Watermark audits can be erased at inference: Logit Laundering lifts detection p-values 513-fold on average while keeping 56.8% of the copyright-knowledge gain
Synopsis
Data-use watermark auditing relies on statistical signals detectable in model outputs, and a non-compliant model provider can weaken those signals by rewriting the output distribution at inference time. Logit Laundering subtracts an estimated watermark-related logit bias from the base model's logits without changing model parameters, and scales the suppression by next-token prediction confidence. Across five foundation models and several mainstream auditing methods, detection p-values rise 513-fold on average and the signals become statistically insignificant, while a 56.8% average relative gain in copyrighted knowledge over the original foundation models is retained.
Interpretation
The adversarial side of data-use auditing is now on the table: a provider need not touch training data, only rewrite the output distribution at inference, to leave auditors without evidence. Prior auditing work mostly designed stronger watermark embedding and detection, with limited systematic study of providers actively evading; directly perturbing training text distorts the text and undercuts its training value. Across five foundation models (LLAMA-3.2-1B-Instruct, LLAMA-3.2-3B-Instruct, Qwen-2.5-1.5B-Instruct, Gemma-2-2B-IT, Stable-LM-2-1.6B-Chat), detection p-values rise 513-fold on average against both STAMP-like and radioactive detection, leaving the signals statistically insignificant.
The evasion happens at decoding: the model first continues pretraining on watermarked copyrighted corpora as usual, then subtracts an estimated bias signal from its output logits. The statistical bias estimator reuses the base model's first eight Transformer layers and language-model head, fitting distributional bias on local segments of the copyrighted corpus while being regularized toward a near-uniform distribution on public corpora. The experiments build a copyrighted corpus from 50 books published in 2025, 4,959 paragraphs in total, mix it with public Pile corpora for continual pretraining, and then reshape logits at inference.
Suppression strength is not fixed: each decoding step uses the top-1 versus top-2 logit margin as a confidence signal, suppressing harder when confidence is high and more conservatively when it is low. A fixed-strength logit perturbation can disturb the candidate-token structure at positions where the model is uncertain, producing unnatural text. Removing confidence scaling and using a fixed coefficient drops MAUVE from 0.107 under the adaptive setting to 0.018 on Llama-3.2-1B, making generated text clearly unnatural.
Perspective
Evaluation is confined to a continual-pretraining setting: five 1B-to-3B foundation models are updated for one epoch on a mix of public corpora and watermarked copyrighted book corpora, then tested against STAMP-like and radioactive watermark auditing methods and their variants. The authors also propose an exploratory post-detection check, True-key Placebo Check, and run a small pilot on 10 paragraphs from one book.
Evasion performance varies with watermark strength, estimator size, segment ratio and JS budget; under stronger watermarks the radioactive-detection margin narrows slightly, though p-values stay above 0.05. True-key Placebo Check was piloted on only 10 paragraphs from one book, and the authors frame it as an exploratory false-negative indicator rather than a mature defense.
