Reproduce “Capsule Lens: Locating and Tracking Concept Geometry in Model 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

Yiming Tang, Harshvardhan Saini, Samyak Jha, Huaming Chen, Xufeng Duan, Dianbo Liu

Capsule Lens models each concept’s representation cloud as an angular cone around a central direction with magnitude bounds, then compares these parameters across layers and training checkpoints. Capsules average 0.943 held-out coverage. In Qwen2.5-1.5B RLVR, the layer-27 span for minute narrows from 0.711 to 0.594, while 7 of 18 concepts change substantially; causality and generalization remain unestablished.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.