Reproduce “Reactive Optimizers” 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

Xingjian Bai, Jia-Hao Liang, Jayden Teoh, Vincent Sitzmann

The authors present OPAL, which uses a pretrained language-model agent to inspect training signals and rewrite executable optimizer code during a run, without optimizer-specific meta-training. On ImageNet image generation, its no-guidance FID-50K was 4.335, versus 5.151 for the authors’ retrained AdamW recipe. This selected comparison does not show superiority over careful tuning: a static Amharic sweep beat OPAL’s first run.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.