Reproduce “VidToC: Rethinking Video Token Compression for Multimodal Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zihan Xiao, Boshen Xu, Pei Fu, Jianzhong Ju, Zhenbo Luo, Jian Luan, Qin Jin
The authors present VidToC, a controlled study of video-token compression across tasks and training conditions. They compare representative ways to select or merge tokens and examine how temporal positions are represented. At 16× compression, training-free temporal grounding on Charades scores 29.21 mean IoU with CLS-Attn, versus 17.29 with Uniform Prune; this task-specific result does not establish a universal winner.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.