Title: Linguistics-Aware Non-Distortionary LLM Watermarking

Abstract:

Effective watermarking must enable the identification of language model outputs without compromising content quality or restricting verification capabilities to the original model provider. However, deploying models across multiple languages complicates this task, as variations in morphology, segmentation, and script alter the natural points where watermark signatures can be embedded. To address these challenges, we present LUNA, a linguistically adaptive watermarking framework that integrates model-free detection with single-token non-distortion, operating within the standard random-key paradigm.

LUNA determines the depth of a non-distortionary binary tournament sampler by estimating normalized next-tag entropy derived from part-of-speech contexts within an external corpus. The detection process mirrors this approach, reconstructing the embedding schedule using a tokenizer, a part-of-speech tagger, the secret key, and the generated text. We benchmarked LUNA against eight primary baselines across six typologically diverse languages and two distinct domains.

The results demonstrate that LUNA achieves an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.9959 and maintains the lowest mean absolute median perplexity shift of 0.045 across all twelve experimental settings. Furthermore, its 95% bootstrap interval for perplexity shift [0.022, 0.073] falls below the intervals recorded by all other baselines. LUNA also recorded the smallest mean shifts in Self-BLEU, Distinct-1, surprisal, and entropy metrics. Notably, LUNA is the sole method to simultaneously achieve an AUROC exceeding 0.99 and an absolute median perplexity shift under 0.1 in a majority of cases, meeting these criteria in 9 out of 12 settings, whereas no baseline surpassed this threshold in more than two settings. The source code for this work is publicly available at: https://github.com/Shinwoo-Park/luna_watermark

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC