Reproduce “Escaping Low-Dimensional Overlap: Multi-Task Model Merging via High-Dimensional Sparse Disentanglement” 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
Yihang Zhang, Shengke Sun, Junjie Wen, Feng Zeng
The authors present a model-merging method that maps selected task updates into sparse feature codes before combining them, while using ordinary addition elsewhere. A forward-pass-only procedure selects layers for this processing. On a four-task Qwen2.5-1.5B evaluation, the method reaches a seven-metric average of 36.48, versus 33.70 for DELLA, but scores lower on five of seven metrics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.