Reproduce “SABLE: Scalable Affine Bayesian Latent-space Exploration” 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
Erik Bodin, Lisa Maria Kreusser, Donney Fan, Henry Moss, Geoff Pleiss, Nicolas Durrande
SABLE separates candidate selection within sampled directions from exploration into fresh directions in high-dimensional latent spaces. Its probabilistic affine guide selects within the sampled span, while a finite-horizon schedule sets the norm of fresh directions. On Calorimeter DS3, after 3,000 evaluations over ten runs, its mean best observed value was 0.66 versus 0.55 for separable CMA-ES; this benchmark result assumes isotropic geometry and useful affine signal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.