SafeWork-R1：在 AI-45° 定律下共进化安全与智能
人工智能
2025-08-08 v3 计算与语言
计算机视觉与模式识别
摘要
我们介绍 SafeWork-R1，这是一个前沿的多模态推理模型，展示了能力与安全的共进化。它由我们提出的 SafeLadder 框架开发，该框架包含大规模、渐进式、以安全为导向的强化学习后训练，支持一套多原则性验证器。与以往的对齐方法如 RLHF 仅学习人类偏好的不同，SafeLadder 使 SafeWork-R1 能够发展出内在的安全推理和自省能力，产生安全的“啊哈”时刻。值得注意的是，SafeWork-R1 在安全相关基准测试中相对于其基础模型 Qwen2.5-VL-72B 平均提升了 ，在安全性能上领先于 GPT-4.1 和 Claude Opus 4 等领先专有模型。为进一步增强其可靠性，我们实施了两种不同的推理时干预方法和一种 deliberative 搜索机制，以强制执行逐步级别的验证。最后，我们进一步开发了 SafeWork-R1-InternVL3-78B、SafeWork-R1-DeepSeek-70B 和 SafeWork-R1-Qwen2.5VL-7B。所有所得模型均表明，安全与能力可以协同进化，凸显了该框架在构建稳健、可靠且值得信赖的通用 AI 方面的通用性。
引用
@article{arxiv.2507.18576,
title = {SafeWork-R1: Coevolving Safety and Intelligence under the AI-45$^{\circ}$ Law},
author = {Shanghai AI Lab and : and Yicheng Bao and Guanxu Chen and Mingkang Chen and Yunhao Chen and Chiyu Chen and Lingjie Chen and Sirui Chen and Xinquan Chen and Jie Cheng and Yu Cheng and Dengke Deng and Yizhuo Ding and Dan Ding and Xiaoshan Ding and Yi Ding and Zhichen Dong and Lingxiao Du and Yuyu Fan and Xinshun Feng and Yanwei Fu and Yuxuan Gao and Ruijun Ge and Tianle Gu and Lujun Gui and Jiaxuan Guo and Qianxi He and Yuenan Hou and Xuhao Hu and Hong Huang and Kaichen Huang and Shiyang Huang and Yuxian Jiang and Shanzhe Lei and Jie Li and Lijun Li and Hao Li and Juncheng Li and Xiangtian Li and Yafu Li and Lingyu Li and Xueyan Li and Haotian Liang and Dongrui Liu and Qihua Liu and Zhixuan Liu and Bangwei Liu and Huacan Liu and Yuexiao Liu and Zongkai Liu and Chaochao Lu and Yudong Lu and Xiaoya Lu and Zhenghao Lu and Qitan Lv and Caoyuan Ma and Jiachen Ma and Xiaoya Ma and Zhongtian Ma and Lingyu Meng and Ziqi Miao and Yazhe Niu and Yuezhang Peng and Yuan Pu and Han Qi and Chen Qian and Xingge Qiao and Jingjing Qu and Jiashu Qu and Wanying Qu and Wenwen Qu and Xiaoye Qu and Qihan Ren and Qingnan Ren and Qingyu Ren and Jing Shao and Wenqi Shao and Shuai Shao and Dongxing Shi and Xin Song and Xinhao Song and Yan Teng and Xuan Tong and Yingchun Wang and Xuhong Wang and Shujie Wang and Xin Wang and Yige Wang and Yixu Wang and Yuanfu Wang and Futing Wang and Ruofan Wang and Wenjie Wang and Yajie Wang and Muhao Wei and Xiaoyu Wen and Fenghua Weng and Yuqi Wu and Yingtong Xiong and Xingcheng Xu and Chao Yang and Yue Yang and Yang Yao and Yulei Ye and Zhenyun Yin and Yi Yu and Bo Zhang and Qiaosheng Zhang and Jinxuan Zhang and Yexin Zhang and Yinqiang Zheng and Hefeng Zhou and Zhanhui Zhou and Pengyu Zhu and Qingzi Zhu and Yubo Zhu and Bowen Zhou},
journal= {arXiv preprint arXiv:2507.18576},
year = {2025}
}
备注
47 pages, 18 figures, authors are listed in alphabetical order by their last names; v3 modifies minor issues