dgMARK: Decoding-Guided Watermarking for Diffusion Language Models

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. Project website: https://dgmark-watermarking.github.io

AI systems can now write text that looks human, making it harder to tell whether a passage came from a person or a machine. This matters because AI-generated text can be misused for misinformation, plagiarism, or other hard-to-trace content. A common defense is to add a hidden “watermark” to AI-generated text, so it can later be detected. Most watermarking methods are designed for older language models that write from left to right. Newer diffusion language models work differently: they begin with blanks and gradually fill them in. We introduce dgMARK, a watermarking method for these models. Instead of directly changing which words the model thinks are likely, dgMARK changes the order in which the model fills in the blanks, following a hidden rule. Across a full passage, this creates a pattern that people usually cannot notice but that can be detected statistically. Experiments show that dgMARK keeps the text natural while the watermark remains detectable even after editing or paraphrasing. This suggests that generation order itself can help identify AI-written text from newer language models and make their use more accountable.