![[bcafd7ef-7307-4e83-b748-c5b73c965037.mp3]]

AI智能体Zochi独立完成论文被顶会ACL录用，评分位列前8.2%

简报：

• Intology AI开发的博士级智能体Zochi独立完成的研究论文被自然语言处理领域顶级会议ACL 2025主会议录用
• Zochi是首个能独立完成从假设提出到论文发表全流程的AI系统，论文平均得分7.67分
• 论文提出突破大模型安全的Tempest框架，通过"多轮对话树搜索"实现97%的越狱攻击成功率
• ACL主会议录用率仅约20%，Zochi论文评分位列投稿前8.2%，达到博士级科研水平
• Intology已开放Zochi的Beta测试注册

相关链接：

• 博士级AI智能体写的论文，首次登上顶会ACL！人类作者只是监工
• 刚刚，AI科学家Zochi在ACL「博士毕业」，Beta测试今日上线
• AI写的论文首次被顶会ACL录用，评分位列投稿前8.2%

AI Agent Zochi Independently Completes Paper Accepted by Top Conference ACL, Scoring in Top 8.2%

Newsletter:

• Zochi, a PhD-level AI agent developed by Intology AI, has independently completed a research paper accepted by the main conference of ACL 2025, a top-tier event in natural language processing 🌟
• Zochi is the first AI system capable of independently handling the entire process from hypothesis formulation to paper publication, achieving an average score of 7.67
• The paper introduces the groundbreaking Tempest framework for large model security, achieving a 97% success rate in jailbreak attacks through "multi-turn dialogue tree search" 🔍
• With an acceptance rate of only about 20% at the ACL main conference, Zochi's paper ranks in the top 8.2% of submissions, demonstrating PhD-level research capability 🎓
• Intology has opened Beta testing registration for Zochi

Related Links:

• PhD-Level AI Agent Writes Paper, Makes Debut at Top Conference ACL! Human Authors Merely Supervise
• Breaking News: AI Scientist Zochi "Graduates PhD" at ACL, Beta Testing Launches Today
• AI-Written Paper Accepted by Top Conference ACL for the First Time, Scoring in Top 8.2% of Submissions

阿里巴巴开源自主搜索AI智能体WebAgent，具备多步推理能力

简报：

• 阿里巴巴于2025年5月30日在GitHub开源自主搜索AI智能体WebAgent，具备端到端自主信息检索与多步推理能力
• WebAgent能模拟人类网络行为，主动搜索多个学术数据库，筛选相关文献并进行深入分析和总结
• 系统包含WebDancer（智能体训练框架）和WebWalker（LLM基准测试）两大组件
• WebDancer框架采用创新的数据合成方法，通过短推理和长推理两种方式生成训练数据
• 该技术可整合不同文献观点，为用户提供全面精准的研究报告

相关链接：

• 阿里巴巴开源自主搜索 AI 智能体 WebAgent

Alibaba Open-Sources Self-Developed Search AI Agent WebAgent with Multi-Step Reasoning Capabilities

Newsletter:

• On May 30, 2025, Alibaba open-sourced its self-developed search AI agent, WebAgent, on GitHub, featuring end-to-end autonomous information retrieval and multi-step reasoning capabilities. 🌐
• WebAgent can mimic human online behavior, proactively search multiple academic databases, filter relevant literature, and conduct in-depth analysis and summarization. 📚
• The system comprises two main components: WebDancer (an agent training framework) and WebWalker (an LLM benchmark).
• The WebDancer framework employs an innovative data synthesis method, generating training data through both short and long reasoning approaches.
• This technology can integrate diverse perspectives from literature, providing users with comprehensive and accurate research reports. 🚀

Related Links:

• Alibaba Open-Sources Self-Developed Search AI Agent WebAgent

大模型在加密数据评测中表现不佳，Qwen3准确率不足10%

简报：

• 上海AI Lab等机构联合推出CipherBank评测，测试大语言模型在密码学解密任务中的表现
• 评测结果显示当前大模型在解密任务上整体表现不佳，最优模型准确率未过半
• Claude-3.5-Sonnet和o1表现最佳，DeepSeek系列略优于通用模型
• GPT-4o、Gemini等模型表现平庸，Qwen2.5、Llama3.1等开源模型表现较差
• 最新Qwen3系列模型表现不佳，30B和32B模型准确率均未超过10%

相关链接：

• SOTA大模型遇上加密数据评测：Qwen3未破10%，o1也栽了

Large Models Underperform in Encrypted Data Evaluation, Qwen3 Accuracy Below 10%

Newsletter:

• Shanghai AI Lab and other institutions jointly launch CipherBank evaluation, testing the performance of large language models in cryptographic decryption tasks
• Evaluation results show that current large models generally perform poorly in decryption tasks, with the best model's accuracy rate not exceeding 50%
• Claude-3.5-Sonnet and o1 perform the best, while DeepSeek series slightly outperforms general models
• GPT-4o, Gemini, and other models show mediocre results, while open-source models like Qwen2.5 and Llama3.1 perform poorly
• The latest Qwen3 series models underperform, with both 30B and 32B models achieving accuracy rates below 10% 😕

Related Links:

• SOTA Large Models Face Encrypted Data Evaluation: Qwen3 Below 10%, o1 Also Struggles 😓🧠

华为发布7180亿参数Pangu Ultra MoE模型，全流程在昇腾NPU训练

简报：

• 华为盘古团队发布参数规模7180亿的Pangu Ultra MoE模型，全流程在昇腾NPU上训练
• 模型采用256个路由专家，每个token激活8个专家，总参数量718B，激活量39B
• 团队提出DSSN稳定架构和TinyInit小初始化方法，实现10T tokens数据长期稳定训练
• 采用EP group loss负载优化方法，保证专家负载均衡并提升领域特化能力
• 预训练阶段昇腾Atlas 800T A2万卡集群MFU提升至41%，后训练阶段单节点吞吐达35K Tokens/s

相关链接：

• Pangu Ultra MoE模型架构与训练方法技术报告
• 华为发布准万亿模型盘古Ultra MoE

Huawei Releases 718 Billion Parameter Pangu Ultra MoE Model, Fully Trained on Ascend NPU

Newsletter:

• Huawei's Pangu team unveils the Pangu Ultra MoE model with a staggering 718 billion parameters, fully trained on Ascend NPU. 🚀
• The model features 256 routing experts, activating 8 experts per token, with a total of 718B parameters and 39B active parameters.
• The team introduces the DSSN stable architecture and TinyInit small initialization method, achieving stable long-term training on 10T tokens data. 📈
• Utilizes the EP group loss load optimization method to ensure balanced expert load and enhance domain-specific capabilities.
• During the pre-training phase, the Ascend Atlas 800T A2 cluster with 10,000 cards achieves an MFU of 41%, while the post-training phase reaches a single-node throughput of 35K Tokens/s. 💻

Related Links:

• Technical Report on Pangu Ultra MoE Model Architecture and Training Methods
• Huawei Releases Near-Trillion Parameter Model Pangu Ultra MoE

百度飞桨推出AIGC应用开发实战营，助力传媒行业AI创新

简报：

• 百度飞桨和文心大模型联合推出「AIGC应用开发实战营」项目，聚焦游戏、营销、影视三大开发方向
• 活动提供技术指导、创新陪跑、资源支持和孵化推广等权益，帮助开发者实现AI应用落地
• 3月25日至27日每晚19:00举办直播分享，揭秘AIGC应用开发技巧和行业解决方案
• 参与者可获得官方技术指导、大模型开发文档和产业课程集锦包等资源
• 活动面向企业和开发者，获奖应用将获得品牌包装和流量推广支持

相关链接：

• 官方技术大牛手把手答疑支持，飞桨AIGC应用开发实战营来啦！

Baidu PaddlePaddle Launches AIGC Application Development Bootcamp to Boost AI Innovation in Media Industry

Newsletter:

• Baidu PaddlePaddle and Wenxin Large Model jointly launch the "AIGC Application Development Bootcamp" project, focusing on three key development areas: gaming, marketing, and film & television 🎮🎥
• The event offers technical guidance, innovation support, resource assistance, and incubation promotion benefits to help developers bring AI applications to life 🚀
• Live sharing sessions will be held from March 25 to 27 at 19:00 each evening, unveiling AIGC application development tips and industry solutions 📡
• Participants can access official technical guidance, large model development documentation, and a comprehensive industry course package
• Open to enterprises and developers, winning applications will receive brand packaging and traffic promotion support

Related Links:

• Official Tech Experts Provide Hands-On Support, PaddlePaddle AIGC Application Development Bootcamp is Here!

波士顿动力Atlas机器人升级3D感知能力，精准完成工业任务

简报：

• 波士顿动力Atlas机器人完成重磅升级，新增3D空间感知和实时物体追踪能力，可自主执行复杂工业任务
• 在汽车工厂演示中，Atlas能360度转动头部识别掉落零件，即使人类故意移动装置也能精准完成装配
• 技术升级包括2D与3D感知融合、物体位姿追踪及基于物理特性的精确校准方案
• 官方YouTube演示视频获十余万观看量，网友称赞其环境感知和纠错能力
• 机器人需解决金属零件反光、低对比度物体识别等工业环境特有挑战

相关链接：

• 波士顿动力机器人进厂打工现逆天操作！3D感知+实时追踪，人类捣乱完全不带怕的
• 3D感知升级！波士顿动力机器人进厂打工
• 波士顿动力机器人进厂打工现逆天操作！3D感知+实时追踪，人类捣乱完全不带怕的

Boston Dynamics Atlas Robot Upgrades 3D Perception, Precisely Completes Industrial Tasks

Brief:

• Boston Dynamics' Atlas robot undergoes a major upgrade, featuring enhanced 3D spatial perception and real-time object tracking, enabling autonomous execution of complex industrial tasks. 🤖
• In a car factory demonstration, Atlas can rotate its head 360 degrees to identify fallen parts and accurately complete assembly tasks, even when humans deliberately move components. 🔧
• Technological advancements include the fusion of 2D and 3D perception, object pose tracking, and precise calibration based on physical properties.
• The official YouTube demonstration video has garnered over 100,000 views, with netizens praising its environmental awareness and error-correction capabilities. 👍
• The robot still needs to overcome industrial environment challenges, such as reflections from metal parts and low-contrast object recognition.

Related Links:

• Boston Dynamics Robot Works in Factory with Mind-Blowing Moves! 3D Perception + Real-Time Tracking, Unfazed by Human Interference
• 3D Perception Upgrade! Boston Dynamics Robot Works in Factory
• Boston Dynamics Robot Works in Factory with Mind-Blowing Moves! 3D Perception + Real-Time Tracking, Unfazed by Human Interference

DeepSeek-R1 0528版本开源，编程能力直逼Claude 4

简报：

• DeepSeek于5月28日凌晨在Hugging Face开源R1-0528版本，模型权重已公开但模型卡暂未更新
• 新版本在LiveCodeBench基准测试中表现接近OpenAI o3-mini（High）和o4-mini（Medium）
• 编程能力显著提升，实测显示在前端编码等任务上超越Claude 4 Sonnet
• 支持单任务30-60分钟的长时思考，网友实测思考时间可达25分钟
• 在AIME 2025数学测试中正确率从70%提升至87.5%，接近OpenAI o3的88.9%
• 采用MIT开源协议，允许商用和模型蒸馏，API价格仅为o3的3.3%

相关链接：

• DeepSeek-R1-0528模型权重
• 日经中文网报道

DeepSeek-R1 0528 Version Released as Open Source, Coding Skills Rival Claude 4

Newsletter:

• DeepSeek released the R1-0528 version on Hugging Face at midnight on May 28, with model weights made public, though the model card has not yet been updated.
• The new version performs close to OpenAI's o3-mini (High) and o4-mini (Medium) in the LiveCodeBench benchmark test.
• Significant improvement in coding capabilities, with real-world tests showing it surpasses Claude 4 Sonnet in tasks like front-end coding. 💻
• Supports long-duration thinking of 30-60 minutes for single tasks, with user tests confirming thinking time up to 25 minutes. ⏳
• Accuracy in the AIME 2025 math test improved from 70% to 87.5%, nearing OpenAI o3's 88.9%.
• Released under the MIT open-source license, allowing commercial use and model distillation, with API pricing at just 3.3% of o3's cost. 💰

Related Links:

• DeepSeek-R1-0528 Model Weights
• Nikkei Chinese Report