1 paper
Yifan Wang, Hao Cheng, Xiaomin Li +13
Long horizon agent self-evolution without model weight updates is essential for enabling deployed agents to accumulate reusable skills and improve over time. Prior self-evolution w…