本文作者分别来自新加坡国立大学、北京大学与清华大学。第一作者王宇豪与共同第一作者屈文杰来自新加坡国立大学，研究方向聚焦于大语言模型中的安 全与隐私风险。共同通讯作者为北京大学翟胜方博士，指导教师为新加坡国立大学张嘉恒助理教授。 本研究聚焦于当前广泛应用的 RAG (Retrieval-Augmented Generation) 系统，提出了一种全新的黑盒攻击方法： 隐式知识提取攻击 (IKEA) 。不同于以 往依赖提示注入 (Prompt Injection) 或越狱操作 (Jailbreak) 的 RAG 提取攻击手段， IKEA 不依赖任何异常指令，完全通过自然、常规的查询，即可高效 引导系统暴露其知识库中的私有信息。 在基于多个真实数据集与真实防御场景下的评估中，IKEA 展现出超过 91% 的提取效率与 96% 的攻击成功率，远超现有攻击基线；此外，本文通过多项 实验证实了隐式提取的 RAG 数据的有效性。本研究揭示了 RAG 系统在表面「无异常」交互下潜在的严重隐私风险。 论文题目：Silent Leaks: Implicit Knowledge Extraction Attack on RAG S ...

Core Insights - The memory capacity of GPT series models is approximately 3.6 bits per parameter, indicating a limit beyond which models stop memorizing and begin to generalize [1][4][27]. Group 1: Memory and Generalization - The research distinguishes between two types of memory: unexpected memory (specific dataset information) and generalization (understanding of the real data generation process) [5][7]. - A new method was proposed to estimate a model's understanding of specific data points, which helps measure the capacity of modern language models [2][8]. Group 2: Model Capacity and Measurement - The study defines model capacity as the total amount of memory that can be stored across all parameters of a specific language model [17][18]. - The maximum memory capacity is reached when the model no longer increases its memory with larger datasets, indicating saturation [19][28]. - Experiments showed that the memory capacity of models scales with the number of parameters, with a stable memory of 3.5 to 3.6 bits per parameter observed [27][28]. Group 3: Experimental Findings - The research involved training hundreds of transformer language models with parameters ranging from 500,000 to 1.5 billion, leading to insights on scaling laws related to model capacity and data size [6][25]. - Results indicated that even with different dataset sizes, the memory bits remained consistent, reinforcing the relationship between model capacity and parameter count [28][29]. - The impact of precision on capacity was analyzed, revealing that increasing precision from bfloat16 to float32 slightly improved capacity, with average values rising from 3.51 bits/parameter to 3.83 bits/parameter [31][32].

Core Viewpoint - AReaL-boba² is a significant upgrade to the asynchronous reinforcement learning (RL) training system, enhancing efficiency, usability, and performance in coding RL tasks, while fully supporting Agentic RL [2][3][39]. Group 1: Efficiency and Performance - AReaL-boba² achieves a training speed improvement of up to 2.77 times compared to the previous version, while maintaining model performance [8]. - The system has set new state-of-the-art (SOTA) benchmarks in coding tasks, with the AReaL-boba²-14B model scoring 69.1 on LiveCodeBench and achieving a Codeforce rating of 2044 [5][4]. - The asynchronous RL framework allows for continuous data generation and model training, significantly improving GPU resource utilization and reducing idle time [14][15]. Group 2: User Accessibility - The upgrade includes comprehensive tutorials and documentation, making it easier for both beginners and experienced users to customize datasets, algorithms, and agent logic without modifying the underlying code [3][8]. - AReaL-boba² is designed to be user-friendly, with a simplified environment setup and experiment initiation process [3][8]. Group 3: Technical Innovations - The system employs a fully asynchronous RL training approach, decoupling data generation from model training, which addresses inefficiencies found in traditional synchronous RL systems [14][15]. - AReaL-boba² introduces two key algorithmic improvements: Staleness Control to manage data freshness and Decoupled PPO Objective to mitigate distribution discrepancies between old and new model versions [24][28]. Group 4: Future Developments - The AReaL team is continuously updating the Agentic RL capabilities, allowing developers to customize agents and environments for multi-turn interactions [39][40]. - The project is built on years of technical accumulation from various research teams and aims to make AI training accessible and customizable for everyone [41].

Core Insights - The article discusses the potential of Prolonged Reinforcement Learning (ProRL) in enhancing reasoning capabilities in language models, suggesting that it can lead to significant improvements in model performance rather than merely optimizing existing knowledge retrieval [1][15]. Group 1: ProRL Framework - ProRL framework significantly increases the training steps from hundreds to over 2000, unlocking the hidden potential of smaller models [3]. - The framework incorporates a diverse set of verifiable rewards from various domains, providing reliable supervision signals for RL training [5]. - The combination of GRPO and DAPO algorithms enhances training efficiency by avoiding policy update imbalances and filtering ineffective samples [7]. Group 2: Performance Improvements - The Nemotron-Research-Reasoning-Qwen-1.5B model demonstrates remarkable performance across various tasks, outperforming larger models in specific areas [9][10]. - ProRL leads to a 14.7% improvement in mathematical tasks, surpassing 7B models, and a 6.5% lead in code generation over DeepCoder-1.5B [12]. - In logical reasoning, accuracy improves by 54.8%, showcasing the model's enhanced capabilities [12][13]. Group 3: Creativity and Reasoning Expansion - ProRL enables models to solve problems that base models could not, achieving a pass@k of 100% in previously unsolvable tasks [13]. - The training process fosters creativity, allowing models to generate new problem-solving paths rather than relying on rote answers [6][14]. - The longer the training, the stronger the model's ability to deviate from pre-training data, resulting in richer and more creative reasoning strategies [14]. Group 4: Future Implications - The research indicates that ProRL could be the key to developing small language models with strong reasoning capabilities, low deployment costs, and high generalization abilities [16][17].

Core Viewpoint - The article discusses the breakthrough of the CrossFlow framework developed by Meta and Johns Hopkins University in the field of cross-modal generation, moving from a noise-dependent approach to a more efficient and flexible modality-to-modality mapping method [1][4][30]. Group 1: Innovation and Methodology - CrossFlow represents a new paradigm in cross-modal generation, allowing direct mapping between modalities without relying on noise distributions or complex conditional mechanisms [4][30]. - The framework utilizes flow matching to create a regularized distribution, enabling smooth and semantically coherent cross-modal paths [8]. - By employing a variational encoder, the model encodes input modalities into a regularized latent space, facilitating effective mapping between text and image spaces [8][12]. Group 2: Performance and Comparisons - CrossFlow demonstrates superior performance in various tasks, including image generation and depth estimation, achieving results comparable to or exceeding state-of-the-art algorithms while using a simpler transformer architecture [7][28]. - In text-to-image generation, CrossFlow outperforms mainstream methods that rely on cross-attention, showcasing better scaling properties [14][15]. - The model significantly reduces training resource requirements compared to models like DALL-E 2, with training time reduced from thousands of GPU days to as low as 208 A100 GPU days [23]. Group 3: Flexibility and Applications - The dual mapping property of flow matching allows CrossFlow to be utilized for both text-to-image generation and image captioning, achieving state-of-the-art results on the COCO dataset [23][28]. - The model's design enables it to adapt to multiple tasks without task-specific configurations, promoting a unified framework for various applications [28][30]. - CrossFlow's approach to customizable source distributions enhances flexibility in image generation and significantly accelerates generation speed [23].

Core Viewpoint - The article discusses the advancements in diffusion language models (dLLMs), particularly focusing on Google's Gemini Diffusion and its implications for AI development, highlighting the speed and performance improvements over traditional autoregressive models [1][8][35]. Group 1: Gemini Diffusion and Its Features - Gemini Diffusion is noted for its impressive generation speed, being five times faster than previous models, and its ability to handle programming tasks effectively [2][8]. - The underlying mechanism of diffusion models allows for rapid iteration and error correction during the generation process, distinguishing it from autoregressive models [2][3]. - Gemini Diffusion's sampling speed can reach an astonishing 1479 tokens per second, showcasing its potential in various benchmarks [8][9]. Group 2: Development of Diffusion Language Models - Prior to Gemini Diffusion, several research teams explored the feasibility of diffusion-based LLMs, including Stanford's Diffusion-LM and Fudan University's DiffusionBERT [3][4]. - The introduction of LLaDA, the first 8 billion parameter diffusion language model, marked a significant milestone in the field, achieving performance comparable to LLaMA 3 [4][21]. - Following LLaDA, other models like d1 and LaViDa have emerged, further establishing LLaDA as a foundational model in dLLM research [20][21]. Group 3: Multimodal Diffusion Language Models - The emergence of diffusion multimodal language models (dMLLMs) is highlighted, with LLaDA-V and MMaDA being prominent examples that integrate visual and language processing capabilities [10][31]. - LLaDA-V combines visual instruction fine-tuning with the diffusion mechanism, demonstrating strong performance in multimodal understanding tasks [26][27]. - MMaDA showcases innovations in text reasoning and multimodal understanding, solidifying its position as a leading research outcome in the dMLLM space [31][32]. Group 4: Future Directions and Implications - The article emphasizes the shift from autoregressive models to diffusion models as a significant paradigm change in AI, suggesting broader implications for future research and applications [35][36]. - The ongoing evolution of models like LLaDA and Gemini Diffusion indicates a growing ecosystem around dLLMs and dMLLMs, with potential applications extending into quantum computing [35][36].

前几天，谷歌在 I/O 2025 大会上正式发布了其最新一代 AI 视频生成模型 Veo 3，在生成高质量视频的同时首次实现了音画同步。对于 Veo 3 的震撼效果，有人高 度评价称，「它会是不亚于 OpenAI Sora 的跨时代产品」，标志着 AI 视频进入到了真正的「有声时代」。 从中可以发现，虽然当前 AI 社区已有的大模型已经足够惊艳，但得益于架构的创新、算力集群的投入，仍然会「卷」出一些新东西来。比如视频生成领域，从最 初的无声进化到如今的有声，提升明显；再比如多模态领域，逐渐朝着理解与生成大一统的方向演进。 因此，为让从业者全面了解 AI 社区涌现的最新创新成果和发展趋势，机器之心计划 6 月 8 日在北京举办「CVPR 2025 论文分享会」，围绕着多模态、视频生成等 热门主题邀请顶级专家、论文作者与现场参会观众共同交流。 作为计算机视觉领域中最重要的国际会议之一，CVPR 具有极高的含金量，每年都会吸引大量研究机构和高校参会。今年，CVPR 2025 共收到 13008 份论文投 稿，最终接收 2878 篇论文，整体接收率为 22.1%。 作为一场为国内 AI 人才打造的盛会，本次论文分享会 ...

在数字化时代，视觉信息在知识传递和决策支持中的重要性日益凸显。然而，传统的检索增强型生成（RAG）方法在处理视觉丰富信息时面临着诸多挑战。 一方面，传统的基于文本的方法无法处理视觉相关数据；另一方面，现有的视觉 RAG 方法受限于定义的固定流程，难以有效激活模型的推理能力。 来自阿里巴巴通义实验室的最新研究成果 ——VRAG-RL（Empower Vision-Perception-Based RAG for Visually Rich Information Understanding via Iterative Reasoning with Reinforcement Learning），将强化学习算法引入多模态智能体训练，借助迭代推理和视觉感知空间，全方位提升视觉语言 模型（VLMs）在检索、推理和理解视觉信息方面的能力，为纯视觉检索增强生成任务提供有效解决方案，代码、模型全面开源！ Paper 地址：arxiv.org/pdf/2505.22019 Github 地址：https://github.com/Alibaba-NLP/VRAG 为了解决现有 RAG 方法在处理视觉丰富文档时面临的挑战，尤其 ...

机器之心报道 机器之心编辑部 不用换模型、不用堆参数，靠 SUGAR 模型性能大增！ 在深度学习领域中，对激活函数的探讨已成为一个独立的研究方向。例如 GELU、SELU 和 SiLU 等函数凭借其平滑梯度与卓越的收敛特性，已成为热门选择。 尽管这一趋势盛行，经典 ReLU 函数仍因其简洁性、固有稀疏性及其他优势拓扑特性而广受青睐。 然而 ReLU 单元易陷入所谓的「死亡 ReLU 问题」， 一旦某个神经元在训练中输出恒为 0，其梯度也为 0，无法再恢复。 这一现象最终制约了其整体效能，也是 ReLU 网络的重大缺陷。 正是死亡 ReLU 问题催生了大量改进的线性单元函数，包括但不限于：LeakyReLU、PReLU、GELU、SELU、SiLU/Swish 以及 ELU。这些函数通过为负预激活值 引入非零激活，提供了不同的权衡。 本文，来自德国吕贝克大学等机构的研究者引入了一种新颖的方法：SUGAR（Surrogate Gradient for ReLU），在不牺牲 ReLU 优势的情况下解决了 ReLU 的局限 性。即前向传播仍使用标准 ReLU（保持其稀疏性和简单性），反向传播时替换 ReLU 的导数为 ...

在大语言模型（LLM）飞速发展的今天，Chain-of-Thought（CoT）技术逐渐成为提升复杂推理能力的关键范式，尤 其是在数学、逻辑等结构化任务中表现亮眼。 本文的共同第一作者是徐皓雷和颜聿辰。徐皓雷是浙江大学的一年级硕士生，主要研究兴趣集中在大模型推理和可解释 性研究；颜聿辰是浙江大学博士三年级研究生，主要研究兴趣集中在大模型推理和智能体。本文通讯作者是浙江大学鲁 伟明教授和沈永亮研究员。 但你是否注意到：即使是精心构建的 CoT 数据，也可能存在 "跳跃式" 推理，缺失关键中间步骤。对人类专家来说这 些步骤或许 "理所当然"，但对模型而言，却可能是无法逾越的鸿沟。 为了解决这一问题，浙江大学联合微软亚洲研究院、香港中文大学提出了 Thought Leap Bridge 任务，并开发了思维 链修复方法：CoT-Bridge。实验显示，该方法显著提升了多个数学与逻辑任务中的推理准确率，并能作为 "即插即用" 的模块嵌入到知识蒸馏、强化学习等流程中。 CoT 不等于 Coherent-of-Thought 思维跳跃是如何破坏推理链的？ CoT 的设计初衷是让大模型像人一样 "按步骤思考"，然而研究团队发 ...