Reproduce “Geometric Signatures of Compositionality Across a Language Model's Lifetime” 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
Jin Hwa Lee, Thomas Jiralerspong, Lei Yu, Yoshua Bengio, Emily Cheng
The authors compare two measures of representation geometry to test how they respond to composition and surface variation. They use nonlinear TwoNN intrinsic dimension and PCA dimension on coherent and word-shuffled inputs. On a controlled grammar, Pythia-410m’s PCA dimension correlates with gzip size at 0.96, while TwoNN’s correlation is −0.05 and not significant. These associations do not establish what features encode.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.