generalfactneutralStandard factorized training in discrete flow matching exposes models to targets of varying difficulty, mixing well-conditioned tokens with high-entropy onesMachine Learning29 Jul 2026http://arxiv.org/abs/2607.21427v1