A new paper lays out the math for generating discrete data - things like categorical labels or symbolic sequences, not pixels - in one step instead of many.
The authors build what they call Discrete Wasserstein Flows: a gradient flow, grounded in discrete optimal transport, that moves probability mass across the transitions of a reversible Markov kernel. During training, they simulate that flow at the particle level using Markov jumps, then compress the resulting transport updates into a latent-conditioned neural generator. The payoff is that the iterative part only happens during training - at inference, the generator produces a sample in a single step. To check the math actually works, they ran it in a controlled setting where the correct transport dynamics can be computed exactly, and confirmed the predicted KL divergence drops and numerical scaling held up.
Most generative models for discrete data - text tokens, molecular graphs, other symbolic structures - lean on many-step diffusion or autoregressive sampling, both of which are slow at inference. This is the same bet continuous-domain diffusion models made a few years ago, when one-step distillation methods arrived to speed up image generation. If it holds beyond toy problems, it points to faster discrete generators without giving up the convergence guarantees diffusion-style training provides.
The catch: this was verified only on small, exactly-solvable toy problems designed to check the authors' own math, not on real text or molecule datasets - a reasonable first step, but not yet evidence the approach scales beyond proof of concept.