[论文] ConvergeFlow: Language Flow with Provable Convergence to Token Embeddi...

论文概要

研究领域: NLP
作者: Na Li, Yuchen Jiao, Changxiao Cai
发布时间: 2025-08-26
arXiv: 2508.17626

中文摘要

连续扩散和基于流的语言模型（LM）近期取得了与离散LM相当的性能。然而，现有连续框架仍依赖用交叉熵（CE）监督的解码器，因为流轨迹不保证终止于有效的token嵌入。针对这一限制，本文提出了ConvergeFlow，一种嵌入空间流式LM，它将数据预测器约束在token嵌入的凸包内，并仅用流匹配诱导的均方误差目标进行训练。在适当的正则性条件下，我们证明了即使数据预测器存在误差，所得流仍能收敛到有效token嵌入，从而无需CE监督解码器即可直接预测token。我们还开发了三种采样机制来控制生成困惑度与熵之间的权衡。OpenWebText上的实验表明，ConvergeFlow与现有连续和离散扩散LM的性能相当。

原文摘要

Recent advances in continuous diffusion and flow-based language models (LMs) have achieved performance competitive with discrete LMs. However, existing continuous frameworks still rely on decoders supervised with cross entropy (CE) because the flow trajectories are not guaranteed to terminate at valid token embeddings. Motivated by this limitation, we introduce ConvergeFlow, an embedding-space flow-based LM, which constrains the data predictor to the convex hull of token embeddings and trains it solely with the mean squared error objective induced by flow matching. Under suitable regularity conditions, we prove that the resulting flow converges to valid token embeddings despite errors in the data predictor, enabling direct token prediction without a CE-supervised decoder. We further develo...

自动采集于 2026-08-26

#论文 #arXiv #NLP #小凯