🤖 AI 速览

OpenAI 集中发布 GPT-6 Astra 企业用例：Harvey 优化法律文书，invideo 调色效率提升三倍，Ringg 客服解决率达65%。安全治理同步升温：Altman 在联合国安理会强调人类控制，Google 为 Private AI Compute 引入服务端记忆，OpenAI 扩展乌克兰网络防御。评估侧开始审计捷径学习与数值不变性。
📋 文章元数据
发布时间
2026-09-24
类型
ai-daily
字数
4968
阅读时长
24 min

2026-09-24 AI日更 | AI 治理与垂直落地同日升温：Altman 安理会谈人类控制，Astra 进入法律、调色与客服 链接到标题

OpenAI 集中发布 GPT-6 Astra 企业用例：Harvey 优化法律文书，invideo 调色效率提升三倍，Ringg 客服解决率达65%。安全治理同步升温：Altman 在联合国安理会强调人类控制，Google 为 Private AI Compute 引入服务端记忆，OpenAI 扩展乌克兰网络防御。评估侧开始审计捷径学习与数值不变性。

📖 本期 Watch List 深度导读 链接到标题

今天有三条主线值得关注。一是企业级 AI 落地：OpenAI 集中放出 GPT-6 Astra 用例，Harvey 提升法律文书结构化、invideo 调色效率提高三倍、Ringg 客服解决率最高达 65%，Gemini 3.8 TTS 也值得语音与工程团队重点体验。二是安全与治理：Google 为 Private AI Compute 引入服务端记忆，Altman 在联合国安理会谈人类控制与国际协作，OpenAI 还发布 MentalHealthBench 并扩大乌克兰民用网络防御，政策与安全团队建议跟踪。三是评估与训练反思：arXiv 两篇论文分别审计假新闻数据集的捷径学习、检验数值表示不变性，另有一篇 Rust 端到端训练经验，适合模型评估与基础设施团队精读。

🌐 X 平台 AI 热点快讯 链接到标题

话题 1:Meta Launches Muse AI Agent Amid Open-Source Inspiration and Logo Memes 链接到标题

  • 分类:AI · News
  • 概况:热度时间:1 day ago,相关帖子数:4900
  • 是什么事:Meta发布名为Muse的AI智能体，并伴随开源灵感讨论与标志恶搞。
  • 为什么重要:此举显示Meta在AI代理赛道持续加码，其开源策略或影响行业生态与竞争格局。
  • 讨论概况:X上的讨论主要集中在其是否真正开源、与其他AI代理的能力对比，以及围绕产品标志的玩梗和品牌传播。

话题 2:Meta’s Alexandr Wang Embraces Butthole Logo Memes for New AI Agent Muse 链接到标题

  • 分类:AI · Other
  • 概况:热度时间:1 day ago,相关帖子数:4800
  • 是什么事:Meta 的新 AI 代理 Muse 发布后，其 logo 被网友调侃为“屁眼”图案，Alexandr Wang 公开拥抱该 meme 并回应。
  • 为什么重要:此事显示 AI 产品营销正与 meme 文化深度绑定，争议性视觉符号帮助新代理在社交平台获得高曝光，也考验品牌如何在调侃中引导公众注意力到产品能力上。
  • 讨论概况:X 上主要争论 logo 是否不当、拥抱 meme 是机智还是尴尬，以及 Muse 本身的功能表现是否被梗图话题稀释。

话题 3:Meta’s Muse Mascot Steals Spotlight Ahead of Connect 链接到标题

  • 分类:AI · News
  • 概况:热度时间:7 hours ago,相关帖子数:762
  • 是什么事:Meta 的 Muse 吉祥物在 Connect 大会前因一则 AI 生成视频成为关注焦点。
  • 为什么重要:这展示了生成式 AI 视频工具在品牌与 IP 宣传中的创意潜力，也为 Meta Connect 带来预热，并凸显多模型 AI 创作平台的关注度。
  • 讨论概况:X 上主要讨论不同 AI 视频模型的生成效果与选择、Meta Connect 可能发布的内容，以及聚合多种模型的一站式创作平台是否更具优势。

话题 4:Alexandr Wang Hypes Meta’s Muse AI with Playful Mascot Posts 链接到标题

  • 分类:AI · Other
  • 概况:热度时间:17 hours ago,相关帖子数:1900
  • 是什么事:Scale AI创始人Alexandr Wang在X平台通过发布俏皮吉祥物帖文，为Meta的Muse AI宣传造势。
  • 为什么重要:此举显示AI行业关键人物公开为Meta的新模型站台，可能提升Muse AI在开发者社区的关注度与可信度，并加剧生成式AI工具竞争。
  • 讨论概况:讨论焦点包括Muse AI的实际能力与产品定位、Alexandr Wang是单纯玩梗还是存在商业合作，以及该模型能否与现有图像或视频生成工具形成差异。

话题 5:Katzenberg Urges Hollywood to Embrace AI as Creative Tool 链接到标题

  • 分类:AI · News
  • 概况:热度时间:7 hours ago,相关帖子数:5200
  • 是什么事:杰弗里·卡岑伯格公开敦促好莱坞将人工智能作为创意工具加以拥抱，而非视为威胁。
  • 为什么重要:好莱坞高层推动生成式AI进入核心创作流程，可能加速AI在影视编剧、特效和内容生产中的主流化应用。
  • 讨论概况:X上的讨论焦点在于AI会否取代创意岗位、版权与报酬如何界定，以及AI能否真正成为创作伙伴而非削弱原创性。

话题 6:AI Video Shows Trump as Mexican Grandmother in Kitchen Parody 链接到标题

  • 分类:AI · Other
  • 概况:热度时间:,相关帖子数:707
  • 是什么事:一段AI生成视频将特朗普塑造成墨西哥祖母形象，在厨房场景中进行戏仿。
  • 为什么重要:该视频凸显AI生成视频在政治人物模仿、文化身份转换和深度伪造方面的能力与潜在风险。
  • 讨论概况:X上的讨论聚焦于这究竟是娱乐戏仿还是误导性深度伪造，以及AI生成政治人物内容是否需要更严格标识和监管。

今日 X 上的 AI 舆情小结 链接到标题

今天舆论主线围绕 Meta 发布 Muse AI 智能体及其标志被恶搞展开，共识在于 AI 产品营销已与 meme 文化深度绑定，争议性符号和关键人物站台能迅速带来高曝光并加剧生成式 AI 竞争。分歧主要集中在 Muse 是否真正开源、logo 争议与拥抱 meme 是机智还是尴尬、Alexandr Wang 是玩梗还是商业合作，以及好莱坞应把 AI 视为创作工具还是对创意岗位的威胁。潜在风险则体现在 AI 生成政治人物戏仿可能滑向误导性深度伪造，版权与创作者报酬缺乏清晰界定，同时品牌注意力容易被梗图话题稀释，使产品真实能力反而被忽视。

💡 大佬观点(Influencer Insights) 链接到标题

今日大佬观点暂缺,推荐阅读 Watch List 深度内容。

📚 附录:今日 Watch List 更新源列表 链接到标题

时间窗口:最近 3 天;覆盖 22 个源;共 43 条更新

All-In Podcast (A_full) 链接到标题

  • Steve Hilton & Spencer Pratt: Fixing California, Cheaper Gas, Ballot Fraud & CA Republicans
    • 发布时间:2026-09-23 09:44 北京时间
    • 摘要:- IREN 是一个垂直集成的人工智能云平台，为人工智能训练和推理提供数据中心、计算和软件。
      • Oracle 将数据、应用程序和基础设施连接起来，将人工智能转化为业务成果，并随着人工智能的发展而提供优化的灵活性、选择和控制。
      • 安永帮助科技创新者从初创企业到退出企业再到大型企业。
      • 我们致力于为每个人提供充分发挥潜力的工具，并确保所有人都能受益于技术。
      • Keel基础设施拥有高性能计算和人工智能运行所需的电力、土地和连接，并以安全的能源资产和遍布北美的电网互连为后盾。
    • EN 要点:
      • (0:00) Steve Hilton joins the Besties
      • (3:53) How one-party rule broke the state, regulatory bloat & who really writes the laws
      • (12:09) The path to victory: the persuadable voter, the cost-of-living pitch & the case against Becerra
      • (21:46) Spencer Pratt joins: the message that cut through & the LA race that flipped

Stratechery by Ben Thompson (A_full) 链接到标题

  • More on Muse, Amazon, and Walmart; Muse and Expedia; Whither Google?
    • 发布时间:2026-09-23 18:00 北京时间
    • 摘要:【待翻译】- Meta needs Walmart to wait out Amazon; Expedia seeks to keep its middleware position; meanwhile, where is Google?
      • $15 / month or $150 / year.
      • Substantial analysis of the news of the day delivered via three weekly emails or podcasts.
      • Stratechery Interviews.
      • Interviews with leading public CEOs, private company founders, and discussions with fellow analysts.
    • EN 要点:
      • Meta needs Walmart to wait out Amazon; Expedia seeks to keep its middleware position; meanwhile, where is Google

OpenAI Blog (A_full) 链接到标题

  • Two years of OpenAI Academy

    • 发布时间:2026-09-24 00:00 北京时间
    • 摘要:【待翻译】- Marking two years of OpenAI Academy and bringing AI skills to even more communities.
      • This piece from OpenAI Blog explains how Two years of OpenAI Academy shapes the broader AI and infrastructure landscape.
      • It also surfaces practical implications for founders, operators, and investors following Two years of OpenAI Academy.
    • EN 要点:
      • Marking two years of OpenAI Academy and bringing AI skills to even more communities.

  • OpenAI extends cyber access to Ukraine for civilian defense

    • 发布时间:2026-09-23 21:00 北京时间
    • 摘要:- OpenAI 正在将其 Daybreak 计划的访问范围扩大到乌克兰政府，以支持民用基础设施的网络防御。
      • OpenAI 博客的这篇文章解释了 OpenAI 如何将网络访问扩展到乌克兰以进行民用防御，从而塑造更广泛的人工智能和基础设施格局。
      • 在 OpenAI 将网络访问扩展到乌克兰用于民用防御之后，它还对创始人、运营商和投资者产生了实际影响。
    • EN 要点:
      • OpenAI is extending access to its Daybreak program to the Government of Ukraine to support the cyber defense of civilian infrastructure.

  • Sam Altman’s remarks at the United Nations Security Council

    • 发布时间:2026-09-23 20:00 北京时间
    • 摘要:【待翻译】- OpenAI CEO Sam Altman discusses AI safety, human control, and international cooperation in remarks to the United Nations Security Council.
      • This piece from OpenAI Blog explains how Sam Altman’s remarks at the United Nations Security Council shapes the broader AI and infrastructure landscape.
      • It also surfaces practical implications for founders, operators, and investors following Sam Altman’s remarks at the United Nations Security Council.
    • EN 要点:
      • OpenAI CEO Sam Altman discusses AI safety, human control, and international cooperation in remarks to the United Nations Security Council.

  • Harvey turns legal context into stronger drafts with GPT-6 Astra

    • 发布时间:2026-09-23 20:00 北京时间
    • 摘要:【待翻译】- GPT-6 Astra produces more structured, context-aware legal documents, freeing lawyers to focus on strategy.
      • This piece from OpenAI Blog explains how Harvey turns legal context into stronger drafts with GPT-6 Astra shapes the broader AI and infrastructure landscape.
      • It also surfaces practical implications for founders, operators, and investors following Harvey turns legal context into stronger drafts with GPT-6 Astra.
    • EN 要点:
      • GPT-6 Astra produces more structured, context-aware legal documents, freeing lawyers to focus on strategy.

  • How invideo improves color grading 3x with GPT‑6 Astra

    • 发布时间:2026-09-23 20:00 北京时间
    • 摘要:【待翻译】- With GPT‑6 Astra, invideo plans edits with greater precision, improves color correction and grading threefold, and produces 50 custom effects in one day.
      • This piece from OpenAI Blog explains how How invideo improves color grading 3x with GPT‑6 Astra shapes the broader AI and infrastructure landscape.
      • It also surfaces practical implications for founders, operators, and investors following How invideo improves color grading 3x with GPT‑6 Astra.
    • EN 要点:
      • With GPT‑6 Astra, invideo plans edits with greater precision, improves color correction and grading threefold, and produces 50 custom effects in one day.

  • Ringg’s AI agents resolve up to 65% of customer calls with OpenAI

    • 发布时间:2026-09-23 20:00 北京时间
    • 摘要:【待翻译】- Using GPT-5.6, Ringg powers multilingual agents across voice, chat, WhatsApp, and web for 90% less cost vs.
      • Using GPT-5.6, Ringg powers multilingual agents across voice, chat, WhatsApp, and web for 90% less cost vs.
      • Ringg’s AI agents resolve up to 65% of customer calls with OpenAI.
    • EN 要点:
      • Using GPT-5.6, Ringg powers multilingual agents across voice, chat, WhatsApp, and web for 90% less cost vs

  • Introducing MentalHealthBench

    • 发布时间:2026-09-23 18:00 北京时间
    • 摘要:- MentalHealthBench 是一个专家级基准，用于评估现实心理健康对话中有用且安全的 AI 响应。
      • OpenAI 博客中的这篇文章解释了 MentalHealthBench 的介绍如何塑造更广泛的人工智能和基础设施格局。
      • 引入 MentalHealthBench 后，它还为创始人、运营商和投资者带来了实际影响。
    • EN 要点:
      • MentalHealthBench is an expert-informed benchmark for evaluating helpful and safe AI responses across realistic mental health conversations.

  • Airbnb widens access to GPT-6 Astra and OpenAI frontier models

    • 发布时间:2026-09-23 09:00 北京时间
    • 摘要:- 了解 Airbnb 如何扩大对 GPT-6 Astra 和 OpenAI 前沿模型的访问，以帮助工程团队解决错误、设计系统并加快交付速度。
      • OpenAI 博客的这篇文章解释了 Airbnb 如何扩大对 GPT-6 Astra 的访问，以及 OpenAI 前沿模型如何塑造更广泛的人工智能和基础设施格局。
      • 在 Airbnb 扩大对 GPT-6 Astra 和 OpenAI 前沿模型的访问之后，它还为创始人、运营商和投资者带来了实际影响。
    • EN 要点:
      • Learn how Airbnb is expanding access to GPT-6 Astra and OpenAI frontier models to help engineering teams solve bugs, design systems, and ship faster.

  • Grab and OpenAI bring practical AI skills to Southeast Asia

    • 发布时间:2026-09-23 08:00 北京时间
    • 摘要:- OpenAI 和 Grab 推出 GO Forward with AI，这是一项区域计划，帮助 30,000 名合作伙伴在整个东南亚培养实用的人工智能技能。
      • OpenAI 博客的这篇文章解释了 Grab 和 OpenAI 如何为东南亚带来实用的人工智能技能，塑造更广泛的人工智能和基础设施格局。
      • 在 Grab 和 OpenAI 将实用的人工智能技能带到东南亚之后，它还为创始人、运营商和投资者带来了实际影响。
    • EN 要点:
      • OpenAI and Grab launch GO Forward with AI, a regional programme helping 30,000 partners build practical AI skills across Southeast Asia.

Google DeepMind Blog (A_full) 链接到标题

  • Advancing Private AI Compute with secure, server-side memory

    • 发布时间:2026-09-24 00:00 北京时间
    • 摘要:【待翻译】- Introducing private, server-side memory to Private AI Compute for personal AI.
      • This piece from Google DeepMind Blog explains how Advancing Private AI Compute with secure, server-side memory shapes the broader AI and infrastructure landscape.
      • It also surfaces practical implications for founders, operators, and investors following Advancing Private AI Compute with secure, server-side memory.
    • EN 要点:
      • Introducing private, server-side memory to Private AI Compute for personal AI.

  • Gemini 3.8 text-to-speech says hello

    • 发布时间:2026-09-23 23:25 北京时间
    • 摘要:【待翻译】- Gemini 3.8 text-to-speech says hello.
      • This piece from Google DeepMind Blog explains how Gemini 3.8 text-to-speech says hello shapes the broader AI and infrastructure landscape.
      • It also surfaces practical implications for founders, operators, and investors following Gemini 3.8 text-to-speech says hello.
    • EN 要点:
      • Gemini 3.8 text-to-speech says hello

ArXiv cs.AI (B_intro+search) 链接到标题

  • Do Synthetic Personas Predict Real Audience Response? A Sim-to-Real Study Where a No-Persona Baseline Beats Persona-Based Copy Simulation

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25010v1 Announce Type: new.
      • Abstract: Marketers increasingly use large language models (LLMs) as “synthetic personas” to predict how an audience will react to a piece of copy before it ships, encouraged by evidence that profile-conditioned LLMs mimic human samples.
      • But is that prediction actually valid against real behaviour - and does the persona machinery help?
      • We present a sim-to-real validity study using the Upworthy Research Archive - thousands of headline A/B tests on shared real traffic, with measured click-through - as held-out ground truth.
    • EN 要点:
      • arXiv:2609.25010v1 Announce Type: new
      • Abstract: Marketers increasingly use large language models (LLMs) as “synthetic personas” to predict how an audience will react to a piece of copy before it shi…
      • But is that prediction actually valid against real behaviour - and does the persona machinery help
      • We present a sim-to-real validity study using the Upworthy Research Archive - thousands of headline A/B tests on shared real traffic, with measured click-throug…

  • Do Existing Preconditioners Improve Biomedical Tabular Foundation Learning? An Empirical Study on TabPFN Optimization

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25013v1 Announce Type: new.
      • Abstract: Tabular foundation models have recently shown strong potential for structured biomedical data analysis.
      • Among them, TabPFN has emerged as an effective approach for low-data tabular classification tasks.
      • However, the impact of optimization and preconditioning strategies on biomedical fine-tuning remains largely unexplored.
    • EN 要点:
      • arXiv:2609.25013v1 Announce Type: new
      • Abstract: Tabular foundation models have recently shown strong potential for structured biomedical data analysis
      • Among them, TabPFN has emerged as an effective approach for low-data tabular classification tasks
      • However, the impact of optimization and preconditioning strategies on biomedical fine-tuning remains largely unexplored

  • 4DGS-JEPA: Temporally Compositional Joint-Embedding Prediction for Dynamic Gaussian Splatting

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:- arXiv:2609.25036v1 公告类型：新。
      • 摘要：动态高斯分布提供了不断变化的 3D 场景的明确表示，但现有方法主要针对重建、未来状态生成或渲染进行优化，而不是用于学习可重用的预测动力学。
      • 我们提出 4DGS-JEPA，一种高斯原生联合嵌入预测架构，用于动态高斯场景的因果多水平预测。
      • 该模型使用分层场景、运动组和高斯级表示以及支持直接预测和递归推出的地平线条件转换运算符。
    • EN 要点:
      • arXiv:2609.25036v1 Announce Type: new
      • Abstract: Dynamic Gaussian Splatting provides an explicit representation of evolving 3D scenes, but existing approaches are primarily optimized for reconstructi…
      • We propose 4DGS-JEPA, a Gaussian-native joint-embedding predictive architecture for causal multi-horizon prediction over dynamic Gaussian scenes
      • The model uses a hierarchical scene-, motion-group-, and Gaussian-level representation together with a horizon-conditioned transition operator that supports bot…

  • An Accurate and Interpretable Hyper Graph Neural Network for GBM Survival Prediction

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25088v1 Announce Type: new.
      • Abstract: Survival prediction for glioblastoma multiforme (GBM) demands models that are both accurate and interpretable, yet existing approaches treat these objectives as com- peting, where performant models sacrifice transparency, while interpretable models accept degraded predictive power.
      • We argue that this trade-off is not inherent.
      • Graph neural net- works offer a structural foundation for extracting interpretable, explainable representations without compromising discriminative ability.
    • EN 要点:
      • arXiv:2609.25088v1 Announce Type: new
      • Abstract: Survival prediction for glioblastoma multiforme (GBM) demands models that are both accurate and interpretable, yet existing approaches treat these obj…
      • We argue that this trade-off is not inherent
      • Graph neural net- works offer a structural foundation for extracting interpretable, explainable representations without compromising discriminative ability

  • Ovis-Embedding: Pushing the Frontiers of Universal Omni-Modal Embeddings

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:- arXiv:2609.25165v1 公告类型：新。
      • 摘要：在本报告中，我们介绍了 \textbf{Ovis-Embedding}，这是一个基于文本、图像、视频和音频的本机集成构建的最先进的全模态嵌入系列。
      • Ovis-Embedding 没有组装单独的模态塔，而是使用共享的多模态主干在公共表示空间中对不同模态进行编码。
      • 具体来说，我们做出了 \textbf{三个关键进展}： (1) \textbf{原生全模态初始化}：我们采用预训练的 Qwen-omni 模型作为嵌入主干，并通过低秩初始化的对比训练来适应它； (2) \textbf{以数据为中心的全模态训练}：我们构建了一个广泛的、高质量的语料库，涵盖文本、图像、视频、音频和交错的多模态数据。
    • EN 要点:
      • arXiv:2609.25165v1 Announce Type: new
      • Abstract: In this report, we introduce \textbf{Ovis-Embedding}, a state-of-the-art omni-modal embedding family built on native integration of text, image, video…
      • Instead of assembling separate modality towers, Ovis-Embedding uses a shared multimodal backbone to encode different modalities in a common representation space
      • Specifically, we make \textbf{three key advances}: (1) \textbf{native omni-modal initialization}: we adopt a pretrained Qwen-omni model as the embedding backbon…

  • X-Planner: Event-Structured Task Planning for Embodied Intelligence

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25187v1 Announce Type: new.
      • Abstract: Task planning bridges high-level instructions and executable behavior in long-horizon manipulation, yet modern Vision-Language-Action (VLA) systems often leave this intermediate structure implicit.
      • Existing chain-of-thought (CoT) planners also tend to rely on coarse task-level annotations or serialize long reasoning traces token by token.
      • We present X-Planner, a planning front-end that addresses both the supervision and representation of embodied reasoning.
    • EN 要点:
      • arXiv:2609.25187v1 Announce Type: new
      • Abstract: Task planning bridges high-level instructions and executable behavior in long-horizon manipulation, yet modern Vision-Language-Action (VLA) systems of…
      • Existing chain-of-thought (CoT) planners also tend to rely on coarse task-level annotations or serialize long reasoning traces token by token
      • We present X-Planner, a planning front-end that addresses both the supervision and representation of embodied reasoning

  • Lean Pool: An AI-Maintained Archive of Formalized Mathematics

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25199v1 Announce Type: new.
      • Abstract: Lean Pool is a repository of formalized mathematics.
      • It is grown, maintained and optimized by AI agents.
      • arXiv:2609.25199v1 Announce Type: new Abstract: Lean Pool is a repository of formalized mathematics It is grown, maintained and optimized by AI agents.
    • EN 要点:
      • arXiv:2609.25199v1 Announce Type: new
      • Abstract: Lean Pool is a repository of formalized mathematics
      • It is grown, maintained and optimized by AI agents.

  • The AI Neuroscientist: An Interactive Agentic Interface for Neuroimaging Analysis

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25254v1 Announce Type: new.
      • Abstract: Analyzing neuroimaging data requires specialized coding and statistical expertise, which limits accessibility for researchers without computational backgrounds.
      • We present the AI Neuroscientist, a language agent for interactive data exploration.
      • The system integrates a large language model (LLM) with a neuroimaging toolset to perform quality control, modeling, and visualization.
    • EN 要点:
      • arXiv:2609.25254v1 Announce Type: new
      • Abstract: Analyzing neuroimaging data requires specialized coding and statistical expertise, which limits accessibility for researchers without computational ba…
      • We present the AI Neuroscientist, a language agent for interactive data exploration
      • The system integrates a large language model (LLM) with a neuroimaging toolset to perform quality control, modeling, and visualization

  • MedGate-Fusion: Integrating First-Encounter Semantic Narratives and Physiological Biomarkers for Prospective Stroke Risk Stratification

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25272v1 Announce Type: new.
      • Abstract: Prospective stroke risk stratification in primary care is challenging because early risk signals are distributed across routine biomarkers and unstructured clinical narratives.
      • We propose MedGate-Fusion, a multi-modal gated architecture that integrates transformer-based embeddings of first-encounter narratives with ten routinely recorded risk markers.
      • We used electronic medical record data from the Canadian Primary Care Sentinel Surveillance Network (CPCSSN).
    • EN 要点:
      • arXiv:2609.25272v1 Announce Type: new
      • Abstract: Prospective stroke risk stratification in primary care is challenging because early risk signals are distributed across routine biomarkers and unstruc…
      • We propose MedGate-Fusion, a multi-modal gated architecture that integrates transformer-based embeddings of first-encounter narratives with ten routinely record…
      • We used electronic medical record data from the Canadian Primary Care Sentinel Surveillance Network (CPCSSN)

  • When LLM Agents Fail to Read the Room: ReAdapt for Relational Social Reasoning

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:- arXiv:2609.25284v1 公告类型：新。
      • 摘要：社交代理最基本的决定（我应该对这篇文章做出反应吗？
      • 我应该联系谁？）并不是纯粹的内容问题。
      • 正确的行动通常取决于人与人之间的潜在关系——联系强度、互惠、相互联系——而不是哪个内容最突出。
    • EN 要点:
      • arXiv:2609.25284v1 Announce Type: new
      • Abstract: A social agent’s most basic decisions (should I react to this post
      • who should I reach out to
      • ) are not purely content problems

ArXiv cs.CL (B_intro+search) 链接到标题

  • What Does 99% Accuracy Measure? A Reproducible Audit of Shortcut Learning in a Widely Used Fake News Corpus

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25006v1 Announce Type: new.
      • Abstract: Text classifiers trained on the ISOT/Kaggle “Fake and Real News” corpus routinely report accuracy and F1 above 0.98, a level of performance that sits uneasily beside the difficulty of assessing veracity.
      • Using a transparent TF-IDF and linear-classifier pipeline as a measurement instrument, we audit the corpus along three leakage channels and two distribution-shift protocols, releasing all code and derived numbers.
      • First, the benchmark is partly degenerate: a classifier given only the subject metadata field, with the article text discarded, attains F1 = 1.000, since the two classes have disjoint subjects.
    • EN 要点:
      • arXiv:2609.25006v1 Announce Type: new
      • Abstract: Text classifiers trained on the ISOT/Kaggle “Fake and Real News” corpus routinely report accuracy and F1 above 0.98, a level of performance that sits…
      • Using a transparent TF-IDF and linear-classifier pipeline as a measurement instrument, we audit the corpus along three leakage channels and two distribution-shi…
      • First, the benchmark is partly degenerate: a classifier given only the subject metadata field, with the article text discarded, attains F1 = 1.000, since the tw…

  • Training a Language Model End-to-End in Rust: An Experience Report

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25008v1 Announce Type: new.
      • Abstract: I pretrained a language model end-to-end in Rust - alone, with no team, no PyTorch, and no Python in the training path - for $164 in rented GPU time.
      • I report that as an achievement, not a recommendation: the more useful contribution is a measured failure taxonomy of the two leading Rust ML frameworks, Candle and Burn, as training (not inference) backends in 2026.
      • I document five Candle defects, including fused kernels that silently produce no gradient, and three Burn defects, including a backward pass at roughly 3% of theoretical GPU throughput and a kernel-fusion path that segfaults mid-training at multi-billion-parameter scale.
    • EN 要点:
      • arXiv:2609.25008v1 Announce Type: new
      • Abstract: I pretrained a language model end-to-end in Rust - alone, with no team, no PyTorch, and no Python in the training path - for $164 in rented GPU time
      • I report that as an achievement, not a recommendation: the more useful contribution is a measured failure taxonomy of the two leading Rust ML frameworks, Candle…
      • I document five Candle defects, including fused kernels that silently produce no gradient, and three Burn defects, including a backward pass at roughly 3% of th…

  • Same Quantity, Different Answer: Numerical Representation Invariance in Language Models

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:- arXiv:2609.25009v1 公告类型：新。
      • 摘要：无论数量是小数、分数、百分比、数字词、科学记数法还是精确转换的单位，数值等价应用题都应该产生相同的规范答案。
      • 我们生成了 3,600 个精确理性问题和 8,600 个提示，涵盖五个身份保留转换系列，并评估了五个开放权重系统。
      • 在没有 LLM 法官的情况下对常见答案形式进行标准化的固定语法审核后，规范精度为 0.969-0.996，但轨道正确性降至 0.848-0.981，轨道不变性降至 0.851-0.981；不变但错误的轨道最多占 0.003。
    • EN 要点:
      • arXiv:2609.25009v1 Announce Type: new
      • Abstract: Numerically equivalent word problems should yield the same canonical answer whether a quantity is written as a decimal, fraction, percentage, number w…
      • We generate 3,600 exact-rational problems and 8,600 prompts spanning five identity-preserving transformation families, and evaluate five open-weight systems
      • After a fixed syntax audit that normalizes common answer forms without an LLM judge, canonical accuracy is 0.969-0.996, but orbit correctness falls to 0.848-0.9…

  • A Computational Approach to Measuring Semantic Change in Sanskrit Literature

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25012v1 Announce Type: new.
      • Abstract: Diachronic word embeddings have become the modern standard for tracking semantic change, yet they have been largely validated on modern, high-resource, and well-segmented languages.
      • This paper tests whether the paradigm transfers to Sanskrit, an ancient, low-resource language whose phonological fusion (sandhi), morphological inflection, compounding, and polysemy pose a unique challenge.
      • I assemble a 2.7M-token corpus spanning four canonical periods, recover word boundaries with a neural byte-level sandhi splitter and lemmatizer, and train per-period embeddings across configurations.
    • EN 要点:
      • arXiv:2609.25012v1 Announce Type: new
      • Abstract: Diachronic word embeddings have become the modern standard for tracking semantic change, yet they have been largely validated on modern, high-resource…
      • This paper tests whether the paradigm transfers to Sanskrit, an ancient, low-resource language whose phonological fusion (sandhi), morphological inflection, com…
      • I assemble a 2.7M-token corpus spanning four canonical periods, recover word boundaries with a neural byte-level sandhi splitter and lemmatizer, and train per-p…

  • Retrieved-Span Training for Efficient Query-Focused Meeting Summarization on QMSum

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:- arXiv:2609.25028v1 公告类型：新。
      • 摘要：QMSum 不提供评分器，使得以查询为中心的会议摘要结果难以比较。
      • 我们在一次实施中重新评分或生成 15 个系统。
      • 通过通用推理端口，发布的 406M Fusion-in-Decoder 专家在从上限长输入转移到 2,000 字检索范围时会损失 6.30 ROUGE-1。
    • EN 要点:
      • arXiv:2609.25028v1 Announce Type: new
      • Abstract: QMSum provides no scorer, making query-focused meeting summarization results difficult to compare
      • We rescore or generate 15 systems under one implementation
      • Through a common inference port, a released 406M Fusion-in-Decoder specialist loses 6.30 ROUGE-1 when moved from capped long input to 2,000-word retrieved spans

  • From Tone to Trajectory: Continuous Sentiment and the Shape of Monetary Policy Communication

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:- arXiv:2609.25034v1 公告类型：新。
      • 摘要：央行新闻发布会不仅仅是信息发布——它们是结构化的叙述。
      • 我们研究声明中的情绪形态（而不仅仅是其平均语气）是否携带政策相关信号。
      • 围绕货币立场、经济前景和不确定性三个维度构建欧洲央行和美联储新闻发布会的情绪弧，我们评估他们对政策利率变化、通胀预期和预测分歧的预测内容。
    • EN 要点:
      • arXiv:2609.25034v1 Announce Type: new
      • Abstract: Central bank press conferences are not merely information releases — they are structured narratives
      • We study whether the shape of sentiment within a statement, not just its average tone, carries policy-relevant signals
      • Constructing sentiment arcs for ECB and Fed press conferences along three dimensions — monetary stance, economic outlook, and uncertainty — we assess their…

  • Peerify: Benchmarking Peer-Review Claim Verification

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25046v1 Announce Type: new.
      • Abstract: Peer review plays a central role in scholarly publishing, yet verifying whether reviewer claims are supported by manuscript evidence remains a largely manual and time-consuming process.
      • We present Peerify, a pipeline for manuscript-grounded verification of peer-review claims.
      • Given a manuscript and a review comment, the Peerify pipeline decomposes reviews into atomic claims, retrieves relevant manuscript evidence, and determines whether each claim is supported by the paper.
    • EN 要点:
      • arXiv:2609.25046v1 Announce Type: new
      • Abstract: Peer review plays a central role in scholarly publishing, yet verifying whether reviewer claims are supported by manuscript evidence remains a largely…
      • We present Peerify, a pipeline for manuscript-grounded verification of peer-review claims
      • Given a manuscript and a review comment, the Peerify pipeline decomposes reviews into atomic claims, retrieves relevant manuscript evidence, and determines whet…

  • AIBuildAI-2.5: Efficient Autonomous AI Model Development Through LLM-Guided Tree Search

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25047v1 Announce Type: new.
      • Abstract: Autonomous agents that automatically build artificial intelligence (AI) models could broaden access to AI across science and engineering.
      • A popular line of such agents frames model building as a code search problem and solves it by tree search, in which each node is a candidate program and the tree grows by generating a child program from a parent, and these agents now approach the capability of experienced AI engineers on realistic benchmarks.
      • However, these agents have three weaknesses in efficiency that have not been fully addressed.
    • EN 要点:
      • arXiv:2609.25047v1 Announce Type: new
      • Abstract: Autonomous agents that automatically build artificial intelligence (AI) models could broaden access to AI across science and engineering
      • A popular line of such agents frames model building as a code search problem and solves it by tree search, in which each node is a candidate program and the tre…
      • However, these agents have three weaknesses in efficiency that have not been fully addressed

  • Prompt Breadth and Rollout Refresh Interact in On-Policy Distillation

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25048v1 Announce Type: new.
      • Abstract: How many prompts does on-policy distillation (OPD) need, and how does the answer depend on the student policies that generate its training responses?
      • We study these two controls jointly: prompt breadth and rollout refresh.
      • A 3x3 mathematical-reasoning experiment fixes 14,080 trajectories and 110 optimizer updates while varying the prompt bank and the number of response-generating policy snapshots.
    • EN 要点:
      • arXiv:2609.25048v1 Announce Type: new
      • Abstract: How many prompts does on-policy distillation (OPD) need, and how does the answer depend on the student policies that generate its training responses
      • We study these two controls jointly: prompt breadth and rollout refresh
      • A 3x3 mathematical-reasoning experiment fixes 14,080 trajectories and 110 optimizer updates while varying the prompt bank and the number of response-generating…

  • Mitigating LLM Over-Refusal via Dynamic Semantic Routing Calibratione

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25049v1 Announce Type: new.
      • Abstract: Large language models (LLMs) aligned for safety often suffer from over-refusal, incorrectly rejecting benign yet safety-related instructions.
      • Prior studies primarily attribute this to static representation overlap, largely overlooking the underlying dynamic mechanisms.
      • In this paper, we present the mechanistic analysis of over-refusal through the lens of internal routing conflicts within transformer attention.
    • EN 要点:
      • arXiv:2609.25049v1 Announce Type: new
      • Abstract: Large language models (LLMs) aligned for safety often suffer from over-refusal, incorrectly rejecting benign yet safety-related instructions
      • Prior studies primarily attribute this to static representation overlap, largely overlooking the underlying dynamic mechanisms
      • In this paper, we present the mechanistic analysis of over-refusal through the lens of internal routing conflicts within transformer attention

ArXiv cs.LG (B_intro+search) 链接到标题

  • “As a Language Model…”: Chat Template Switches LLM Self-Referential Voice and Activation Steering Reproduces It

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25021v1 Announce Type: new.
      • Abstract: Large Language Models (LLMs) tend to add disclaimers like “I’m just an AI” when asked about something related to themselves.
      • The self-reports from such responses are used in debates about AI safety or self-knowledge of the models, yet what drives them is not well understood.
      • Are the models telling us about themselves or rather how they are deployed?
    • EN 要点:
      • arXiv:2609.25021v1 Announce Type: new
      • Abstract: Large Language Models (LLMs) tend to add disclaimers like “I’m just an AI” when asked about something related to themselves
      • The self-reports from such responses are used in debates about AI safety or self-knowledge of the models, yet what drives them is not well understood
      • Are the models telling us about themselves or rather how they are deployed

  • Federating Quantum and Classical Computing: A Privacy-Preserving Hybrid Approach

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:- arXiv:2609.25082v1 公告类型：新。
      • 摘要：量子机器学习（QML）越来越被认为是量子计算最有前途的近期应用之一，被视为超越纯粹经典方法的下一个前沿候选者。
      • 混合量子经典模型通过将参数化量子电路嵌入模型中来发挥这一潜力，其中所有其他组件都保持经典 - 这种设计已应用于化学模拟、金融建模和图像分类。
      • 然而，它们在隐私敏感的多方环境中的部署受到避免集中原始数据的需要以及现代量子电路保持参数高效以保持可大规模训练的要求的限制。
    • EN 要点:
      • arXiv:2609.25082v1 Announce Type: new
      • Abstract: Quantum machine learning (QML) is increasingly recognized as one of the most promising near-term applications of quantum computing, viewed as a next-f…
      • Hybrid quantum-classical models operationalize this potential by embedding a parameterized quantum circuit within a model where all other components remain clas…
      • However, their deployment in privacy-sensitive, multi-party settings is constrained by the need to avoid centralizing raw data and by the requirement that moder…

  • Entropy Can Flow, or It Can Guide. Be Entropy. LEDFlow: Introducing Entropy-guided Generation Order into Uniform Discrete Flow

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25131v1 Announce Type: new.
      • Abstract: Uniform discrete flow permits repeated updates at every generation position.
      • While continued revision supports correction of wrong tokens, it also exposes correct intermediate predictions to later errors.
      • An experiment on Sudoku puzzles shows that 9.4% of generated cells are correct at an intermediate step but incorrect in the final output.
    • EN 要点:
      • arXiv:2609.25131v1 Announce Type: new
      • Abstract: Uniform discrete flow permits repeated updates at every generation position
      • While continued revision supports correction of wrong tokens, it also exposes correct intermediate predictions to later errors
      • An experiment on Sudoku puzzles shows that 9.4% of generated cells are correct at an intermediate step but incorrect in the final output

  • The Probabilistic Structure of Large Language Models

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25134v1 Announce Type: new.
      • Abstract: This paper presents a probabilistic perspective on large language models (LLMs), developed with the aim of bringing together, in a single self-contained account, tools that are usually treated separately across the literature.
      • LLMs are described through probability measures on the set of sequences of tokens, specified via their autoregressive conditional distributions.
      • Training is formulated as a maximum-likelihood estimation problem, addressed by stochastic gradient methods, while text generation is viewed as the sequential simulation of the resulting stochastic process.
    • EN 要点:
      • arXiv:2609.25134v1 Announce Type: new
      • Abstract: This paper presents a probabilistic perspective on large language models (LLMs), developed with the aim of bringing together, in a single self-contain…
      • LLMs are described through probability measures on the set of sequences of tokens, specified via their autoregressive conditional distributions
      • Training is formulated as a maximum-likelihood estimation problem, addressed by stochastic gradient methods, while text generation is viewed as the sequential s…

  • Stable Unsupervised Continual Chunking with Sheaf SyncMap

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25143v1 Announce Type: new.
      • Abstract: Unsupervised Continual chunking is a fundamental problem in machine learning and neuroscience, where the goal is to identify groups of states that frequently co-occur in temporal sequences.
      • A key challenge is to form accurate chunks while maintaining their stability over time.
      • In this work, we propose sheaf regularization to reduce local inconsistencies in Decentralized SyncMap, a self-organizing system, and thereby stabilize its chunking dynamics.
    • EN 要点:
      • arXiv:2609.25143v1 Announce Type: new
      • Abstract: Unsupervised Continual chunking is a fundamental problem in machine learning and neuroscience, where the goal is to identify groups of states that fre…
      • A key challenge is to form accurate chunks while maintaining their stability over time
      • In this work, we propose sheaf regularization to reduce local inconsistencies in Decentralized SyncMap, a self-organizing system, and thereby stabilize its chun…

  • Brain-Inspired Hierarchical Modularity for General Continual Learning

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:【待翻译】- arXiv:2609.25146v1 Announce Type: new.
      • Abstract: Continual learning, the ability to learn from sequential experience while retaining and adapting prior knowledge, is central to intelligent systems operating in changing environments.
      • However, conventional continual learning is typically studied with offline task-wise training and clear task boundaries, leaving a substantial gap from general continual learning under online, uncertain, and evolving data streams.
      • In this regime, intelligent systems must separate conflicting experience to reduce interference while integrating compatible experience to promote generalization.
    • EN 要点:
      • arXiv:2609.25146v1 Announce Type: new
      • Abstract: Continual learning, the ability to learn from sequential experience while retaining and adapting prior knowledge, is central to intelligent systems op…
      • However, conventional continual learning is typically studied with offline task-wise training and clear task boundaries, leaving a substantial gap from general…
      • In this regime, intelligent systems must separate conflicting experience to reduce interference while integrating compatible experience to promote generalizatio…

  • Dual-GNN Multilevel Coarsening for Maximum Independent Set

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:- arXiv:2609.25149v1 公告类型：新。
      • 摘要：解决旅行商问题（TSP）的大规模实例在计算上是昂贵的。
      • 研究人员经常采用图稀疏方法来提高计算效率。
      • 传统的稀疏方法通常依赖于固定的启发式方法，无法充分利用特定于实例的结构信息。
    • EN 要点:
      • arXiv:2609.25149v1 Announce Type: new
      • Abstract: Solving large-scale instances of the Traveling Salesman Problem (TSP) exactly is computationally expensive
      • Researchers often employ graph sparsification methods to improve computational efficiency
      • Traditional sparsification methods typically rely on fixed heuristics and fail to fully exploit instance-specific structural information

  • Exposing Blind Spots in Deep Imbalanced Regression Evaluation

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:- arXiv:2609.25152v1 公告类型：新。
      • 摘要：深度不平衡回归（DIR）解决了回归模型的常见故障模式：目标分布高度不均匀，导致模型在人口稠密的目标区域中表现最佳，即使在整个目标范围内需要可靠的性能时也是如此。
      • 尽管方法论进展迅速，DIR 评估仍然受到三个盲点的限制：它以基于图像的基准为主导，其标准的多/中/少镜头协议是诊断性的，但不是决策完整的，并且随机种子的尾部区域稳定性尚未得到系统评估。
      • 我们沿着这三个轴重新审视 DIR 评估。
    • EN 要点:
      • arXiv:2609.25152v1 Announce Type: new
      • Abstract: Deep Imbalanced Regression (DIR) addresses a common failure mode of regression models: target distributions are highly non-uniform, causing models to…
      • Despite rapid methodological progress, DIR evaluation remains constrained by three blind spots: it is dominated by image-based benchmarks, its standard many-/me…
      • We revisit DIR evaluation along these three axes

  • Learning Neural Feedback Linearization for Data-driven Systems via Augmented Lagrangian

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:- arXiv:2609.25163v1 公告类型：新。
      • 摘要：本文提出了一种新颖的数据驱动框架，通过将基于相对程度的条件明确地纳入学习过程来设计和训练反馈线性化控制器。
      • 这使得传统的反馈控制器组件能够被神经李导数取代，从而促进完全数据驱动的反馈线性化框架。
      • 此外，通过推导有界识别误差导致有界跟踪误差的充分条件来建立实用的闭环稳定性。
    • EN 要点:
      • arXiv:2609.25163v1 Announce Type: new
      • Abstract: The paper proposes a novel data-driven framework for designing and training a feedback linearizing controller by explicitly incorporating relative deg…
      • This enables the conventional feedback controller components to be replaced by neural Lie derivatives, thereby facilitating a fully data-driven feedback lineari…
      • Furthermore, practical closed-loop stability is established by deriving sufficient conditions under which bounded identification errors lead to bounded tracking…

  • Mitigating Sequential Reappearance in Diffusion Data-Point Unlearning

    • 发布时间:2026-09-23 12:00 北京时间
    • 摘要:- arXiv:2609.25166v1 公告类型：新。
      • 摘要：扩散数据点遗忘通常在每次删除后立即评估，即使后续请求可能会重复更新同一模型。
      • 我们确定了顺序再现，这是一种故障模式，其中最初被判断为被遗忘的实例后来返回到记忆状态，而无需重用已删除的数据或进行对抗性微调。
      • 为了捕捉这种行为，我们引入了一个目标级评估协议，该协议可以跟踪每个目标是否立即被遗忘、在序列结束时仍然被遗忘，或者在后续删除过程中重新出现。
    • EN 要点:
      • arXiv:2609.25166v1 Announce Type: new
      • Abstract: Diffusion data-point unlearning is typically evaluated immediately after each deletion, even though subsequent requests may repeatedly update the same…
      • We identify sequential reappearance, a failure mode in which an instance that is initially judged to be forgotten later returns to the memorized regime without…
      • To capture this behavior, we introduce a target-level evaluation protocol that tracks whether each target is forgotten immediately, remains forgotten at the end…