这项研究打破了深度学习十年来的反向传播依赖，做大规模模型训练或非可微分任务（如强化学习、神经架构搜索）的团队可以直接关注 EGGROLL，它可能改变你构建模型的方式。
NVIDIA 与牛津大学联合发表论文，提出 EGGROLL 方法，将进化策略（Evolution Strategies）扩展到十亿参数模型，无需反向传播即可训练。该方法用两个低秩矩阵替代密集随机扰动矩阵，大幅降低内存消耗，达到纯推理吞吐量的 91%。EGGROLL 支持数十万并行变异，可与不可微分组件配合，在推理任务上与 GRPO 竞争。团队还训练了纯 8 位整数循环语言模型 EGG，非线性来自整数溢出裁剪，而非激活函数，预训练时种群规模超过百万。
NVIDIA just proved AI doesn't need backpropagation…
NVIDIA just proved AI doesn't need backpropagation to learn anymore.
Training giant neural networks has meant one thing for a decade.
You run backpropagation, compute gradients, and update weights with calculus. A new paper from NVIDIA and Oxford breaks that assumption. EGGROLL scales Evolution Strategies to billion-parameter models.
The trick replaces dense random perturbation matrices with two skinny low-rank ones. This cuts memory and hits 91% of pure inference throughput. The practical upside changes what you can build: > No gradients or backprop required > Hundreds of thousands of parallel mutations > Works with non-differentiable components > Competitive with GRPO on reasoning tasks The team also built EGG, a recurrent language model trained from scratch.
It runs entirely on 8-bit integers. No floating-point math appears anywhere in the network. Nonlinearity comes from integer overflow clipping, not activation functions. Population sizes exceeded one million workers during pretraining.