Authors: Bidipta Sarkar and Mattie Fellows and Juan Agustin Duque and Alistair Letcher and Antonio León Villares and Anya Sims and Clarisse Wibault and Dmitry Samsonov and Dylan Cope and Jarek Liesen and Kang Li and Lukas Seier and Theo Wolf and Uljad Berdica and Valentin Mohl and Alexander David Goldie and Aaron Courville and Karin Sevegnani and Shimon Whiteson and Jakob Nicolaus Foerster
Cite as: \cite{sarkar2025evolution}
arXiv: https://arxiv.org/abs/2511.16652
Summary in three sentences
Evolution Strategies at the Hyperscale argues that ES is attractive for noisy, non-differentiable objectives but expensive at scale when perturbations are unstructured. The paper introduces EGGROLL, which structures perturbations as rank- matrices to improve GPU arithmetic intensity and make large-population ES more practical. For this project, the paper is the ES scaling and theory anchor, while the proposed delta is to change the adapter/search parameterisation using DoRA-like weight decomposition.
Main contribution
The main contribution is an efficient low-rank perturbation scheme for ES at very large model scales, together with theory for high-dimensional Gaussian ES and empirical demonstrations across language modelling, post-training, and reinforcement-learning settings.
Key equations and definitions
The project-relevant object is a rank- structured perturbation of a parameter matrix used inside an ES estimator. The paper also analyses high-dimensional Gaussian ES and a linearising effect; those details should be re-read before writing the theory section.
Empirical results
Every number here must name the section or figure it comes from.
If you cannot locate it in the paper, do not write it.
The arXiv abstract reports that EGGROLL improves training speed for large-population ES on billion-parameter models and compares with GRPO on reasoning post-training. No project note numbers are recorded here until the relevant tables or figures are read directly.
Relation to our project
- what we reuse: low-rank structured perturbations for scalable ES, the high-dimensional ES framing, and the comparison discipline against gradient-based post-training methods.
- what we dispute: not the method, but whether rank-structured ES alone is the best search geometry for adapter-based LLM agent adaptation.
- what we must reproduce: a small matched comparison of plain low-rank ES against decomposed low-rank ES before attempting long-context agent tasks.
Criticism and limitations
The paper is broader than this project and is not specifically about DoRA-style adapter geometry for long-context LLM agents. Its most useful role here is to provide the scalable ES baseline and theoretical language; the new paper must justify why decomposition changes the search behaviour rather than merely renaming the perturbation space.