Reproduce “A Margin Perspective on LoRA: Robustness to Catastrophic Forgetting and Adapter Merging (MaLoRA)” 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
Ziqing Xu, Hancheng Min, Lachlan Ewen MacDonald, Salma Tarmoun, Enrique Mallada, Weijie Su, Rene Vidal
MaLoRA links LoRA adaptation and adapter merging through margins. It models LoRA as a low-rank update to weights and analyzes linear classification, using margins to characterize adaptation and balance task adapters. In four frozen CIFAR-100 experiments, predicted mixture weights were close to grid-search optima. Across six benchmarks, normalized accuracy was 93.32%, versus 93.15% for Task Arithmetic; theory assumes strict orthogonality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.