Discrete diffusion and flow models have promised an alternative to autoregressive generation, but shrinking them to one step has usually meant distilling a pretrained teacher and inheriting its ceiling. A new paper, Discrete Beckmann Transport Models for One-Step Language Modeling and Reasoning, proposes a different route: a time independent flow whose autonomous transport map provably carries any point in the ambient space to a fixed point on the vocabulary simplex, enabling one step sampling without a two stage teacher student pipeline.
The authors frame generation as Beckmann transport rather than time dependent interpolation, which removes the need to learn a schedule of intermediate distributions. Because the map is autonomous and fixed point seeking, training can directly optimize the one step objective instead of compressing many steps after the fact. The claim matters for shipping models, where distillation doubles training cost and caps student quality, and for reasoning where single step fidelity determines whether chain of thought can be collapsed or verified quickly.
Why one step still needs better optimization and memory
Two companion results help explain why this direction is timely. Same Flow, Different Paths: Variance Reduction in Flow Matching shows that not all paths that induce the same marginal distributions and marginal velocity field are equal for optimization. The paper analyzes the class G(p_t, v*_t) of paths that share the same flow matching objective and finds that path choice directly controls stochastic gradient variance, offering a principled lever to stabilize training for both continuous and discrete flows. If DBTM changes the map, this work suggests how to choose the path that makes that map learnable.
Register Tokens for Bounded-State Reasoning in Diffusion Language Models tackles the other bottleneck: what happens after you generate. Masked diffusion language models normally keep all prior text in context to continue reasoning across chunks. The paper asks whether a model can continue after that text is cleared using only a fixed size carried state, implemented as a small number of register tokens whose continuous hidden states are trained to carry reasoning. Together, the three papers point to a shift from scaling steps and context to designing transport, variance, and state that make single pass or bounded state generation reliable enough to deploy.