SAC-Copula: Quality-Preserving Watermarking for Diffusion Language Models via Smooth Correlated Gumbel Fields
The episode details SAC-Copula, a method for watermarking text generated by diffusion language models. It uses Gaussian copulas to inject smooth, correlated noise into masked positions. This process embeds a hidden signature that maintains high text quality while providing verifiable provenance against post-generation edits.