Language models can conceal watermarked training data by altering outputs

Hu and colleagues report that changing a model’s next-word scores can weaken evidence of watermarked training data while preserving downstream task performance.

Academic

Ruihan Hu · Wei Luo · Yu-Ming Shang · Jiakai Wang · Chuxuan Zhang · Mei Li · +2 more

Beijing University of Posts and Telecommunications · Zhongguancun Laboratory · Augusta University

Research Digest·9 November 2026·2 min read

Illustration: Language models can conceal watermarked training data by altering outputs

In one line

Logit Laundering reshapes inference-time logits to remove detectable watermark signals, enabling evasion of data-use auditing without model retraining.

Before you read this

Three ideas this paper builds on

large language modelA neural network trained on massive text corpora to generate coherent text and perform diverse language tasks.290 papers →audit workflowA systematic process for verifying data integrity, model behavior, and evaluation validity through checks for errors, confounds, and hidden artifacts.104 papers →activation steeringActivation steering modifies a model's internal vector representations during inference to shift behavior, such as editing outputs, removing capabilities, or controlling trust.66 papers →

Read it as

Auditors use invisible statistical watermarks to detect if a (LLM) was trained on copyrighted data. A hidden signal in the model's output reveals the unauthorized use.

Logit Laundering is a method that suppresses this signal at inference time, the moment the model generates a reply, by subtly reshaping the model's output probabilities. It works without altering the model's weights or its useful learned knowledge.

In extensive tests across several LLMs, Logit Laundering made the watermarks undetectable, showing the audit evidence was statistically insignificant. This highlights a critical weakness in current copyright enforcement: providers can use the data and then easily hide the trace.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

data watermarking large language model audit workflow activation steering confidence calibration continual learning