ForkLeft: Entropy-First Rollouts for Prefix-Aligned Autoregressive-to-Diffusion Distillation
ForkLeft, a distillation framework that resolves a fundamental mismatch between autoregressive teacher predicts from a left prefix and NTP teacher under the same context, is introduced, showing that DLMs can learn NTP-style reasoning without sacrificing native parallel generation.
Jun-Ming Liu, Ji-Cheng Wang, Yifeng He et al.
· 0 citations