arXivDaily arXiv每日学术速递 周一至周五更新

大厂专区

Google(谷歌)

2026-06-19 至 2026-06-19 共收录 4
2605.22748 2026-06-19 cs.RO cs.AI cs.LG cs.MA 版本更新

Superhuman Safe and Agile Racing through Multi-Agent Reinforcement Learning

通过多智能体强化学习实现超人类安全且敏捷的赛车

Ismail Geles, Leonard Bauersfeld, Markus Wulfmeier, Davide Scaramuzza

机构 * Robotics and Perception Group, University of Zurich(苏黎世大学机器人与感知组) Google DeepMind(谷歌DeepMind) Nomagic

AI总结 本文提出通过多智能体强化学习在高速四旋翼赛车中实现安全且敏捷的性能,展示了多智能体交互对真实世界交互安全性的关键作用,同时在高速赛车中超越人类飞行员并减少碰撞率。

Comments 12 pages (+4 supplementary). Website: https://rpg.ifi.uzh.ch/marl

详情

展开后加载摘要…

URL PDF HTML 收藏
2511.22283 2026-06-19 cs.LG 版本更新

The Hidden Cost of Approximation in Online Mirror Descent

在线镜像下降中近似的隐藏代价

Ofir Schlisselberg, Uri Sherman, Tomer Koren, Yishay Mansour

机构 * Tel Aviv University(特拉维夫大学) Google Research(谷歌研究)

AI总结 研究在线镜像下降(OMD)在近似误差下的鲁棒性,发现正则子光滑度与误差容忍度密切相关:均匀光滑正则子有紧界,而负熵在单纯形上需指数小误差,对数障碍和Tsallis正则子仅需多项式误差。

详情

展开后加载摘要…

URL PDF HTML 收藏
2309.15769 2026-06-19 math.ST cs.LG stat.ME stat.TH 版本更新

Benign overfitting beyond prediction: The ordinary least squares interpolator

超越预测的良性过拟合:普通最小二乘插值器

Dennis Shen, Dogyoon Song, Peng Ding, Jasjeet S. Sekhon

机构 * Department of Data Sciences & Operations, University of Southern California(数据科学与运营系,南加州大学) Department of Statistics, University of California, Davis(统计学系,加州大学戴维斯分校) Department of Statistics, University of California, Berkeley(统计学系,加州大学伯克利分校) Google DeepMind(谷歌DeepMind)

AI总结 本文研究过参数化线性模型中最小ℓ2范数OLS插值器的参数估计与推断性质,推导了留k法、遗漏变量偏误公式和Frisch-Waugh-Lovell定理的过参数化版本,并扩展了高斯-马尔可夫定理。

Comments This work is accepted for publication in Biometrika

详情

展开后加载摘要…

URL PDF HTML 收藏
2402.14035 2026-06-19 cs.LG cs.AI 版本更新

Wisdom of Committee: Diverse Distillation from Large Foundation Models and Domain Experts

委员会智慧:来自大型基础模型和领域专家的多样化蒸馏

Zichang Liu, Qingyun Liu, Yuening Li, Liang Liu, Anshumali Shrivastava, Shuchao Bi, Lichan Hong, Ed H. Chi, Zhe Zhao

机构 * Rice University(Rice大学) Google DeepMind(谷歌DeepMind) Google Inc(谷歌公司) University of California, Davis(加州大学戴维斯分校)

AI总结 针对基础模型向紧凑领域模型蒸馏时能力、架构和模态差异大的问题,提出DiverseDistill框架,通过可学习的问答机制和对齐异构教师输出,在推荐和视觉任务上恢复73-114%的性能差距。

Comments Accepted at the 1st Workshop on Resource-Efficient Learning and Knowledge Discovery (RelKD), KDD 2026

Journal ref Proceedings of the International Workshop on Resource-Efficient Learning and Knowledge Discovery (RelKD), KDD 2026

详情

展开后加载摘要…

URL PDF HTML 收藏