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

Researchers introduced Trajectory-Shaped Discrete Flow Matching (TS-DFM), a method that guides discrete flow matching training with an energy compass to improve text generation quality and speed. TS-DFM achieves 32% lower perplexity than a 1,024-step teacher model using only 8 steps, making it 128 times faster. This approach outperforms other discrete-generation baselines even with smaller models and less training data.