Standard Low-Rank Adaptation (LoRA) scales linearly with model rank and weight matrix dimensions, incurring significant transmission costs during distributed edge adaptation. SOLAR reparameterizes low-rank projections into orthogonal latent subspaces, achieving 35% smaller memory footprint and transmission payloads across Vision Transformers (ViT), LLaMA-3, and GPT-2 without degradation on downstream benchmarks.
@article{sajjadi2026solar,
title = {{SOLAR}: Communication-Efficient Model Adaptation via Subspace-Oriented Latent Adapter Reparametrization},
author = {Sajjadi Mohammadabadi, Seyed Mahmoud and Ma, Xiaolong and Yang, Lei and Yan, Feng and Zhang, Junshan},
journal = {arXiv preprint arXiv:2604.08368},
year = {2026}
}