Time-Agnostic Predictors Compete with Time-Conditioned Discrete Diffusion

Discrete diffusion models for language may not require explicit time conditioning in finite-data settings. In practice, time-agnostic predictors remain competitive with or outperform time-conditioned models by exploiting the fact that corrupted sequences stay close to clean sequences during training.