中文

Magistral

计算与语言 2025-06-13 v1

摘要

我们介绍 Magistral,即 Mistral 首个推理模型及其可扩展强化学习 (RL) 管线。我们不依赖现有实现或从先前模型中提炼的 RL 轨迹,而是采用从零开始的方法,仅依赖自身模型和基础设施。值得注意的是,我们展示了一个使我们能够探索纯 RL 训练大型语言模型极限的堆栈,提出了一种简单方法以强制模型的推理语言,并表明仅用文本数据进行 RL 仍能保持大部分初始检查点的能力。我们发现,针对文本进行 RL 可保持或提高多模态理解、指令遵循和函数调用能力。我们介绍 Magistral Medium,基于 Mistral Medium 3 采用 RL 进行推理训练,并开源 Magistral Small(Apache 2.0),其中包括来自 Magistral Medium 的冷启动数据。

关键词

引用

@article{arxiv.2506.10910,
  title  = {Magistral},
  author = {Mistral-AI and : and Abhinav Rastogi and Albert Q. Jiang and Andy Lo and Gabrielle Berrada and Guillaume Lample and Jason Rute and Joep Barmentlo and Karmesh Yadav and Kartik Khandelwal and Khyathi Raghavi Chandu and Léonard Blier and Lucile Saulnier and Matthieu Dinot and Maxime Darrin and Neha Gupta and Roman Soletskyi and Sagar Vaze and Teven Le Scao and Yihan Wang and Adam Yang and Alexander H. Liu and Alexandre Sablayrolles and Amélie Héliou and Amélie Martin and Andy Ehrenberg and Anmol Agarwal and Antoine Roux and Arthur Darcet and Arthur Mensch and Baptiste Bout and Baptiste Rozière and Baudouin De Monicault and Chris Bamford and Christian Wallenwein and Christophe Renaudin and Clémence Lanfranchi and Darius Dabert and Devon Mizelle and Diego de las Casas and Elliot Chane-Sane and Emilien Fugier and Emma Bou Hanna and Gauthier Delerce and Gauthier Guinet and Georgii Novikov and Guillaume Martin and Himanshu Jaju and Jan Ludziejewski and Jean-Hadrien Chabran and Jean-Malo Delignon and Joachim Studnia and Jonas Amar and Josselin Somerville Roberts and Julien Denize and Karan Saxena and Kush Jain and Lingxiao Zhao and Louis Martin and Luyu Gao and Lélio Renard Lavaud and Marie Pellat and Mathilde Guillaumin and Mathis Felardos and Maximilian Augustin and Mickaël Seznec and Nikhil Raghuraman and Olivier Duchenne and Patricia Wang and Patrick von Platen and Patryk Saffer and Paul Jacob and Paul Wambergue and Paula Kurylowicz and Pavankumar Reddy Muddireddy and Philomène Chagniot and Pierre Stock and Pravesh Agrawal and Romain Sauvestre and Rémi Delacourt and Sanchit Gandhi and Sandeep Subramanian and Shashwat Dalal and Siddharth Gandhi and Soham Ghosh and Srijan Mishra and Sumukh Aithal and Szymon Antoniak and Thibault Schueller and Thibaut Lavril and Thomas Robert and Thomas Wang and Timothée Lacroix and Valeriia Nemychnikova and Victor Paltz and Virgile Richard and Wen-Ding Li and William Marshall and Xuanyu Zhang and Yunhao Tang},
  journal= {arXiv preprint arXiv:2506.10910},
  year   = {2025}
}