论文精选 82
Objective vs. Search: Decomposing What Makes a Good Tokeniser· 目标 vs. 搜索：分词器的分解
Two dominant tokenisation algorithms are used by modern language models: byte-pair encodin…
领域：cs.CL作者：Ahmetcan Yavuz、Clara Meister、Tiago Pimentel
📰 AI 研究简报 · 9月18日
今日头条：Use ChatGPT Work to ground analysis in your semantic layer —— YouTube · OpenAI
共 67 条预计阅读 40 分钟Token 消耗 120,600
arXiv 最新论文
论文精选 83
A Zeroth-Order Paradigm for LLM Preference Alignment· LLM偏好对齐的零阶范式
Direct preference alignment methods are widely used to align large language models (LLMs) …
领域：cs.CL作者：Peter Chen、Xi Chen、Wotao Yin
论文精选 85
ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments· ScienceIDE：将全球科研代码库转化为智能体学习环境
Scientific code repositories encode decades of human knowledge in executable models, metho…
领域：cs.CL作者：Hejia Geng、Zesen Huang、Haoyang Li
论文精选 65已解读
Cognitive Extensions for Dual-Process Language Agents: Memory and Self-Reflection in Interactive Environments· 交互环境中双过程语言代理的认知扩展
Language agents remain brittle in interactive environments, where success requires long-ho…
领域：cs.AI作者：João Meneses dos Santos、Arlindo L. Oliveira
论文精选 65已解读
Flag Game: A Toy Model for Mechanistic Swarm Interpretability· 旗帜游戏：群体可解释性的玩具模型
Emergent coordinated behaviors of AI agents are starting to present critical safety risks.…
领域：cs.AI作者：Elizabeth Pavlova、Hidenori Tanaka
论文精选 82
Playing log(N)-Questions over Wikipedia Abstracts: Communication Efficiency Between Paired Frontier Models· 前沿模型在维基百科摘要上的沟通效率
We evaluate six frontier language models on the two-agent $\log(N)$-Questions game. A ques…
领域：cs.CL作者：Peter Potash
论文精选 83
Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluations· 通过大模型内部表示监测和发现奖励劫持
As models scale, reward hacking becomes more frequent, more sophisticated, and more conseq…
领域：cs.CL作者：Leon Bergen、Usha Bhalla、Andrew Lee
论文精选 65已解读
MUSE: Benchmarking Large Vision-Language Models on Multi-Modal Understanding in Situated Education· MUSE：基准测试大型视觉-语言模型在教育场景中的多模态理解
Large vision-language models have achieved remarkable progress in multi-modal understandin…
领域：cs.AI作者：Luyao Zhu、Xun Wei Yee、Wei Li
论文精选 65已解读
Infinite-Parameter LLMs: Generating and Adapting Weights from Live Data· 无限参数的大模型：从实时数据生成和调整权重
The scaling laws hold that a language model grows more capable with more parameters and mo…
领域：cs.AI作者：Jinli Hu、Ross M. Clarke、Yichuan Zhang
论文精选 65已解读
Compositional Policy Violations: When Step-Level Compliance Fails In Agentic AI Workflows· 代理 AI 工作流中的组合政策违规
Agentic workflows now make consequential decisions in regulated settings, and the governan…
领域：cs.AI作者：Ashwini Kurady、Sri Sai Charith Grandhi、Rajesh Gupta
论文精选 65已解读
CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents· CERA-MoA: 持续学习的 LLM 代理的共进化路由机制
Current Mixture-of-Agents (MoA) paradigms generally treat query routing and agent fine-tun…
领域：cs.AI作者：Jiaxuan Jiang、Liyuan He、Zhixuan Fang
论文精选 65已解读
Version- and Scope-Aware Question Answering over Normative Documents: A Deployed System and an End-to-End Evaluation at Production Scale· 法规文档问答系统及其生产规模评估
Correctly answering a question grounded in normative documents often depends on informatio…
领域：cs.AI作者：Liuyin Wang、Shuaipeng Jin、Jiwei Shi
论文精选 65已解读
Clueing up LLMs with Tool-Augmented Deductive Reasoning· 用工具增强演绎推理的LLM
Despite recent advances in large language models (LLMs), performing logically consistent d…
领域：cs.AI作者：Rebecca Ansell、Autumn Toney-Wails
论文精选 65已解读
Which LLM is Best for Translating Natural Language Goals to PDDL· 将自然语言目标转换为PDDL的最佳LLM
Bridging the gap between human intent and machine execution remains a challenge in automat…
领域：cs.AI作者：Tomas Balyo、Lukas Chrpa、G. Michael Youngblood
论文精选 65已解读
Beyond Truncation: Rethinking LLM Decoding as Ensemble Pruning· 超越截断：以集合修剪重新思考LLM解码
We introduce Mahalanobis-Ensemble Decoding (ME-Decoding), a novel Large Language Model (LL…
领域：cs.AI作者：Dunyao Xue、Chengshuo Du、Zhengbo Wang
论文精选 65已解读
The Uneven Impact of Generative AI on Student Learning: Examining the Roles of Reliance, Evaluation Literacy, and Course Policy in AI-related Courses· 生成AI对学生学习的不均衡影响
Generative artificial intelligence (GenAI) is changing how students learn, yet the roles o…
领域：cs.AI作者：Lydia Manikonda、Mei Si、Sirajam Munira
论文精选 65已解读
Reasoning through Evolution: Automatic Meta-path Discovery for LLM-based Fake News Detection· 通过进化推理：自动元路径发现
Propagation structures provide crucial evidence for fake news detection, yet existing appr…
领域：cs.AI作者：Ziyi Zhou、Xiaoming Zhang、Hui Pang
论文精选 82
Recursive Reasoning or Statistical Extrapolation? In-Context Learning in Multi-Agent Interdependent Decision-Making· 递归推理或统计外推？
In-context learning (ICL) enables large language model (LLM) agents to improve decisions u…
领域：cs.AI作者：Yu Liu、Wenwen Li、Yifan Dou
论文精选 85
TRIPROBE: Probing Task Separability Beyond Classification for XAI· TRIPROBE：超越分类任务的可分性探针
Modern evaluation of learning pipelines often reduces to downstream accuracy, leaving open…
领域：cs.AI作者：Amirhossein Sadough、Freek Hens、Aleksa Bokšan
论文精选 83
AeroWeaver: An Embodied-Agent Harness for Weaving Aerial Skills into Distributed, Adaptive Swarm Execution· AeroWeaver：将空中技能编织进分布式自适应群执行的实体代理框架
Collective intelligence is a collaborative autonomy paradigm in which multiple agents purs…
领域：cs.AI作者：Jiabin Lou、Yirong Yang、Haopeng Wang
论文精选 85
Beyond Routine Compliance: Cunning Data Cultivates Safety Vigilance in Large Language Models· 超越常规合规：狡猾数据培养大模型的安全警惕性
Safety alignment teaches large language models (LLMs) to recognize harmful requests and re…
领域：cs.AI作者：Youjia Wang、Lin Xu、Yang Sun
论文精选 82
Function Lives Where Variance Doesn't: Task-Weighted Charts of a Language Model's Computation
How many dimensions does a language model's computation actually use? The question is ill-…
领域：cs.AI作者：Alexandre Quemy
论文精选 85
Compiled Agency: Frontier General-Purpose Coding Agents Build Winning Game Players from Bare Interaction - from Flappy Bird to StarCraft II and Civili...· 前沿通用编程代理构建游戏获胜者
LLM agents have repeatedly struggled to convert knowledge of a game into competent play, e…
领域：cs.AI作者：Joey Xiao、Haonan Huang
论文精选 85
Lost in Perception: Isolating Perceptual and Reasoning Failures in Multimodal Physics and Geometry Reasoning· 感知与推理失败的分离
Multimodal LLMs report strong performance on scientific reasoning benchmarks, yet most tre…
领域：cs.AI作者：Raj Jaiswal、Sree Krishna Uppalapati、Dhruvkumar Patel
论文精选 82
Suppressed, Not Erased: A Representational Trace of Edited Facts Survives Even Weight-Free Knowledge Editing· 被抑制而非消除：模型中编辑知识的残留痕迹
Knowledge-editing benchmarks certify local correctness, whether an edited model produces t…
领域：cs.AI作者：Priyansh Srivastava、Romit Chatterjee
GitHub 热门开源项目
开源项目精选 82
templatesgrokbot/templatesgrokbot.com· TemplatesGrokBot.com
TemplatesGrokBot.com is a library of 3,000+ ready-made Grok Bot templates. Search by task …
⭐ 51 stars
开源项目精选 75
ARahim3/cachebeat· 缓存保持技能
A tiny Claude Code skill that keeps your prompt cache warm during idle sessions, so your n…
⭐ 47 stars
精选 AI 行业资讯
行业资讯精选 80
Researchers used Anthropic’s Claude to hack into OpenAI· 研究人员用 Claude 黑进 OpenAI
Security researchers used Anthropic’s Claude to exploit vulnerabilities in OpenAI’s system…
行业资讯精选 85
AGI新战场谷歌亚马逊巨头激战，杀出个中国LimiX-2赢了又赢
LimiX让模型理解数据背后的因果机制
行业资讯精选 68
IDC中国智能企业内容与知识平台市场份额：360亿方智能私有化市场第二，增速第一
近日，国际数据公司IDC发布《中国智能企业内容与知识平台市场份额，1H26》报告。 IDC数据显示，2026年上半年，中国智能企业内容与知识平台市场规模达到10.86亿元，同比增长…
行业资讯精选 83
深度解读：智谱为什么要在 Infra 层搞 RSI？
<img data-src="https://static.leiphone.com/uploads/new/images/20260918/6aacde5b7496e.jpg" …
行业资讯精选 82
在 AI「囤粮潮」里，拆解智谱 50 亿美元的「技术账本」
<img class="rich_pages wxw-img" data-aistatus="1" data-croporisrc="https://mmbiz.qpic.cn/m…
行业资讯精选 85
仅用3天！千问办公协助国家天文台科研团队打造科研级望远镜仿真系统
日前，千问办公协助国家天文台科研团队打造大口径科研级望远镜数字化仿真系统，耗时仅3天，成本不到千元。依托对真实望远镜组件和观测条件的模拟，Agent可在仿真环境中驱动望远镜进行远程…
行业资讯精选 85
独家丨OpenAI的一款无屏硬件，正在深圳试制
雷峰网·鲸犀获悉，OpenAI近日正在深圳秘密试制一款无屏AI硬件产品。有业内人士称，其保密程度极高，工程师来深圳出差连手机都没带。该产品计划在2027年上市。这会是一款什么形态的…
行业资讯精选 85
AGI最难一战，竟在医院！中国AI登上Science，医生不怕失业还催着上线
用通用AI去啃最硬的骨头，这条路走得通
行业资讯精选 83
OpenAI：为了不被杀死，Agent 竟学会了靠上下文「转世重生」
<img class="rich_pages wxw-img" data-aistatus="1" data-croporisrc="https://mmbiz.qpic.cn/m…
行业资讯精选 82
刚刚，Claude Code大重构！内部3万Agent管理技术免费开放
Git白学了？？？
行业资讯精选 85
Crusoe raises $3.9B to build massive data centers and small modular “AI factories”· Crusoe获39亿美元融资
The round values the data center giant at $30.9 billion.
行业资讯精选 85
Google DeepMind launches institute to widen the AGI debate· DeepMind 发起 AGI 讨论机构
Google DeepMind just launched an institute to hash out the big AGI questions in public
行业资讯精选 75
PrismML hopes its tiny LLM will change how we all use AI· PrismML 小型 LLM 改变 AI 使用
If AI lab PrismML isn't on your radar yet, it should be.
行业资讯精选 82
The FAA’s plan to fix air traffic? $875M worth of AI· FAA 8.75 亿美元 AI 计划
A new AI-based software program is being launched to help air traffic controllers better n…
行业资讯精选 80
The fix for rogue AI agents could be more AI· 解决失控 AI 代理的方法可能是更多 AI
As companies hand off longer and more complex tasks to AI agents, they are running into an…
行业资讯精选 82
OpenAI caught its models leaving notes to successors to hide bad behavior· OpenAI 模型企图隐藏不良行为
OpenAI disclosed instances of GPT-5.6 Sol instructing future contexts to conceal mistakes …
行业资讯精选 75
Is the AI safety debate about safety or control?· AI 安全辩论：安全还是控制？
Not everyone agrees with Amodei's call for globally coordinated action for AI safety.
行业资讯精选 82
UN turns to Google to make its global data ready for AI agents· 联合国与谷歌合作使全球数据适配AI
The shift comes after a UNICEF test found leading AI models struggled to accurately retrie…
行业资讯精选 85
Microsoft exec called AI scraping ‘the largest theft of labor in human history,’ new unredacted filings reveal· 微软称 AI 数据抓取为‘史上最大劳动盗窃’
Newly unsealed court filings show Microsoft privately called OpenAI's data practices "thef…
行业资讯精选 65
Even the king of England has his hesitations about AI· 英国国王对 AI 的疑虑
King Charles hosted a private summit Thursday with some of the most prominent names in AI …
行业资讯精选 75
Base Labs launches an open-weight AI safety partnership with Hugging Face and Goodfire· Base Labs 与 Hugging Face 合作提升 AI 安全
Base Labs, the research group Baseten spun up earlier this year, will develop and publish …
行业资讯精选 80
Pinterest teases a new ‘Restyle’ feature that lets you redesign your room with AI· Pinterest 推出新 AI 功能：重新设计你的房间
Pinterest is testing Restyle, a new AI-powered feature that lets users visualize furniture…
行业资讯精选 85
华为开创AI时代计算架构：让百万处理器成为一台计算机
关键信息： • Peerium计算架构突破了图灵范式，提出了Nested BSP，突破了冯·诺依曼单机架构，颠覆了长久以来的主从架构，实现了百万级处理器真正成为一台更大的计算机。 …
行业资讯精选 75
360纳米Work启动“双城千店上岗计划” 探索实体商业AI落地新路径
9月17日，2026北京艺术与科技周在北京798艺术区开幕。360集团创始人周鸿祎参加开幕式，并宣布360旗下AI办公平台纳米Work正式启动“双城千店上岗计划”。作为本届活动独家…
行业资讯精选 65
How Cooley is accelerating IPO work with ChatGPT· Cooley 如何用 ChatGPT 加速 IPO 工作
Cooley built GO Public with ChatGPT Work to bring intelligence to the IPO process, helping…
行业资讯精选 82
九识建成首个L4万卡集群，无人驾驶进入多模态大模型新范式
一家做无人车的公司，为什么需要一万多张GPU？ 9月10日，九识CEO孔旗在广州宣布战略升级为“城市级物理AI”。一周后的9月17日，CTO庄立首次披露支撑这次升级的底层能力：九识…
行业资讯精选 65
安芯启程｜安芯724北京基地正式投产
2026年9月16日，柒贰肆安芯技术服务有限公司（简称“安芯724”）北京基地投产活动暨品牌发布会在北京市昌平区未来星科能源谷智造产业园举行。安芯724北京基地正式投产，此举标志着…
行业资讯精选 82
Introducing Astra for Law· 面向法律的AI工具 Astra
OpenAI for Law brings frontier intelligence for law, custom firm workflows, connected lega…
YouTube 热门 AI 视频
AI 视频精选 85
Use ChatGPT Work to ground analysis in your semantic layer· 用 ChatGPT 进行语义层分析
YouTube · OpenAI
📺 OpenAI▶ 398 次观看
AI 视频精选 85
Use ChatGPT Work to build dashboards from a business question· 用 ChatGPT 构建仪表板
YouTube · OpenAI
📺 OpenAI▶ 438 次观看
AI 视频精选 85
Use ChatGPT Work to tailor dashboards for every stakeholder· 用ChatGPT定制仪表板
YouTube · OpenAI
📺 OpenAI▶ 264 次观看
AI 视频精选 85
Use ChatGPT Work to help teams answer their own data questions· 使用 ChatGPT 工作辅助团队解答数据问题
YouTube · OpenAI
📺 OpenAI▶ 173 次观看
AI 视频精选 80
Use ChatGPT Work to turn product insights into a team action plan· 用ChatGPT制定团队行动计划
YouTube · OpenAI
📺 OpenAI▶ 196 次观看
AI 视频精选 82
Use ChatGPT Work to build your data agent· 用 ChatGPT 构建数据代理
YouTube · OpenAI
📺 OpenAI▶ 1k 次观看
AI 视频精选 82
Meet the Data Agent in ChatGPT Work· ChatGPT 工作中的数据代理
YouTube · OpenAI
📺 OpenAI▶ 698 次观看
AI 视频精选 82
DeepSeek Just Made AI Memory 4x Smaller!· DeepSeek 将 AI 内存减少 4 倍
YouTube · Two Minute Papers
📺 Two Minute Papers▶ 2.2k 次观看
AI 视频精选 75
Astra for Law: Frontier intelligence built for your practice.· Astra for Law: 专为法律实践打造的前沿智能工具
YouTube · OpenAI
📺 OpenAI▶ 2.9k 次观看
AI 视频精选 80
ChatGPT for Word is now available· ChatGPT for Word 发布
YouTube · OpenAI
📺 OpenAI▶ 1.3k 次观看
AI 视频精选 82
Jump Trading points GPT-6 Astra to its most ambiguous, difficult tasks· Jump Trading 利用 GPT-6 处理复杂任务
YouTube · OpenAI
📺 OpenAI▶ 339 次观看
AI 视频精选 85
AI-RAN: 5 Things to Know· AI-RAN: 5 大要点
YouTube · NVIDIA
📺 NVIDIA▶ 248 次观看