EGGROLL
EGGROLL stands for Evolution Guided GeneRal Optimisation via Low-rank Learning. It is the named method introduced by Evolution Strategies at the Hyperscale.
Role
EGGROLL makes evolution strategies more hardware-efficient by replacing full-rank perturbation matrices with low-rank factors and evaluating many perturbations through batched inference-style computation.
Mechanism
For a matrix parameter, EGGROLL samples factors A and B, forms AB^T / sqrt(r), evaluates the perturbed model, and weights the perturbation by scalar fitness. The per-sample perturbation is low-rank, but the population-weighted update can still be full-rank.
Why It Matters
The method connects ES to the same low-rank systems intuition that made LoRA practical for gradient-based fine-tuning, but it uses low-rank structure for black-box perturbation evaluation rather than learned adapters.