作为Mistral推出的首个基于纯强化学习（RL）训练的推理大模型，Magistral采用改进的Group Relative Policy Optimization（GRPO）算法。 通过消除KL散度惩罚、动态调整探索阈值和基于组归一化的优势计算，在AIME-24数学基准上实现从26.8%到73.6%的准确率跃升。

Hefei is also moving quickly to grow its low-altitude economy. In June 2024, city officials rolled out a policy package to ...