2025 WAIC 云帆奖全球征集已进入倒数 3天，小伙伴们赶快 on board！点击文末「阅读原 文」完成报名或提名！ 在全球 AI 创新格局加速重塑之际，2025 WAIC 云帆奖以「聚智・共进」为主题，继续发掘和表彰 AI 青年 先锋。从基础理论突破到产业实践创新，从跨学科交叉到开源生态建设，我们寻找能够在通往 AGI 的征途 上贡献中国智慧的青年才俊，并将为奖项得主提供学术影响、科研资助、产业生态全方位赋能。 现在，三位重量级奖项召集人向你发出诚挚邀请： WAIC 云帆奖得主赋能计划 学术影响力跃升 ：在 WAIC 揭晓获奖名单并颁发证书，提升获奖者国际知名度与学术影响力。为有意愿者 向知名高校、科研机构推荐工作；提供与国际顶尖学术机构、学者交流合作机会，如受邀参加国际会议作 汇报。 奖项设置 璀璨明星 明日之星 百万级科研加速 ：联合科研机构提供重点科研项目资助内推，资助金额最高可达数百万；提供算力资源， 助力突破算力瓶颈；多渠道推广研究成果，提升影响力与应用价值。 产业生态赋能 ：组织与知名企业、科研机构对接，提供项目实践机会，推动产学研融合。为有创业意愿者 提供创业培训、资金及上下游企业对接等 ...

机器之心报道 机器之心编辑部 未来的世界级研究，可能会出自这些年轻科学家。 本周四， 2025 年谷歌研究学者计划（Research Scholar Program）获奖名单公布了。 研究学者计划是谷歌为了支持学术界研究工作而设立的一个专门项目，旨在通过提供资金支持的方式鼓励 与科研人员的合作，建立长期合作关系，以推动科学和技术的进步。申请人在提交申请时必须是大学或授 予学位的研究机构的全职助理、副教授或教授。 地址：https://research.google/programs-and-events/research-scholar-program/recipients/ 机器之心将获奖华人学者名单整理如下（排名不分先后）： 应用科学 Julia Yang：佐治亚理工学院 获奖研究：LLM-GUAL：基于 LLM 的用户定义原子模拟生成 Julia Yang 现在是佐治亚理工学院助理教授，主要研究方向包括电池材料、分子模拟、机器学习以及电化 学。 Julia Yang 于卡内基梅隆大学获得材料科学与工程学士学位，辅修物理学，后于加州大学伯克利分校获得材 料科学与工程博士学位。 每年，谷歌会评选出多个领域有 ...

Core Viewpoint - vLLM is a high-performance, open-source LLM inference and service engine developed by the University of California, Berkeley, aimed at enhancing inference speed and resource utilization, particularly memory efficiency, while being compatible with popular model libraries like Hugging Face [2][3]. Group 1: vLLM and Nano-vLLM - vLLM enables mainstream models like GPT, Mistral, and LLaMA to run faster and consume fewer resources through its innovative attention mechanism called PagedAttention [3]. - A lightweight implementation of vLLM, named Nano-vLLM, was developed by DeepSeek AI researcher Yu Xingkai, simplifying the code to under 1200 lines [4][7]. - Nano-vLLM has gained over 200 stars on GitHub, indicating community interest and engagement [5]. Group 2: Features of Nano-vLLM - Nano-vLLM offers three core functionalities: 1. Fast offline inference with performance comparable to vLLM [6]. 2. A readable codebase with a simplified implementation [7]. 3. An optimization suite that includes features like prefix caching, Torch compilation, and CUDA computation graphs [8]. Group 3: Benchmarking Results - Benchmark tests showed that Nano-vLLM produced the same output tokens as vLLM but took slightly longer, resulting in a throughput of 1314.65 tokens/s compared to vLLM's 1353.86 tokens/s [9][11]. - The testing configuration included using an RTX 4070 GPU, with a model size of Qwen3-0.6B, and random sampling of input and output lengths between 100 and 1024 tokens [10].

Core Viewpoint - OpenAI aims to achieve "organizational-level" intelligence as the ultimate goal in the five stages towards AGI (Artificial General Intelligence), where AI can manage complex processes, make high-level decisions, and coordinate large-scale operations [1] Group 1: MASLab Introduction - MASLab is a collaborative initiative launched by ten institutions, including Shanghai Jiao Tong University and Oxford University, to accelerate the healthy development of Multi-Agent Systems (MAS) [2] - MASLab provides a unified, comprehensive, and research-friendly codebase for large model multi-agent systems, facilitating ease of use and reproducibility [4] Group 2: Features of MASLab - MASLab integrates over 20 mainstream MAS methodologies, covering results from major conferences over the past two years across various fields and task types [6] - The platform ensures evaluation fairness and reproducibility through standardized input preprocessing, LLM configuration, and evaluation protocols [8] Group 3: Experimental Analysis - Researchers have conducted extensive experiments using MASLab, covering over ten evaluation benchmarks, including MATH and GPQA, and analyzing the performance of eight major models [11] - The results demonstrate the current state of MAS methods, highlighting their strengths and weaknesses [14] Group 4: Innovations and Future Directions - MASLab-ReAct, a more efficient MAS method, supports various tools and has shown superior results on the GAIA validation set, indicating significant potential for real-world applications [16] - MASLab is an open-source platform aimed at community contributions, with plans to continuously release more methods and benchmarks to foster a sustainable MAS research community [22][23]

Core Viewpoint - The article discusses the evolving job market for AI graduates, emphasizing the shift from model parameters and training techniques to defining valuable problems and creating evaluation systems that fit real-world scenarios [2][6][11]. Group 1: Industry Trends - The AI job market is rapidly changing, with companies of all sizes actively recruiting AI talent [2]. - The focus of AI competition is shifting from merely improving model performance to understanding how to apply AI effectively in real-world contexts [6][11]. - The saturation of benchmark tests is occurring faster, indicating diminishing returns from traditional model development approaches [6][11]. Group 2: Company Selection Criteria - Graduates should consider companies that can sustain AI development, focusing on user engagement and the ability to create a complete cycle from technology development to commercial application [11][12]. - The strength of the coupling between technology and business is crucial; AI should be a core driver rather than a supplementary feature [12]. - Companies must demonstrate commercial validation of AI capabilities, such as having revenue-generating AI applications and clients willing to pay for AI features [13][14]. Group 3: Tencent as a Case Study - Tencent exemplifies a company with a broad and deep engagement in various fields, providing a rich environment for AI development [15][16]. - Tencent's AI technologies are integrated into its core business operations, enhancing user engagement and driving revenue growth [17][18]. - The company has clear AI monetization cases, with significant revenue contributions from AI-driven advertising and gaming sectors [18][19]. Group 4: Talent Development Programs - Tencent's "Qingyun Plan" is a high-priority initiative aimed at nurturing top technical talent, offering competitive compensation and a supportive environment for innovation [21][22]. - Participants in the Qingyun Plan have opportunities for significant contributions to AI projects and can publish research in top conferences [23][24]. - The program emphasizes a non-traditional management culture, allowing for exploration and creativity in research [24][25].

Core Insights - The article discusses the necessity of world models for general agents in achieving flexible, goal-directed behavior, emphasizing that any AI capable of generalizing to multi-step tasks must learn a predictive model of its environment [4][9][20]. Group 1: Importance of World Models - World models are essential for agents to generalize across complex, long-term tasks, as they allow for the prediction of future states based on current actions [4][5][9]. - Google DeepMind's research indicates that learning world models is not just beneficial but necessary for achieving human-level artificial intelligence [9][20]. Group 2: Theoretical Framework - The authors developed a mathematical framework consisting of four components: environment, goals, agents, and world models, to formalize the relationship between these elements [24][30]. - The framework posits that any agent capable of handling simple goal-directed tasks must learn a predictive model of its environment, which can be extracted from the agent's policy [20][30]. Group 3: Algorithm for World Model Recovery - The article outlines an algorithm that allows for the recovery of world models from bounded agents by querying them with carefully designed composite goals [37][39]. - Experiments demonstrated that even when agents deviated from theoretical assumptions, the algorithm successfully recovered accurate world models, confirming the link between agent capabilities and the quality of the world model [40][46]. Group 4: Implications for AI Development - The findings suggest that the race for superintelligent AI may actually be a competition to build more complex world models, transitioning from a "human data era" to an "experience era" [49][52]. - The development of foundational world models like Genie 2, which can generate diverse 3D environments from a single image, represents a significant advancement in training and evaluating embodied agents [51][52].

本工作于 2024 年 11 月完成，目前已经被 CVPR 2025 接收并评选为 Highlight，第一作者为龙宇星，导师为北京大学董豪老师。课题组致力于研究统一的物体表征 操作研究，以实现具有可解释性和泛化能力的物体操作策略。 自 19 世纪末爱迪生发明电灯以来，电器的发展和革新不断提升人类的生活水平。如今，电器已经走进千家万户，成为我们的得力助手，与我们的生活密不可分。 赋予机器人使用家电的能力具有重要的学术价值和广阔的应用前景。 目前在机器人操作领域，一般物体（如刚体和铰接物体）的操作研究已经取得一定进展，但是现有操作策略主要执行单步原子操作。对于设备（如家电）而言， 必须按照正确顺序和方式进行多步操作，才能正确完成高层次任务。因此，参照说明书进行长程操作规划对于家电操作而言十分必要。 然而，受限于以下三大挑战，基于说明书的长程家电操作探索几乎处于空白状态： 为应对上述挑战， 北京大学联合智元机器人团队提出了全新的家用电器操作评测基准 CheckManual，这是首个专为研究基于说明书的家电操作而设计的评测框架 论文标题：CheckManual: A New Challenge and Benchm ...

Core Viewpoint - The article discusses the significance of representation learning in generative models, particularly through the introduction of a new method called Dispersive Loss, which integrates self-supervised learning into diffusion-based generative models without requiring additional pre-training or external data sources [6][9][43]. Group 1: Diffusion Models and Representation Learning - Diffusion models excel in modeling complex data distributions but are largely disconnected from the representation learning field [2]. - The training objectives of diffusion models typically focus on reconstruction tasks, such as denoising, lacking explicit regularization for learned representations [3]. - Representation learning, particularly self-supervised learning, is crucial for learning general representations applicable to various downstream tasks [4]. Group 2: Introduction of Dispersive Loss - Dispersive Loss is a flexible and general plug-in regularizer that integrates self-supervised learning into diffusion-based generative models [9]. - The core idea of Dispersive Loss is to introduce a regularization target for the model's internal representations, encouraging them to spread out in the latent space [10][13]. - This method does not require additional layers or parameters, making it a simple and independent approach [15][16]. Group 3: Comparison with Existing Methods - Dispersive Loss operates without the need for pre-training, external data, or additional model parameters, unlike the REPA method, which relies on pre-trained models [7][41][43]. - The new method demonstrates that representation learning can benefit generative modeling without external information sources [13][43]. - In practical applications, introducing Dispersive Loss requires minimal adjustments, such as specifying the intermediate layers for regularization [29]. Group 4: Performance Evaluation - Experimental results show that Dispersive Loss consistently outperforms corresponding contrastive losses while avoiding the complexities of dual-view sampling [33]. - The method has been tested across various models, including DiT and SiT, showing improvements in all scenarios, particularly in larger models where effective regularization is crucial [36][37]. - The article highlights that Dispersive Loss can be generalized for one-step diffusion-based generative models, indicating its versatility [44].

作者介绍： 本文作者来自通义实验室 RAG 团队，致力于面向下一代 RAG 技术进行基础研究。该团队 WebWalker 工作近期也被 ACL 2025 main conference 录 用。 它得能看懂网页，能做多步决策； 它得能适应开放动态环境； 它得能自主提问、自主行动、自主修正…… 一、背景：信息检索的新需求与挑战 在当今信息爆炸的时代，解决复杂问题不再仅仅是简单的知识检索，而是需要深入的信息挖掘和多步推理。从医学研究到科技创新，从商业决策到学术探索，每 一个领域都呼唤着能够自主思考、自主决策的智能体。Deep Research 等系统已经为我们展示了自主多步研究的巨大潜力，但构建这样的智能体并非易事。它们需 要在复杂的网络环境中感知、决策、行动，还要面对任务复杂度高、泛化能力弱等诸多挑战。 但打造这样一个 Deep Research 类智能体智能体，并不简单！ 在这种背景下，WebDancer 的出现，走出了一条复现 Deep Research 类智能体的可行路径。 自主信息检索智能体的构建，或者如何复现 Deep Research 类的模型一直面临着两大棘手难题：高质量训练数据的稀缺与开放环境训 ...

Core Insights - The article discusses the significant upgrades and new product releases by Volcano Engine at the Force 2025 conference, highlighting the advancements in AI models and their capabilities [1][2][3]. Group 1: Product Releases and Upgrades - Volcano Engine launched several new products, including Doubao Model 1.6, Seedance 1.0 Pro, and an AI cloud-native platform, showcasing a comprehensive suite of AI capabilities [2][3]. - Doubao Model 1.6 features three versions: Standard, Deep Thinking Enhanced, and Flash, with notable improvements in performance and capabilities [3][4]. - Doubao Model 1.6 achieved a high score of 144 in the national college entrance examination, indicating its advanced reasoning and understanding capabilities [4][6]. Group 2: Performance and Capabilities - Doubao Model 1.6 is the first domestic model to support a 256K context window and has demonstrated significant advancements in multimodal understanding and GUI operations [4][6]. - The Seedance 1.0 Pro model outperformed leading competitors in video generation, showcasing its ability to create seamless narratives and realistic motion [6][35]. - Volcano Engine emphasized the concept of "AI cloud-native," focusing on optimizing cloud infrastructure for AI workloads, which is expected to drive future developments [8][70]. Group 3: AI Infrastructure and Development Kits - Volcano Engine introduced three development kits: AgentKit, TrainingKit, and ServingKit, aimed at enhancing AI application development and deployment [8][66]. - The company is focusing on the integration of intelligent agents capable of executing complex tasks, moving beyond simple generative AI [52][70]. - The new AI-native data infrastructure aims to support enterprises in building robust data foundations for AI model training and decision-making [64][66]. Group 4: Market Position and Future Outlook - Volcano Engine's approach contrasts with the industry norm of "model first, application later," as it emphasizes practical applications and productization [71][72]. - The company is committed to long-term investments to establish itself as a trusted cloud service platform, with a focus on real-world AI applications [72].