Reproduce “Grounding Functional Similarity by Invariance-Aware Model Stitching” 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

Ioannis Athanasiadis, Anmar Karmush, Michael Felsberg

The paper tests whether models have functionally interchangeable representations by stitching a learned affine bridge between one model’s early layers and another’s later layers. It trains this bridge on altered inputs preserving the front representation, then tests the end model. On CIFAR-10, this exposes shortcut-based agreement and cross-bias agreement falling toward chance, but remains a diagnostic for image-classification models, not a definitive similarity measure.

Summary by alphaXiv.

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