University of Maryland, College Park（马里兰大学帕克分校）

Going Down Memory Lane: Scaling Tokens for Video Stream Understanding with Dynamic KV-Cache Memory

回溯记忆：通过动态KV缓存内存扩展令牌实现视频流理解

Vatsal Agarwal, Saksham Suri, Matthew Gwilliam, Pulkit Kumar, Abhinav Shrivastava

机构 * University of Maryland, College Park（马里兰大学 College Park 分校） 纠错

AI总结 MemStream通过扩展令牌预算和自适应选择策略，提升视频流理解的准确性和细粒度信息保留能力。

Comments Project page: see https://vatsalag99.github.io/memstream/

ROCKET: Residual-Oriented Multi-Layer Alignment for Spatially-Aware Vision-Language-Action Models

ROCKET：基于残差的多层对齐用于空间感知的视觉-语言-动作模型

Guoheng Sun, Tingting Du, Kaixi Feng, Chenxiang Luo, Xingguo Ding, Zheyu Shen, Ziyao Wang, Yexiao He, Ang Li

机构 * University of Maryland, College Park ； University of Wisconsin, Madison ； City University of Hong Kong ； St.\ Paul's School 纠错

AI总结 ROCKET通过共享投影器和稀疏激活方案实现多层对齐，提升3D空间理解能力，在LIBERO等任务中达到98.5%的高成功率。