FlowLM: Few-Step Language Modeling via Diffusion-to-Flow Adaptation
FlowLM is a new language model that adapts pre-trained diffusion models for efficient few-step text generation. It achieves high-quality results with significantly fewer training epochs compared to traditional methods. The model demonstrates improved performance by aligning sampling trajectories and using a refined training objective.
- ▪FlowLM transforms pre-trained diffusion language models through efficient fine-tuning.
- ▪It enables high-quality few-step generation that can outperform 2,000-step diffusion sampling.
- ▪Finetuned FlowLM reaches performance saturation with half as many training epochs as training from scratch.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computation and Language arXiv:2605.20199 (cs) [Submitted on 6 Apr 2026] Title:FlowLM: Few-Step Language Modeling via Diffusion-to-Flow Adaptation Authors:Runzhe Zhang, Letian Chen, Wenpeng Zhang, Zhouhan Lin, Peilin Zhao View a PDF of the paper titled FlowLM: Few-Step Language Modeling via Diffusion-to-Flow Adaptation, by Runzhe Zhang and 4 other authors View PDF Abstract: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.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
