arXivDaily arXiv每日学术速递 周一至周五更新

高校专区

University of Michigan(密歇根大学安娜堡分校)

共收录 1281
2605.17887 2026-05-19 cs.LG cs.AI

Attention Sinks and Outliers in Attention Residuals

注意力沉底与注意力残差中的异常值

Haozheng Luo, Haoran Dai, Shaoyang Zhang, Xi Chen, Eric Hanchen Jiang, Yijiang Li, Jingyuan Huang, Chenghao Qiu, Chenwei Xu, Zhenyu Pan, Haotian Zhang, Binghui Wang, Yan Chen

机构 * Department of Computer Science, Northwestern University(西北大学计算机科学系) Department of Computer Science and Engineering, University of Michigan(密歇根大学计算机科学与工程系) Department of Statistics and Data Science, University of California Los Angeles(加州大学洛杉矶分校统计与数据科学系) Department of Electrical and Computer Engineering, University of California San Diego(加州圣地亚哥大学电气与计算机工程系) Department of Computer Science, Rutgers University-New Brunswick(新泽西州立大学鲁特学院计算机科学系) Department of Computer Science and Engineering, Texas A&M University(德克萨斯农工大学计算机科学与工程系) Department of Computer Science, Columbia University(哥伦比亚大学计算机科学系)

AI总结 本文提出OASIS技术,通过层间空信号来解决注意力残差架构中注意力沉底、激活异常值以及推理稳定性下降的问题,通过双归一化设计和实验验证提升了模型的结构鲁棒性和量化鲁棒性。

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.17653 2026-05-19 cs.LG cs.AI

LLMForge: Multi-Backend Hardware-Aware Neural Architecture Search with Infinite-Head Attention for Edge Language Models

LLMForge: 多后端硬件感知的神经架构搜索与无限头注意力用于边缘语言模型

Xinting Jiang, Junyi Luo, Ruichen Qi, Kauna Lei, Ben Laurie, Gregory Kielian, Mehdi Saligane

机构 * Brown University(布朗大学) University of Michigan(密歇根大学) Google Research(谷歌研究)

AI总结 本文提出LLMForge,一种多后端硬件感知的神经架构搜索框架,通过无限头注意力扩展了每层注意力配置空间,并结合Forge-Former和Forge-DSE实现了高效的边缘语言模型架构搜索,最终在不同硬件子系统上获得了不同形状的架构,展示了在不同性能指标上的优化效果。

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.17590 2026-05-19 cs.LG math.OC

Form and Function: Machine Unlearning as a Problem of Misaligned States

形式与功能:将机器去学习视为不一致状态的问题

Kennon Stewart

机构 * Second Street Labs, Detroit, MI, USA(第二街实验室,密歇根州底特律) Department of Statistics, University of Michigan, Ann Arbor, MI, USA(密歇根大学统计系,密歇根州安阿伯)

AI总结 本文提出将在线L-BFGS的机器去学习问题建模为反事实状态对齐问题,通过引入状态感知度量和反事实 oracle 模型,证明去学习不仅仅是参数修正问题,还需要与可实现的反事实优化器状态对齐。

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.17564 2026-05-19 cs.CV

A Conditional U-Net Pipeline with Pre- and Post-Processing for Aerial RGB-to-Thermal Image Translation

具有预处理和后处理的条件U-Net管道用于航空RGB到热图像转换

Tseten Sherpa, Sikandar Ali, Shubham Parab, Haoyun Feng, Matthew Dennis, Keenan Gibbons, Verrah Otiende, Geoffrey H. Siwo

机构 * Department of Data Science, University of Michigan, Ann Arbor, MI, USA(数据科学系,密歇根大学,安阿伯,MI,美国) Department of Information Science, University of Michigan, Ann Arbor, MI, USA(信息科学系,密歇根大学,安阿伯,MI,美国) Department of Computer Science, University of Michigan, Ann Arbor, MI, USA(计算机科学系,密歇根大学,安阿伯,MI,美国) Arcknow, New York, USA(Arcknow,纽约,美国) School of Environmental Sustainability, University of Michigan, Ann Arbor, MI, USA(可持续环境学院,密歇根大学,安阿伯,MI,美国) SmithGroup, Ann Arbor, MI, USA(SmithGroup,安阿伯,MI,美国) Michigan Institute for Data and AI in Society (MIDAS), University of Michigan, Ann Arbor, MI, USA(密歇根数据与人工智能社会研究院(MIDAS),密歇根大学,安阿伯,MI,美国) United States International University (USIU), Nairobi, Kenya(美国国际大学(USIU),内罗毕,肯尼亚) Department of Learning Health Sciences, University of Michigan Medical School, Ann Arbor, MI, USA(学习健康科学系,密歇根大学医学院,安阿伯,MI,美国) Department of Pharmacology, University of Michigan Medical School, Ann Arbor, MI, USA(药理学系,密歇根大学医学院,安阿伯,MI,美国) Center for Global Health Equity, University of Michigan, Ann Arbor, MI, USA(全球健康公平中心,密歇根大学,安阿伯,MI,美国)

AI总结 本文提出了一种基于条件U-Net的简单架构,结合天气数据和针对性预处理与后处理技术,以提高航空RGB到热图像转换的性能,实验结果显示其在PSNR、SSIM和LPIPS指标上优于现有方法。

Comments 8 pages, 7 figures, NeurIPS 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2603.11276 2026-05-19 stat.ML cs.LG

RIE-Greedy: Regularization-Induced Exploration for Contextual Bandits

RIE-Greedy: 基于正则化的探索策略用于上下文老虎机

Tong Li, Thiago de Queiroz Casanova, Eric M. Schwartz, Victor Kostyuk, Dehan Kong, Joseph J. Williams

机构 * University of Toronto(多伦多大学) University of Michigan(密歇根大学)

AI总结 本文提出了一种基于正则化的探索策略(RIE-Greedy),利用模型拟合过程中的随机性作为内在探索源,理论证明其在两臂老虎机情况下等价于Thompson Sampling,并在大规模商业环境中优于epsilon-greedy等基准方法。

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.17284 2026-05-19 cs.CV cs.AI cs.LG cs.RO

CLAP: Contrastive Latent-space Prompt Optimization for End-to-end Autonomous Driving

CLAP:用于端到端自动驾驶的对比潜在空间提示优化

Ruiyang Zhu, Yuehan He, Boyuan Zheng, Zesen Zhao, Ahmad Chalhoub, Qingzhao Zhang, Z. Morley Mao

机构 * University of Michigan(密歇根大学) University of Arizona(亚利桑那大学)

AI总结 本文提出CLAP方法,通过对比潜在空间提示优化解决自动驾驶中罕见但安全关键的长尾场景问题,利用V2X通信获取数据并优化提示,从而提升规划性能。

Comments 9 pages + appendix

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.16785 2026-05-19 cs.CV cs.AI

Encoding Robust Topological Signatures for Hyperdimensional Computing

为超维计算编码鲁棒的拓扑特征

Arpan Kusari

机构 * University of Michigan Transportation Research Institute(密歇根大学交通研究院) University of Michigan(密歇根大学)

AI总结 本文提出了一种基于拓扑特征的超维计算方法,通过提取离散拓扑原始特征并结合RTS不变的形状签名,提高了超维计算在旋转、噪声和遮挡等扰动下的鲁棒性,实验表明其在多个数据集上优于传统方法。

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.16644 2026-05-19 eess.SY cs.LG cs.SY math.OC stat.ML

The Score Kalman Filter

分数卡尔曼滤波器

Kaito Iwasaki, Anthony Bloch, Taeyoung Lee, Maani Ghaffari

机构 * Department of Mathematics University of Michigan(数学系密歇根大学) Department of Mechanical and Aerospace Engineering George Washington University(机械与航空航天工程系乔治华盛顿大学) Department of Naval Architecture & Marine Engineering and Department of Robotics University of Michigan(海军建筑与海洋工程系和机器人系密歇根大学)

AI总结 本文提出分数卡尔曼滤波器,通过结合分数匹配与斯蒂恩恒等式,避免了分区函数的计算,实现了非线性系统的高效滤波,适用于高维问题。

Comments 56 pages, 27 figures

详情

展开后加载摘要…

URL PDF HTML 收藏
2411.17917 2026-05-19 cs.CV cs.RO

DECODE: Domain-aware Continual Domain Expansion for Motion Prediction

DECODE:面向领域的持续领域扩展用于运动预测

Boqi Li, Haojie Zhu, Henry X. Liu

机构 * Department of Civil and Environmental Engineering, University of Michigan(密歇根大学土木与环境工程系)

AI总结 DECODE提出一种持续学习框架,通过预训练模型逐步扩展领域专用模型,结合超网络和流机制实现高效模型选择与不确定性估计,有效降低遗忘率并提升预测精度。

Comments This work has been published in IEEE TPAMI Early Access

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.16588 2026-05-19 cs.RO cs.SY eess.SY

Policy Library CBF: Finite-Horizon Safety at Runtime via Parallel Rollouts

策略库CBF:通过并行滚动预测实现有限时间范围内的运行时安全

Taekyung Kim, Hideki Okamoto, Bardh Hoxha, Georgios Fainekos, Dimitra Panagou

机构 * Department of Robotics(机器人学系) Department of Aerospace Engineering(航空航天工程系) University of Michigan(密歇根大学) Toyota Motor North America, Research & Development(丰田美国北美洲研发部门)

AI总结 本文提出PL-CBF,通过并行有限时间滚动预测评估备用策略库,选择最安全模式并最小修改名义策略以确保安全,实验显示在保持毫秒级运行时间的同时提升了安全覆盖率。

Comments Project page: https://www.taekyung.me/plcbf

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.16278 2026-05-19 cs.CY cs.AI cs.HC

Keeping an Eye on AI: A Framework for Effective Human Oversight of AI Systems

关注人工智能:一种有效的人工智能系统人类监督的框架

Susanne Gaube, Markus Langer, Tim Miller, Kevin Baum, Raimund Dachselt, Anna Maria Feit, Ujwal Gadiraju, Harmanpreet Kaur, Mark T. Keane, Richard Landers, Johann Laux, Q. Vera Liao, Brian Lim, Linda Onnasch, Tim Schrills, Liz Sonenberg, Chenhao Tan, Nava Tintarev, Ziang Xiao, Hanwei Zhang

机构 * University College London(伦敦大学) University of Freiburg(弗赖堡大学) University of Queensland(昆士兰大学) Saarland University(萨尔兰大学) TU Dresden(德累斯顿技术大学) Delft University of Technology(代尔夫特理工大学) University of Minnesota(明尼苏达大学) University College Dublin(都柏林大学) University of Oxford(牛津大学) University of Michigan(密歇根大学) National University of Singapore(新加坡国立大学) Technische Universität Berlin(柏林技术大学) University of Lübeck(吕贝克大学) University of Melbourne(墨尔本大学) University of Chicago(芝加哥大学) Maastricht University(马斯特里赫特大学)

AI总结 本文提出一个跨学科框架,用于有效的人工智能系统人类监督,定义了监督架构和流程,并探讨了该领域需要考虑的开放性研究挑战。

Comments The conceptual analysis for this work was undertaken by the authors at Dagstuhl seminar 25272 'Challenges of Human Oversight: Achieving Human Control of AI-Based Systems' (https://www.dagstuhl.de/25272), held at Schloss Dagstuhl (June 29th-July 4th, 2025)

详情

展开后加载摘要…

URL PDF HTML 收藏
2508.01608 2026-05-19 cs.CV

From Pixels to Places: A Systematic Benchmark for Evaluating Image Geolocalization Ability in Large Language Models

从像素到地点:一个系统性基准,用于评估大语言模型中的图像地理定位能力

Lingyao Li, Runlong Yu, Qikai Hu, Bowei Li, Min Deng, Yang Zhou, Xiaowei Jia

机构 * University of South Florida Tampa USA University of Alabama Tuscaloosa USA University of Michigan Ann Arbor USA Texas Tech University Lubbock USA Texas A \& M University College Station USA University of Pittsburgh Pittsburgh USA University of South Florida University of Alabama University of Michigan Texas Tech University Texas A \& M University University of Pittsburgh

AI总结 本文提出IMAGEO-Bench基准,系统评估大语言模型在图像地理定位中的准确性、距离误差、地理偏见和推理过程,揭示闭源模型在高资源区域表现优于欠代表区域。

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.09403 2026-05-18 cs.LG cs.AI cs.NE

Sparsity Moves Computation: How FFN Architecture Reshapes Attention in Small Transformers

稀疏性推动计算:FFN架构如何重塑小规模Transformer中的注意力

Gabriel Smithline, Chris Mascioli

机构 * University of Michigan(密歇根大学)

AI总结 研究通过单层Transformer在数字加法、模运算和直方图计数中发现,稀疏MoE路由将计算从FFN转移到注意力,且GLU门控旋转任务相关傅里叶结构至分布式子空间。

Comments Preprint

详情

展开后加载摘要…

URL PDF HTML 收藏
2512.00417 2026-05-18 cs.CL

CryptoBench: A Dynamic Benchmark for Expert-Level Evaluation of LLM Agents in Cryptocurrency

CryptoBench: 一种动态基准,用于评估LLM代理在加密货币领域的专家级能力

Jiacheng Guo, Suozhi Huang, Zixin Yao, Yifan Zhang, Yifu Lu, Jiashuo Liu, Zihao Li, Nicholas Deng, Qixin Xiao, Jia Tian, Kanghong Zhan, Tianyi Li, Xiaochen Liu, Jason Ge, Chaoyang He, Kaixuan Huang, Lin Yang, Wenhao Huang, Mengdi Wang

机构 * Princeton University(普林斯顿大学) Zenith Lab(Zenith实验室) University of California, Los Angeles(加州大学洛杉矶分校) University of California, Berkeley(加州大学伯克利分校) University of Michigan(密歇根大学)

AI总结 本文提出CryptoBench,首个专家 curated 的动态基准,用于严格评估LLM在加密货币领域的真实能力。通过50题/月的动态任务,细分子类评估数据获取与预测能力,揭示LLM在检索与预测上的不平衡问题。

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.15604 2026-05-18 cs.LG cs.CL

VSPO: Vector-Steered Policy Optimization for Behavioral Control

VSPO:用于行为控制的向量引导策略优化

Xuechen Zhang, Zijian Huang, Kai Yang, Weijia Zhang, Jiasi Chen, Samet Oymak

机构 * University of Michigan(密歇根大学)

AI总结 VSPO通过引入与目标行为关联的引导向量,控制生成轨迹的行为强度,解决多目标优化中的稀疏奖励问题,提升策略优化效率。

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.15565 2026-05-18 cs.LG cs.AI

AstraFlow: Dataflow-Oriented Reinforcement Learning for Agentic LLMs

AstraFlow:面向代理大语言模型的数据流强化学习

Haizhong Zheng, Yizhuo Di, Jiahui Wang, Shuowei Jin, Xueshen Liu, Yongji Wu, Z. Morley Mao, Ion Stoica, Jiawei Zhao, Beidi Chen

机构 * Carnegie Mellon University(卡内基梅隆大学) University of Michigan(密歇根大学) UC Berkeley(加州大学伯克利分校) Meta

AI总结 AstraFlow通过数据流导向的强化学习系统,实现复杂多策略协作训练和高效利用异构计算资源,提升代理LLM的推理与工具使用能力。

详情

展开后加载摘要…

URL PDF HTML 收藏
2507.01909 2026-05-15 cs.CV

Modality-agnostic, patient-specific digital twins modeling temporally varying digestive motion

模态无关、患者特异性的数字双胞胎建模时变消化运动

Jorge Tapias Gomez, Nishant Nadkarni, Lando S. Bosma, Jue Jiang, Ergys D. Subashi, William P. Segars, James M. Balter, Mert R Sabuncu, Neelam Tyagi, Harini Veeraraghavan

机构 * Computer and Information Science, Cornell University(康奈尔大学计算机与信息科学系) Department of Medical Physics, Memorial Sloan Kettering Cancer Center(纪念斯隆凯特琳癌症中心医学物理系) University Medical Center Utrecht(乌得勒支大学医学中心) Department of Radiation Physics, University of Texas MD Anderson Cancer Center(德克萨斯大学MD安德森癌症中心放射物理系) Carl E. Ravin Advanced Imaging Laboratories and Center for Virtual Imaging Trials, Duke University Medical Center(杜克大学医学中心卡尔·E·拉文高级影像实验室和虚拟影像试验中心) Department of Radiation Oncology, University of Michigan(密歇根大学放射肿瘤学系)

AI总结 本文提出一种模态无关的患者特异性数字双胞胎建模方法,用于评估变形图像配准方法的准确性,通过生成21个运动阶段的4D序列,评估DIR方法的性能。

Comments This work is still review, it contains 7 Pages, 6 figures, and 4 tables

Journal ref Phys. Med. Biol. 71 (2026) 015029

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.14301 2026-05-15 cs.LG stat.ML

Language-Induced Priors for Domain Adaptation

领域适应中的语言诱导先验

Qiyuan Chen, Jiayu Zhou, Raed Al Kontar

机构 * University of Michigan(密歇根大学)

AI总结 本文提出语言诱导先验(LIP)用于领域适应,通过利用目标领域的专家描述,结合预训练大语言模型学习偏好,改进源域选择并提升性能。

详情

展开后加载摘要…

URL PDF HTML 收藏
2511.05159 2026-05-15 stat.ML cs.LG

A New Framework for Convex Clustering in Kernel Spaces: Finite Sample Bounds, Consistency and Performance Insights

核空间中凸聚类的新框架:有限样本界、一致性和性能洞察

Shubhayan Pan, Kushal Bose, Debolina Paul, Saptarshi Chakraborty, Swagatam Das

机构 * Indian Statistical Institute, Kolkata(印度统计研究院,加尔各答) Electronics and Communication Sciences Unit, Indian Statistical Institute(印度统计研究院电子与通信科学单位) Department of Statistics, University of Oxford(牛津大学统计系) Department of Statistics, University of Michigan(密歇根大学统计系)

AI总结 本文提出核化凸聚类方法,通过映射数据到RKHS空间,解决线性不可分和非凸数据的聚类问题,理论分析和实验验证显示其优于现有方法。

详情

展开后加载摘要…

URL PDF HTML 收藏
2508.14950 2026-05-15 eess.IV cs.LG

Potential and challenges of generative adversarial networks for super-resolution in 4D Flow MRI

生成对抗网络在4D流体磁共振成像超分辨率中的潜力与挑战

Oliver Welin Odeback, Arivazhagan Geetha Balasubramanian, Jonas Schollenberger, Edward Ferdiand, Alistair A. Young, C. Alberto Figueroa, Susanne Schnell, Outi Tammisola, Ricardo Vinuesa, Tobias Granberg, Alexander Fyrdahl, David Marlevi

机构 * Surgery, Karolinska Institutet , addressline= Karolinska Universitetssjukhuset Solna (L1:00) , city= Stockholm , postcode= 171 76 , country= Sweden organization= FLOW, Engineering Mechanics, KTH Royal Institute of Technology , addressline= Osquars Backe 18 , city= Stockholm , postcode= 100 44 , country= Sweden organization= Department of Radiology Biomedical Imaging, University of California San Francisco , addressline= 505 Parnassus Avenue , city= San Francisco , postcode= 94143 , state= CA , country= USA organization= Faculty of Informatics, Telkom University , addressline= Jl.Telekomunikasi No. 1, Terusan Buahbatu , city= Bandung , postcode= 40257 , state= West Java , country= Indonesia organization= Auckland Bioengineering Institute, University of Auckland , addressline= Bioengineering House, 70 Symonds St , city= Grafton , postcode= 1010 , country= New Zealand organization= School of Biomedical Engineering \& Imaging Sciences, King's College London , addressline= 1 Lambeth Palace Rd, South Bank , city= London , postcode= SE1 7EU , country= UK organization= Department of Biomedical Engineering, University of Michigan , addressline= 1107 Carl A. Gerstacker Bldg 2200 Bonisteel Blvd. , city= Ann Arbor , postcode= 48109-2099 , state= MI , country= USA organization= Department of Physics, University of Greifswald , addressline= Felix-Hausdorff-Str. 6 , city= Greifswald , postcode= 174 89 , country= Germany organization= Department of Aerospace Engineering, University of Michigan , addressline= 1320 Beal Avenue , city= Ann Arbor , postcode= 48109-2140 , state= MI , country= USA organization= Department of Neuroradiology, Karolinska University Hospital , addressline= Hälsovägen 13, O42 , city= Stockholm , postcode= 141 86 , country= Sweden organization= Department of Clinical Physiology, Karolinska University Hospital , addressline= Eugeniavägen 3, A8:01 , city= Solna , postcode= 171 64 , country= Sweden organization= Institute for Medical Engineering Science, Massachusetts Institute of Technology , addressline= 45 Carleton St , city= Cambridge , postcode= 02142 , state= MA , country= USA

AI总结 本文研究了生成对抗网络在4D流体磁共振成像超分辨率中的应用,通过对比不同对抗损失函数,发现Wasserstein GAN在稳定性和性能上表现最佳,提升了近壁速度恢复效果。

Comments 26 pages, 10 figures

Journal ref Computers in Biology and Medicine 211 (2026) 111745

详情

展开后加载摘要…

URL PDF HTML 收藏
2602.03429 2026-05-14 cs.AI cs.CL cs.HC cs.LG

DiscoverLLM: From Executing Intents to Discovering Them

DiscoverLLM:从执行意图到发现意图

Tae Soo Kim, Yoonjoo Lee, Jaesang Yu, John Joon Young Chung, Juho Kim

机构 * University of Michigan(密歇根大学)

AI总结 DiscoverLLM通过引入新的用户模拟器,帮助用户形成和发现意图,提升任务性能并减少对话长度。

Comments Accepted at ICML 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.13163 2026-05-14 cs.CR cs.CV cs.LG

LoREnc: Low-Rank Encryption for Securing Foundation Models and LoRA Adapters

LoREnc:低秩加密用于保护基础模型和LoRA适配器

Beomjin Ahn, Jungmin Kwon, Chanyong Jung, Jaewook Chung

机构 * Samsung Research(三星研究院) Samsung Electronics(三星电子) Amazon Web Services(亚马逊网络服务) University of Michigan(密歇根大学)

AI总结 LoREnc通过谱截断和补偿技术,在不重新训练的情况下保护基础模型和LoRA适配器,防止模型恢复攻击和知识产权泄露,实验表明其在1%计算开销下有效。

Comments Accepted to ICIP 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.13156 2026-05-14 cs.CV

Dual-Pathway Circuits of Object Hallucination in Vision-Language Models

视觉语言模型中物体幻觉的双路径电路

Jiaxin Liu, Ding Zhong, Yue Wang, Zhidong Yang, Zhaolu Kang, Guangyuan Dong, Qishi Zhan, Pengcheng Fang, Aofan Liu

机构 * UIUC(伊利诺伊大学香槟分校) UMich(密歇根大学) Stanford(斯坦福大学) HKUST(香港科技大学) PKU(北京大学) NUS(新加坡国立大学) Marquette(马quette大学) Southampton(南安普顿大学)

AI总结 研究通过双路径电路分析框架揭示视觉语言模型中物体幻觉的机制,发现视觉接地路径与幻觉路径的交互特性,并通过抑制幻觉路径组件降低幻觉发生率。

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.13117 2026-05-14 cs.RO cs.AI

SECOND-Grasp: Semantic Contact-guided Dexterous Grasping

SECOND-Grasp:语义接触引导的灵巧抓取

Han Yi Shin, Heeju Ko, Jaewon Mun, Qixing Huang, Jaehyeok Lee, Sung June Kim, Honglak Lee, Sujin Jang, Sangpil Kim

机构 * Korea University(韩国大学) University of Texas at Austin(德克萨斯大学奥斯汀分校) University of Michigan(密歇根大学) Hanyang University(翰阳大学)

AI总结 本文提出SEmantic CONtact-guided Dexterous Grasping框架,通过语义推理与物理可行性结合,提升机器人手部抓取的稳定性和语义理解能力。

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.12668 2026-05-14 stat.ML cs.LG

Online Conformal Prediction: Enforcing monotonicity via Online Optimization

在线置信预测:通过在线优化强制单调性

Eduardo Ochoa Rivera, Ambuj Tewari

机构 * University of Michigan(密歇根大学)

AI总结 本文提出两种新的在线置信预测方法,生成不同置信水平下的嵌套预测集,实现整个风险范围的不确定性量化。方法通过在线优化控制后悔,提高统计效率。

详情

展开后加载摘要…

URL PDF HTML 收藏
2604.23887 2026-05-14 cs.CR cs.AI

Evaluation of Prompt Injection Defenses in Large Language Models

对大型语言模型中提示注入防御措施的评估

Priyal Deep, Shane Emmons, Amy Fox, Kyle Bacon, Kelley McAllister, Peter Ortiz, Krisztian Flautner

机构 * Swept AI University of Michigan(密歇根大学)

AI总结 研究通过构建自适应攻击者测试了九种防御配置,发现依赖模型自我保护的防御均失效,而输出过滤通过硬编码规则在应用代码中检查响应,实现了零泄露。

Comments 14 pages, 9 figures

详情

展开后加载摘要…

URL PDF HTML 收藏
2603.22364 2026-05-14 cs.LG cs.AI cs.CV

MCLR: Improving Conditional Modeling via Inter-Class Likelihood-Ratio Maximization and Unifying Classifier-Free Guidance with Alignment Objectives

MCLR:通过类间似然比最大化和统一分类器免费引导与对齐目标改进条件建模

Xiang Li, Yixuan Jia, Xiao Li, Jeffrey A. Fessler, Rongrong Wang, Qing Qu

机构 * University of Michigan(密歇根大学) Michigan State University(密歇根州立大学)

AI总结 本文提出MCLR,通过训练时最大化类间似然比来改进扩散模型的条件建模,实现无引导条件生成的提升,并理论证明CFG是MCLR目标的最优解。

详情

展开后加载摘要…

URL PDF HTML 收藏
2602.02977 2026-05-14 cs.CV cs.AI cs.LG

Aligning Forest and Trees in Images & Long Captions for Visually Grounded Understanding

对图像与长描述中的森林和树木进行对齐以实现视觉基础理解

Byeongju Woo, Zilin Wang, Byeonghyun Pak, Sangwoo Mo, Stella X. Yu

机构 * Agency for Defense Development(国防发展局) University of Michigan(密歇根大学) POSTECH

AI总结 本文提出CAFT模型,通过分层视觉语言学习原理,解决长描述中细节丰富的场景理解问题,实现图像与文本的细粒度对齐,取得六个长文本检索基准的最优性能。

Comments Preprint

详情

展开后加载摘要…

URL PDF HTML 收藏
2506.03120 2026-05-14 stat.AP cs.LG

Validating remotely sensed biomass estimates with forest inventory data in the western US

利用地面调查数据验证美国西部地区遥感生物量估计

Xiuyu Cao, Joseph O. Sexton, Panshi Wang, Dimitrios Gounaridis, Neil H. Carter, Kai Zhu

机构 * School for Environment and Sustainability, University of Michigan(环境可持续发展学院,密歇根大学) terraPulse, Inc.(terraPulse公司)

AI总结 本研究利用美国森林调查数据对terraPulse公司的生物量数据进行独立验证,发现两者在不同尺度上均表现出高度一致性,为全球生物量监测提供了新的基准。

Comments 32 pages, 5 figures

Journal ref Science of Remote Sensing, Volume 13, June 2026, 100441

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.12361 2026-05-13 cs.CL cs.AI cs.IR

MedHopQA: A Disease-Centered Multi-Hop Reasoning Benchmark and Evaluation Framework for LLM-Based Biomedical Question Answering

MedHopQA: 一种以疾病为中心的多跳推理基准和评估框架,用于基于大语言模型的生物医学问答

Rezarta Islamaj, Robert Leaman, Joey Chan, Nicholas Wan, Qiao Jin, Natalie Xie, John Wilbur, Shubo Tian, Lana Yeganova, Po-Ting Lai, Chih-Hsuan Wei, Yifan Yang, Yao Ge, Qingqing Zhu, Zhizheng Wang, Zhiyong Lu

机构 * National Library of Medicine, Division of Intramural Research(国家医学图书馆,院内研究部) University of Illinois at Urbana-Champaign, Department of Computer Science(伊利诺伊大学厄巴纳-香槟分校计算机科学系) University of Michigan Medical School(密歇根大学医学院)

AI总结 MedHopQA旨在通过多跳推理评估大语言模型在生物医学领域的表现,提供了一个以疾病为中心的基准和框架,强调组合推理、抗饱和性和抗污染性。

详情

展开后加载摘要…

URL PDF HTML 收藏