New watermarking method cuts LLM detection overhead to constant time
Researchers propose Stateless Bernoulli Watermarking, a computationally efficient alternative to existing LLM watermarking schemes that cuts membership detection from O(n) to O(1) per token. The method uses independent random trials rather than vocabulary permutations, enabling inference-time watermarking without intermediate memory allocations and unlocking full-vocabulary self-salting capabilities over 6000 times faster than prior approaches. This addresses a critical bottleneck in deploying watermarks at scale, making provenance tracking more practical for production systems while maintaining statistical rigor under null hypothesis testing.
Modelwire context
ExplainerThe paper's core contribution is architectural, not just algorithmic: by replacing vocabulary permutations with independent Bernoulli trials, the method eliminates the memory allocation overhead that made prior watermarking schemes impractical at inference time. This is a systems-level insight, not a theoretical one.
This work sits adjacent to the privacy and evaluation concerns surfaced in recent coverage. The membership inference attack paper from September 1st exposed token-level memorization vulnerabilities in diffusion models, creating urgency around provenance tracking. Watermarking addresses the inverse problem: not detecting what a model memorized, but proving what it generated. The speed gains here matter because prior watermarking schemes were too expensive to deploy alongside real inference pipelines, leaving that vulnerability window open. If watermarking becomes cheap enough to run by default, it shifts the cost-benefit calculation for production systems.
If major inference providers (Anthropic, Together, or similar) announce watermarking as a standard output option within the next six months, that signals the 6000x speedup crossed a practical deployment threshold. If they don't, the method remains a research artifact despite its efficiency gains.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsStateless Bernoulli Watermarking · KGW · SynthID
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “Flip, Don't Shuffle: Watermarking LLMs at the Speed of Inference”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.