Researchers have developed a new method for watermarking diffusion language models called SAC-Copula. This method preserves the quality of generated text while making it detectable by using smooth, correlated Gumbel perturbation fields. The approach is designed to work with iterative parallel unmasking and has been tested on several datasets, showing improved performance over existing methods in terms of both detection and generation quality. --- Why it matters: This matters because watermarking diffusion language models is a crucial step towards ensuring the integrity and ownership of generated text. SAC-Copula's ability to balance quality and detectability makes it an attractive solution for researchers and developers working with these models. Source: https://arxiv.org/abs/2608.20839