EggRoll

EGGROLL is an Evolution Strategies (ES) algorithm that enables training/tuning of billion-parameter models with massive populations without backprop. The bottleneck in traditional ES is that each population member requires full-rank perturbations and forward passes, causing memory/computation to explode. EGGROLL addresses this by applying low-rank perturbations to each layer using LoRA-inspired ideas to reduce costs. Experimental results show: (1) performance similar to or better than OpenES across various RL environments, (2) competitive/superior results compared to GRPO in RWKV-based LLM reasoning fine-tuning, and (3) demonstration of pure integer RNN LM pretraining feasibility.