arXiv：cs.CL· Yue Wu, Qinghe Zhang, Yu Zhang, Jian Huang·· 3 小时前AI 评分36
CoDR：面向扩散语言模型的无训练置信度漂移重掩码方法
CoDR: Training-Free Confidence-Drift Remasking for Diffusion Language Models
AI 导读
针对掩码扩散语言模型（MDLM）中已提交 token 不可逆、早期错误会传播的问题，研究者提出无训练且与采样器无关的细化方法 CoDR（置信度漂移重掩码），仅用 k 次前向传播估计所有已提交位置的置信度漂移，再对模型不再认可的 token 重新掩码并生成。
正文
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. 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. 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. Code is available at this https URL.
| Comments: | 17 pages, 6 figures, and 17 tables |
| Subjects: | Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.08833 [cs.CL] |
| (or arXiv:2610.08833v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.08833 arXiv-issued DOI via DataCite |
Submission history
From: Jian Huang [view email]
[v1]
Mon, 28 Sep 2026 16:43:09 UTC (864 KB)
来源：arXiv：cs.CL · arxiv.org