可信但需验证：为离散扩散设计区域非均匀前向动力学
原标题：Trust but Verify: Designing Region-Inhomogeneous Forward Dynamics for Discrete Diffusion
AI 摘要
该论文重新将离散扩散的前向过程视为可设计空间，提出区域非均匀前向动力学：空洞位置采用标准吸收过程，用户提供的上下文以低速率均匀混合噪声进行腐蚀，使模型在训练中见过“已观测但不确定”的状态。作者对170M参数的预训练掩码扩散模型进行8000步微调（单张6GB笔记本GPU不到五小时），在明显随机腐蚀下检测召回率达0.96、精确率0.91，10%上下文腐蚀时填充token-F1提升3.8分，且干净上下文上的误编辑率低于0.4%。
正文节选
AI for Stochastic Dynamics Trust but Verify: Designing Region-Inhomogeneous Forward Dynamics for Discrete Diffusion Abstract In continuous diffusion, the forward process is a design surface: noise schedules, SDE variants, and per-position noise levels are all engineered to match the task. Discrete diffusion for language has collapsed this surface to a single point, the absorbing process, under which conditioning context sits at infinite signal-to-noise ratio for the entire reverse trajectory. Th