FlowLM turns diffusion models into fast, few-step text generators

FlowLM, introduced by Runzhe Zhang and colleagues, transforms any pre-trained diffusion language model into a flow matching model via lightweight fine-tuning. The key innovation is re-aligning the curved sampling trajectories typical of diffusion models into straight-line flows. This straightening allows the model to produce high-quality text in just a few sampling steps, matching or even exceeding the quality achieved by the original diffusion model after 2,000 steps. Remarkably, the fine-tuned FlowLM reaches performance saturation with only half as many training epochs as training from scratch, validating its efficiency.

The paper also identifies a more effective training objective for flow matching: directly predicting clean data instead of the conventional velocity field. This guides the sampling process consistently toward the true data distribution, further improving quality. Empirical results across multiple text generation benchmarks show that FlowLM delivers practical, high-quality generation with dramatically reduced inference cost. This approach makes diffusion-based language models viable for real-time applications where step count matters.