CoDR: Training-Free Confidence-Drift Remasking for Diffusion Language Models

查看原文 / View Original 编辑 / Edit

Computer Science > Computation and Language arXiv:2610.08833 (cs) [Submitted on 28 Sep 2026] Title:CoDR: Training-Free Confidence-Drift Remasking for Diffusion Language Models Authors:Yue Wu, Qinghe Zhang, Yu Zhang, Jian Huang

计算机科学 > 计算与语言 arXiv:2610.08833 (cs) [提交于 2026 年 9 月 28 日] 标题：CoDR：面向扩散语言模型的无需训练的置信度漂移重掩码技术作者：Yue Wu, Qinghe Zhang, Yu Zhang, Jian Huang

Abstract: Masked diffusion language models (MDLMs) decode by repeatedly committing tokens to masked positions, but these commitments are usually irreversible. A token chosen under sparse, partial context is kept fixed, even when later context no longer supports it. Existing samplers mainly decide when to commit a token, but rarely check whether an already committed token should still be kept, allowing early mistakes to propagate.

摘要：掩码扩散语言模型（MDLM）通过反复向掩码位置提交标记来进行解码，但这些提交通常是不可逆的。在稀疏、部分上下文下选择的标记会被固定下来，即使后续的上下文不再支持它。现有的采样器主要决定何时提交标记，但很少检查已经提交的标记是否仍应保留，这导致早期的错误得以传播。

We trace this issue to confidence drift, where the model’s confidence in a committed token drops from its sparse commit-time context to the denser context available later. Based on this signal, we propose CoDR (Confidence Drift Remasking), a training-free and sampler-agnostic refinement pass. CoDR estimates drift for all committed positions in only k forward passes via k-partition probing, then remasks and regenerates only the tokens the model no longer endorses.

我们将此问题归因于置信度漂移，即模型对已提交标记的置信度从其稀疏的提交时上下文下降到后续可用的更密集上下文。基于这一信号，我们提出了 CoDR（置信度漂移重掩码），这是一种无需训练且与采样器无关的优化过程。CoDR 通过 k 分区探测，仅需 k 次前向传递即可估计所有已提交位置的漂移，然后仅对模型不再认可的标记进行重掩码和重新生成。

Across two backbones, four reasoning and coding tasks, and three base samplers, CoDR improves average accuracy across all evaluated model-sampler configurations and improves most individual task settings with modest overhead. Controlled experiments show that the gains come from targeted confidence-drift remasking rather than extra compute alone, and that CoDR uses far fewer forward passes than prior remasking methods.

在两个骨干网络、四个推理和编码任务以及三个基础采样器上，CoDR 提高了所有评估模型-采样器配置的平均准确率，并以适度的开销改善了大多数单独的任务设置。对照实验表明，这些收益来自于有针对性的置信度漂移重掩码，而非仅仅是额外的计算，且 CoDR 使用的前向传递次数远少于先前的重掩码方法。