Evolution Strategies at the Hyperscale

Bidipta Sarkar, Mattie Fellows, Juan Agustin Duque, Alistair Letcher, Antonio León Villares, Anya Sims, Clarisse Wibault, Dmitry Samsonov, Dylan Cope, Jarek Liesen, Kang Li, Lukas Seier, Theo Wolf, Uljad Berdica, Valentin Mohl, Alexander David Goldie, Aaron Courville, Karin Sevegnani, Shimon Whiteson, Jakob Nicolaus Foerster

Open source

Type: other
Venue: arXiv / University of Oxford / MILA / NVIDIA

Summary

EGGROLL samples rank-r adapters AB^T instead of full-rank noise, reaching ~91% of batch-inference throughput and ~100× naïve ES at large populations. Individual perturbations are low-rank but the population average is full-rank (O(1/r) to Gaussian ES). Pretrains an int8 nonlinear RNN (EGG, no activations) on MiniPile to 3.40 bits/byte vs a backprop Transformer 3.58 at matched data batch (pop 2^20). Competitive with GRPO on countdown/GSM8K (RWKV-7) and with OpenES on 16 tabula-rasa RL envs. Code https://github.com/ESHyperscale/HyperscaleES; project https://eshyperscale.github.io/.

Keywords

eggroll · evolution-strategies · lora · int8 · rwkv · grpo · oxford · mila

Topics

evolution strategies, zeroth-order optimization, LLM post-training

Research notes

Primary: arxiv abs (cs.LG; also cs.AI). CC BY 4.0 on HTML. Equal contrib Sarkar/Fellows/Duque; equal senior Whiteson/Foerster. FLAIR/WhiRL Oxford, MILA, NVIDIA. Correspondence bidipta.sarkar/matthew.fellows/jakob.foerster@eng.ox.ac.uk. Code https://github.com/ESHyperscale/HyperscaleES (339 stars at check); nano-egg 173 stars; eggroll-vllm 13–15 stars. Project https://eshyperscale.github.io/. HF paper page 5 upvotes; 3 linked models not copied into hf_* fields. Discord posted PDF. Uses public MiniPile/GSM8K/countdown/DeepScaleR; no new corpus, so no datasets_local row. License field left blank per catalog convention.