arXivDaily arXiv每日学术速递 周一至周五更新

高校专区

Peking University(北京大学)

2026-08-18 至 2026-08-18 共收录 13
2608.12385 2026-08-18 cs.AI 版本更新

Decode-Branch Transformers: Decoupling the Primary Prefill Path from Additional Decode Computation

双流式Transformer:将主预填路径与额外解码计算解耦

Liming Liu, Mingze Wang, Tuo Zhao

机构 * Georgia Institute of Technology(佐治亚理工学院) Peking University(北京大学)

AI总结 该研究提出双流式Transformer,将主预填路径与额外解码计算解耦,通过共享权重与缓存降低推理成本,在多架构与数据配置下实现更低验证损失,且可灵活分配MoE专家预算。

Comments 19 pages

详情

展开后加载摘要…

URL PDF HTML 收藏
2608.03025 2026-08-18 cs.AI 版本更新

DiffImaginE: Imagine to Verify Entity Types with Diffusion

DiffImaginE:通过想象验证实体类型

Feng Zhang, Feiyu Han, Rongxin Yang, Yang Liu, Yancheng Chen, Rui Wang, Yingguang Yang, Tian Xueyun, Chongyang Zhang, Hao Zheng, Xu Kefu, Congjing Ran, Fuhai Chen, Bin Chong

机构 * Fuzhou University(福州大学) Chinese Academy of Sciences(中国科学院) Peking University(北京大学) Alibaba Group(阿里巴巴集团) University of Science and Technology of China(中国科学技术大学) Fullive Innovation (Beijing) AI Technology Co., Ltd.(福莱创新(北京)人工智能科技有限公司) Baidu(百度) Wuhan University(武汉大学)

AI总结 DiffImaginE将多模态命名实体识别的类型验证建模为条件潜在扩散推理,在Twitter-2015和Twitter-2017数据集上,相比确定性对照模型取得了一致的性能提升。

详情

展开后加载摘要…

URL PDF HTML 收藏
2608.02474 2026-08-18 cs.CV 版本更新

EchoCache: Energy-Guided Cross-Modal Caching for Efficient Audio-Driven Video Generation

EchoCache:面向高效音频驱动视频生成的能量引导跨模态缓存

Jiayu Chen, Xiaoyu Wu, Rongshan Gao, Maoliang Li, Zihao Zheng, Xinhao Sun, Hailong Zou, Guojie Luo, Xiang Chen

机构 * Peking University(北京大学) Taiyuan University of Technology(太原理工大学)

AI总结 EchoCache是一种能量引导跨模态缓存框架,通过利用音频时频能量锚点与动态缓存机制,在保持音频驱动视频生成质量的同时,实现了2.46倍的推理加速,优化了延迟-质量权衡。

Comments EchoCache is honored to be accepted by ACM MM 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.12485 2026-08-18 cs.LG cs.AI 版本更新

Speculative Rollback Correction for Quality-Diverse Web Agent Imitation

面向质量多样性的Web智能体模仿的推测性回滚修正

Longkun Hao, Hongyu Lin, Hao Li, Zhuowen Liu, Zhichao Yang, Haojie Hao, Dongshuo Huang, Haitao Yang, Hongyu Ge, Ming jie Xie, Yanjun Wu, Zi Hao Yin, Yan Bai, Yihang Lou

机构 * Beihang University(北京航空航天大学) Institute of Software, Chinese Academy of Sciences(中国科学院软件研究所) The Hong Kong University of Science and Technology(香港科技大学) Northwestern Polytechnical University(西北工业大学) Tsinghua University(清华大学) The Hong Kong University of Science and Technology (Guangzhou)(香港科技大学(广州)) Peking University(北京大学)

AI总结 提出推测性回滚修正(SRC)框架,通过固定视野分支审查和回滚机制,在减少教师查询的同时保持轨迹多样性,在WebArena-Infinity上收集了977条通过验证的轨迹和9183个下一步动作示例。

详情

展开后加载摘要…

URL PDF HTML 收藏
2510.08759 2026-08-18 cs.CV cs.RO 版本更新

Dissecting Embodied Abilities in Multimodal Language Models through Skill-level Evaluation and Diagnosis

通过技能级评估与诊断解构多模态语言模型的具身能力

Yu Qi, Haibo Zhao, Ziyu Guo, Siyuan Ma, Ziyan Chen, Yaokun Han, Renrui Zhang, Zitiantao Lin, Yizhe Zhu, Shiji Xin, Yijian Huang, Boce Hu, Kai Cheng, Peiheng Wang, Jiazheng Liu, Jiayi Zhang, Yizhe Zhu, Wenqing Wang, Yiran Qin, Haojie Huang, Lawson L.S. Wong

机构 * Northeastern University, Boston, MA, USA The Chinese University of Hong Kong, Hong Kong, China Peking University, Beijing, China Westlake University, Hangzhou, China Harvard University, Cambridge, MA, USA Purdue University, West Lafayette, IN, USA University of Oxford, Oxford, United Kingdom

AI总结 本文提出BEAR基准,通过分解具身任务为14个原子技能进行细粒度评估,发现感知能力是推理失败的主要瓶颈,并提出BEAR-Agent多模态对话代理,显著提升具身技能性能。

Comments Accepted to ICML 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2601.21351 2026-08-18 cs.LG cs.AI 版本更新

Analytical Provisioning for Attention-FFN Disaggregated LLM Serving under Stochastic Workloads

注意力-前馈网络解耦大语言模型服务的分析资源配置

Chendong Song, Meixuan Wang, Hang Zhou, Hong Liang, Yuan Lyu, Zixi Chen, Yuwei Fan, Zijie Zhou

机构 * Dept. of Industrial Engineering and Decision Analytics HKUST(工业工程与决策分析系香港科技大学) Dept. of Computer Science and Technology Tsinghua University(计算机科学与技术系清华大学) IIIS Tsinghua University(清华大学信息学院) Huawei Hong Kong Research Center(华为香港研发中心) School of Mathematical Sciences Peking University(北京大学数学科学学院)

AI总结 本文提出在随机负载下,针对注意力-前馈网络解耦架构的分析资源配置框架,通过考虑工作负载统计量θ,确定最优注意力与前馈网络比例,减少阻塞和设备空闲时间。

Comments Submitted to Neurips 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2510.22819 2026-08-18 cs.LG 版本更新

Last-Iterate Analyses of FTRL with the 1/2-Tsallis Entropy in Stochastic Bandits

FTRL在随机老虎机中使用1/2-Tsallis熵的最后迭代分析

Jingxin Zhan, Yuze Han, Zhihua Zhang

机构 * School of Mathematical Sciences, Peking University(北京大学数学科学学院) Center for Applied Statistics and School of Statistics, Renmin University of China(中国人民大学统计学院)

AI总结 本文研究了使用1/2-Tsallis熵正则化器的FTRL算法,证明了其最后迭代收敛率为t^{-1/2},并验证了对数遗憾与该收敛率的对应关系。

Comments Substantially revised; adds $\mathcal{O}(t^{-1})$ simple-regret upper and lower bounds and allows multiple optimal arms in the upper bound

详情

展开后加载摘要…

URL PDF HTML 收藏
2603.27959 2026-08-18 cs.CV 版本更新

MathGen: Revealing the Illusion of Mathematical Competence through Text-to-Image Generation

MathGen:通过文本到图像生成揭示数学能力的幻觉

Ruiyao Liu, Hui Shen, Ping Zhang, Yunta Hsieh, Yifan Zhang, Jing Xu, Qi Han, Junchen Li, Jiawei Lu, Jianing Ma, Jiaqi Mo, Sicheng Chen, Zhen Zhang, Zhongwei Wan, Jing Xiong, Xin Wang, Ziyuan Liu, Hangrui Cao, Ngai Wong

机构 * University of Pennsylvania(宾夕法尼亚大学) University of Michigan(密歇根大学) The Ohio State University(俄亥俄州立大学) USTC(中国科学技术大学) City University of Hong Kong(香港城市大学) University of Wisconsin(威斯康星大学) UCSB(加州大学圣塔芭芭拉分校) University of Hong Kong(香港大学) Peking University(北京大学) Carnegie Mellon University(卡内基梅隆大学)

AI总结 MathGen通过900道跨七个核心领域的数学问题,评估生成模型在视觉化数学解答中的准确性,发现现有模型在数学 fidelity 上存在显著瓶颈。

详情

展开后加载摘要…

URL PDF HTML 收藏
2603.20253 2026-08-18 physics.comp-ph cs.AI cs.DC cs.LG 版本更新

SimulCost: A Cost-Aware Benchmark and Toolkit for Automating Physics Simulations with LLMs

SimulCost: 一个用于自动化物理模拟的代价感知基准与工具包

Yadi Cao, Sicheng Lai, Jiahe Huang, Yang Zhang, Zach Lawrence, Rohan Bhakta, Izzy F. Thomas, Mingyun Cao, Chung-Hao Tsai, Zihao Zhou, Yidong Zhao, Hao Liu, Alessandro Marinoni, Alexey Arefiev, Rose Yu

机构 * University of California San Diego(加州大学圣地亚哥分校) The Chinese University of Hong Kong, Shenzhen(香港中文大学(深圳)) Peking University(北京大学) University of California, Los Angeles(加州大学洛杉矶分校) California Institute of Technology(加州理工学院) ETH Zurich(苏黎世联邦理工学院)

AI总结 针对现有LLM评估忽略工具使用代价的问题,提出SimulCost基准,通过单轮和多轮参数调优任务比较LLM与传统扫描方法在准确性和计算代价上的表现,发现LLM在高精度任务中初始猜测不可靠且多轮模式效率更低。

Comments post conference revision version at ICML; update: removed CGYRO due to bug in cases search. Will add back soon; Make the title consistent w/ pdf

详情

展开后加载摘要…

URL PDF HTML 收藏
2603.16959 2026-08-18 cond-mat.mtrl-sci cs.AI 版本更新

Data-knowledge dual-driven intelligent framework for full-chain, experiment-efficient synthesis of 2D dendrites

机器智能支持二维树突合成的全流程

Wenqiang Huang, Xuhang Gu, Susu Fang, Shen'ao Xue, Huanhuan Xing, Junjie Jiang, Junying Zhang, Shen Zhou, Zheng Luo, Jin Zhang, Fangping Ouyang, Shanshan Wang

机构 * School of Physics, Hunan Key Laboratory for Super-Microstructure and Ultrafast Process, and Hunan Key Laboratory of Nanophotonics and Devices, Central South University(物理学院,湖南超级微结构与超快工艺重点实验室,湖南纳米光子学与器件重点实验室,中南大学) College of Aerospace Science and Engineering, National University of Defense Technology(航空科学与工程学院,国防科技大学) School of Advanced Materials, Guangdong Provincial Key Laboratory of Nano-Micro Materials Research, Peking University Shenzhen Graduate School(先进材料学院,广东省纳米-微米材料研究重点实验室,北京大学深圳研究生院) College of Science, National University of Defense Technology(科学学院,国防科技大学) School of Physics and Technology, and Xinjiang Key Laboratory of Solid-State Physics and Devices, Xinjiang University(物理与技术学院,新疆固态物理与器件重点实验室,新疆大学) State Key Laboratory of Powder Metallurgy, and Powder Metallurgy Research Institute, Central South University(粉末冶金国家重点实验室,中南大学粉末冶金研究所)

AI总结 本文提出基于机器学习的材料合成全流程框架,通过主动学习优化实验参数,结合数据增强策略提升预测精度,并构建双驱动机制模型揭示多参数对产物形态的协同作用。

Comments 57 pages, 30 figures

详情

展开后加载摘要…

URL PDF HTML 收藏
2603.12478 2026-08-18 cs.CV cs.LG 版本更新

Less Data, Faster Convergence: Goal-Driven Data Optimization for Multimodal Instruction Tuning

数据更少,收敛更快:面向多模态指令微调的目标驱动数据优化

Rujie Wu, Haozhe Zhao, Hai Ci, Yizhou Wang

机构 * Peking University(北京大学) University of Illinois at Urbana-Champaign(伊利诺伊大学厄巴纳-香槟分校) National University of Singapore(新加坡国立大学)

AI总结 本文提出目标驱动数据优化框架GDO,通过优化训练样本实现更快收敛和更高精度,适用于多模态指令微调任务。

Comments Accepted to ECCV 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2512.22983 2026-08-18 cs.RO 版本更新

Embodied Robot Manipulation in the Era of Foundation Models: Planning and Learning Perspectives

基础模型时代中的具身机器人操作:规划与学习视角

Shuanghao Bai, Wenxuan Song, Jiayi Chen, Yuheng Ji, Zhide Zhong, Jin Yang, Han Zhao, Wanqi Zhou, Zhe Li, Pengxiang Ding, Cheng Chi, Chang Xu, Xiaolong Zheng, Donglin Wang, Haoang Li, Shanghang Zhang, Badong Chen

机构 * Xi’an Jiaotong Univeristy(西安交通大学) Hong Kong University of Science and Technology (Guangzhou)(香港科技大学(广州)) Chinese Academy of Sciences(中国科学院) Westlake University(西湖大学) Zhejiang University(浙江大学) University of Sydney(悉尼大学) BAAI(百度人工智能研究院) Peking University(北京大学)

AI总结 本文探讨了基础模型时代机器人操作的规划与学习方法,分析了高层推理与低层控制的统一框架,并提出了未来研究方向。

Comments This work is a re-architected core derived from the full survey ( arXiv:2510.10903 (https://arxiv.org/abs/2510.10903) ), refined to highlight the most central themes and representative studies

详情

展开后加载摘要…

URL PDF HTML 收藏
2502.14424 2026-08-18 stat.ML cs.AI cs.LG stat.ME 版本更新

Bringing Generative Learning to Representation Learning: Self-Supervised Transfer Learning as Distribution Matching

将生成式学习引入表征学习:作为分布匹配的自监督迁移学习

Yuling Jiao, Wensen Ma, Defeng Sun, Hansheng Wang, Yang Wang

机构 * School of Artificial Intelligence and School of Mathematics and Statistics, Wuhan University(人工智能学院和数学与统计学学院,武汉大学) School of Mathematics and Statistics, Wuhan University(数学与统计学学院,武汉大学) Department of Applied Mathematics, The Hong Kong Polytechnic University(应用数学系,香港理工大学) Guanghua School of Management, Peking University(光华管理学院,北京大学) Department of Mathematics, The Hong Kong University of Science and Technology(数学系,香港科技大学)

AI总结 该研究将表征学习建模为分布匹配,采用Mallows距离作为差异度量,关联总体目标与类中心分离及分类误差并证明非渐近神经筛保证,经模拟和图像基准验证其具备流形校正等优势。

Comments 70 pages, 5 figures, and 6 tables. Substantially revised version with a new title, an explicit distribution-matching formulation linking generative learning and representation learning, expanded theoretical treatment, additional transfer experiments, and appendices integrated into the main file. Code is available at this https URL (https://github.com/vincen-github/DM)

详情

展开后加载摘要…

URL PDF HTML 收藏