FlowLM: Few-Step Language Modeling via Diffusion-to-Flow Adaptation
This work addresses the need for efficient, high-quality text generation by reducing the number of sampling steps in diffusion language models, benefiting applications requiring fast inference.
FlowLM transforms pre-trained diffusion language models into flow matching models via efficient fine-tuning, enabling high-quality few-step generation that rivals 2,000-step diffusion sampling with very few training epochs. Finetuned FlowLM reaches performance saturation with half the training epochs of training from scratch, both outperforming the original diffusion model.
We present FlowLM, a flow matching language model transformed from pre-trained diffusion language models via efficient fine-tuning. By re-aligning the curved sampling trajectories of diffusion models into straight-line flows, FlowLM enables high quality few-step generation that rivals or even outperforms the quality of 2,000-step diffusion sampling with very few training epochs. Remarkably, finetuned FlowLM reaches performance saturation with only half as many training epochs as training from scratch, both approaches greatly outperforming the original diffusion model, thereby validating our method. Furthermore, we validate a more effective training objective for flow matching: predicting clean data to consistently guide the sampling process towards the true data distribution. Empirical results demonstrate that our approach is highly effective for high-quality, few-step text generation.