Tag
VideoMaMa++ 提出一种掩码引导的扩散式视频抠图框架，通过分离的 RGB/掩码 token 流以及 preserve-and-refine 可学习嵌入，实现跨片段边界的滑窗推理，从而在长视频中保持时间一致性并提升细节精度。
This paper provides theoretical guarantees for distinguishing LLM responses by modeling token embeddings as trajectories of a dynamical system, proving exponential decay of misclassification probability and characterizing cross-embedding generalization.
This paper introduces TIDE, a method that addresses the Rare Token and Contextual Collapse problems in LLMs by injecting token identity into every layer via Embedding Memory. The authors demonstrate theoretical and empirical improvements across language modeling and downstream tasks.