arXivDaily arXiv每日学术速递 周一至周五更新

高校专区

Harvard University(哈佛大学)

2026-06-24 至 2026-06-24 共收录 8
2606.24855 2026-06-24 cs.AI 新提交

OpenThoughts-Agent: Data Recipes for Agentic Models

OpenThoughts-Agent: 智能体模型的数据配方

Negin Raoof, Richard Zhuang, Marianna Nezhurina, Etash Guha, Atula Tejaswi, Ryan Marten, Charlie F. Ruan, Tyler Griggs, Alexander Glenn Shaw, Hritik Bansal, E. Kelly Buchanan, Artem Gazizov, Reinhard Heckel, Chinmay Hegde, Sankalp Jajee, Daanish Khazi, Emmanouil Koukoumidis, Xiangyi Li, Hange Liu, Shlok Natarajan, Harsh Raj, Nicholas Roberts, Ethan Shen, Nishad Singhi, Michael Siu, Ashima Suvarna, Hanwen Xing, Patrick Yubeaton, Robert Zhang, Leon Liangyu Chen, Xiaokun Chen, Steven Dillmann, Saadia Gabriel, Xunyi Jiang, Anurag Kashyap, Boxuan Li, Yein Park, Minh Pham, Sujay Sanghavi, Lin Shi, Ke Sun, Yixin Wang, Zhiwei Xu, Erica Zhang, Siyan Zhao, Wanjia Zhao, Jenia Jitsev, Alex Dimakis, Benjamin Feuer, Ludwig Schmidt

机构 * UC Berkeley(加州大学伯克利分校) Stanford University(斯坦福大学) JSC(于利希超级计算中心) LAION University of Texas at Austin(德克萨斯大学奥斯汀分校) Bespoke Labs Laude Institute UCLA(加州大学洛杉矶分校) Harvard University & Harvard Medical School(哈佛大学与哈佛医学院) TU Munich & Munich Center for Machine Learning(慕尼黑工业大学与慕尼黑机器学习中心) New York University(纽约大学) Medical University of South Carolina(南卡罗来纳医科大学) The LLM Data Company BenchFlow Independent Researcher(独立研究员) Northeastern University(东北大学) University of Wisconsin–Madison(威斯康星大学麦迪逊分校) University of Washington(华盛顿大学) TU Darmstadt(达姆施塔特工业大学) University of Southern California(南加州大学) UC San Diego(加州大学圣地亚哥分校) Amazon(亚马逊) Microsoft(微软) Korea University(高丽大学) Cornell Tech(康奈尔科技) University of Michigan(密歇根大学)

AI总结 提出全开放数据筛选流水线,通过100多次消融实验研究任务来源与多样性,构建10万样本训练集,在7个智能体基准上平均44.8%准确率,较最强开源模型提升3.9个百分点。

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.24418 2026-06-24 cs.LG stat.ML 新提交

Data Augmentation: A Fourier Analysis Perspective

数据增强:傅里叶分析视角

Behrooz Tahmasebi, Melanie Weber, Stefanie Jegelka

机构 * Harvard John A. Paulson School of Engineering and Applied Sciences, Harvard University(哈佛大学约翰·A·保尔森工程与应用科学学院) Technical University of Munich (CIT, MCML, MDSI) and MIT Computer Science and Artificial Intelligence Laboratory (CSAIL)(慕尼黑工业大学(CIT、MCML、MDSI)和麻省理工学院计算机科学与人工智能实验室(CSAIL))

AI总结 通过傅里叶分析和有限群表示论,研究部分数据增强能否达到与完全增强相同的统计收益,证明在广泛学习问题中部分增强可达到极小极大最优率,并给出精确不变性需要全群平均的不可行性结果。

Comments 42 pages, 1 figure. Published at COLT 2026

Journal ref Conference on Learning Theory (COLT) 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.24093 2026-06-24 cs.CL cs.AI 新提交

Predicting Poets' Origins from Verse: A Computational Analysis of Regional Linguistic Fingerprints in the Complete Tang Poems

从诗句预测诗人籍贯:全唐诗中地域语言指纹的计算分析

Chi-Sheng Chen, Hung-Yun Liu

机构 * Harvard University(哈佛大学) University of Washington(华盛顿大学)

AI总结 通过全唐诗和CBDB数据,使用字符n-gram TF-IDF和领域特征,以多类分类预测诗人籍贯,发现语言距离随地理距离衰减、南北可分性随时间变化等历史意义。

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.23913 2026-06-24 cs.LG 新提交

Closing the Loop: Formally Verified Law as a Reward Signal for Self-Improving Legal AI

闭环:形式化验证的法律作为自我改进法律AI的奖励信号

Armin Heydari, Torben Leowald

机构 * Harvard University(哈佛大学) Columbia University(哥伦比亚大学)

AI总结 提出一种架构,通过LLM驱动的形式化转换和验证内核,为法律AI提供可验证的奖励信号,实现闭环强化学习,并在多个法律领域展示其优势。

Comments 14 pages, no figures

详情

展开后加载摘要…

URL PDF HTML 收藏
2501.02378 2026-06-24 cs.LG q-bio.NC stat.ML

A ghost mechanism: An analytical model of abrupt learning in recurrent networks

鬼机制:递归网络中突然学习的分析模型

Fatih Dinc, Ege Cirakman, Bariscan Kurtkaya, Mert Yuksekgonul, Yiqi Jiang, Mark J. Schnitzer, Hidenori Tanaka

机构 * Kavli Institute for Theoretical Physics, University of California, Santa Barbara, CA 93106, USA Geometric Intelligence Lab, University of California, Santa Barbara, CA 93106, USA CNC Program, Stanford University, Stanford, CA 94305, USA Physics of Artificial Intelligence Group, NTT Research Inc., Sunnyvale, CA 94085, USA KUIS AI, Department of Computer Engineering, Koc University, Istanbul, Turkey Computer Science, Stanford University, Stanford, CA 94305, USA James H. Clark Center for Biomedical Engineering \& Sciences, Stanford University, Stanford Howard Hughes Medical Institute, Stanford University, Stanford, CA 94305, USA CBS-NTT Program in Physics of Intelligence, Harvard University, Cambridge, MA 94305, USA

AI总结 研究揭示递归网络中突然学习现象的机制,通过鬼机制分析动态系统在鞍点消失 bifurcation 附近出现的短暂减速,推导出一维规范形式,揭示学习过程受单尺度参数控制,并提出通过增加可训练秩和降低输出置信度来解决学习困难的方法。

Comments to appear in Physical Review X

详情

展开后加载摘要…

URL PDF HTML 收藏
2603.10044 2026-06-24 cs.AI cs.CL cs.CY cs.LG

Safety Under Scaffolding: How Evaluation Conditions Shape Measured Safety

脚手架下的安全性:评估条件如何影响测量的安全性

David Gringras

机构 * Harvard University(哈佛大学) MIT(麻省理工学院)

AI总结 本研究通过62,808次盲法预注册评估,测试了六种前沿模型在四种部署配置下的安全性,发现脚手架架构对安全性影响较小,而格式转换(如选择题与开放式问题)可导致5-20个百分点的测量差异,且模型-脚手架间存在显著异质性,质疑了单一综合安全性分数的实用性。

Comments 74 pages including appendices. 6 frontier models, 62,808 primary observations (~89k total). Pre-registered: OSF DOI 10.17605/OSF.IO/CJW92. Code and data: https://github.com/davidgringras/safety-under-scaffolding

详情

展开后加载摘要…

URL PDF HTML 收藏
2507.11768 2026-06-24 stat.ML cs.LG 版本更新

LLMs are Bayesian, In Expectation, Not in Realization

LLM是贝叶斯的:在期望中,而非在实现中

Leon Chlon, Fatima Sheaib, Zein Khamis, Maggie Chlon, Mahdi El Zein, MarcAntonio M. Awada

机构 * Hassana Labs(哈萨纳实验室) Harvard University(哈佛大学)

AI总结 针对上下文学习的贝叶斯解释面临顺序不变性反驳,本文通过预序编码长度分解和实验证明,Transformer在期望上接近贝叶斯后验预测,但具体实现不要求交换性。

详情

展开后加载摘要…

URL PDF HTML 收藏
2510.04033 2026-06-24 cs.AI 版本更新

A global log for medical AI

医疗AI的全局日志

Ayush Noori, Aaron E. Boussina, Hai Ho Bich, James Anibal, Julia Maslinski, Manuel Burger, Martin Faltys, Adam Rodman, Alan Karthikesalingam, Alessandro Blasimme, Annelia Itwaru, Ben Kaplan, Bilal A. Mateen, Christopher A. Longhurst, Daniel Yang, Dave deBronkart, Effy Vayena, Fedor Sergeev, Gauden Galea, Ha Thi Hai Duong, Harold F. Wolf, Jacob Waxman, Joerg C. Schefold, Joshua C. Mandel, Juliana Rotich, Kenneth D. Mandl, Lily Poursoltan, Maryam Mustafa, Melissa Miles, Nigam H. Shah, Noa Dagan, Pavan Bodanki, Peter Lee, Philipp Koralus, Prathamesh Parchure, Prem Timsina, Ran D. Balicer, Robert Korom, Scott Mahoney, Seth Hain, Tien Yin Wong, Trevor Mundel, Vivek Natarajan, Ankit Sakhuja, Benjamin Glicksberg, C. Louise Thwaites, Gunnar Rätsch, Karandeep Singh, David A. Clifton, Isaac S. Kohane, Marinka Zitnik

机构 * Harvard Medical School(哈佛医学院) University of Oxford(牛津大学) Institute for Ethics in AI(人工智能伦理研究所) Cosmos Institute(宇宙研究所) Harvard Medical School and Clalit Research Institute(哈佛医学院和Clalit研究机构) University of California, San Diego(加州大学圣地亚哥分校) Joan and Irwin Jacobs Center for Health Innovation(乔安和伊万·雅各布健康创新中心) Oxford University Clinical Research Unit(牛津大学临床研究中心) Icahn School of Medicine at Mount Sinai(辛格纳医学中心) The Hasso Plattner Institute for Digital Health at Mount Sinai(辛格纳医学中心数字健康研究所) ETH Zurich(苏黎世联邦理工学院) University Hospital, University of Bern(伯恩大学医院) Beth Israel Deaconess Medical Center(贝斯以色列医疗中心) Google DeepMind(谷歌DeepMind) The Mount Sinai AI Assurance Lab(辛格纳医学中心人工智能保证实验室) University of Birmingham(伯明翰大学) PATH(PATH组织) Seattle Children’s Hospital(西雅图儿童医院) Department of Pediatrics, University of California, San Diego(加州大学圣地亚哥分校儿科部) Kaiser Foundation Health and Hospitals(凯撒基金会健康与医院) e-Patient Dave, LLC(e-Patient Dave公司) Regional Office for Europe, World Health Organization(世界卫生组织欧洲地区办公室) University of Malta(马耳他大学) Centre for Tropical Medicine and Global Health, University of Oxford(牛津大学热带医学与全球健康中心) Healthcare Information and Management Systems Society(医疗信息与管理系统协会) Clalit Research Institute, Innovation Division, Clalit Health Services(Clalit研究机构创新部门,Clalit健康服务) Microsoft Research(微软研究院) Gates Foundation(比尔及梅琳达·盖茨基金会) Computational Health Informatics Program, Boston Children’s Hospital(波士顿儿童医院计算健康信息学项目)

AI总结 提出MedLog协议,为医疗AI系统提供事件级日志记录,包含九个核心字段,并在多个部署中验证其监测模型行为、工作流交互及下游结果的能力。

Comments MedLog website: https://medlogprotocol.ai

详情

展开后加载摘要…

URL PDF HTML 收藏