Does the lack of frontier expansion with increased NFE persist at larger scales?

Ascertain whether the observed absence of entropy–perplexity frontier expansion as the number of function evaluations (NFE) increases for discrete diffusion language models is merely an artifact of the model scale used in the experiments, or whether this behavior persists for larger diffusion models.

Background

In extended OWT experiments, the authors observe that increasing NFE does not significantly expand the entropy–perplexity frontier; temperature tuning appears to recover most benefits that additional steps would otherwise provide.

They explicitly state uncertainty about whether this phenomenon stems from the relatively small scale of their models or represents a general property that would also hold at larger scales.