提出新方法让离散扩散语言模型生成更准，训练时自动修正采样偏差。
Drifting Objectives for Refining Discrete Diffusion Language Models
- 将连续扩散的漂移思想转化为离散文本生成，通过软令牌特征实现梯度传递。
- 在4次迭代下，模型生成困惑度降低89%（MDLM）和86%（DUO），显著提升质量。
- 适合想优化离散扩散语言模型生成效果的研究者或工程师。
离散扩散语言模型（DDLMs）通过迭代去噪类别化标记序列生成文本。近期针对连续生成器的漂移方法表明，采样阶段的修正可融入训练过程，采用反对称固定点目标。本文研究如何将该原理应用于DDLMs，主要挑战在于离散文本接口：硬标记样本不可微，类别预测不直接提供连续样本用于漂移。为此提出TokenDrift，将类别预测升维为软标记特征，在冻结语义空间中应用反对称漂移，并将结果作为停止梯度的目标反向传播至DDLMlogits。在掩码和均匀状态扩散主干的受控持续训练实验中，TokenDrift在固定NFE条件下优于匹配的延续基线，使MDLM和DUO在4次NFE下的生成困惑度分别降低89%和86%。结果表明，漂移可为DDLMs提供实用的精炼目标。
原文摘要 · Abstract (English)
Discrete diffusion language models (DDLMs) generate text by iteratively denoising categorical token sequences, while recent drifting methods for continuous generators suggest that part of this sampling-time correction can instead be absorbed into training through an anti-symmetric fixed-point objective. We study how to transfer this principle to DDLMs, where the main challenge is the interface with discrete text: hard token samples are non-differentiable, and categorical predictions do not directly provide continuous samples to drift. We formulate TokenDrift, a drifting objective that lifts categorical predictions to soft-token features, applies anti-symmetric drifting in a frozen semantic space, and backpropagates the resulting stop-gradient feature target to DDLM logits. In controlled continual-training experiments with masked and uniform-state diffusion backbones, TokenDrift improves fixed-NFE generation quality over matched continuation baselines, reducing Gen.-PPL at 4 NFEs by 89% on MDLM and 86% on DUO. These results suggest that drifting can provide a practical refinement objective for DDLMs.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。