Reproduce “The Evolution Strategies Paradox: Better Models from Worse Perturbations” 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
Roberto Dailey, Kajetan Schweighofer, Jamieson Warner, Conor F. Hayes, Elliot Meyerson, Babak Hodjat, Risto Miikkulainen, Xin Qiu
The paper shows how evolution strategies can improve a model even when nearly all sampled candidates score below the parent. ES weights parameter perturbations by standardized relative rewards and averages them into an update; the proposed explanation links curvature and directional score differences with aggregation that reduces off-direction noise. Parent reward rises in tested settings, but this is not universal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.