首次获取该候选人的完整档案及 StartLux 评估共 20 积分；每位用户单独解锁，之后查看免费。主动刷新 20 积分。
实时查找电子邮箱，费用不超过20积分；未找到邮箱不扣费。
Yingbo HAO · Hanyong Shao · Ting Song · Yan Xia · Di Zhang · Shaohan Huang · Xun Wu · Songchen Xu · Le Xu · Li Dong · Zewen Chi · Yi Zou · Furu Wei
Di Zhang · xun wu · Shaohan Huang · Lingjie Jiang · Yaru Hao · Li Dong · Zewen Chi · Zhifang Sui · Furu Wei
Reinforcement learning with verifiable rewards (RLVR) has emerged as a powerful paradigm for improving reasoning capabilities. However, training RLVR with Mixture-of-Experts (MoE) policies remains fragile and is often prone to reward collapse.We identify a MoE-specific source of instability, refe...