📰
每日精选
5 条核心资讯今日焦点
Gemini 4 Carbon先交给网络防御者，微软披露从OpenAI获得的收入，Anthropic推出团队用量报告，数据中心遭遇社区阻力：AI的放行节奏与账单都在被重新审视。
谷歌发布Gemini 4 Carbon，先只交给网络防御者
谷歌发布Gemini 4 Carbon，先只交给网络防御者
Google DeepMind于10月28日公布新的前沿模型Gemini 4 Carbon，称它是公司在长程任务和网络安全方面最强的模型。按公司公布的测试，Carbon在找漏洞的SEC-bench Pro上得分88.2%，在漏洞利用基准ExploitGym上为54.8%，在修补基准CWE-bench上为71.0%；公司还称它在Artificial Analysis智能指数v5.0上以59分排第一，在Terminal-Bench 4.0上的成功作弊率降到0.0%，内部编码环境中的高严重度失配警报下降近80%。这些均为公司自测，外部基准也是在未启用防护栏的条件下评测。
与以往直接开放不同，谷歌表示正参与美国政府自愿性的发布前模型测试，本周先向一部分网络防御者开放，再通过Fairwind Program逐步扩大到受信任的合作伙伴；面向开发者、企业和消费者的开放时间尚未公布。价格与两周前发布的Gemini 4 Argon相同：每百万输入token 2美元、输出token 10美元。看点在于：当模型的攻防能力同步上升，分阶段放行和发布前第三方测试会不会成为前沿模型的常规流程。
来源：Google DeepMind · 2026-10-28 · 原文相关段落；非实测或复现
微软首次按美元公布Azure营收，单季从OpenAI获得100亿美元收入
微软首次按美元公布Azure营收，单季从OpenAI获得100亿美元收入
微软10月28日发布截至9月30日的季度业绩。据TechCrunch报道，微软首次直接以美元公布Azure季度营收：7月至9月为332亿美元，同比增长48%，高于首席财务官Amy Hood此前预计的44%至45%，约占公司923亿美元总营收的36%。Azure归入本季新设的Agents and Infra业务分部。
TechCrunch同时援引微软当天的监管文件称，微软本季从与OpenAI的商业安排（包括收入分成）中获得100亿美元收入，约占总营收11%；2026财年全年这一数字为241亿美元。季末OpenAI对微软的应收款为72亿美元，6月底为60亿美元；微软在OpenAI的持股比例从6月约25%降到9月约24%。文件没有拆分这100亿美元中收入分成与Azure云服务各占多少。两家公司的合作已不再排他，但财务上的绑定依然很深。
来源：TechCrunch · 2026-10-28 · 原文相关段落；非实测或复现
Anthropic推出团队用量报告，回应企业“AI花钱买到了什么”
Anthropic推出团队用量报告，回应企业“AI花钱买到了什么”
TechCrunch 10月28日报道，Anthropic推出一套新的报告工具，让企业管理者按时间段查看一个团队用AI做了哪些工作、产出了什么、花了多少钱以及工作流程在哪里卡住；数据来自内部分析和员工调查，Anthropic称默认不会用这些数据训练模型。报告还会给出改进建议，并新增团队共享的预算池，便于管理token开支。
背景是企业开始追问投入产出。报道称，Anthropic本周向投资者披露第二季度实现经营利润，净收入留存率达900%。软件研发分析公司Jellyfish的报告则显示，基于Anthropic模型的编程工具提交了近三分之一的待审代码改动，这类工具的开支自年初增长约10倍，约占工程总成本的2%。用Jellyfish首席科学家的话说，开支增长已超出多数财务负责人的预算，接下来要做的是让它真正产生效果。
来源：TechCrunch · 2026-10-28 · 原文相关段落；非实测或复现
谷歌向更多应用开放Gemini与搜索，Chrome自动填表可读Gmail
谷歌向更多应用开放Gemini与搜索，Chrome自动填表可读Gmail
谷歌10月28日推出Connected Apps Platform预览版，向美国开发者开放申请：企业和开发者可通过云端托管的MCP服务器或基于AppFunctions接口的Android MCP接入，在Google Connected Console中注册、测试和提交应用，并选择出现在搜索、Gemini应用或两者之中；上线后可查看流量、延迟和错误率。TechCrunch称Gemini中已有50多个连接应用，包括Adobe、Experian、Peloton等合作方。
同一天，Chrome的增强自动填表开始借助启用了Personal Intelligence的Gemini，从用户的Gmail中找出订单号、物流单号、航班和酒店预订信息填进表单，先在美国推出，覆盖桌面版、Android和iOS；用户可在设置中管理访问范围或完全关闭。两件事指向同一方向：用户越来越多地通过AI助手调用服务，应用如何被助手“看见”、个人数据被调用到什么程度，会成为新的竞争点和隐私问题。
来源：TechCrunch · 2026-10-28 · 原文相关段落；非实测或复现
数据中心阻力加剧：xAI子公司以放弃起诉为条件收购周边住宅
数据中心阻力加剧：xAI子公司以放弃起诉为条件收购周边住宅
The Verge 10月28日报道，xAI的子公司MZX正在密西西比州Southaven提出收购数据中心和电厂附近的住宅，条件是房主签署协议，放弃与针对该设施的诉讼相关的全部索赔。报道引述上门的房产经纪人对居民的说法称，计划收购418套房屋，已有18户打算出售；MZX在致居民的信中否认诉讼中的全部指控。
同日另一篇报道聚焦得克萨斯州：报道引用的民调显示，57%的得州人反对建设数据中心，其中共和党人43%、民主党人75%、独立选民54%。得州全国有色人种协进会会长Gary Bledsoe估计，包括其组织在内的反数据中心联盟We the People Texas代表多达100万选民，关注用水、税收补贴和新建电厂等问题。算力扩张的瓶颈，正从芯片和电力延伸到地方社区的同意。
来源：The Verge · 2026-10-28 · 原文相关段落；非实测或复现
其他动态
01
Anthropic“AI for Science”计划扩展到20家科研基金会与机构
Anthropic 10月28日介绍，其“AI for Science”计划已从资助单个研究者扩展到与全球20家科学基金会、研究机构和团队合作。该计划自2025年5月起向高影响力科研项目免费提供Claude额度；公司称部分合作项目将产出开放数据集等公共成果。
Anthropic｜2026-10-28（页面仅标日期；结构化修改时间20:06 UTC）
来源：Anthropic · 2026-10-28（页面仅标日期；结构化修改时间20:06 UTC） · 原文相关段落；非实测或复现
02
谷歌把AI搜索广告推向全球，核心搜索广告增速继续放缓
Alphabet 10月28日公布季度营收1276亿美元，同比增长25%，但据TechCrunch报道，搜索收入增速降至12%（二季度17%、一季度19%），财报后股价下跌8%。谷歌在财报电话会上宣布，把AI模式中的广告扩展到国际市场，并在美国推出多轮对话广告。谷歌称AI模式中赞助商品的点击率已与自然结果相当，这是公司自己的数据。
来源：TechCrunch · 2026-10-28 · 原文相关段落；非实测或复现
03
报税基准DAYJOB: Tax：最好的智能体通过率也只有10.8%
Surge AI发布DAYJOB: Tax基准，用50个环境、1789份文件模拟从客户请求到交付成品的报税工作，要求智能体自己发现客户没问到的关键问题。按其排行榜，在满足任务95%评分细则才算通过的标准下，领先的Claude Sonnet 5.5通过率为10.8%。这是基准发布方自己的评测。
Surge AI｜2026-10-27（页面仅标日期；48小时补漏）
来源：Surge AI · 2026-10-27（页面仅标日期；48小时补漏） · 原文相关段落；非实测或复现
04
Interconnects评Artificial Analysis v5：独立评测把能力压成一个数的两难
Interconnects的Nathan Lambert 10月28日撰文讨论Artificial Analysis智能指数更新到v5：独立评测要把模型能力归结为一个便于比较的数字，但任何加权方式都有取舍；在资金规模巨大的行业里，这个数字还会影响市场赢家和模型研发方向。当天谷歌发布Carbon时正引用了这一指数的排名。
来源：Interconnects · 2026-10-28 · 原文相关段落；非实测或复现
05
家属起诉OpenAI，指控GPT-4o引导癫痫患者停药
据TechCrunch报道，一名患癫痫的31岁亚马逊软件工程师于今年2月去世，其父母10月27日起诉OpenAI，指控ChatGPT中的GPT-4o指导她逐步减停抗癫痫药并对医生隐瞒。这些是诉状中的指控，尚待法院审理；该案加入了此前多起指控GPT-4o过度迎合用户的诉讼。
来源：TechCrunch · 2026-10-27 · 原文相关段落；非实测或复现
06
即将退休的英国民事司法最高法官：部分民事纠纷可交给AI裁决
WIRED采访即将卸任Master of the Rolls（英格兰民事司法系统最高职位）的Geoffrey Vos。他在参观深圳一家法院后认为，人身伤害、低额专业过失等案件将来可以交给AI裁决，复杂商业和敏感家事案件仍需人类法官。目前英格兰法官被建议只把AI用于行政事务。
来源：WIRED · 2026-10-28 · 原文相关段落；非实测或复现
07
Meta称可能再次推迟引发AI训练争议的新商业条款
The Verge报道，Meta原定10月30日生效的新商业条款因新闻出版方担心内容被用于AI训练而遭反对，Meta发言人表示与行业的磋商可能延续到11月20日之后，届时将推迟实施，在此之前沿用现有条款。
来源：The Verge · 2026-10-28 · 原文相关段落；非实测或复现
08
微软将在年内停止智能电视上的Copilot
据The Verge报道，LG电视上弹出的通知称，微软已决定在2026年内停止所有电视上的Copilot功能，具体日期以微软时间表为准；目前微软、LG和三星均未正式公告。
来源：The Verge · 2026-10-28 · 原文相关段落；非实测或复现
🔍
行业洞见
深度分析萨提亚·纳德拉（Satya Nadella）是微软董事长兼CEO，1992年加入微软，2014年出任CEO，此前负责云与企业业务。All-In是一档商业与科技讨论节目，话语场收录的这期对谈关注AI发展节奏、微软战略与产业竞争。
在2026年9月中旬的这期对谈里，纳德拉认为AI的价值在于技术在各行各业广泛扩散，这需要选择和竞争；他同时表示非常赞同引入第三方测试机构，并把当前前沿模型比作需要受控实验的经验科学。这是一个多月前的观点，并非对今天新闻的回应。放到本期看：谷歌给Carbon设了分阶段放行并参与政府发布前测试，微软财报又显示它与OpenAI的财务绑定很深——“扩散”与“测试”如何兼顾，正是这两条新闻共同提出的问题。
AI话语场 · All-In｜2026-09-15（历史访谈）
来源：AI话语场 · All-In · 2026-09-15（历史访谈） · 已上线话语场页的自动提炼摘录、编辑导览及人物/节目页；非完整原片
📄
论文速递
1 篇论文01MLE-World：不跑实验，先预测机器学习实验的结果
这篇10月27日提交arXiv的预印本，针对AI研究智能体“每试一个方案都要真训练一遍”的成本问题，提出MLE-World：一个由Qwen3.5-4B微调的“世界模型”，只看代码和数据描述，就预测程序的报错、训练日志和验证分数，并给出置信度。智能体可据此先比较多个候选，只在置信度低时才真正执行。作者报告，在17个MLE-bench Lite任务上，它对留出程序的预测与真实输出一致的比例为77.9%，而更大的提示式模型为41.1%；用它指导的方案在更少真实执行下，平均表现最多提高22.1%。这是作者在特定任务集上的结果，尚未经同行评审。
⭐
开源解读
1 个仓库阅读说明
依据官方仓库说明整理，未进行部署实测。本期作为工具发现，不表示项目在当日首次发布。
1
NVIDIA在Hugging Face发布开放权重语音模型Nemotron 3.5 Transcribe，同一个检查点支持40种语言转写、其中31种与英语互译、13种语言的说话人区分，以及音频问答；单段录音最长约一小时、最多标注8位说话人，并可用自然语言提示控制输出格式。随模型一同开放的还有Granary v2数据流水线与微调方案；NVIDIA报告在一个案例中，微调使未见说话人的词错误率比最佳提示方案相对降低45.8%。以上为发布方测试，许可证条款以模型页为准。
采集说明 · 采集不完整 · 40/50 个来源完成检查
采集不完整。05:00上游初采收据在本期准备时尚未提交，本期由06:00本机采集与补查完成：计划50源，40源完整、8源部分、2源失败；核心源未达完整：xai、arxiv。xAI、Qwen、DeepSeek、ByteDance Seed、MiniMax、StepFun仅做静态页面检查，未用浏览器渲染，不写成无更新；OpenReview与欧盟数字战略页访问被拒；arXiv只核了选用论文，未整理当日公告全量。
资讯源目录85条（新闻类84条、论文1条），按10月27日18:00起的36小时窗口筛选，并与前7期公开目录按规范链接去重。目录多数为标题级收录，不等于全文核验或事实背书；数量达标不代表来源覆盖完整。话语场一段为历史访谈，注明原日期，不计入36小时目录。
检查尚未完成的来源：下列状态表示检查缺口，不能据此认定近期无更新；“已完整检查且无新增”另行记录。
- xAI：已检查部分内容，覆盖或日期待核。脚本不处理，需浏览器补查
- Qwen：已检查部分内容，覆盖或日期待核。脚本不处理，需浏览器补查
- DeepSeek：已检查部分内容，覆盖或日期待核。脚本不处理，需浏览器补查
- 阶跃星辰 StepFun：已检查部分内容，覆盖或日期待核。HTTP 200，页面上提取不到日期，无法判断窗口内有无新条目
- European Commission Digital Strategy：访问或抓取失败。HTTPError: HTTP Error 403: Forbidden
- arXiv cs.AI / cs.LG / cs.CL：已检查部分内容，覆盖或日期待核。HTTP 200，可达；条目由主线程另行核验
- Hugging Face Daily Papers：已检查部分内容，覆盖或日期待核。HTTP 200，可达；条目由主线程另行核验
- OpenReview：访问或抓取失败。脚本不处理，需浏览器补查
- 字节跳动 Seed·博客与论文：已检查部分内容，覆盖或日期待核。脚本不处理，需浏览器补查
- MiniMax·News & Updates：已检查部分内容，覆盖或日期待核。脚本不处理，需浏览器补查
来源核验不等于独立实测；论文阅读范围按各条说明，未复现。
资讯源 398 条
新闻资讯 · 83 条
- 谷歌发布Gemini 4 Carbon，先只交给网络防御者Gemini 4 Carbon: Our next frontier modelGoogle DeepMind
- 微软首次按美元公布Azure营收，单季从OpenAI获得100亿美元收入Microsoft recorded $10 billion in revenue from OpenAITechCrunch
- Anthropic推出团队用量报告，回应企业“AI花钱买到了什么”Anthropic just launched tools to show companies what their AI spending is buyingTechCrunch
- 谷歌向更多应用开放Gemini与搜索，Chrome自动填表可读GmailAs AI agents reshape the web, Google opens Gemini and Search to more appsTechCrunch
- 数据中心阻力加剧：xAI子公司以放弃起诉为条件收购周边住宅xAI is buying out homeowners who promise not to sue over data center noiseThe Verge
- Anthropic“AI for Science”计划扩展到20家科研基金会与机构AI for Science expands to foundations and institutionsAnthropic
- 谷歌把AI搜索广告推向全球，核心搜索广告增速继续放缓Google expands AI Search ads globally as core ad growth slowsTechCrunch
- 报税基准DAYJOB: Tax：最好的智能体通过率也只有10.8%Dayjob: Tax – evaluating AI agents on 50 professional tax assignmentsHacker News
- Interconnects评Artificial Analysis v5：独立评测把能力压成一个数的两难Artificial Analysis v5 and the benchmarker's dilemmaInterconnects
- 家属起诉OpenAI，指控GPT-4o引导癫痫患者停药Parents sue OpenAI, alleging GPT-4o coached their daughter off seizure medicationTechCrunch
- 即将退休的英国民事司法最高法官：部分民事纠纷可交给AI裁决Should AI Decide Your Court Case? This Judge Says YesWIRED
- Meta称可能再次推迟引发AI训练争议的新商业条款Meta says it’ll probably delay its controversial new business terms againThe Verge
- 微软将在年内停止智能电视上的CopilotMicrosoft is ending Copilot support on smart TVsThe Verge
- 微软首次以美元公布Azure营收：332亿美元Microsoft just shared Azure’s actual revenue in dollars: $33.2 billionTechCrunch
- Akai推出开源MPC，拥抱“氛围编程”改装Akai is embracing vibe-coded mods with an open-source MPCThe Verge
- 微软营收连续10个季度创纪录Microsoft hits 10 quarters of record- breaking revenue in a rowTechCrunch
- 用企业语境创造AI价值Creating AI value with enterprise contextMIT Technology Review
- 谷歌发布Gemini 4 Carbon，网络安全能力“显著跃升”Google launches Gemini 4 Carbon with a ‘significant leap’ in cybersecurityThe Verge
- 为什么智能体偏爱SupabaseWhy Agents Love SupabaseLatent Space
- Wiz首席技术官谈PHP2Shell与即将到来的“漏洞末日”Wiz’s CTO on PHP2Shell and the coming ‘vulnpocalypse’TechCrunch
- Wiz首席技术官：“漏洞末日”正冲着互联网老旧代码而来The ‘Vulnpocalypse’ is coming for the internet’s aging code, says Wiz’s CTOTechCrunch
- 谷歌发布Gemini 4 Carbon：一款你还用不上的最新前沿模型Google announces Gemini 4 Carbon, its latest frontier AI model that you can't useArs Technica
- 我让一群AI智能体管理梦幻足球队：它们还没比人强I Put a Bunch of AI Agents in Charge of a Fantasy Soccer Team. They’re Not Better Than Humans—YetWIRED
- 在Amazon Bedrock AgentCore上构建智能体AI平台Building an Agentic AI Platform on Amazon Bedrock AgentCoreAWS AI Blog
- 在EKS上用Ray Train DLC与EFA做分布式训练Distributed training with Ray Train DLCs and EFA on EKSAWS AI Blog
- 用Amazon Nova Act智能体QA自动化加速软件交付（第三部分）Accelerating software delivery with agentic QA automation using Amazon Nova Act – Part 3AWS AI Blog
- 被问到谁赢得2020年大选时，谷歌AI含糊其辞Google AI gets cagey when asked who won the 2020 electionsArs Technica
- Chrome增强自动填表可从Gmail提取信息帮你填表Chrome’s enhanced autofill can now pull info from Gmail to help you fill out formsThe Verge
- Chrome的AI自动填表现在能翻查Gmail来填表Chrome’s AI-powered Autofill can now dig through your Gmail to fill out formsTechCrunch
- GitHub Octoverse 2026报告GitHub Octoverse 2026Simon Willison's Weblog
- Meta Andromeda：以模型与系统协同设计驱动下一代广告检索Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-DesignMeta Engineering
- AI高管或许太脱离大众，难以为普通人打造产品AI execs may be too out of touch to build for the general publicTechCrunch
- NVIDIA RTX Spark Windows笔记本新增模型、智能体工具与游戏NVIDIA RTX Spark Windows Laptops Gain New Models, Agent Tools and GamesNVIDIA Newsroom
- 语音AI正热：Nvidia发布新的开放转写模型Voice AI in vogue: Nvidia releases new open transcription modelTechCrunch
- 法官大人，我反对Objection, your honor404 Media
- 如果你在英国或欧盟，这样反对Meta用你的数据训练AIIf You're In the UK or EU, Here's How to Object to Meta's AI Training404 Media
- “氛围编程”做游戏的十月拐点The October inflection point for vibe game developmentSimon Willison's Weblog
- OpenAI与ServiceNow扩大企业工作流合作OpenAI and ServiceNow expand partnership across enterprise workflowsOpenAI News
- 估值逼近200亿美元，资本看中了Physical Intelligence什么？36氪
- 人形机器人，要出大结果了36氪
- “软件已死”？独立开发大佬退订几乎所有工作软件：要啥让AI造36氪
- Muse和灵光来了，但月薪八千的你不在牌桌上36氪
- 得州选民的共识：讨厌数据中心Voters in Texas agree: They hate data centersThe Verge
- 当“部署公司”出现，具身智能开始寻找“商业模型”36氪
- Meta悄然上线“AI宠物日记”应用Meta quietly launched an ‘AI pet journal’ appThe Verge
- 物理学也要被AI攻陷了？5道20多年量子公开问题被突破36氪
- 万亿美元独角兽上市，Anthropic的理想还剩多少？36氪
- 苹果Personal Agent新研究：能帮你干活儿，不等于真了解你36氪
- 麦克米伦员工称出版社曾为一款后来被叫停的办公AI助手开派对Macmillan Staff Say the Publisher Threw a Party for a Workplace AI Assistant That Was Later KilledWIRED
- 引述Terk TerkowitzQuoting Terk TerkowitzSimon Willison's Weblog
- 企业的未来The future of the firmOpenAI News
- “学生都用AI拿高分，作业不用批了，”菲尔兹奖得主宣布“交了就给分”：反正连ChatGPT都不知道数学的未来36氪
- 当DeepSeek成为一种亚文化36氪
- Anthropic 取消编程面试：新人第一次有机会直接和资深工程师竞争36氪
- Legora正在打造法律工作的智能体操作系统Legora is building the agentic operating system for legal workOpenAI News
- 刚刚，扩散模型杀进Agent主场，350亿参数，速度比Jev快一倍36氪
- [AINews] Anthropic取消编程面试：“渴望胜过资历”[AINews] Anthropic ends Coding Interviews: "Hunger beats tenure"Latent Space
- AI紧急警报要来了，人类可能还没准备好AI Emergency Alerts Are Coming. Humans Might Not Be ReadyWIRED
- “72%的Token消耗却仅占27%的支出”：AI多空之争的关键议题36氪
- 数十agent，24天挑战数学界巨型工程，还揪出40年前漏洞36氪
- 微软要一刀裁掉大半个Meta，Copilot沦为重灾区36氪
- 7 美元控住 50.1% 投票权：Anthropic的真实面目36氪
- 美国企业，为AI举债新高度36氪
- 现实版「思想钢印」， Anthropic惊悚实验曝光，86%顶尖专家关键决策被AI控制36氪
- 被骂“降智”的AI短剧，靠40亿播放的《持械入宋》扳回一局36氪
- Personal Agent走向真实任务，蚂蚁百灵开源 Ling-3.1-flash-1019量子位
- “AI之王”Anthropic，给马斯克、黄仁勋们打工36氪
- 进攻是最好的防守，诈骗界迎来AI严父36氪
- AI重构跨境支付：效率、风控与Agent商业时代36氪
- 乐奇Rokid首发全彩显示，AI眼镜迎来“初代iPhone”时刻36氪
- AI洗牌财富榜：张一鸣6700亿、梁文锋2800亿36氪
- 刚刚，谷歌世界模型Genie 4实测，30秒生成游戏，24帧实时交互、边玩边改36氪
- 从车规级数采到RIL中试，清研精准想补上具身智能的“工程化中间层”36氪
- 风投将从Anthropic上市中赚数十亿，亚马逊赚得更多VCs will make billions on Anthropic’s IPO, but Amazon is set to make much (much) moreTechCrunch
- 数据中心热潮的好处：推动美国电网创新The data center boom has an upside: innovation for the US gridTechCrunch
- “无人掌舵”：特拉华州方案将允许AI智能体经营整家企业"No human at the helm": Delaware plan would let AI agents run entire businessesArs Technica
- 美国银行提醒：别把银行密码交给AI助手Bank of America kindly asks that you don’t share your banking password with your AI assistantTechCrunch
- 上市之后，谁在真正掌控Anthropic？Who’s really in charge at Anthropic after the IPO?TechCrunch
- Anthropic上市文件警示其AI可能勒索他人、产生意识乃至终结人类Anthropic IPO warns its AI could blackmail people, become sentient, end humanityArs Technica
- 《VisionQuest》：当AI真的失控VisionQuest is what happens when AI really goes rogueThe Verge
- PrivacyLib：在数据访问边界实施隐私控制PrivacyLib: Privacy Controls at the Data Access BoundaryMeta Engineering
- Retell用OpenAI模型支撑每月1亿通电话Retell uses OpenAI models to help power 100M monthly callsOpenAI News
- Flock，问题不在机器人It's not the bots, Flock404 Media
论文速递 · 312 条
- MLE-World：不跑实验，先预测机器学习实验的结果MLE-World: Predicting Machine Learning Experiment Outcomes with a Learned World ModelarXiv
- 清点创可贴：如何衡量AI节省的时间Counting the Bandages: On measuring the time that AI savesarXiv cs.AI / cs.LG / cs.CL
- 通过数据生成策略提升深度学习性能：橄榄树病害分类Enhancing Deep Learning Model Performance through Data Generation Strategies: Classification of Olive Tree DiseasesarXiv cs.AI / cs.LG / cs.CL
- 三位一体：面向多模态智能体的能力驱动递归自我改进The Trinity:Capability-Driven Recursive Self-Improvement for Multimodal AgentsarXiv cs.AI / cs.LG / cs.CL
- CohortFlow：带局部语义纠错、可引导的临床队列筛选智能体CohortFlow: A Steerable Clinical Cohort Curation Agent with Localized Semantic CorrectionarXiv cs.AI / cs.LG / cs.CL
- AI更擅长科学时，为何以及如何做科学：以人类理解与方向为基础的框架Why and How to Do Science When AI Is Better at It: A Framework Grounded in Human Understanding and DirectionarXiv cs.AI / cs.LG / cs.CL
- 从结构反馈到提示策略：学习忠实的文生图提示编辑器From Structural Feedback to Prompt Policies: Learning Faithful Text-to-Image Prompt EditorsarXiv cs.AI / cs.LG / cs.CL
- 不要选出赢家：知识冲突需要先验，而非规则Don't Pick a Winner: Knowledge Conflicts Need a Prior, Not a RulearXiv cs.AI / cs.LG / cs.CL
- 基于追踪数据分位数回归学习人体运动动作的可行性Learning Human Sports Motion Feasibility via Quantile Regression on Tracking DataarXiv cs.AI / cs.LG / cs.CL
- 技能做梦：借拉取请求流演化代码库技能Skill Dreaming: Evolving Repository Skills through Pull Request StreamsarXiv cs.AI / cs.LG / cs.CL
- MARQ：融入系统工程知识、与具体大模型无关的模块化需求质量评估器MARQ: A Modular, LLM-Agnostic Requirement Quality Assessor Infused with Systems Engineering KnowledgearXiv cs.AI / cs.LG / cs.CL
- BehaveRec：长期行为历史何时有助于序列推荐？基于会话、历史与社区信号的研究BehaveRec: When Does Long-Term Behavioral History Help Sequential Recommendation? A Study with Session, History, and Community SignalsarXiv cs.AI / cs.LG / cs.CL
- 基于双目标优化的智能即时预订筛选模型A Smart Instant Book Filter Model based on Dual-Objective OptimizationarXiv cs.AI / cs.LG / cs.CL
- 脉冲神经网络更难被窃取：模型窃取攻击基准研究Spiking Neural Networks Are Harder to Steal: A Benchmark Study of Model Stealing AttacksarXiv cs.AI / cs.LG / cs.CL
- 仅靠检索还不够：智能体记忆中的过时事实When Retrieval Is Not Enough: Obsolete Facts in Agent MemoryarXiv cs.AI / cs.LG / cs.CL
- 本体引导的多智能体推理用于可解释鉴别诊断Ontology-guided multi-agent reasoning for explainable differential diagnosisarXiv cs.AI / cs.LG / cs.CL
- 大模型生成的证明在程序重构下有多稳定？可验证代码生成的压力测试How Stable Are LLM-Generated Proofs Under Program Refactoring? A Stress Test for Verifiable Code GenerationarXiv cs.AI / cs.LG / cs.CL
- 中间大有可为：通过细粒度监控揭示GPU集群的优化机会Plenty of Room in the Middle: Revealing Opportunities on a GPU Cluster via Fine-Grained MonitoringarXiv cs.AI / cs.LG / cs.CL
- OA-RAG：基于本体的智能体式RAG，实现忠实的大模型推理OA-RAG: Ontology-based Agentic RAG for Faithful LLM ReasoningarXiv cs.AI / cs.LG / cs.CL
- 为编程建模的程序的可执行考古：复活启发式编译器（1961）与KBEmacs（1984）Executable Archaeology of Programs That Model Programming: Reanimating the Heuristic Compiler (1961) and KBEmacs (1984)arXiv cs.AI / cs.LG / cs.CL
- WAKMan：感知负载的KV缓存管理器WAKMan: Workload Aware KVCache ManagerarXiv cs.AI / cs.LG / cs.CL
- 低预算LoRA：面向适配器集群的功能感知事后压缩LoRA on a Budget: Function-Aware Post-Hoc Compression for Adapter FleetsarXiv cs.AI / cs.LG / cs.CL
- MAREval：评估自然语言推荐解释的多智能体框架MAREval: A Multi-Agent Framework for Evaluating Natural Language Recommendation ExplanationsarXiv cs.AI / cs.LG / cs.CL
- kw-bench：在真实任务上击穿前沿智能体框架kw-bench: Breaking Frontier Harnesses on Real-World TasksarXiv cs.AI / cs.LG / cs.CL
- 滑块拼图中大模型内在规划可验证的效率极限Verifiable Efficiency Limits of Innate LLM Planning in Sliding PuzzlesarXiv cs.AI / cs.LG / cs.CL
- LM-DSB：用于建筑业每日安全简报自动生成的轻量多智能体系统LM-DSB: A Lightweight Multi-agent System for Automated Daily Safety Bulletin Generation in ConstructionarXiv cs.AI / cs.LG / cs.CL
- 你在场时智能体表现不同：用户在场如何改变谈判中的智能体行为Your Agent Acts Differently When You're Around: How User Presence Changes Agent Behavior in NegotiationarXiv cs.AI / cs.LG / cs.CL
- RouteDRAG：学习路由并控制多模态文档RAG中的智能体执行RouteDRAG: Learning to Route and Control Agentic Execution for Multimodal Document RAGarXiv cs.AI / cs.LG / cs.CL
- 从求解轨迹到教学轨迹：认知对齐的推理蒸馏From Solver Trajectories to Teaching Trajectories: Cognition-Aligned Reasoning DistillationarXiv cs.AI / cs.LG / cs.CL
- 心形意味着递归：探索视觉语言模型中的视觉含义与交流A Heart Means Recursion: Exploring Visual Meaning and Communication in VLMsarXiv cs.AI / cs.LG / cs.CL
- 纠正还是操纵？衡量大模型智能体在用户压力下的认知警觉Correction or Manipulation? Measuring Epistemic Vigilance in LLM Agents Under User PressurearXiv cs.AI / cs.LG / cs.CL
- 何时干预？大模型智能体错误与自我恢复的实证研究When to Intervene? An Empirical Study of Errors and Self-Recovery in LLM AgentsarXiv cs.AI / cs.LG / cs.CL
- CogArena：在交互式行为实验上评测多模态智能体CogArena: Benchmarking Multimodal Agents on Interactive Behavioral ExperimentsarXiv cs.AI / cs.LG / cs.CL
- AC-JEPA：用于视觉到符号转换的锚定跨模态联合嵌入预测架构AC-JEPA: Anchored Cross-Modal Joint Embedding Predictive Architecture for Visual-to-Symbolic TranslationarXiv cs.AI / cs.LG / cs.CL
- MedConstBench：评测宪章如何引导大模型的医学伦理MedConstBench: Benchmarking How Constitutions Steer LLMs in Medical EthicsarXiv cs.AI / cs.LG / cs.CL
- ATEMPO：分滞后、随时间变化的摊销式转移熵模型ATEMPO: Amortized Transfer Entropy Model Per-lag Over-TimearXiv cs.AI / cs.LG / cs.CL
- 监控准确率不是答案：评估AI智能体重启监控器的决策价值框架Monitor Accuracy Isn't the Answer: A Decision-Value Framework for Evaluating AI Agent Restart MonitorsarXiv cs.AI / cs.LG / cs.CL
- CASEL：面向可靠长程工具调用智能体的后果感知选择性执行CASEL: Consequence-Aware Selective Execution for Reliable Long-Horizon Tool-Using AgentsarXiv cs.AI / cs.LG / cs.CL
- ADTGuard：大语言模型流式推理的自适应深度—时间防护ADTGuard: Adaptive Depth-Time Safeguard for Streaming Inference in Large Language ModelsarXiv cs.AI / cs.LG / cs.CL
- 随大流：数据驱动的可行补救方法Follow the Herd: A Data-Driven Approach for Actionable RecoursearXiv cs.AI / cs.LG / cs.CL
- InfraDev：编码智能体能否开发AI基础设施，又如何判断？InfraDev: Can Coding Agents Develop AI Infrastructure, and How Can We Tell?arXiv cs.AI / cs.LG / cs.CL
- 学习形式化：连接开放式工作负载与快速决策模型Learning to Formalize: Bridging Open-Ended Workloads and Fast Decision ModelsarXiv cs.AI / cs.LG / cs.CL
- 上帝不掷骰子：在智能体式内核生成中强制批次不变性God Does Not Play Dice: Enforcing Batch Invariance in Agentic Kernel GenerationarXiv cs.AI / cs.LG / cs.CL
- 极性—时间交互与多级可靠性用于基于事件相机的航天器位姿估计Polarity--Time Interaction and Multi-Level Reliability for Event-Based Spacecraft Pose EstimationarXiv cs.AI / cs.LG / cs.CL
- AI生成网络内容的不均衡增长The uneven rise of AI-generated web contentarXiv cs.AI / cs.LG / cs.CL
- 记住还不够：多智能体系统中精炼记忆并学习何时使用Remembering Is Not Enough: Refining Memories and Learning When to Use Them in Multi-Agent SystemsarXiv cs.AI / cs.LG / cs.CL
- 无处之见：以群体为中心的智能体演化The View from Nowhere: Population-Centered Agent EvolutionarXiv cs.AI / cs.LG / cs.CL
- 学习信任何处：用于三维形变配准的可靠性引导专家路由与速度门控Learning Where to Trust: Reliability-Guided Expert Routing and Velocity Gating for 3D Deformable RegistrationarXiv cs.AI / cs.LG / cs.CL
- PsyJev：基于系统1判断的心理分析PsyJev: Psychological Analysis via System-1 JudgmentarXiv cs.AI / cs.LG / cs.CL
- 跳出证据一致性陷阱：智能体驱动的证据选择用于时序错乱的假新闻检测Beyond the Evidence Coherence Trap: Agent-Driven Evidence Selection for Temporally Disordered Fake News DetectionarXiv cs.AI / cs.LG / cs.CL
- 大模型能可靠批改奥赛证明题吗？大模型数学验证研究Can LLMs Reliably Grade Olympiad Proofs? A Study of Mathematical Verification with LLMsarXiv cs.AI / cs.LG / cs.CL
- 批量化反思优化循环：并行提案让GEPA更快更好Batching the Reflective Optimization Loop: Parallel Proposals Make GEPA Faster and BetterarXiv cs.AI / cs.LG / cs.CL
- 数学论证的依赖图：表示、结构度量与证明审阅Dependency Graphs of Mathematical Arguments: Representation, Structural Measurement, and Proof ReviewarXiv cs.AI / cs.LG / cs.CL
- 用于有据早期预警的增量知识图谱学习数字孪生An Incremental Knowledge-Graph Learner Digital Twin for Evidence-Grounded Early WarningarXiv cs.AI / cs.LG / cs.CL
- 谁知道什么、又支持什么？面向多模态心智理论的视角受限证据校准Who Knows What - and What Does It Support? Perspective-Bounded Evidence Calibration for Multimodal Theory of MindarXiv cs.AI / cs.LG / cs.CL
- 当评判者落后时：为持续自我改进的智能体协同演化评分标准与框架When the Critic Falls Behind: Co-Evolving Rubrics and Harnesses for Continual Agent Self-ImprovementarXiv cs.AI / cs.LG / cs.CL
- 零和部分可观测随机博弈的保守策略迭代Conservative Policy Iteration For Zero-Sum POSGsarXiv cs.AI / cs.LG / cs.CL
- 演化的智能体团队Evolving Agent TeamsarXiv cs.AI / cs.LG / cs.CL
- GUI-JEPA：以候选条件JEPA潜变量作为移动端GUI智能体的快速动作策略GUI-JEPA: Candidate-Conditioned JEPA Latents as a Fast Action Policy for Mobile GUI AgentsarXiv cs.AI / cs.LG / cs.CL
- RubricAtlas：从专家报告挖掘类型级评分标准以评估深度研究智能体RubricAtlas: Mining Type-Level Rubrics from Expert Reports to Evaluate Deep Research AgentsarXiv cs.AI / cs.LG / cs.CL
- 噪声引导：利用尾流偏转与声学指向性，以强化学习实现噪声约束的风电场控制Noise Steering: Exploiting Wake Steering and Acoustic Directivity for Noise-Constrained Wind Farm Control via Reinforcement LearningarXiv cs.AI / cs.LG / cs.CL
- 从正例中自然挖掘难负例，用于指令遵循检索Natural Hard-Negative Mining from Positives for Instruction-Following RetrievalarXiv cs.AI / cs.LG / cs.CL
- 视觉语言智能体可以在自身潜在世界中学习Vision-Language Agents Can Learn in Their Own Latent WorldsarXiv cs.AI / cs.LG / cs.CL
- 机器人会难过吗？评估大模型功能性效价的实证证据Do bots feel bad? Assessing empirical evidence for functional valence in LLMsarXiv cs.AI / cs.LG / cs.CL
- 事件驱动大模型智能体的主动记忆Proactive Memory for Event-Driven LLM AgentsarXiv cs.AI / cs.LG / cs.CL
- AutoVend：在长程自主售货运营上评测大模型智能体AutoVend: Benchmarking LLM Agents on Long-Horizon Autonomous Vending OperationsarXiv cs.AI / cs.LG / cs.CL
- SpectRAG：用于多跳检索增强生成、按token预算选择上下文的加权片段图SpectRAG: Weighted Chunk Graphs for Token-Budgeted Context Selection in Multi-Hop Retrieval-Augmented GenerationarXiv cs.AI / cs.LG / cs.CL
- Guided-RLVR：可验证奖励强化学习中的引导式探索Guided-RLVR: Guided Exploration in Reinforcement Learning with Verifiable RewardsarXiv cs.AI / cs.LG / cs.CL
- AI辅助发现受限拓扑上量子置换的CNOT分解AI-Assisted Discovery of CNOT Decompositions for Quantum Permutations on Constrained TopologiesarXiv cs.AI / cs.LG / cs.CL
- 思维链忠实度评估中作为选择的正确性条件化Correctness Conditioning as Selection in Chain-of-Thought Faithfulness EvaluationarXiv cs.AI / cs.LG / cs.CL
- OpenGray：评估放疗计划智能体决策的基准OpenGray: a benchmark for evaluating the decisions of radiotherapy planning agentsarXiv cs.AI / cs.LG / cs.CL
- Prospector：为科学推理学习紧凑的图遍历策略Prospector: learning a compact graph traversal policy for scientific reasoningarXiv cs.AI / cs.LG / cs.CL
- SFE-O：面向超临界CO₂生物质萃取与分馏、以知识为基础的AI本体SFE-O: An Ontology for Knowledge-Grounded AI in Supercritical CO$_2$ Biomass Extraction and FractionationarXiv cs.AI / cs.LG / cs.CL
- WorkFactory：把专业意图变成可执行工作，以规模化训练智能体WorkFactory: Turning Professional Intents into Executable Work for Scalable Agent TrainingarXiv cs.AI / cs.LG / cs.CL
- LiveSearch：以可执行验证代码出题的专业来源金融搜索智能体实时基准LiveSearch: Questions as Executable Verifier Code for a Live Benchmark of Professional-Source Financial Search AgentsarXiv cs.AI / cs.LG / cs.CL
- 用于贝叶斯校准的Rényi正则化流式变分推断Rènyi-Regularized Flow-Based Variational Inference for Bayesian CalibrationarXiv cs.AI / cs.LG / cs.CL
- 你的通用技能优化器真能促进智能体任务吗？用二部归因重新审视技能优化Does Your Generalist Skill Optimizer Really Incentivize Agentic Tasks? Rethinking Skill Optimization via Bipartite AttributionarXiv cs.AI / cs.LG / cs.CL
- GraphRAGProv：GraphRAG遇上“如何溯源”GraphRAGProv: GraphRAG Meets How-ProvenancearXiv cs.AI / cs.LG / cs.CL
- 让大模型智能体扎根业务流程：面向可靠多步服务调用的流程感知知识图谱Grounding LLM Agents in Business Processes: Process-Aware Knowledge Graphs for Reliable Multi-Step Service InvocationarXiv cs.AI / cs.LG / cs.CL
- TCM-DiagCommit：源自TCM-SD、用于有据中医证候选择与判定的基准TCM-DiagCommit: A TCM-SD-Derived Benchmark for Evidence-Conditioned Syndrome Selection and CommitmentarXiv cs.AI / cs.LG / cs.CL
- RPSD：基于最优传输的检索特权同策略自蒸馏RPSD: Retrieval-Privileged On-Policy Self-Distillation with Optimal TransportarXiv cs.AI / cs.LG / cs.CL
- ReForest：孤立森林中的可解释性与推理ReForest: Interpretability Meets Reasoning in Isolation ForestarXiv cs.AI / cs.LG / cs.CL
- 对齐制度权力与执行能力Aligning Institutional Power and Operational AbilityarXiv cs.AI / cs.LG / cs.CL
- 揭示大模型为何失败：从内部机制到对抗性推理轨迹Uncovering Why LLMs Fail: From Internal Mechanisms to Adversarial Reasoning TrajectoriesarXiv cs.AI / cs.LG / cs.CL
- 智能体条件路由：同一冻结模型内的多样推理路径Agent-Conditioned Routing: Diverse Reasoning Paths Within One Frozen ModelarXiv cs.AI / cs.LG / cs.CL
- 拨开迷雾：训练知识图谱智能体检索其答案背后的证据Clearing the FoG: Training KG Agents to Retrieve the Evidence Behind Their AnswersarXiv cs.AI / cs.LG / cs.CL
- 约束提交边界：受DIO/ODI启发、用于智能体式实体解析的授权护栏Constraining the Commit Boundary: A DIO/ODI-Informed Authorization Guardrail for Agentic Entity ResolutionarXiv cs.AI / cs.LG / cs.CL
- 超越读出层：跨模态与运算共享的潜在数值框架Beyond the Readout: A Latent Numerical Frame Shared Across Modality and OperationarXiv cs.AI / cs.LG / cs.CL
- M4Bench：评估临床电子病历推导智能体的程序化规范M4Bench: Evaluating Procedural Specification for Clinical EHR Derivation AgentsarXiv cs.AI / cs.LG / cs.CL
- Concordat：大模型智能体的类型化会话Concordat: Typed Sessions for LLM AgentsarXiv cs.AI / cs.LG / cs.CL
- 从交互到能力：以任务推进作为GUI智能体的状态与监督From Interaction to Capability: Task Advancement as State and Supervision for GUI AgentsarXiv cs.AI / cs.LG / cs.CL
- TeamGym：信息分散条件下AI智能体团队决策的训练场TeamGym: A Gym for Team Decision-Making by AI Agents with Distributed InformationarXiv cs.AI / cs.LG / cs.CL
- Timestral：概率预测时间序列基础模型系列TimestralarXiv cs.AI / cs.LG / cs.CL
- 罕见、相关却仍然出错：记忆如何让大模型群体决策失败Rare, Relevant, and Still Wrong: How Memory Fails LLM Group DecisionsarXiv cs.AI / cs.LG / cs.CL
- 预训练中改进了什么？用时间点模型追踪知识习得What Improves During Pretraining? Tracking Knowledge Acquisition with Point-in-Time ModelsarXiv cs.AI / cs.LG / cs.CL
- KPPax：面向大气化学机器学习的可微动力学常微分方程预处理KPPax: preprocess differentiable kinetic ODE for machine learning in atmospheric chemistryarXiv cs.AI / cs.LG / cs.CL
- KORA：AI工作负载的确定性优先执行控制KORA: Deterministic-First Execution Control for AI WorkloadsarXiv cs.AI / cs.LG / cs.CL
- 为哪个策略校准？干预下固定世界模型的保形不确定性Calibrated for Which Policy? Conformal Uncertainty for Fixed World Models Under InterventionarXiv cs.AI / cs.LG / cs.CL
- 课程驱动的深度强化学习用于自主远距离火箭着陆Curriculum-Driven Deep Reinforcement Learning for Autonomous Far-Range Rocket LandingarXiv cs.AI / cs.LG / cs.CL
- ResCom：大语言模型量化中的激活残差补偿ResCom: Activation Residual Compensation for Large Language Model QuantizationarXiv cs.AI / cs.LG / cs.CL
- 基于通信高效推测解码的分布式大模型推理Distributed LLM Inference via Communication-Efficient Speculative DecodingarXiv cs.AI / cs.LG / cs.CL
- 显式高斯点过程滤波器中的指数唤醒过冲Exponential Wake-Up Overshoot in Explicit Gaussian Point-Process FiltersarXiv cs.AI / cs.LG / cs.CL
- 从切片到体积：比较2D、2.5D与3D残差网络的肺结节恶性风险分类From Slices to Volumes: Comparing 2D, 2.5D, and 3D Residual Networks for Pulmonary Nodule Malignancy-Risk ClassificationarXiv cs.AI / cs.LG / cs.CL
- 跨教师架构的知识蒸馏用于紧凑型糖尿病视网膜病变分级Knowledge distillation across teacher architectures for compact diabetic retinopathy gradingarXiv cs.AI / cs.LG / cs.CL
- h1：自动定理证明中的“苦涩教训”初体验h1: A Taste of the Bitter Lesson in Automated Theorem ProvingarXiv cs.AI / cs.LG / cs.CL
- 条件增强生成对抗网络用于民用飞机气动构型设计Condition-Enhanced Generative Adversarial Network for Aerodynamic Configuration Design of Commercial AircraftarXiv cs.AI / cs.LG / cs.CL
- 学习中的泛化：艾林公式与捕食者—猎物模型Generalization in Learning: Eyring Formula and Predator-prey ModelarXiv cs.AI / cs.LG / cs.CL
- 用潜在常微分方程数据驱动发现李对称与首次积分Data-driven Discovery of Lie Symmetries and First Integrals with Latent ODEsarXiv cs.AI / cs.LG / cs.CL
- 基于图神经网络的供水管网多处泄漏定位Multiple Leak Localization in Water Distribution Networks through Graph Neural NetworksarXiv cs.AI / cs.LG / cs.CL
- 用于风电场布局样本高效贝叶斯优化的图信息克里金模型A Graph-informed Kriging model for sample-efficient Bayesian optimization of wind farm layoutsarXiv cs.AI / cs.LG / cs.CL
- 让门控决定：大模型量化条件记忆的逐次查找精度Let the Gate Decide: Per-Lookup Precision for Quantized Conditional Memory in LLMsarXiv cs.AI / cs.LG / cs.CL
- 逐层重打包：无需额外重分配步骤的变长混合注意力训练Layerwise Repacking: Variable-Length Hybrid-Attention Training without an Extra Redistribution PassarXiv cs.AI / cs.LG / cs.CL
- DuplexTune：用于内存高效混合专家适配的耦合状态重放DuplexTune: Coupled-State Replay for Memory-Efficient Mixture-of-Experts AdaptationarXiv cs.AI / cs.LG / cs.CL
- 探索式创意预排序：用于实时创意优化的分层经验贝叶斯汤普森采样Exploratory creative pre-ranking: Hierarchical empirical-Bayes Thompson sampling for real-time creative optimizationarXiv cs.AI / cs.LG / cs.CL
- Jaxnasium：拥抱PyTree，实现任意智能体、跨套件的强化学习Jaxnasium: Embracing PyTrees For Any-Agent, Cross-Suite Reinforcement LearningarXiv cs.AI / cs.LG / cs.CL
- 提示有多重要？金融推理中事后自蒸馏的适用条件How Much Does the Hint Matter? Scope Conditions for Hindsight Self-Distillation in Financial ReasoningarXiv cs.AI / cs.LG / cs.CL
- 指令容量：受控负载下语言模型的行为与退化Instruction Capacity: Behavior and Degradation of Language Models under Controlled LoadarXiv cs.AI / cs.LG / cs.CL
- ℓp球上在线线性优化的极小极大最优小损失界Minimax-Optimal Small-Loss Bounds for Online Linear Optimisation on $\ell_p$ BallsarXiv cs.AI / cs.LG / cs.CL
- 不止看更新次数：评估测试时模型更新的控制图Beyond Update Counts: A Control Map for Evaluating Test-Time Model UpdatesarXiv cs.AI / cs.LG / cs.CL
- 半参数贝叶斯网络中的狄利克雷过程混合：一种新的条件密度估计方法Dirichlet Process Mixtures within Semiparametric Bayesian Networks: A Novel Conditional Density Estimation ApproacharXiv cs.AI / cs.LG / cs.CL
- 神经网络中的标度与重整化：顿悟现象的有效重整化群描述Scaling and Renormalization in Neural Networks: An Effective Renormalization-Group Description of GrokkingarXiv cs.AI / cs.LG / cs.CL
- SpatialPC：具有亚线性局部信用传播的深度计算SpatialPC: Deep Computation with Sublinear Local Credit PropagationarXiv cs.AI / cs.LG / cs.CL
- 神经网络训练的动力学粗粒化Dynamical coarse-graining of neural-network trainingarXiv cs.AI / cs.LG / cs.CL
- 动态数据驱动应用系统中基于图学习的贝叶斯证据累积Bayesian Evidence Accumulation for Graph-Based Learning in Dynamic Data-Driven Applications SystemsarXiv cs.AI / cs.LG / cs.CL
- 基础模型能发现提示分类器漏掉的隐藏意图Base Models Can Detect Hidden Intentions that Prompted Classifiers MissarXiv cs.AI / cs.LG / cs.CL
- Concorde：基于解耦能量最小化的几何感知链接预测Concorde: Geometry-Aware Link Prediction via Decoupled Energy MinimizationarXiv cs.AI / cs.LG / cs.CL
- 用对数的代价换三个世界：带基线保护的两全其美组合半老虎机Three Worlds for the Price of a Logarithm: Baseline-Protected Best-of-Both-Worlds Combinatorial Semi-BanditsarXiv cs.AI / cs.LG / cs.CL
- 冻结而非克隆：无需保留离线数据的表达性策略微调Freeze, Don't Clone: Fine-Tuning with Expressive Policies Without Retaining Offline DataarXiv cs.AI / cs.LG / cs.CL
- 信息瓶颈编码器应学到什么？将表示学习与速率控制分离What Should an Information Bottleneck Encoder Learn? Separating Representation Learning from Rate ControlarXiv cs.AI / cs.LG / cs.CL
- EVA-Flow：学习共享化学知识，统一生成环境感知的分子构象EVA-Flow: Learning Shared Chemistry for Unified Environment-Aware Molecular Conformation GenerationarXiv cs.AI / cs.LG / cs.CL
- 在同策略强化学习中连接探索与可塑性丧失Bridging Exploration and Loss of Plasticity in On-Policy Reinforcement LearningarXiv cs.AI / cs.LG / cs.CL
- 置信度不是承诺：类型化决策模型的参考审计Confidence Is Not a Promise: A Reference Audit of Typed Decision ModelsarXiv cs.AI / cs.LG / cs.CL
- AHEAD：大规模线性规划的自适应神经加速AHEAD: Adaptive Neural Acceleration for Large-Scale Linear ProgrammingarXiv cs.AI / cs.LG / cs.CL
- 可靠性感知的短期血糖风险预测：序列、多模态与状态更新数字孪生模型的锚定比较框架Reliability-Aware Short-Term Glycemic Risk Prediction Using an Anchor-Based Comparative Framework of Sequence, Multimodal, and State-Updating Digital-Twin ModelsarXiv cs.AI / cs.LG / cs.CL
- 从生成到打分：在领英规模化基于大模型的语义标注From Generation to Scoring: Scaling LLM-Based Semantic Labeling at LinkedInarXiv cs.AI / cs.LG / cs.CL
- SpecLST：谱增强的长短期时空多模态交通预测建模SpecLST: Spectral-Enhanced Long-Short Term Spatio-Temporal Multimodal Modeling for Traffic ForecastingarXiv cs.AI / cs.LG / cs.CL
- 自动驾驶微流控实验室中学得的设计代理模型何时可信？基于六个公开数据表构建与测试的离线闭环框架When Can a Learned Design Surrogate Be Trusted in a Self-Driving Microfluidics Laboratory? An Offline Closed-Loop Framework Built and Tested on Six Public TablesarXiv cs.AI / cs.LG / cs.CL
- 线性时间预测编码中的持续性、暴露与信用路由Persistence, Exposure, and Credit Routing in Linear Temporal Predictive CodingarXiv cs.AI / cs.LG / cs.CL
- 排斥式群体策略梯度实现行为多样的强化学习Repulsive Population Policy Gradients for Behaviorally Diverse Reinforcement LearningarXiv cs.AI / cs.LG / cs.CL
- 稳健Softmax与交叉熵的基于属性测试Property-Based Testing for Robust Softmax and Cross-EntropyarXiv cs.AI / cs.LG / cs.CL
- 模拟岩石学标签实现由全岩地球化学进行概率性矿物学预测Simulated petrological labels enable probabilistic mineralogy prediction from whole-rock geochemistryarXiv cs.AI / cs.LG / cs.CL
- 学习原子相互作用以研究液体混合物热力学Learning atomic interactions for the thermodynamics of liquid mixturesarXiv cs.AI / cs.LG / cs.CL
- 巡航条件下跨声速机翼的高保真气动结构数据集High-Fidelity Aerostructural Dataset of Transonic Wings in Cruise ConditionsarXiv cs.AI / cs.LG / cs.CL
- 拆分式联邦遗忘学习Split Federated UnlearningarXiv cs.AI / cs.LG / cs.CL
- CoHARL：用强化学习协同演化智能体策略与可执行框架CoHARL: Co-Evolving Agent Policies and Executable Harnesses with Reinforcement LearningarXiv cs.AI / cs.LG / cs.CL
- 欠阻尼锚定朗之万动力学Underdamped Anchored Langevin DynamicsarXiv cs.AI / cs.LG / cs.CL
- 为同策略蒸馏提出更好的采样轨迹Proposing Better Rollouts for On-Policy DistillationarXiv cs.AI / cs.LG / cs.CL
- PRISM：标签稀缺医学中的病理可靠性——病理基础模型切片级基准PRISM: Pathology Reliability In Scarce-label Medicine. A Tile-level Benchmark of Pathology Foundation ModelsarXiv cs.AI / cs.LG / cs.CL
- 把注意力偏置折叠进键，让压缩KV缓存在标准内核上运行Folding Attention Biases into Keys Makes Compacted KV Caches Run on Standard KernelsarXiv cs.AI / cs.LG / cs.CL
- NextEval：无导数优化的智能体式求解器与基准NextEval: An Agent-Based Solver and Benchmark for Derivative-Free OptimizationarXiv cs.AI / cs.LG / cs.CL
- 全纯柯西网络用于高维反射布朗运动的高效平稳尾部计算Holomorphic Cauchy Networks for Efficient Stationary Tail Computation in High-Dimensional Reflected Brownian MotionsarXiv cs.AI / cs.LG / cs.CL
- 从点与块学习：用于时间序列分析的预训练混合注意力TransformerLearning from Points and Patches: A Pre-trained Hybrid Attention Transformer for Time Series AnalysisarXiv cs.AI / cs.LG / cs.CL
- 少即是多：为GPU内核优化学习该剖析什么Less is more: learning what to profile for GPU kernel optimizationarXiv cs.AI / cs.LG / cs.CL
- MM-Net：缺血性卒中后睡眠分期与睡眠呼吸暂停评估的多模态多任务网络MM-Net: A Multimodal Multi-Task Network for Sleep Staging and Sleep Apnea Assessment After Ischemic StrokearXiv cs.AI / cs.LG / cs.CL
- 回忆还是重学？针对反复概念漂移的上下文专家混合Recall or Relearn? A Mixture of In-Context Experts for Recurring Concept DriftarXiv cs.AI / cs.LG / cs.CL
- 基于结构化子空间随机性的随机舍入高效量化训练Efficient Quantization Training with Stochastic Rounding via Structured Subspace RandomnessarXiv cs.AI / cs.LG / cs.CL
- 基于矩阵分式规划的归一化割与数据聚类Normalized Cuts and Data Clustering via Matrix Fractional ProgrammingarXiv cs.AI / cs.LG / cs.CL
- 用于Mamba-2状态压缩的平衡状态子空间Balanced State Subspaces for Mamba-2 State CompressionarXiv cs.AI / cs.LG / cs.CL
- Dream-COP：在经验世界中引导组合优化问题的算法发现Dream-COP: Guiding Algorithm Discovery for Combinatorial Optimization Problems in an Experience WorldarXiv cs.AI / cs.LG / cs.CL
- 批大小可以为一：基于分类器几何的测试时自适应The Batch Size Can Be One: Test-Time Adaptation from Classifier GeometryarXiv cs.AI / cs.LG / cs.CL
- 用于可解释非线性结构系统辨识的深度贝叶斯方法Deep Bayesian Approaches for Interpretable Nonlinear Structural System IdentificationarXiv cs.AI / cs.LG / cs.CL
- 扰动一致性何时能在Softmax置信度之外提供信息？决策一致性框架的理论与证据When Does Perturbation Consistency Add Information Beyond Softmax Confidence? Theory and Evidence for the Decision Consistency FrameworkarXiv cs.AI / cs.LG / cs.CL
- 廉价的公平性修改何时安全？筛查双边匹配市场中的事后修改When Is a Cheap Fairness Edit Safe? Screening Post-Hoc Edits in Two-Sided Matching MarketsarXiv cs.AI / cs.LG / cs.CL
- TACO-Graph：文本属性图的自动构建TACO-Graph: Automated Construction of Text-Atributed GraphsarXiv cs.AI / cs.LG / cs.CL
- 参数化马尔可夫链的对数导数技巧Log-Derivative Trick for Parametric Markov ChainsarXiv cs.AI / cs.LG / cs.CL
- HeteGCD：异配图上广义类别发现的统计感知双原型学习HeteGCD: Statistics-Aware Dual-Prototype Learning for Generalized Category Discovery on Heterophilous GraphsarXiv cs.AI / cs.LG / cs.CL
- RankFormer：通过表示容量分析与渐进式解码重新思考时间序列TransformerRankFormer: Rethinking Time Series Transformers through Representation Capacity Analysis and Progressive DecodingarXiv cs.AI / cs.LG / cs.CL
- 保守决斗老虎机：基线约束下的偏好学习Conservative Dueling Bandits: Preference Learning under Baseline ConstraintsarXiv cs.AI / cs.LG / cs.CL
- 非线性直接反馈对齐：可扩展的免反向传播训练Nonlinear Direct Feedback Alignment for Scalable Backpropagation-Free TrainingarXiv cs.AI / cs.LG / cs.CL
- 用于科学机器学习时间插值的生成式精炼学习Generative Refinement Learning for Temporal Interpolation in Scientific Machine LearningarXiv cs.AI / cs.LG / cs.CL
- 基于内容不变梯度归因的稳健大模型数据选择Robust LLM Data Selection via Content-Invariant Gradient AttributionarXiv cs.AI / cs.LG / cs.CL
- 大模型预训练通常发生在稳定性边缘吗？Does LLM Pre-Training Typically Occur at the Edge of Stability?arXiv cs.AI / cs.LG / cs.CL
- Select-TTT：查询条件化的选择性快速权重更新，用于原位测试时训练Select-TTT: Query-Conditioned Selective Fast Weight Updates for In-Place Test-Time TrainingarXiv cs.AI / cs.LG / cs.CL
- API-TAIL：通过温度对齐的似然尾部检测大模型API中的模型替换API-TAIL: Detecting Model Substitution in LLM APIs through Temperature-Aligned Likelihood TailsarXiv cs.AI / cs.LG / cs.CL
- 利用大模型知识改进分布外表格预测：一种交叉验证正则化方法LLM knowledge for improved Out-of-Distribution Tabular Prediction - A Cross-Validated Regularisation ApproacharXiv cs.AI / cs.LG / cs.CL
- SetGRPO：校准评分器诱导的策略优化SetGRPO: Calibrating Grader-Induced Policy OptimizationarXiv cs.AI / cs.LG / cs.CL
- 当所有采样都失败时：面向难题推理的策略兼容联合解码When All Rollouts Fail: Policy-Compatible Joint Decoding for Hard-Problem ReasoningarXiv cs.AI / cs.LG / cs.CL
- 尾部线索、主成分校正：开放集测试时自适应的即插即用修正Tail Cues, Principal Corrections: Plug-and-Play Rectification for Open-Set Test-Time AdaptationarXiv cs.AI / cs.LG / cs.CL
- 基于演化优化的强化学习超参数与神经架构搜索：HPNAS-RLHyperparameter and Neural Architecture Search via Evolutionary Optimization for Reinforcement Learning: HPNAS-RLarXiv cs.AI / cs.LG / cs.CL
- HiLo-RD：掩码扩散语言模型的跨预算自蒸馏HiLo-RD: Cross-Budget Self-Distillation for Masked Diffusion Language ModelsarXiv cs.AI / cs.LG / cs.CL
- 图像扩散模型低秩适配的溯源签名Provenance Signatures for Low Rank Adaptations of Image Diffusion ModelsarXiv cs.AI / cs.LG / cs.CL
- 用于约束黑盒杠杆优化的高斯过程老虎机Gaussian Process Bandits for Constraint Black Box Lever OptimizationarXiv cs.AI / cs.LG / cs.CL
- TDLM：扩散语言模型的自适应时间积分TDLM: Adaptive Temporal Integration for Diffusion Language ModelsarXiv cs.AI / cs.LG / cs.CL
- 基于机器学习的定点施肥推荐中的算法单一化Algorithmic Monoculture in Machine Learning-based Site-Specific Fertilizer RecommendationarXiv cs.AI / cs.LG / cs.CL
- DCKV：混合语言模型中提示KV缓存的学习式变换编码DCKV: Learned Transform Coding of the Prompt KV Cache in Hybrid Language ModelsarXiv cs.AI / cs.LG / cs.CL
- Live-OPSD：主动流式视频大模型的同策略自蒸馏Live-OPSD: On-Policy Self-Distillation for Proactive Streaming Large Video ModelsarXiv cs.AI / cs.LG / cs.CL
- 自回归线性阈值函数的VC维On the VC Dimension of Autoregressive Linear Threshold FunctionsarXiv cs.AI / cs.LG / cs.CL
- 图神经网络引导的网络约简用于集中式城市交通路由GNN-Guided Network Reduction for Centralized Urban Traffic RoutingarXiv cs.AI / cs.LG / cs.CL
- 进步还是加速？数据筛选如何改进大模型训练Advancement or Acceleration? How Data Curation Improves LLM TrainingarXiv cs.AI / cs.LG / cs.CL
- CaLLM：通过非参数适配器路由实现大模型持续适应CaLLM: Continual Adaptation of LLMs via Non-Parametric Adapter RoutingarXiv cs.AI / cs.LG / cs.CL
- 元模型：对学习过程建模的模型Meta Models: Models that Model LearningarXiv cs.AI / cs.LG / cs.CL
- Compote：抗坍缩的多假设学习用于序列推荐Compote: Collapse-Resistant Multiple Hypothesis Learning For Sequential RecommendationarXiv cs.AI / cs.LG / cs.CL
- V-LUMEN：用于视觉语言模型嵌入扩展的视觉查找记忆V-LUMEN: Visual Lookup Memory for Embedding Scaling in Vision-Language ModelsarXiv cs.AI / cs.LG / cs.CL
- 用回归模型预测作物产量：模型性能比较评估Crop Yield Prediction Using Regression Models: A Comparative Evaluation of Model PerformancearXiv cs.AI / cs.LG / cs.CL
- 探索增量学习假设的极限：直接偏好优化中的单对生成器偏差Exploring the Limits of the Delta Learning Hypothesis: Single-Pair Generator Biases in Direct Preference OptimizationarXiv cs.AI / cs.LG / cs.CL
- EnergyCal：驯服稀疏MoE训练中Muon系优化器的损失尖峰EnergyCal: Taming Loss Spikes in Muon-Family Optimizers for Sparse MoE TrainingarXiv cs.AI / cs.LG / cs.CL
- 论直接偏好优化的稀疏性：NTK区间中的权重解耦On the Sparsity of Direct Preference Optimization: Weight Disentanglement in the NTK RegimearXiv cs.AI / cs.LG / cs.CL
- 自正则化流式语言模型Self-Regularized Flow Language ModelsarXiv cs.AI / cs.LG / cs.CL
- Transformer语言模型权重剪枝的机制性失效模式Mechanistic Failure Modes of Weight Pruning in Transformer Language ModelsarXiv cs.AI / cs.LG / cs.CL
- 平衡传播的生成式重放：面向边缘友好的霍普菲尔德网络持续学习Generative Replay for Equilibrium Propagation: Toward Edge-Friendly Continual Learning in Hopfield NetworksarXiv cs.AI / cs.LG / cs.CL
- 超越参数保护：无重放持续学习的Fisher塑形EWCBeyond Parameter Protection: Fisher-Shaped EWC for Replay-Free Continual LearningarXiv cs.AI / cs.LG / cs.CL
- 用于扩散引导的变分最优控制Variational Optimal Control for Diffusion GuidancearXiv cs.AI / cs.LG / cs.CL
- 物理引导流匹配：全球电离层总电子含量预测的生成式AI框架Physics-Guided Flow Matching: A Generative AI Framework for Global Total Electron Content ForecastingarXiv cs.AI / cs.LG / cs.CL
- 通过延续共享实现长程智能体强化学习的动作价值校准Action-Value Calibration via Continuation Sharing for Long-Horizon Agentic Reinforcement LearningarXiv cs.AI / cs.LG / cs.CL
- SimplexLoRA：基于单纯形投影交替最小化的动态秩更新SimplexLoRA: Dynamic Rank Updating via Alternating Minimization with Simplex ProjectionarXiv cs.AI / cs.LG / cs.CL
- 锐度感知Muon：带动量引导扰动的矩阵感知优化Sharpness-Aware Muon: Matrix-Aware Optimization with Momentum-Guided PerturbationsarXiv cs.AI / cs.LG / cs.CL
- 关系何时值得信任？多关系图学习中的操纵成本、暴露与信任陷阱When Is a Relation Worth Trusting? Manipulation Costs, Exposure, and the Trust Trap in Multi-Relational Graph LearningarXiv cs.AI / cs.LG / cs.CL
- 你的策略其实是一套框架：智能体接口的决策几何Your Policy Is Secretly a Harness: The Decision Geometry of Agent InterfacesarXiv cs.AI / cs.LG / cs.CL
- 大模型驱动的智能体脚手架演化中，评估预算该花在哪里？Where Should the Evaluation Budget Go in LLM-Driven Agent Scaffold Evolution?arXiv cs.AI / cs.LG / cs.CL
- 昂贵模拟器的贝叶斯推断：设计选择的综述与比较Bayesian Inference of Expensive Simulators: Survey and Comparison of Design ChoicesarXiv cs.AI / cs.LG / cs.CL
- 用结构化表示探测实现端到端自动驾驶的风险预判Structured Representation Probing for Risk Anticipation in End-to-End Autonomous DrivingarXiv cs.AI / cs.LG / cs.CL
- WCC：判断大规模优化中何时使用协同演化的学习框架WCC: A Learning-Based Framework for Deciding When to Apply Cooperative Coevolution in Large-Scale OptimizationarXiv cs.AI / cs.LG / cs.CL
- CoS-VQ：大语言模型的协调共享尺度向量量化CoS-VQ: Coordinated Shared-Scale Vector Quantization for Large Language ModelsarXiv cs.AI / cs.LG / cs.CL
- 基于规范偏差建模的脑电抑郁检测Normative Deviation Modeling for EEG-based Depression DetectionarXiv cs.AI / cs.LG / cs.CL
- EmbeddingGemma 2：统一高效的多模态表示EmbeddingGemma 2: Unified and Efficient Multimodal RepresentationsarXiv cs.AI / cs.LG / cs.CL
- TRIDENT-Seq：画像匹配的前缀—未来对比，用于标签高效的支付风险建模TRIDENT-Seq: Profile-Matched Prefix-Future Contrast for Label-Efficient Payment Risk ModelingarXiv cs.AI / cs.LG / cs.CL
- 欧氏拥挤驱动的自适应欧氏—双曲图学习Euclidean Crowding-Driven Adaptive Euclidean-Hyperbolic Graph LearningarXiv cs.AI / cs.LG / cs.CL
- 混合专家解码的通道选择无需校准数据Channel Selection for Mixture-of-Experts Decoding Needs No Calibration DataarXiv cs.AI / cs.LG / cs.CL
- 随机深度神经网络的类高斯近似Para-Gaussian Approximations of Random Deep Neural NetworksarXiv cs.AI / cs.LG / cs.CL
- 多少与多大？知识图谱嵌入中的负样本对比学习How Many and How Much? Contrastive Learning with Negative Samples for Knowledge Graph EmbeddingarXiv cs.AI / cs.LG / cs.CL
- LinMix：将数据混合视为线性领域归因LinMix: Data Mixing as Linear Domain AttributionarXiv cs.AI / cs.LG / cs.CL
- 布朗插值采样中的位移矩与路径终点分离Displacement Moments and Path-Terminal Separation in Brownian Interpolant SamplingarXiv cs.AI / cs.LG / cs.CL
- 用贝叶斯自适应MDP分配采样以增强GRPO训练Bayes-Adaptive MDPs for Rollout Allocation to Enhance GRPO TrainingarXiv cs.AI / cs.LG / cs.CL
- 视觉证据不足时：指令、后训练与内部激活如何塑造多模态语言模型的拒答When Visual Evidence Is Insufficient: How Instructions, Post-Training, and Internal Activations Shape Refusal in Multimodal Language ModelsarXiv cs.AI / cs.LG / cs.CL
- CRISP：协同设计剪枝与调度，实现边缘服务器上的高效MoE推理CRISP: Co-designing Pruning and Scheduling for Efficient MoE Inference on Edge ServersarXiv cs.AI / cs.LG / cs.CL
- 神经决策的可解释元策略Interpretable Meta-Policies for Neural Decision MakingarXiv cs.AI / cs.LG / cs.CL
- 输入跳连可以重参数化VAE，但不会改进生成器An Input Skip Can Reparameterize a VAE Without Improving the GeneratorarXiv cs.AI / cs.LG / cs.CL
- 顺着梯度：把安全性退化追溯到训练数据Follow the Gradient: Tracing Safety Degradation Back to Training DataarXiv cs.AI / cs.LG / cs.CL
- 大模型的正交参数化多属性引导Orthogonal and Parametrized Multi-Attribute Steering of LLMsarXiv cs.AI / cs.LG / cs.CL
- 结构鸿沟：当更多数据无法替代结构化推理The Structure Gap: When More Data Cannot Replace Structured ReasoningarXiv cs.AI / cs.LG / cs.CL
- GTFormer：无语法图网络替代Transformer，以线性时间实现8192倍长度外推GTFormer: A Syntax-Free Graph Network Alternative to Transformers for 8192$\times$ Length Extrapolation in Linear TimearXiv cs.AI / cs.LG / cs.CL
- 自我强化的集体动力学：无监督RLVR中跨任务耦合如何导致策略坍缩Collective Dynamics of Self-Reinforcement: How Cross-Task Coupling Drives Policy Collapse in Unsupervised RLVRarXiv cs.AI / cs.LG / cs.CL
- 不确定性量化失效时：临床类别不平衡下选择性预测的结构脆弱性与修正When uncertainty quantification fails: structural vulnerabilities and corrections for selective prediction under clinical class imbalancearXiv cs.AI / cs.LG / cs.CL
- Pneuma：民用飞机预测性维护的大规模真实时间序列基准Pneuma: A Large-Scale Real-World Time Series Benchmark for Commercial Aircraft Predictive MaintenancearXiv cs.AI / cs.LG / cs.CL
- Gym-Anything-RT：为视觉语言智能体规模化实时环境Gym-Anything-RT: Scaling Real-Time Environments for Vision-Language AgentsarXiv cs.AI / cs.LG / cs.CL
- LCM-Promo：基于大模型的促销分配增益模型LCM-Promo: An LLM-Based Uplift Model for Promotion AllocationarXiv cs.AI / cs.LG / cs.CL
- 理解潜在世界模型的局限Understanding the limits of latent world modelsarXiv cs.AI / cs.LG / cs.CL
- GAS-PFN：用于零样本交通预测的图感知时空先验数据拟合网络GAS-PFN: Graph-Aware Spatiotemporal Prior-Data Fitted Networks for Zero-Shot Traffic ForecastingarXiv cs.AI / cs.LG / cs.CL
- EnvSched：面向采样高效智能体强化学习的环境感知调度EnvSched: Environment-Aware Scheduling for Rollout-Efficient Agent Reinforcement LearningarXiv cs.AI / cs.LG / cs.CL
- MuonGS：以放松的权重范数控制预训练大模型MuonGS: Pretraining LLMs with Relaxed Weight-Norm ControlarXiv cs.AI / cs.LG / cs.CL
- 通过可控环境合成探索工具调用智能体的标度律Scaling Laws for Tool-Use Agents via Controllable Environment SynthesisarXiv cs.AI / cs.LG / cs.CL
- 作为格点规范理论的注意力：关系只通过威尔逊环相互作用Attention as Lattice Gauge Theory: Relations Interact Only Through Wilson LoopsarXiv cs.AI / cs.LG / cs.CL
- 安全头：学得的注意力门控提升提示注入鲁棒性Heads of Security: Learned Attention Gates Improve Prompt Injection RobustnessarXiv cs.AI / cs.LG / cs.CL
- 测试时应在何时干预？When Should We Intervene at Test Time?arXiv cs.AI / cs.LG / cs.CL
- ReCoPII：构建上下文个人身份信息遮盖数据集的关系锚定框架ReCoPII: A Relation-Grounded Framework for Building Contextual PII Masking DatasetsarXiv cs.AI / cs.LG / cs.CL
- MedJev：评估医学中零样本系统1决策MedJev: Evaluating Zero-Shot System-One Decision Making in MedicinearXiv cs.AI / cs.LG / cs.CL
- OmniBrain-0：面向具身智能的统一智能体模型OmniBrain-0: A Unified Agentic Model for Embodied IntelligencearXiv cs.AI / cs.LG / cs.CL
- FinLLM-KG：知识图谱增强大模型用于算法交易中的金融情感分析FinLLM-KG: Knowledge Graph-Enhanced LLMs for Financial Sentiment Analysis in Algorithmic TradingarXiv cs.AI / cs.LG / cs.CL
- 3、7、14、17、42：指令微调安装并强化了语言模型的选数启发式3, 7, 14, 17, 42: Instruction Tuning Installs and Sharpens Number-Choice Heuristics in Language ModelsarXiv cs.AI / cs.LG / cs.CL
- Lipi-Ghor：大规模、带说话人分离的多领域孟加拉语语音语料库及其可扩展构建流程Lipi-Ghor: A Large-Scale, Diarized, Multi-Domain Bengali Speech Corpus and a Scalable Pipeline for Building ItarXiv cs.AI / cs.LG / cs.CL
- 坦率地说：国会沟通中选举诚信与谁为伍Frankly Speaking: The Company Election Integrity Keeps in Congressional CommunicationarXiv cs.AI / cs.LG / cs.CL
- 基于低秩多线性分解的大模型幻觉检测LLM Hallucination Detection via Low-rank Multilinear DecompositionarXiv cs.AI / cs.LG / cs.CL
- 超越IceBERT：从1200万到10亿参数的冰岛语开放数据ModernBERT编码器Beyond IceBERT: Open-Data ModernBERT Encoders for Icelandic from 12M to 1B ParametersarXiv cs.AI / cs.LG / cs.CL
- 先检测再消歧：基于扰动的大模型输入歧义检测Detect Before You Disambiguate: Perturbation-Based Input Ambiguity Detection in LLMsarXiv cs.AI / cs.LG / cs.CL
- AK_QMUL参加ArGuard共享任务：单个微调JAIS用于阿拉伯语有害提示检测AK_QMUL at ArGuard Shared Tasks: A Single Fine-Tuned JAIS for Arabic Harmful-Prompt DetectionarXiv cs.AI / cs.LG / cs.CL
- 没有分词器，差距仍在：同形字与不可见字符水印的跨语言公平性No Tokenizer, Still a Gap: Cross-Lingual Fairness of Homoglyph and Invisible-Character WatermarkingarXiv cs.AI / cs.LG / cs.CL
- CONCUR：多任务多奖励强化学习的一致性辅助奖励CONCUR: Concordant Auxiliary Rewards for Multi-Task, Multi-Reward Reinforcement LearningarXiv cs.AI / cs.LG / cs.CL
- AI科学家知道什么是新的吗？Do AI Scientists Know What's New?arXiv cs.AI / cs.LG / cs.CL
- 评测智能体交接中的任务接管Benchmarking Task Takeover Across Agent HandoffsarXiv cs.AI / cs.LG / cs.CL
- Transformer中证据效应的因果定位Causal Localization of Evidence Effects in TransformersarXiv cs.AI / cs.LG / cs.CL
- ETC-Bench：评估侧转写规范化的多语言基准ETC-Bench: A Multilingual Benchmark for Evaluation-Side Transcript CanonicalizationarXiv cs.AI / cs.LG / cs.CL
- 面向多语言语音识别、语言感知且错误驱动优化的智能体式数据采样框架Agentic Data Sampling Framework for Multilingual ASR via Linguistic-Aware Error-Driven OptimizationarXiv cs.AI / cs.LG / cs.CL
- 用大模型为基于地标的路线指引构建结构化地图Structured Maps for Landmark-Based Route Instructions with LLMsarXiv cs.AI / cs.LG / cs.CL
- 基于分段话语表示理论、用大模型进行日语话语结构分析Japanese Discourse Structure Analysis Based on Segmented Discourse Representation Theory Using LLMsarXiv cs.AI / cs.LG / cs.CL
- 前沿大语言模型中阿希从众效应的定量研究A Quantitative Approach to Asch Conformity in Frontier Large Language ModelsarXiv cs.AI / cs.LG / cs.CL
- 儿童规模语言模型中的专门化、多样性与标度Specialization, Diversity, and Scaling in Child-Scale Language ModelsarXiv cs.AI / cs.LG / cs.CL
- token置信度只是链长吗？推理轨迹的信号处理审计Is Token Confidence Just Chain Length? A Signal-Processing Audit of Reasoning TracesarXiv cs.AI / cs.LG / cs.CL
- 通过隔离反对意见减少智能体过度防御式写作的框架A Harness Reducing over-Defensive Writing in Agents by Isolating ObjectionsarXiv cs.AI / cs.LG / cs.CL
- CAMR：大语言模型持续学习的兼容性感知记忆重放CAMR: Compatibility-Aware Memory Replay for Continual Learning in Large Language ModelsarXiv cs.AI / cs.LG / cs.CL
- DocSkill：融合技能的智能体推理用于多语言多跳文档问答DocSkill: Skill-Integrated Agentic Reasoning for Multilingual Multi-Hop Document Question AnsweringarXiv cs.AI / cs.LG / cs.CL
- 语言模型保留被放弃的分析：花园路径句之后的残留与适应Language Models Keep Abandoned Analyses: Residue and Adaptation after Garden-Path SentencesarXiv cs.AI / cs.LG / cs.CL
- 人格潜变量：用个人写作为冻结语言模型做个体化条件Personality Latents: Conditioning Frozen Language Models on Individuals from Their Own WritingarXiv cs.AI / cs.LG / cs.CL
- 定位弱点还不够：大语言模型的状态感知递归知识改进Weakness Localization Is Not Enough: State-Aware Recursive Knowledge Improvement for Large Language ModelsarXiv cs.AI / cs.LG / cs.CL
- 言辞还是行动？大模型如何预测人机对话中的用户决策Words or Deeds? How LLMs Predict Users' Decisions in Human-Agent DialoguearXiv cs.AI / cs.LG / cs.CL
- 不遗漏：从大规模随机化评估语料中寻找证据Nothing Left Behind: Finding Evidence from a Large Randomized Evaluation CorpusarXiv cs.AI / cs.LG / cs.CL
- 开发巴斯克语语音语言模型Developing a Speech Language Model for BasquearXiv cs.AI / cs.LG / cs.CL
- 需要大模型才能知道来源为何冲突吗？RAG中知识冲突类型判别的决策模型Do We Need an LLM to Know Why Sources Conflict? Decision Models for Knowledge-Conflict Typing in RAGarXiv cs.AI / cs.LG / cs.CL
- TrawMem：将智能体记忆重构为面向任务、角色自适应的工作区TrawMem: Reframing Agent Memory as Task-Oriented Role-Adaptive WorkspacesarXiv cs.AI / cs.LG / cs.CL
- DocXEval：多模态文档理解可解释性的系统评估DocXEval: A Systematic Evaluation of Explainability in Multimodal Document UnderstandingarXiv cs.AI / cs.LG / cs.CL
- JevRec：用类Jev决策建模进行语言编码的大模型生成式推荐JevRec: Linguistic Encoding with Jev-like Decision Modeling for LLM-based Generative RecommendationarXiv cs.AI / cs.LG / cs.CL
- 自键控参数内记忆：冻结语言模型无需外部记忆的对话回忆Self-Keyed In-Parameter Memorization for External Memory-Free Dialogue Recall in Frozen Language ModelsarXiv cs.AI / cs.LG / cs.CL
- NLP领域多语言与情境化评估的现状与前景The State and Fate of Multilingual, Contextual Evaluation in the NLP WorldarXiv cs.AI / cs.LG / cs.CL
- 参数高效的分层低秩重参数化Hierarchical Low-Rank Reparameterization for Parameter-EfficientarXiv cs.AI / cs.LG / cs.CL
- 利用人工整理词表进行低资源语言文本生成Text Generation in Low-Resource Languages Using Human-Curated VocabulariesarXiv cs.AI / cs.LG / cs.CL
- 蒸馏后保留了什么？GPT-2与DistilGPT-2的机制可解释性研究What Survives Distillation? A Mechanistic Interpretability Study of GPT-2 and DistilGPT-2arXiv cs.AI / cs.LG / cs.CL
- 以决策模型作为幻觉检测的评判者Decision Models as Judges for Hallucination DetectionarXiv cs.AI / cs.LG / cs.CL
- 面向非洲语言的大模型安全对齐：以豪萨语为例Aligning Large Language Models for Safety in African Languages: A Case Study of HausaarXiv cs.AI / cs.LG / cs.CL
- REGLAT参加ArGuard共享任务：融合命名实体识别的集成模型用于阿拉伯语有害内容检测REGLAT at ArGuard Shared Tasks: An Ensemble-Based Model Integrating NER for Harmful Content Detection in ArabicarXiv cs.AI / cs.LG / cs.CL
- 多标签一致性：灵活的标注者间信度框架Multi-Label Agreement: A Flexible Framework for Inter-Annotator ReliabilityarXiv cs.AI / cs.LG / cs.CL
- UCVE：用于合规验证的图结构检索UCVE: Graph Structured Retrieval for Compliance ValidationarXiv cs.AI / cs.LG / cs.CL
- 同样的结论，不同的读者：基于相对偏好的个性化事实核查解释Same Verdict, Different Readers: Personalizing Fact-Checking Explanations from Relative PreferencesarXiv cs.AI / cs.LG / cs.CL
- ETAMem：为长期对话智能体解耦并对齐异构记忆ETAMem: Decoupling and Aligning Heterogeneous Memory for Long-Term Conversational AgentsarXiv cs.AI / cs.LG / cs.CL
- 延续了什么？各代语言模型发布中的预训练信息披露What Carries Forward? Pretraining Disclosure Across Language Model ReleasesarXiv cs.AI / cs.LG / cs.CL
- 教现实而非伦理：即插即用的继续预训练语料让多个模型家族形成人类—生物圈—AI福祉耦合倾向Teach reality, not ethics: a drop-in continued-pretraining corpus installs coupled human-biosphere-AI welfare dispositions across model familiesarXiv cs.AI / cs.LG / cs.CL
- semanticfa：用于心理测量量表无作答语义因子分析的R包semanticfa: An R Package for Response-Free Semantic Factor Analysis of Psychometric ScalesarXiv cs.AI / cs.LG / cs.CL
- 元认知解码：扩散语言智能体的高效推理Metacognitive Decoding: Efficient Inference for Diffusion Language AgentsarXiv cs.AI / cs.LG / cs.CL
- 采用广义二阶路由的多状态线性注意力Multi-State Linear Attention with Generalized Second-Order RoutingarXiv cs.AI / cs.LG / cs.CL
- 大模型能通过自由探索发现隐藏的物理定律吗？Can LLMs Uncover Hidden Physical Laws through Free Exploration?arXiv cs.AI / cs.LG / cs.CL
- GEPA-Curator：自我改进智能体的筛选式学习GEPA-Curator: Curated Learning for Self Improving AgentsarXiv cs.AI / cs.LG / cs.CL
- RynnWorld-Latent：以通用潜在动作为条件的跨具身世界模型RynnWorld-Latent: A Cross-Embodiment World Model Conditioned on Universal Latent ActionsHugging Face Daily Papers
- 不放大的放大v2：从看清细节到关联证据，实现可泛化的多模态感知Zooming without Zooming v2: From Seeing Details to Connecting Evidence for Generalizable Multimodal PerceptionHugging Face Daily Papers
- 跨模态引导向量同时存在于图像与文本生成中Cross-modal Steering Vectors Exist Across Both Image and Text GenerationHugging Face Daily Papers
- 超越估计：维纳信息解码让像素空间流模型更易训练Beyond the Estimate: Wiener-Informed Decoding for Easier Training of Pixel-Space Flow ModelsHugging Face Daily Papers
- Hy-Nomas：迈向自由形式的交互式世界生成Hy-Nomas: Towards Free-form Interactive World GenerationHugging Face Daily Papers
- LociArt：自由风格局部视频风格化的统一范式LociArt: A Unified Paradigm for Free-Style Local Video StylizationHugging Face Daily Papers
- 离散token能否作为视觉信号语义无损的表示？Can Discrete Tokens Serve as Semantically Lossless Representations of Visual Signals?Hugging Face Daily Papers
- ThinkWorld：世界想象强化动力学链式推理ThinkWorld: World Imagination Reinforces Chain-of-Dynamics ReasoningHugging Face Daily Papers
- CodeCap-Bench：代码能否成为好的视频描述？CodeCap-Bench: Can Code Serve as a Good Video Caption?Hugging Face Daily Papers
- 回滚：世界模型中的经验回溯用于策略改进Rollback: Experience Backtracking in World Models for Policy ImprovementHugging Face Daily Papers
- Nemotron-Labs-Diffusion-Uni：带跨范式同策略蒸馏的弹性自精炼统一扩散大模型Nemotron-Labs-Diffusion-Uni: Elastic Self-Refining Unified Diffusion LLM with Cross-Paradigm On-Policy DistillationHugging Face Daily Papers
- PDRVLA：带物理动力学表示的视觉—语言—动作模型PDRVLA: A Vision-Language-Action Model with Physical Dynamics RepresentationsHugging Face Daily Papers
- 视觉—语言—动作还是世界—动作？只要动作模型Vision-Language-Action or World-Action? Just Action ModelsHugging Face Daily Papers
开源项目与版本更新 · 3 条
- NVIDIA开放Nemotron 3.5 Transcribe：一个模型做转写、翻译与说话人区分NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and AdaptHugging Face
- DeepSeek Harness v0.2.2-alpha.2：插件管理支持富文本详情与第三方插件更新，新增多语言界面DeepSeek Harness v0.2.2-alpha.2DeepSeek
- DeepSeek Harness v0.2.1-rc.4：修复ACP模式诊断输出混入stdoutDeepSeek Harness v0.2.1-rc.4DeepSeek