南京大学博士生蒋俊鹏 投稿 量子位 | 公众号 QbitAI 在AI应用中，表格数据的重要性愈发凸显，广泛应用于金融、医疗健康、教育、推荐系统及科学研究领域。 深度神经网络（DNN）凭借其强大的表示学习能力，在表格数据建模上展现出令人瞩目的潜力。 南京大学团队系统介绍了表格表示学习这一研究领域，他们将现有方法按泛化能力划分为三大类： 专用模型 （Specialized）、可迁移模型（Transferable）和通用模型（General） 。 除此之外，他们还比较了DNN与传统方法——树模型的优劣，并剖析表格数据学习中的核心挑战，讨论了集 成学习方法以及开放环境下的表格学习和多模态表格任务等扩展方向。同时，考虑到不同数据集之间方法表现 差异显著，研究团队还探讨了数据集收集、评估与分析的系统策略，旨在建立跨数据集的稳健评估体系。 背景 表格数据本质上是一种 结构化的信息表示方式 ，在组织与表达复杂数据关系方面具有天然优势。 此研究聚焦于 有监督的表格机器学习任务 ，主要包括分类与回归两类常见问题。 除了结构化的组织形式外，表格数据通常还具有 属性类型异质性 ，即包含数值型、类别型或混合型等多种数 据类型，且这些数 ...

明敏 克雷西 发自 凹非寺 量子位 | 公众号 QbitAI 不止于此，寄快递、查快递也更方便了。 对小艺说：给上海市某中心的张三寄快递，他的号码是132-XXXX-XXXX。即可完成下单， 无需再自己手动输入。快递的状态也同样可以在负一屏实时查看，不用再去APP/小程序里自 己查找， 各购物平台购物信息通过手机号聚合到负一屏 ，甚至连 取件码、快递柜位置、待 取件数量等都会在负一屏显示 ，用户无需在短信界面挨个查找，省时省力。 除了上述的点单、查快递的例子外，股市金融、日常出行、生活缴费、运动健康等都能在华 为手机的负一屏上以卡片形式展示出来，各种想要了解的重点信息也都已经被筛选出来。 这就是在HDC 2025上，鸿蒙服务分发展示的最新能力，通过一张张小卡片，就让用户能最 低成本获取到自己需要的信息，让手机更聪明、更懂你。 简单来说，就是把过去的人从手机里找服务，变成服务主动帮你做事情。 这看似是小转变，但背后带来体验服务的升级可不止一丁半点，甚至是鸿蒙生态崛起的重要 窗口。 哦？这怎么说？在HDC2025现场，华为已经揭秘了。 鸿蒙再进化，日常用手机的姿势又被重新定义了！ 试想，在周末走进商场时，结合使用 ...

Core Viewpoint - The article discusses the importance of data-driven Agents in enterprises, emphasizing that a truly effective Agent must understand both the business and decision-making processes within a company [1][8]. Group 1: Importance of Data in Agents - A significant amount of business data within enterprises remains underutilized, highlighting the need for an Agent that can effectively leverage this data [1]. - The article raises questions about what kind of Agent can make complex, non-standard, cross-departmental data useful for businesses [1]. Group 2: Expert Insights - The article features insights from industry experts, including Tan Li, co-founder of Shushi Technology, who has experience in building enterprise-level data intelligence analysis AI Agents [3]. - The other expert, Xiao Kang, co-founder of Feilun Technology, has extensive experience in database architecture and big data, contributing to the discussion on how data influences decision-making [5]. Group 3: AI Application Across Industries - The "365 Industry AI Implementation Plan" is introduced, showcasing successful AI applications across various sectors, indicating AI's role in driving industrial upgrades [7]. - The article invites industry professionals to share their experiences and solutions regarding AI technology implementation [7].

PCL团队 投稿 量子位 | 公众号 QbitAI 生成式视角可以对监督学习重新思考乃至重新定义！ 想象你在教一个学生解数学题——你会直接让他交卷对答案，还是会让他参考完整答案来理解解题思路？ 如今，一种全新的监督学习范式正受到关注：标签不应只是用于对照回答的标准答案，更可能成为学习过程中的辅助参考。 受生成式一致性模型的启发，来自上海交大、SII、MIT、港中文深圳等机构的研究团队在ICML 2025最新提出预测一致性学习（PCL， Predictive Consistency Learning）。 PCL通过扩散模型的扩散过程消减标签的信息，将噪声标签（Noised Labels）引入模型的输入，使得模型在数据输入和噪声标签的共同参照 下预测完整标签，实现标签信息的复用和价值挖掘。 训练过程概览 传统监督学习中，输入 通过神经网络预测 ，通过对比和标准答案 之间的关系，来计算损失和反向传播更新模型，对应损失函数 $${\mathcal{L}}_{S L}=d(f_{\theta}(\mathbf{x}),\mathbf{y}),$$ 其中 为具体损失函数， 为神经网络函数。受生成一致性模型中一致性映射思想 ...

不圆 发自 凹非寺 量子位 | 公众号 QbitAI Thinking模式当道，教师模型也该学会" 启发式 "教学了—— 由Transformer作者之一Llion Jones创立的明星AI公司 Sakana AI ，带着他们的新方法来了！ 这个方法要求教师模型像优秀的人类教师一样，根据已知解决方案输出清晰的逐步解释，而不再是从头开始自己解决。 用Sanaka AI的新方法训练出的7B小模型，在传授推理技能方面，比671B的DeepSeek-R1还要有效。 | Teacher | Student | | | Final model AIME 2024 MATH 500 GPQA Diamond Overall | | | | --- | --- | --- | --- | --- | --- | --- | | N.A. | | Owen-7B | 10.00 | 74.20 | 33.30 | 39.17 | | DeepSeek-R1 (671B) Qwen-7B | | Bespoke-7B | 20.00 | 82.00 | 37.80 | 46.60 | | RLT teacher (7B) | | ...

西风 梦晨 发自 凹非寺 量子位 | 公众号 QbitAI A I大牛梅涛坐镇，全新多模态AI问世！ 用 法上堪称： 全能 。 不仅 支持 图 片、视频 生成 ： 奇幻场景、多样视角都能驾驭： 而且 唇形同步 功能上线，社 恐大"i"人也能玩转 播客 ： 划重点： 官方还提供了 上百种可直接套用的趣味特效模版 ，让 用户实现"躺 平创 作"。 人物、 动物、建筑物的"变身"模版通通都有 ： 像下面这种炫酷转换， 操作 简单到只需上传一张图： 另外，生图板块的Image Agent也是官方主打，修图生图只需大白话表述，不会写prompt不是问题，它会自动帮你优化 修改。 不卖关子，这个最新创作工具就是 vivago2.0 （智小象AI） 。 打造出它的团队 智象 未 来 （HiDr eam.a i） ，是圈内鼎鼎有名的大牛——加拿大工程院外籍院士梅涛创立的AI公司，研发团队中挤满了 来自中科大的中坚。 前段时间，团队推出的 开源模型HiDream-I1 曾在文生图模型竞技场一鸣惊人， 开源24小时就拿下了排行榜榜首 ，在国内一众开源大模型 中率先跻身第一梯队。 | CREATOR | NAME | ARENA ...

梦晨 发自 凹非寺 量子位 | 公众号 QbitAI 具身智能融资大爆发！6月还没过完，总额已经突破27亿元人民币，这还只是统计了部分公开披露的数据。 宇树科技 则是近期完成C轮融资交割，同时完成 股份制改造 ，由有限公司变更为股份有限公司， 此举被解读为IPO的前兆 。 宇树科技本轮由中国移动旗下基金、腾讯、锦秋基金、阿里、蚂蚁、吉利资本共同领投，另外有消息称绝大部分老股东都跟投。 5月底宇树科技发布海报，透露新品人形机器人将拥有26个关节自由度，价格则在10000美元以内。 据统计5月、6月具身智能行业还有大大小小50多笔融资事件，藏着具身智能行业几个让人意想不到的新趋势。 北上广不再一家独大 这波融资潮呈现出一个有趣的格局：北上广不再一家独大。 从数据来看，北京依然是具身智能创业的"大本营"，银河通用、加速进化、智在无界等14家公司扎堆于此，占了总数的近四分之一。 上海紧随其后，贡献了10家公司，包括智元新创、开普勒探索等明星项目。 广东省也不甘示弱，深圳、东莞等地共有10家公司获得融资。 但真正的黑马是 江苏 。苏州、南京、无锡三城联手，近期拿下了11家公司的融资纪录。 其中最瞩目的两笔是银河通用的1 ...

Core Viewpoint - The article discusses the recent $2 billion seed funding round for Thinking Machines, led by OpenAI's former CTO, which sets a record in the AI sector and highlights the competitive landscape of talent acquisition and investment in AI companies [1][16]. Funding and Investment - Thinking Machines raised $2 billion in seed funding, surpassing previous records where the largest seed rounds ranged from $200 million to $450 million [12][16]. - The funding round was led by Andreessen Horowitz, with participation from Accel and Conviction Partners, bringing the company's valuation to $10 billion [16]. Competitive Landscape - Prior to the funding, both Apple and Meta attempted to invest in or acquire Thinking Machines but were unsuccessful [3][4]. - Meta's CEO, Mark Zuckerberg, shifted focus to recruiting AI talent after failing to acquire Thinking Machines, likening the recruitment process to an NBA transfer [5][11]. Team Composition - Approximately two-thirds of Thinking Machines' team members are former OpenAI employees, including key figures like John Schulman, who is now the co-founder and CTO of Thinking Machines [7][8]. - The company has also brought on board two prominent advisors from OpenAI, Alec Radford and Bob McGrew [7]. Company Vision and Strategy - Thinking Machines aims to develop highly customizable, multimodal AI systems that enhance human expertise rather than replace it, contrasting with the trend of fully autonomous AI agents [19]. - The company positions itself as an advocate for open science, intending to foster collaboration within the research community through frequent publication of technical papers and open-source initiatives [19]. Future Plans - The company plans to develop customized AI models based on enterprise business KPIs and intends to launch consumer-facing products [20].

Core Viewpoint - OpenAI's acquisition of AI hardware company IO for $6.4 billion has led to a lawsuit from IYO, alleging trademark and product design infringement, raising concerns about potential intellectual property theft and competition in the AI hardware space [2][37]. Group 1: Background of IYO - IYO was founded in 2021, incubated by Google X, focusing on "screenless audio computing" technology [3][14]. - The company has a team of approximately 15 people and has invested over $65 million in developing its core product, IYO ONE, an "ear-worn voice computer" [15][17]. - IYO ONE is designed to allow users to perform tasks like translation and searching through voice commands, with a price point of $999 for the Wi-Fi version and $1,199 for the Wi-Fi+LTE version, expected to launch in September-October 2025 with an initial production of 20,000 units [23][24]. Group 2: Allegations Against OpenAI - IYO claims that OpenAI and the LoveFrom team, led by Jony Ive, engaged in discussions with IYO from 2022 regarding potential collaboration and investment, but ultimately launched IO with similar product features [4][5][31]. - The lawsuit alleges that IO's founders and engineers disguised their identities to purchase IYO products and participated in IYO's product trials, indicating a breach of trust [6][35]. - IYO is seeking damages that include three times the profits earned by IO, actual economic losses, and costs for market correction due to the alleged infringement [37]. Group 3: Current Developments - OpenAI has removed all content related to IO following the lawsuit, but the AI hardware project is reportedly still ongoing [10][47]. - The situation has attracted attention from industry figures, including Elon Musk, who has expressed interest in the unfolding events [11].

Core Viewpoint - The article highlights the recent completion of a significant financing round for Feixi Technology, a company specializing in general-purpose intelligent robots, emphasizing its innovative approach and market potential in the robotics sector [1][3][6]. Financing Details - Feixi Technology has successfully completed a Series C financing round, raising hundreds of millions of dollars, led by Yonggui Fund and GF Xinde, with participation from other investors including Hongtai Fund and Huakong Fund [3][5]. - The funds will primarily be used for production expansion, research and development, and ecosystem development [5]. Company Background - Founded in 2016 by core members from Stanford University's Robotics and AI Lab, Feixi Technology has focused on developing adaptive robots and a comprehensive technology stack for intelligent robotics [2][7]. - The company has established itself in various industries, including automotive, home appliances, consumer electronics, and food medical sectors, addressing automation challenges with scalable applications [7][8]. Technological Innovation - Feixi Technology has pioneered the "adaptive robot" category, focusing on human-like capabilities and advanced control systems to enhance robots' adaptability and operational efficiency in complex environments [3][7]. - The company has developed a full-stack technology system that integrates adaptive robot bodies, hierarchical intelligence, and operating system ecosystems [7]. Market Position and Future Outlook - Feixi has served numerous leading global clients, achieving significant upgrades in manufacturing processes across various industries, and has built a solid foundation for future growth [8]. - The company is actively involved in setting industry standards, having led the development of national standards for robot adaptive capabilities, which further establishes its leadership in the field [7]. Investor Insights - Investors express strong confidence in Feixi Technology's potential, highlighting its ability to address productivity challenges and its innovative approach to robotics [9][12]. - The company is seen as a key player in the robotics sector, with a promising future as it continues to innovate and expand its market presence [10][12].