●builderThis enables more efficient training and fine-tuning of generative models that require global structural constraints.
●researcherYou can implement complex sequence-level objectives in discrete diffusion without the exponential cost of token enumeration.