如果RL可预测，我们还需要把训练跑满吗？中科大揭示参数更新的线性秘密
"训练加速" 相关文章
What Makes Looped Transformers Perform Better Than Non-Recursive Ones (Provably)
What Makes Looped Transformers Perform Better Than Non-Recursive Ones (Provably)
What Makes Looped Transformers Perform Better Than Non-Recursive Ones (Provably)
On Predictability of Reinforcement Learning Dynamics for Large Language Models
SPEC-RL: Accelerating On-Policy Reinforcement Learning via Speculative Rollouts
Improving Consistency Models with Generator-Augmented Flows
在GSM8K上比GRPO快8倍！厦大提出CPPO，让强化学习快如闪电
标点符号成大模型训练神器！KV缓存狂减一半，可处理400万Tokens长序列，来自华为港大等 | 开源
