Reproduce “Rethinking Layer-wise Model Merging through Chain of Merges” 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
Pietro Buzzega, Riccardo Salami, Angelo Porrello, Simone Calderara
CoM merges independently fine-tuned task models into one shared model by updating each layer’s inputs through the partially merged network before fitting the next layer. This addresses Merging Covariate Shift, where earlier merges change the activations later layers receive. On an eight-task ViT-B/32 LoRA vision benchmark, CoM reaches 92.7% normalized accuracy versus 76.1% for Core ISO in the tested static-merge setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.