Safactory：用于训练可信自主智能的可扩展智能体基础设施
人工智能
2026-05-11 v2 分布式、并行与集群计算
摘要
随着大型模型从对话助手演变为自主智能体，挑战越来越多地源于长程决策、工具使用和真实环境交互。现有的智能体基础设施在评估、数据管理和智能体演化方面仍然分散，难以系统地发现风险并在持续的闭环中改进模型。在本报告中，我们提出了Safactory，一个用于可信自主智能的可扩展智能体工厂。Safactory集成了三个紧密耦合的平台：用于轨迹生成的并行仿真平台，用于轨迹存储和经验提取的可信数据平台，以及用于异步强化学习和on-policy蒸馏的自主演化平台。据我们所知，Safactory是第一个为下一代可信自主智能提出统一演化流程的框架。
引用
@article{arxiv.2605.06230,
title = {Safactory: A Scalable Agentic Infrastructure for Training Trustworthy Autonomous Intelligence},
author = {Xinquan Chen and Zhenyun Yin and Shan He and Bin Huang and Shanzhe Lei and Pengcheng Shi and Kun Cai and Bei Chen and Bangwei Liu and Zeyu Kang and Chao Huang and Yang Zhang and Wenjie Li and Ruijun Ge and Yajie Wang and Tianshun Fang and Tianyang Xu and Yiwen Cong and Meng Jin and Gaolei Li and Xuansheng Wu and Linhan Liu and Zijing He and An Li and Yan Teng and Xin Tan and Dongrui Liu and Jing Shao and ChaoChao Lu and Ji He and Jie Li and Chunfeng Song and Jinya Xu and Fan Song and Shujie Wang and Jianmin Qian and Jie Hou and Xuhong Wang and Yingchun Wang and Hui Wang and Xia Hu},
journal= {arXiv preprint arXiv:2605.06230},
year = {2026}
}
备注
50 pages, 21 figures