目标驱动导航，赋予机器人自主完成导航目标 具身导航作为具身智能的核心领域，涉及语言理解、环境感知、路径规划三大技术支柱。目标驱动导航（Goal-Oriented Navigation）通过赋予机器人自主决策能 力，是具身导航中最具代表性的方向。 目标驱动导航要求智能体在陌生的三维环境中，仅凭目标描述（如坐标、图片、自然语言）等，即可自主完成环境探索与 路径规划。 与传统视觉语言导航（VLN）依赖显式指令不同，目标驱动导航系统需要实现从"听懂指令走对路"到"看懂世界自己找路"的跃迁：当人类下达"去厨房拿可乐"的指 令时，机器人需自主完成语义解析（识别厨房空间特征与可乐视觉属性）、环境建模（构建家居场景的空间拓扑）以及动态决策（避开移动的人类或宠物），这 背后凝聚着计算机视觉、强化学习与3D语义理解的交叉突破。 目标驱动导航技术已在多个垂直领域实现产业化落地。在终端配送场景中，该技术与社交导航算法结合，使机器人具备应对动态环境和人际交互的能力：美团无 人配送车通过动态路径重规划在复杂城市环境中执行递送任务，Starship Technologies的园区配送机器人已在欧美高校和社区部署。在医疗、酒店及餐饮场景，嘉 ...

编辑丨王多鱼 排版丨水成文 细胞外囊泡 （ EV） 是由细胞分泌的纳米级颗粒，携带着蛋白质、核酸等重要生物分子，广泛参与细胞间通讯及多种疾 病的发生发展过程。 EV 存在于血液、尿液等易获取的体液中，其稳定的膜结构能保护内部的分子"货物"，EV 已成为 针对癌症和神经退行性疾病等进行"液体活检"的理想研究对象。 然而， EV 群体内部存在的高度异质性，使得传统批量分析技术难以捕捉单个囊泡的分子特征，导致关键疾病信号被掩 盖，严重阻碍了其临床应用的 进展 。在此背景下，单 EV 分析技术应运而生，以其超高灵敏度与分辨能力，为解析 EV 异质性、实现多组学整合提供了全新路径，但技术标准不统一与数据碎片化 限制了 EV 标志物的临床应用。 针对这一挑战，山东大学国家健康医疗大数据研究院 侯庆振 团队构建了 首个 跨疾病、跨体液、跨物种的单个细胞外囊 泡多组学图谱 —— SVAtlas 。该成果以 ： SVAtlas: a comprehensive single extracellular vesicle omics resource 为题 ，发表于 Nucleic Acids Research 期刊。山东大学研究 ...

▲ 戳蓝 色字关注我们！ AI 博士生千万别再做 LLM 了。 —— 杨立昆（ Yann LeCun ） 当国内高校的 AI 学院把 LLM（大语言模型）捧成 "镇院之宝"，课程表恨不得从第一页写到最后一页都印着 "Transformer 架构""大模型微调"； 当企业融资路演时，PPT 里不塞个 "千亿参数""多模态大模型" 都不敢抬头说话； 当博士生们挤破头往 LLM 赛道钻，哪怕只是给模型调个参数都觉得自己踩在了 "技术风口"—— Hugging Face CEO 克莱姆・德朗格的警告像一盆冰水浇下来：" 我们不在 AI 泡沫里，只在 LLM 泡沫里。 " 更狠的是 AI 界 "活化石" 杨立昆，这位 40 年没看走眼的技术大牛直接拍桌子：" AI 博士生？千万别碰 LLM！ " 现在这场全民追 LLM 的闹剧，本质上是一群人围着半块面包狂欢，却忘了后厨里还有鱼翅、燕窝、海参没来得及做。 （图灵奖得主杨立昆 Yann LeCun ） 而国内更要警惕！正踩着这场狂欢的节奏，一步步往 "路径依赖" 的坑里跳，等哪天泡沫破了才会发现：下一个时代的入场券，早被我们在抢面包的时 候弄丢了。 一、AI 百年史 从不 ...

Core Insights - Yann LeCun, a Turing Award winner and a key figure in deep learning, announced his departure from Meta to start a new company focused on Advanced Machine Intelligence (AMI) research, aiming to revolutionize AI by creating systems that understand the physical world, possess persistent memory, reason, and plan complex actions [2][4][11]. Departure Reasons & Timeline - LeCun's departure from Meta was confirmed after rumors circulated, with the initial report coming from the Financial Times on November 11, indicating his plans to start a new venture [10][11]. - Following the announcement, Meta's market value dropped approximately 1.5% in pre-market trading, equating to a loss of about $44.97 billion (approximately 320.03 billion RMB) [11]. - The decision to leave was influenced by long-standing conflicts over AI development strategies within Meta, particularly as the focus shifted towards generative AI (GenAI) products, sidelining LeCun's foundational research efforts [11][12]. Research Philosophy & Future Vision - LeCun emphasized the importance of long-term foundational research, which he felt was being undermined by Meta's shift towards rapid product development under the leadership of younger executives like Alexandr Wang [12][13]. - He expressed skepticism towards large language models (LLMs), viewing them as nearing the end of their innovative potential and advocating for a focus on world models and self-supervised learning to achieve true artificial general intelligence (AGI) [14][15]. - LeCun's vision for AMI includes four key capabilities: understanding the physical world, possessing persistent memory, true reasoning ability, and the capacity to plan actions rather than merely predicting sequences [16][15]. Industry Context & Future Outlook - The article suggests a growing recognition in the industry that larger models are not always better, with a potential shift towards smaller, more specialized models that can effectively address specific tasks [18]. - Delangue, co-founder of Hugging Face, echoed LeCun's sentiments, indicating that the current focus on massive models may lead to a bubble, while the true potential of AI remains largely untapped [18][15]. - Meta acknowledged LeCun's contributions over the past 12 years and expressed a desire to continue benefiting from his research through a partnership with his new company [22].

Core Viewpoint - The departure of Yann LeCun from Meta signifies a major shift in the AI landscape, highlighting internal strategic disagreements and a pivot in Meta's AI development approach [2][3][4]. Group 1: Departure and Strategic Shift - Yann LeCun, a prominent figure in AI and Meta's Chief AI Scientist, is leaving the company after 12 years, marking a formal split with CEO Mark Zuckerberg over AI strategy [2][3]. - The decision to leave was foreshadowed by increasing disagreements with Meta's management regarding the AI development roadmap and company strategy [3][4]. - Meta's internal restructuring has shifted focus from long-term foundational research led by LeCun's FAIR lab to a more agile product development approach, driven by immediate market needs [4][7]. Group 2: Internal Changes and Leadership Dynamics - Meta has made significant changes, including a $100 million compensation package to attract young talent from competitors, and the formation of a new "superintelligence" team led by 28-year-old Alexandr Wang [4]. - LeCun's reporting structure changed, requiring him to report to Wang instead of the Chief Product Officer, which marginalized his FAIR lab and its research initiatives [4][7]. Group 3: Technological Disagreements - LeCun has publicly criticized the current trend of large language models (LLMs), arguing they are inadequate for achieving true reasoning and planning capabilities, which diverges from Zuckerberg's focus on immediate monetization [7][8]. - The emphasis on "world models," which LeCun advocates, contrasts sharply with the short-term goals set by Meta's leadership, leading to his decision to leave [7][8]. Group 4: Future Aspirations - Post-Meta, LeCun aims to fully commit to developing "world models," which he believes will redefine AI by enabling machines to learn from observing the physical world, akin to human cognitive development [8]. - He predicts that within 3-5 years, "world models" will become the mainstream AI architecture, challenging the current dominance of LLMs [8]. Group 5: Legacy and Impact - LeCun's career has been pivotal in the evolution of AI, having co-developed convolutional neural networks (CNNs) and led the FAIR lab to prominence [9]. - His departure is seen as a significant loss for Meta, indicating a potential shift in the AI research landscape and the company's future direction [9].

文 | 新质动能，作者｜沛林，编辑｜沐风 一个离职消息，让AI界震动。 图灵奖得主、Meta首席科学家Yann LeCun（中文名：杨立昆），即将从Meta离职创业。消息曝出， Meta股价下跌1.5%，市值蒸发1400亿元！ 这位大神有多牛？ 他不仅是深度学习领域的奠基人之一——其开创的卷积神经网络（CNN）架构为现代AI发展铺平了道 路，并为他赢得了计算机领域的最高荣誉"图灵奖"；他一手创建FAIR实验室，并帮助Meta奠定其在行 业内的AI地位。 然而，如今他在Meta的处境，却是另一番景象：他坚持的技术路线在内部失势，团队核心被裁，甚至 要向小30岁的年轻高管汇报。 杨立昆与扎克伯格最终走向决裂，核心原因在于对AGI（Artificial General Intelligence，通用人工智能） 路线的选择：后者砸重金追逐当下最热门的LLM（Large Language Model，大语言模型）路线，杨立昆 则视LLM为AGI路上的岔路，认为"世界模型"技术路线才是走向AGI的正确道路。 这场决绝的分手，已然超出个人职业变动，成为一场关乎AI未来发展路线的"道统"之争。 理念冲突，愤然离职 杨立昆要离 ...

Core Insights - The paper titled "UNComp: Can Matrix Entropy Uncover Sparsity?" addresses the paradox of matrix entropy in deep learning models, revealing that traditional matrix entropy increases with depth, contradicting the observed sparsity in deeper models [5][7] - A breakthrough is achieved by introducing Truncated Matrix Entropy, which shows a decreasing trend with increasing layers, explaining the sparsity phenomenon and providing a theoretical basis for compression strategies [7][12] Theoretical Framework - The new theoretical tool allows for a deeper understanding of the internal workings of models, focusing on the information flow patterns rather than merely optimizing attention distributions [8][12] - Key structural insights are identified, linking fluctuations in intermediate layer entropy to retrieval layers and heads, enabling structured pruning based on theoretical guidance [13] Practical Applications - The UNCOMP framework is designed to optimize both computation and memory by compressing hidden states during the prefill phase and KV Cache during decoding, achieving layer-wise and head-wise compression [16][17] - Experimental results indicate a 60% acceleration in the prefill phase and a 6.4 times increase in throughput, with KV Cache compressed to 4.74% [19] Performance Metrics - The framework maintains model performance even under extreme compression rates, with various methods showing high retention rates for Llama2 and Llama3 models, such as Ours-group achieving 98.42% and 84.13% respectively [20] - Merging retrieval layers with final layers shows minimal performance loss, with some tasks surpassing the full-size baseline [21] Conclusion - UNCOMP serves not only as a tool but also as a window into understanding the complex information compression behaviors within large language models [22]

Core Viewpoint - The case involving attorney Steven Schwartz highlights the critical misunderstanding of the capabilities of large language models (LLMs) in legal research, leading to the submission of fabricated court cases and citations [3][4][5]. Group 1: Case Overview - Judge Kevin Castel addressed the submission of six cases by Schwartz, which were later found to be entirely fabricated and non-existent [3][4]. - Schwartz initially believed that LLMs like ChatGPT could serve as reliable legal research tools, equating them to a "super search engine" [4][5]. Group 2: Limitations of LLMs - The case illustrates a fundamental misunderstanding of LLMs' capabilities, particularly in the context of legal research, which requires precise and verifiable information [5][7]. - LLMs are known to produce "hallucinations," or false information, which poses significant risks in fields requiring high accuracy, such as law [5][7][9]. - The architecture of LLMs presents challenges, including lack of transparency, difficulty in updating knowledge, and absence of domain-specific expertise [7][8][9]. Group 3: Knowledge Graphs as a Solution - Knowledge graphs (KGs) are proposed as a solution to enhance the reliability of AI systems by providing structured, verifiable, and up-to-date information [10][12][19]. - KGs support dynamic updates and maintain a clear audit trail, which is essential for accountability in professional environments [12][20]. - The integration of KGs with LLMs can mitigate the risks associated with hallucinations and improve the accuracy of domain-specific applications [19][20]. Group 4: Future of AI in Professional Fields - The future of AI in critical applications, such as legal research, hinges on the development of intelligent advisory systems that combine the strengths of KGs and LLMs [21]. - Professionals deploying AI tools must ensure that their systems support accountability and accuracy, rather than undermine them [21].

Core Insights - The article presents a groundbreaking approach to reinforcement learning (RL) that achieves state-of-the-art (SOTA) performance using a simple, single-stage training method with fixed hyperparameters, resulting in a 50% reduction in computational power [4][14][15] - The findings suggest that a well-scaled, simple baseline can be more powerful than previously thought, challenging the complexity often associated with advanced RL techniques [4][15][27] Background and Context - The research is set against the backdrop of a "technical arms race" in training small models using RL, with various methods evolving rapidly over a few months [6] - Early approaches included hyperparameter tuning, multi-stage progressive training, and curriculum learning, leading to increasingly complex training pipelines [6][8] Methodology - The JustRL approach emphasizes simplicity, utilizing standard GRPO without modifications, a single continuous training phase, and fixed hyperparameters [11] - The training data consists of regular math problem sets without offline difficulty screening or data augmentation, demonstrating effectiveness across different model baselines [11][14] Performance Metrics - JustRL-DeepSeek-1.5B achieved an average accuracy of 54.87% across nine benchmarks, outperforming ProRL-V2, which used a nine-stage training approach [14] - JustRL-Nemotron-1.5B reached an average accuracy of 64.32%, slightly surpassing QuestA, while using significantly fewer tokens [14][15] Training Dynamics - The training process for JustRL-DeepSeek-1.5B was notably stable, with key metrics such as policy entropy and average reward showing healthy fluctuations without typical issues like exploration collapse or premature convergence [17][19] - The training was conducted on 32 A800-80GB GPUs over approximately 15 days, highlighting the reduced engineering complexity and computational overhead compared to multi-stage methods [15] Key Discoveries - The research revealed that adding certain "optimizations" could lead to worse performance, indicating that not all seemingly beneficial techniques are necessary [21][24] - The findings emphasize the importance of establishing a clear, simple baseline to accurately assess the value of complex techniques in RL training [27] Philosophical Implications - The article concludes with a philosophical reflection on the value of simplicity in technology, suggesting that often, simpler methods may yield sufficient results when adequately scaled [26][27][28]

Core Insights - The next frontier for AI is "Spatial Intelligence," which is crucial for understanding and interacting with the physical world [1][4][14] - Current AI systems lack the ability to comprehend spatial relationships and physical interactions, limiting their effectiveness in real-world applications [1][12][26] - The development of a "world model" is essential for achieving true spatial intelligence in AI, enabling machines to perceive, reason, and act in a manner similar to humans [14][15][20] Group 1: Importance of Spatial Intelligence - Spatial intelligence is identified as a missing component in AI, which could lead to significant advancements in capabilities, particularly in achieving Artificial General Intelligence (AGI) [3][12] - The limitations of current AI systems are highlighted, emphasizing their inability to perform basic spatial reasoning tasks, which hinders their application in various fields [12][26] - The potential of spatial intelligence to revolutionize creative industries, robotics, and scientific exploration is underscored, indicating its broad implications for human civilization [1][4][10] Group 2: Development of World Models - The concept of world models is introduced as a new paradigm that surpasses existing AI capabilities, focusing on understanding, reasoning, and generating interactions with the physical world [14][15] - Three core capabilities for effective world models are outlined: generative ability to create realistic environments, multimodal processing of diverse inputs, and interactive capabilities to predict outcomes based on actions [15][16][17] - The challenges in developing these models include creating new training objectives, utilizing large-scale training data, and innovating model architectures to handle complex spatial tasks [18][19][20] Group 3: Applications and Future Prospects - The applications of spatial intelligence span various fields, including creative industries, robotics, and healthcare, with the potential to enhance human capabilities and improve quality of life [21][26][27] - The World Labs initiative is highlighted as a key player in advancing spatial intelligence through the development of tools like the Marble platform, which aims to empower creators and enhance storytelling [20][22] - The long-term vision includes transforming how humans interact with technology, enabling immersive experiences and fostering collaboration between humans and machines [28][29]