FlowLM: Few-Step Language Modeling via Diffusion-to-Flow Adaptation
FlowLM, introduced in arXiv paper 2605.20199, adapts pre-trained diffusion language models to flow matching through efficient fine-tuning. By re-aligning curved sampling trajectories into straight-line flows, FlowLM achieves high-quality generation in just a few steps, matching or exceeding the quality of 2,000-step diffusion sampling. The method requires very few training epochs; fine-tuned FlowLM reaches performance saturation with only half as many epochs as training from scratch, and both approaches outperform the original diffusion model. The paper also validates a more effective training objective for flow matching: predicting clean data to consistently guide sampling toward the true data distribution. Empirical results show FlowLM is highly effective for high-quality, few-step text generation.
Enables fast, high-quality text generation from pre-trained diffusion models with minimal fine-tuning.