A proposed approach to one-step language generation without a teacher model

alphaXiv describes “Discrete Beckmann Transport Models for One-Step Language Modeling and Reasoning” as an approach that maps noise directly to discrete tokens, generating a whole sequence in one step without a teacher model or tracking diffusion time. Its summary contrasts this with few-step diffusion and flow language models, which it says usually require distilling a pretrained teacher, adding training cost and limiting student quality to the teacher’s. With extra steps, alphaXiv says, the model revisits uncertain tokens using tokens it is already confident about.