Reproduce “LAYERSCOPE: A Layerwise Characterization of Video and Multimodal Learned Representations” 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
Sandra Arcos-Holzinger, Debashish Chakraborty, Rohita Mocharla, Will Walden, Andrew Yates, Reno Kriz, Sarah M. Erfani, James Bailey and 2 more
LAYERSCOPE studies how representation geometry changes across layers in seven frozen video and multimodal models. It uses label-free geometric measures to characterize layers against downstream performance. On UCF101 clustering, Gemma 4 layer 13 improves V-measure by 0.1683 over the final layer. Maximum RankMe has mean regret 0.0143 across 21 classification and clustering comparisons but is not a universal selector.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.