AI Training Manager: Bounded Closed-Loop Control of Adaptive Training Recipes

方法与贡献

异步 LLM 监督控制器观察训练遥测数据，对正在执行的训练配置提出受约束的修改，并在刻意引入的监督学习与强化学习失败模式上实验。

研究对正在进行的训练实施自动控制，与 Interactive Training 为人和自动化控制器提供的可复用基础设施互补。

结合主题阅读

让人和 AI 智能体控制正在运行的训练任务 →

引用

Anjali Rao, Nikhil Kamalkumar Advani. AI Training Manager: Bounded Closed-Loop Control of Adaptive Training Recipes. arXiv.

@misc{rao2026aitrainingmanager,
  title = {{AI Training Manager: Bounded Closed-Loop Control of Adaptive Training Recipes}},
  author = {Rao, Anjali and Advani, Nikhil Kamalkumar},
  year = {2026},
  url = {https://arxiv.org/abs/2606.29871},
  eprint = {2606.29871},
  archivePrefix = {arXiv},
  doi = {10.48550/arXiv.2606.29871}
}