Reproduce “Contrastive-Difference CKA Reveals Concept-Specific Structural Alignment Across Language Model Architectures” 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
Xueping Gao
The paper evaluates whether a classifier trained on one language model can detect concept polarity in another. Its CKAΔ diagnostic compares activation changes for positive- and negative-concept prompts, while a learned affine map aligns representations before classification. Across eight traits, aligned transfer averaged 99.9% accuracy versus 51.3% direct transfer, but it was not zero-shot and does not prove identical representations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.