CForce: Boosting Parallel Decoding for dLLMs via Consistency Forcing
This work presents Consistency Forcing (CForce) for dLLMs, a distillation method to force the mask predictions of early stages to align with those of later stages, thereby improving training-inference alignment.