Seq-Drifting for One-Step Parallel Text Generation

Seq-Drifting is a framework that enables diffusion language models to generate entire sequences in a single forward pass. It uses a drift field to guide continuous token representations from Gaussian noise toward contextually plausible targets while using repulsion to prevent sequence repetition.