Reproduce “Model-intrinsic, Weight-level, and Training-free Spectral Fingerprinting for 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
Yanan He, Yanli Chen, Ling Xiong, Gejian Zhao, Hanzhou Wu
The method checks whether a language-model checkpoint resembles a claimed source by fingerprinting static weights, without inputs or forward passes. It summarizes singular-value patterns from GATE projections across model depth and compares the vectors with linear CKA. Across 29 pairs, related and unrelated scores showed a gap of 0.7439; evidence is limited to evaluated gated-FFN models and does not prove attribution.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.