arXiv：cs.CL· Zakhar Kohut, Severyn Shykula, Mykola Vysotskyi, Serhii Dmytryshyn, Dmytro Khamula, Michal Zakrzewski, Damian Rynczak, Jacek Ma{\l}ecki, Taras Rumezhak, Volodymyr Karpiv·· 4 小时前AI 评分32
扩散语言模型的 token 级早停：Just on Time
Just on Time: Token-Level Early Stopping for Diffusion Language Models
AI 导读
针对扩散语言模型迭代去噪中大量 token 提前收敛导致的算力浪费，研究者提出一种免训练、token 级的早停方法，基于模型预测与局部上下文的轻量信号，为每个位置动态判断何时可以定格 token，实现自适应的逐 token 冻结。该方法无需任务特定微调，在数学推理、通用问答和科学理解等多项基准上大幅减少了所需扩散步数，同时保持生成质量。
正文
Abstract:Diffusion language models generate text through iterative refinement, a process that is often computationally inefficient because many tokens reach stability long before the final denoising step. We introduce a training-free, token-level early stopping approach that identifies convergence independently at each position. Our method leverages lightweight signals derived from the model's predictions and local context to dynamically determine when individual tokens can be finalized. This yields adaptive per-token freezing without task-specific fine-tuning, substantially reducing the total number of diffusion steps required. Across diverse benchmarks, spanning mathematical reasoning, general question answering, and scientific understanding, our approach achieves substantial efficiency gains while preserving generation quality.
| Comments: | Under review |
| Subjects: | Machine Learning (cs.LG); Computation and Language (cs.CL) |
| Cite as: | arXiv:2602.11133 [cs.LG] |
| (or arXiv:2602.11133v3 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2602.11133 arXiv-issued DOI via DataCite |
Submission history
From: Mykola Vysotskyi [view email]
[v1]
Wed, 11 Feb 2026 18:44:04 UTC (298 KB)
[v2]
Sun, 2 Aug 2026 20:20:48 UTC (193 KB)
[v3]
Wed, 7 Oct 2026 07:54:48 UTC (190 KB)
来源：arXiv：cs.CL · arxiv.org