T3D: Few-Step Diffusion Language Models via Trajectory Self-Distillation with Direct Discriminative Optimization
This work addresses the inference speed bottleneck for diffusion language models, offering a practical improvement for applications requiring fast text generation, though it is incremental as full-step decoding remains superior.
The paper tackles the problem of diffusion large language models requiring many refinement steps for text generation, which limits inference efficiency, and proposes a trajectory self-distillation framework with Direct Discriminative Optimization to improve few-step decoding, narrowing the gap to full-step performance across benchmarks.
Diffusion large language models (DLLMs) have the potential to enable fast text generation by decoding multiple tokens in parallel. However, in practice, their inference efficiency is constrained by the need for many refinement steps, while aggressively reducing the number of steps leads to a substantial degradation in generation quality. To alleviate this, we propose a trajectory self-distillation framework that improves few-step decoding by distilling the model's own generative trajectories. We incorporate Direct Discriminative Optimization (DDO), a reverse-KL objective that promotes mode-seeking distillation and encourages the student to concentrate on high-probability teacher modes. Across benchmarks, our approach consistently outperforms strong few-step baselines and standard training under tight step budgets. Although full-step decoding remains superior, we substantially narrow the gap, establishing a strong foundation towards practical few-step DLLMs. The source code is available at https://github.com/Tyrion58/T3D.