity storm — one that rippled from a research forum all the way to global stock markets.

Background: What Is KV Cache?

The KV (Key-Value) cache is a large language model's "short-term memory" during text generation. It stores key-value pairs of prior tokens to maintain context, but grows rapidly with conversation length, consuming massive GPU memory and driving up inference costs. TurboQuant targeted exactly this bottleneck.

The Claims of TurboQuant

Google's TurboQuant paper proposed a two-stage compression scheme:

1. Random rotation + PolarQuant: high-dimensional vectors are mapped into polar coordinates for extreme compression. 2. Quantized Johnson-Lindenstrauss (QJL) transform: uses only 1 bit per dimension to correct bias in inner-product estimation.

The paper claimed at least 6x memory reduction, up to 8x inference speedup, and zero accuracy loss for KV cache compression.

The Plagiarism Allegations

A postdoctoral team led by Gao Jianyang at ETH Zurich publicly alleged that TurboQuant closely mirrors RaBitQ, their algorithm published two years earlier. They posted comments on the ICLR OpenReview platform and filed a formal complaint with the conference, citing three specific issues:

Avoidance of prior work: TurboQuant failed to adequately discuss its relationship to RaBitQ, even after reviewers explicitly flagged the similarity.
Mischaracterization of theory: TurboQuant labeled RaBitQ's error bounds as "suboptimal" without evidence, while RaBitQ's extended paper — presented at a top theoretical computer science venue — had already proven asymptotically optimal error bounds.
Unfair benchmarking: TurboQuant benchmarked RaBitQ using an unofficial Python implementation restricted to a single CPU core with multithreading disabled, while running its own method on a full NVIDIA A100 GPU — inflating the apparent performance gap.

Google's first author reportedly promised to revise the paper after the conference, but avoided addressing the core question of technical similarity.

Market Shockwave

If TurboQuant worked as claimed, mature KV cache compression would reduce demand for server memory chips (DRAM/HBM). As the controversy broke, memory-related stocks plunged, with over $90 billion in market capitalization wiped out across multiple companies — a stark illustration of how an algorithm dispute can shake capital markets.

Broader Implications

Tech giants' AI papers carry outsized weight; alleged misconduct threatens not just reputations but the incentive structure for original research.
The incident has renewed calls for mandatory citation of related work, standardized experimental benchmarks, and third-party auditing in AI research.
The RaBitQ team's transparent, platform-public complaint is being held up as a model for defending original work.

References

1. 世间精彩. 谷歌AI论文涉嫌学术造假，TurboQuant压缩算法遭质疑，内存股巨震. 2026-03-30. 2. Google TurboQuant Team. TurboQuant: Efficient KV Cache Compression for LLMs. ICLR Submission, 2026. 3. Gao Jianyang et al. RaBitQ: Random Bit Quantization for Vector Search. Extended Version, Theoretical Computer Science Conference, 2024. 4. ETH Zurich Postdoc Team. OpenReview Comment on TurboQuant Similarity to RaBitQ. ICLR Platform, 2026. 5. AI Market Analysis Group. Impact of LLM Optimization on Memory Chip Demand. Industry Report, 2026.