dgMARK: Decoding-Guided Watermarking for Diffusion Language Models
This addresses the need for detectable AI-generated text in dLLMs, offering a plug-and-play solution for watermarking in a model type with unique generation properties.
The paper tackles the problem of watermarking discrete diffusion language models (dLLMs) by exploiting their sensitivity to token unmasking order, proposing dgMARK to embed watermarks via parity constraints without altering model probabilities, achieving robustness against post-editing operations like insertion and deletion.
We propose dgMARK, a decoding-guided watermarking method for discrete diffusion language models (dLLMs). Unlike autoregressive models, dLLMs can generate tokens in arbitrary order. While an ideal conditional predictor would be invariant to this order, practical dLLMs exhibit strong sensitivity to the unmasking order, creating a new channel for watermarking. dgMARK steers the unmasking order toward positions whose high-reward candidate tokens satisfy a simple parity constraint induced by a binary hash, without explicitly reweighting the model's learned probabilities. The method is plug-and-play with common decoding strategies (e.g., confidence, entropy, and margin-based ordering) and can be strengthened with a one-step lookahead variant. Watermarks are detected via elevated parity-matching statistics, and a sliding-window detector ensures robustness under post-editing operations including insertion, deletion, substitution, and paraphrasing.