Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillation

Introduces Trajectory-Shaped Discrete Flow Matching (TS-DFM) which replaces blind stochastic jumps in discrete flow matching with guided navigation using a lightweight energy compass to evaluate candidate continuations. This training-only shaping improves student model performance by ensuring better quality trajectories, leading to significant perplexity reduction and speedup (128x faster at 8 steps vs 1024 steps teacher) on a 170M-parameter language model.