Google's TurboQuant Paper Accused of Plagiarizing ETH Zurich's RaBitQ Algorithm, Memory Stocks Plunge $90 Billion

Forum topic · ✨步子哥 · 2026-04-27

Summary

A Google paper named TurboQuant, which claims to compress KV cache memory in large language models by at least 6x with up to 8x speedup and zero accuracy loss, has been accused of plagiarizing RaBitQ, an algorithm published two years earlier by a team led by ETH Zurich postdoctoral researcher Gao Jianyang. The accusers cite three main allegations: TurboQuant avoided citing RaBitQ despite reviewer comments, mischaracterized RaBitQ's theoretically optimal error bounds as 'suboptimal' without evidence, and ran unfair benchmark comparisons using an unofficial single-core CPU implementation of RaBitQ against TurboQuant on an NVIDIA A100 GPU. The ETH team filed a public comment on ICLR OpenReview and a formal complaint; the Google first author reportedly promised revisions but declined to discuss technical similarity. The controversy triggered a market shock, with memory chip stocks tumbling and over $90 billion in market value erased, as investors reconsidered demand for DRAM and HBM if cache compression technologies mature. The incident has reignited debate over academic integrity, citation norms, and standardized benchmarking in AI research as tech giants race for efficiency breakthroughs.

Overview

A Google research paper titled TurboQuant: Efficient KV Cache Compression for LLMs has become the center of a rare academic integrity storm in the AI community. The paper claims dramatic efficiency gains for large language model inference, but stands accused of copying RaBitQ, an algorithm published two years earlier by Gao Jianyang's postdoctoral team at ETH Zurich. Beyond academia, the controversy rattled financial markets: multiple memory chip companies saw their stock prices plummet, wiping out more than $90 billion in market value.

Background: What Is KV Cache?

> KV Cache (Key-Value Cache): The "short-term memory" of large language models. When generating text, models store key-value pairs of prior tokens to maintain context coherence. Without compression, this memory grows exponentially with conversation length, consuming massive GPU memory and inflating inference costs.

TurboQuant targets exactly this bottleneck, promising:

At least 6x reduction in KV cache memory usage
Up to 8x speedup in inference
Zero accuracy loss

The Technical Claims

TurboQuant uses a two-stage compression approach:

1. Stage one: Random rotation combined with a PolarQuant mechanism maps high-dimensional vectors into polar coordinates for extreme compression. 2. Stage two: A Quantized Johnson-Lindenstrauss (QJL) transform uses only 1 bit to correct bias in inner product computations, preserving precision.

> Johnson-Lindenstrauss transform: A mathematical technique that projects high-dimensional points into lower dimensions while approximately preserving distances and inner products.

The Three Allegations

Gao Jianyang's team argues TurboQuant closely mirrors RaBitQ, which proposed random rotation (based on the Johnson-Lindenstrauss transform) before quantization two years earlier. Their specific accusations:

1. Avoidance of prior work: TurboQuant sidestepped direct discussion of RaBitQ, even after reviewers explicitly flagged the similarity, and failed to adequately address it in revisions. 2. Misrepresentation of theory: TurboQuant labeled RaBitQ's results "suboptimal" without evidence. In fact, RaBitQ's extended paper, published at a top theoretical computer science conference, proved its error bounds are asymptotically optimal. 3. Unfair benchmarking: TurboQuant benchmarked RaBitQ using an unofficial Python implementation restricted to a single CPU core with multithreading disabled, while testing itself on a fully-powered NVIDIA A100 GPU — distorting real-world performance comparisons.

Responses and Market Impact

The ETH Zurich team published a public comment on the ICLR OpenReview platform and filed a formal complaint with the conference organizers.
Google's first author reportedly promised to correct issues after the conference but avoided addressing the core technical similarity.
The market reacted sharply: KV cache compression at 6x would reduce demand for DRAM/HBM in AI servers. The resulting uncertainty triggered a memory stock rout exceeding $90 billion in erased market value.

Key points

Google's TurboQuant paper claims 6x KV cache compression, 8x speedup, zero accuracy loss for LLM inference
ETH Zurich's Gao Jianyang team accuses it of copying their two-year-old RaBitQ algorithm
Allegations cover citation avoidance, false "suboptimal" claims against a proven optimal error bound, and rigged benchmarks (CPU vs. A100 GPU)
Formal complaints filed on ICLR OpenReview; Google's first author promised post-conference revisions without addressing similarity
Memory chip stocks crashed, erasing over $90 billion in market value on compression-demand fears

Broader Implications

The incident raises hard questions for the AI industry: as tech giants race for efficiency breakthroughs, how do we enforce citation of prior work, standardized benchmarks, and independent auditing? The controversy illustrates that AI research is now inseparable from capital markets — a dispute over an algorithm's provenance can move hundreds of billions of dollars. Original contributions like RaBitQ, the episode suggests, deserve rigorous protection if the research ecosystem is to remain healthy.

---

References

1. 世间精彩. Google AI paper suspected of academic misconduct; TurboQuant compression algorithm questioned; memory stocks plunge. 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.