Reproduce “ENGINE: Endogenous Variational MultiScale Optimization for Zeroth-Order LLM Fine-Tuning” 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
Zhuoli Ouyang, Changxi Chi, Siyuan Li, Tailin Wu
ENGINE combines low-rank LoRA updates with full-space curvature feedback for zeroth-order fine-tuning. It alternates coarse LoRA updates and fine full-parameter updates, correcting mismatch between their directions. On LLaMA-3-8B/CB, it reaches LOREN’s best accuracy with up to 2.61× fewer forward passes and 2.58× less wall-clock time, with unchanged peak GPU memory versus LOREN in evaluated classification settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.