Bio-Inspired Activation Function Discovery
This paper provides a practical guideline for improving the efficiency of neural network training by matching the timescale of evolutionary operators to the evaluation budget.
This paper provides a practical guideline for improving the efficiency of neural network training by matching the timescale of evolutionary operators to the evaluation budget.
ES offers a memory-efficient alternative to GRPO for LLM reasoning, with better Pass@K performance and smaller required population size for larger models.
这篇论文提出了SMC-ES,把进化策略和统计模型检查结合起来,能生成带形式安全证明的控制策略。比普通RL更靠谱。
别浪费算力跑进化循环了——这篇论文实测发现多轮迭代没用,用一次采样加集合选择反而效果翻倍,还省了90%的调用成本。
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。