论文精选 85
STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization· 量化误差何时何地重要
Linear attention replaces growing KV caches with fixed-size recurrent states, yet these pe…
领域：cs.CL作者：Bingchen Yao、Haobo Xu、Haokun Lin
📰 AI 研究简报 · 10月1日
今日头条：Muse爆火，dots入场：Personal Agent开始和互联网平台抢入口 —— <section data-mpa-powered-by="yiban.io" data-pm-slice="0 0 []" style="-webkit-tap-highlight-color: transparent;margin: 0px 0px 24px;padding: 0px;outline: 0px;max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;color: rgba(0, 0, 0, 0.9);font-family: "PingFang SC", system-ui, -apple-system, BlinkMacSystemFont, "Helvetica Neue", "Hiragino Sans GB", "Microsoft YaHei
共 62 条预计阅读 37 分钟Token 消耗 111,600
arXiv 最新论文
论文精选 83
Pretraining Latent Information Feedback Transformers with Teacher Supervision· 预训练带有教师监督的潜在信息反馈Transformer
Transformer language models (LMs) are feed-forward: deep-layer representations are never f…
领域：cs.CL作者：Dor Tirosh、Ido Amos、Mor Geva
论文精选 65已解读
Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning· 思考前的思考：通过元推理扩展智能体推理
As agents take on longer and more complex problems, controlling the execution becomes a ta…
领域：cs.AI作者：Paras Dahal、Anton Bakhtin、Taco Cohen
论文精选 65已解读
Learning Meta-Skills for Agent Harness Design in Test-Time AI4AI· 测试时AI-for-AI的元技能学习
Agent performance depends on both reasoning ability and the environment in which it acts. …
领域：cs.AI作者：Cheng Qian、Kunlun Zhu、Beibin Li
论文精选 65已解读
AdviSD: Learning to Advise Frontier LLMs via Targeted Multi-Turn Self-Distillation· AdviSD：通过定向多轮自我蒸馏学习为前沿大模型提供指导
A small trainable advisor can steer a frozen language-model executor using natural-languag…
领域：cs.AI作者：Rishabh Agrawal、Hejie Cui、Shasha Li
论文精选 85
LongHarness Bench: Stress-Testing Language Model Harnesses for Long-Context Reasoning· 长Harness基准：压力测试语言模型长上下文推理
Language-model (LM) harnesses enable LMs to operate effectively over long contexts using a…
领域：cs.CL作者：Quang Hieu Pham、Thuy Duong Nguyen、Jocelyn Qiaochu Chen
论文精选 65已解读
Stochastic World Models for Verifying Vision-Based Neural Feedback Systems· 验证基于视觉的神经反馈系统的随机世界模型
Verifying a vision-based neural feedback system requires a model of the observations its c…
领域：cs.AI作者：I. Samuel Akinwande、Mykel J. Kochenderfer、Clark Barrett
论文精选 82
From Routing Signals to Selective Review: Visual regrounding in MoE VLMs· 从信号路由到选择性审查：MoE VLM 中的视觉再定位
Vision-language models (VLMs) may accept false visual premises, answering questions about …
领域：cs.CL作者：Hongzhu Guo、Mohsen Fayyaz、Nanyun Peng
论文精选 82
How Local Mixing Encodes Relative Position in Global NoPE Attention· 局部混合如何在全局 NoPE 注意力中编码相对位置
The attention operation is naively position invariant. However, positional information is …
领域：cs.CL作者：Cutter Dawes、Nick Alonso、Tom Figliolia
论文精选 65已解读
Do LLM Agents Execute the Plans They Declare? From Planning-Mode Declaration to Pattern-Specific Execution· LLM 代理执行声明的计划吗？
Large language models (LLMs) enable agents to solve long-horizon tasks by generating a pla…
领域：cs.AI作者：Subba Reddy Oota、Francisco Herrera、Jordi Cabot Sagrera
论文精选 65已解读
NeuronEye: Query-Guided Visual Concept Activation for Vision-Language Reasoning· NeuronEye: VLM中视觉概念激活
Current vision-language models (VLMs) encode visual information in dense hidden states whe…
领域：cs.AI作者：Ruiyu Yan、Bowen Chen、Shaowen Wan
论文精选 65已解读
Character Training for Risk-Averse Agents· 为风险厌恶代理人进行性格训练
Risk aversion in resources could prevent misaligned AI agents from causing catastrophic ha…
领域：cs.AI作者：Arav Dhoot、Punya Syon Pandey、Jamie Johnson
论文精选 65已解读
Probability is Not Enough: Exploring and Counting Divergent Tokens for Reasoning Uncertainty Quantification in LLMs· 概率不足：探索和计数分歧标记
As the chain-of-thought reasoning capabilities of large language models improve, evaluatin…
领域：cs.AI作者：Feiyang Li、Shengjing Liu、Qi Zhan
论文精选 65已解读
UserProxyBench: Evaluating LLM User Simulators for Agent Benchmarks and Training· 用户代理基准：评估LLM用户模拟器
Interactive agent benchmarks and multi-turn reinforcement learning increasingly place a se…
领域：cs.AI作者：Ashish Jain、Armaan Sandhu
论文精选 65已解读
Retrieval-Augmented Skill Optimization via Cross-Harness Adaptation· 通过跨框架适应增强检索技能优化
An agent skill is a reusable, actionable natural-language artifact that guides an agent to…
领域：cs.AI作者：Jaewon Chu、Ji Soo Lee、Jihwan Park
论文精选 65已解读
PE-EK-PINN: Physics Embedding with Evolving Kernel for Scalable Physics-Informed Neural Networks· PE-EK-PINN：用于可扩展物理信息神经网络的物理嵌入与演化内核
Physics-Informed Neural Networks (PINNs) embed governing equations into deep learning, but…
领域：cs.AI作者：Huiwen Zhang、Feng Ye、Chu Ma
论文精选 65已解读
Brain-SAD: A Brain-Inspired Safe Autonomous Driving Control Framework with Dynamic Fear-Oriented Constraint on Dual-Policy· Brain-SAD：受大脑启发的自动驾驶安全控制框架
Constrained Reinforcement Learning has recently gained increasing attention in the field o…
领域：cs.AI作者：Huan Rong、Chao Yin、Anouar Imel
论文精选 82
HARISSA: Inference-Time Self-Checks for Efficient and Safe Local Language Model Deployment· HARISSA: 本地语言模型的高效安全推理部署
Running a language model locally offers advantages in privacy, latency, and cost, but loca…
领域：cs.AI作者：Kenan Alkiek、Moontae Lee、David Jurgens
论文精选 82
Diagnosing and Improving Probabilistic Reasoning in Large Language Models· 大语言模型的概率推理诊断与改进
Large language models (LLMs) are increasingly proposed as decision assistants who must rea…
领域：cs.AI作者：Huaman Sun、Dingcheng Wang、Jason Hartline
论文精选 83
Which Attention Heads are like the Human Head? Not the Ones that Compute· 哪些注意力头像人类大脑？不是那些计算的
Brain-AI alignment is often interpreted as a sign that model and brain perform similar com…
领域：cs.AI作者：Christopher Pinier、Gustaw Opiełka、Hannes Rosenbusch
论文精选 82
KV-Kaizen: Learning Context-Adaptive Cache Compression Choices· KV-Kaizen：学习上下文自适应缓存压缩策略
As the context size of text processed with an LLM grows, the size of KV caches can outstri…
领域：cs.AI作者：Joao Monteiro、Louis Béthune、Anastasiia Filippova
论文精选 83
SelfSearch: Reward-Free Search for Self-Improving Agents· SelfSearch: 无需奖励的自我改进搜索
Advances in the coding capabilities of LLM agents allow them to inspect and modify their o…
领域：cs.AI作者：Jungwoo Yang、In Jin Kong、Yohan Jo
论文精选 82
BrainNet Studio: A Unified Toolkit for Brain Network Construction, Intelligent Analysis, and Visualization· BrainNet Studio: 统一的脑网络工具包
Brain networks characterize structural and functional relationships among brain regions an…
领域：cs.AI作者：Xiwei Zeng、Shengrong Li、Yiheng Liu
论文精选 82
Topological Coherence for Self-evolving Multi-agent Systems· 多智能体系统的拓扑一致性
Complex tasks inherently couple workflow structure, agent responsibility, collaboration, a…
领域：cs.AI作者：Sen Zhao、Ruiqi Kong、Zuyu Zhang
论文精选 83
Video-RSI: Recursive Self-Improvement of Video Understanding Agents via Harness Evolution· 视频理解代理的递归自我改进
Video understanding agents acquire evidence through an executable harness that controls wh…
领域：cs.AI作者：Bingjun Luo、Jialin Guo、Siqi Li
GitHub 热门开源项目
本板块今日暂无内容
精选 AI 行业资讯
行业资讯精选 80
Photon held a funeral for mobile apps. Now it has $4.5M to help replace them with agents· Photon 获投 450 万美元用 AI 代理取代移动应用
The startup helps developers build AI agents that work over iMessage, SMS/RCS, email, and …
行业资讯精选 80
Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment· 英伟达 AI 工厂最大化投资回报
AI factories are built by the megawatt, even by the gigawatt. Each megawatt factory costs …
行业资讯精选 75
Hearing tech startup Legato launches its AI hearing glasses· Legato推出AI听力眼镜
The glasses stem from the startup’s goal of making hearing care more accessible by address…
行业资讯精选 75
Satlyt, founded by a former Google and SpaceX product manager, raises $8M to run AI on satellites· Satlyt获800万美元融资，欲成为轨道计算的‘Android’
Satlyt wants to be the Android of orbital computing, offering open software that works on …
行业资讯精选 82
Google releases Gemini 4 Argon, called its most powerful model yet· 谷歌发布 Gemini 4 Argon 模型
Google has released its latest Gemini model, marketing it as a workhorse for coding and cy…
行业资讯精选 80
Valor, Atreides, and Sequoia back AI startup Flow Engineering at $750M valuation· Valor, Atreides 和 Sequoia 投资 AI 创业公司 Flow Engineering
Flow Engineering, which is bringing AI agents to hardware design, also landed Roelof Botha…
行业资讯精选 80
OpenAI’s Jev clone could help the frontier lab stop its swarming agents· OpenAI 的 Jev 克隆工具
OpenAI's "Decisions API" is a Jev clone that confirms the importance of fast, cheap intell…
行业资讯精选 65
AI voice startup ElevenLabs doubles valuation to $22B· AI语音初创公司估值翻倍
The $300 million employee tender was co-led by Wellington and T. Rowe Price.
行业资讯精选 83
The ugly economics of consumer AI· 消费者AI的丑陋经济学
There’s a reason frontier labs have gotten gunshy about consumer AI — and it’s not because…
行业资讯精选 80
NVIDIA Opens Applications for 2027–2028 Graduate Fellowships With Awards Up to $60,000· NVIDIA 开放 2027-2028 年研究生奖学金申请
Bringing together the world’s brightest minds and the latest accelerated computing technol…
行业资讯精选 75
Meta disputes claim that Muse read a user’s private messages without permission
Meta says its Muse AI agent cannot access a user’s Messages without explicit permission, d…
行业资讯精选 65
DoorDash launches an AI agent you can text to order food· DoorDash 推出 AI 点餐助手
By launching an AI agent for food ordering, DoorDash is looking to gain an edge over rival…
行业资讯精选 80
Destro AI’s secret sauce is getting robots and humans on the same page· Destro AI 的秘诀：使机器人和人类在同一页面上
"One of the biggest reasons we are winning against robotics companies is because we are no…
行业资讯精选 82
From Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AI· NVIDIA和CoreWeave优化代理AI的全流程
Building on nearly a decade of co-engineering, CoreWeave has built NVIDIA compute, network…
行业资讯精选 82
Cerebras Systems’ Andrew Feldman on whether AI can keep scaling at TechCrunch Disrupt 2026· Cerebras Systems CEO 在 TechCrunch Disrupt 2026 探讨 AI 扩展
At TechCrunch Disrupt 2026, Cerebras Systems CEO and co-founder Andrew Feldman will explor…
行业资讯精选 75
Restate lands $20M as the need for durable infrastructure increases with AI agents· Restate获2000万美元融资
Founded by Apache Flink veterans, the startup is taking on workflow heavyweight Temporal.
行业资讯精选 82
OpenAI推理之父最新访谈！数学只是多智能体时代的开胃菜· 数学与多智能体时代的 Interview
千禧年难题的突破，10000个Agent最多占了10%的功劳。
行业资讯精选 82
AI办公进入「上下文战争」，百度如何出牌？
上下文，正成为AI办公新的竞争变量。 “任何事情都可以成为用户的上下文。”在百度AI Day开放日接受雷峰网采访时，百度集团副总裁、个人超级智能事业群组（PSIG）总裁王颖表示， …
行业资讯精选 82
DeepSeek 开源算子工具大礼包，联手华为昇腾，手撕 CUDA 绑定！
<img data-src="https://static.leiphone.com/uploads/new/images/20260930/6abceaf0bcb61.jpg" …
行业资讯精选 82
Disrupting a coordinated model-distillation campaign· 打击协调的模型蒸馏活动
Learn how OpenAI disrupted a campaign to extract protected model reasoning and is strength…
行业资讯精选 83
Helping small businesses put AI to work· 帮助小企业应用 AI
OpenAI is partnering with America’s SBDC to expand hands-on AI training and local support …
行业资讯精选 85
Muse爆火，dots入场：Personal Agent开始和互联网平台抢入口
<section data-mpa-powered-by="yiban.io" data-pm-slice="0 0 []" style="-webkit-tap-highligh…
行业资讯精选 83
实测 Qwen-3.8 与 GLM-5.3 的 Flash 版：谁能让我提前下班半小时？
<section data-mpa-powered-by="yiban.io" data-pm-slice="0 0 []" style="-webkit-tap-highligh…
行业资讯精选 80
从 Codex Harness 开源，看 AI 公司的护城河是什么？
<section data-mpa-powered-by="yiban.io" data-pm-slice="0 0 []" style="-webkit-tap-highligh…
行业资讯精选 82
全网最硬核 OpenClaw 2.0 实测：从 ToC 到 ToB 的试探，Agent 网关能重塑工作流边界吗？
<section data-mpa-powered-by="yiban.io" data-pm-slice="0 0 []" style="-webkit-tap-highligh…
行业资讯精选 82
5500颗光引擎如何撑起4096卡超节点？华为NPO的故事还没讲完
<section data-mpa-powered-by="yiban.io" data-pm-slice="0 0 []" style="-webkit-tap-highligh…
行业资讯精选 80
对比 Codex，免费的 AgnesCode 也能在底层算子优化任务中打得有来有回
<img data-src="https://static.leiphone.com/uploads/new/images/20260930/6abcc3d98887c.jpg" …
行业资讯精选 75
从海拉鲁到现实世界：视频模型如何理解「变化」
<img data-src="https://static.leiphone.com/uploads/new/images/20260930/6abcc0f00d04e.jpg" …
行业资讯精选 75
在三张圆明园鼠首照片里，我们看到了3D AI的Harness时刻
<img data-src="https://static.leiphone.com/uploads/new/images/20260930/6abcc09a2b245.jpg" …
行业资讯精选 82
越大越强，不再是 LLM 的专利：PixVerse R2 攻克实时世界模型 Scaling 难题
<img data-src="https://static.leiphone.com/uploads/new/images/20260930/6abcbfa91d08e.jpg" …
行业资讯精选 85
GPT-6 Astra 上手体验：花了一周，它真在电脑里建了座曼哈顿！
<img data-src="https://static.leiphone.com/uploads/new/images/20260930/6abcbe4407c5c.jpg" …
行业资讯精选 80
Kimi 和 DeepSeek，为什么出现在同一张模型架构图里？
<img data-src="https://static.leiphone.com/uploads/new/images/20260930/6abcbc99d176a.jpg" …
行业资讯精选 82
我们让 GPT‑6 Astra、Fable 5.1 和 Sol 控制同一台机器人：谁真的把活干完了？
<img class="rich_pages wxw-img" data-aistatus="1" data-croporisrc="https://mmbiz.qpic.cn/s…
行业资讯精选 65
拆解 WorkBuddy、千问办公和豆包工作，它们其实不是同一种产品
<img data-src="https://static.leiphone.com/uploads/new/images/20260930/6abcbbcd11533.jpg" …
行业资讯精选 62
实测 Seedance 2.5 vs 可灵 3.0，谁能让我们无痛当上视频导演？· Seedance 2.5 vs 可灵 3.0 视频生成能力
<img class="rich_pages wxw-img" data-aistatus="1" data-croporisrc="https://mmbiz.qpic.cn/m…
YouTube 热门 AI 视频
AI 视频精选 82
What does “fungible” mean for AI infrastructure?· AI 基础设施的可互换性
YouTube · NVIDIA
📺 NVIDIA▶ 399 次观看
AI 视频精选 80
Meet The Startup: Probabl: by the Creators of Scikit-Learn· Probabl：Scikit-Learn 创始团队的新创公司
YouTube · NVIDIA
📺 NVIDIA▶ 168 次观看