论文精选 82
SocietyBench: Forecasting Counterfactual Social-World Evolution· SocietyBench：反事实社会世界演化预测
Large language models (LLMs), and the agents built on top of them, are now benchmarked hea…
领域：cs.CL作者：Zhenran Wang、Zhonghan Bian、Jinsong Li
📰 AI 研究简报 · 8月6日
今日头条：Exclusive: Mirendil inks $100M+ Google Cloud deal to scale self-improving AI —— Mirendil has signed a $100 million-plus Google Cloud partnership to expand its compute infrastructure, powering research into self-improving AI systems designed to accelerate scientific discovery and AI development.
共 59 条预计阅读 35 分钟Token 消耗 106,200
arXiv 最新论文
论文精选 83
WorldCup Arena: Prospective, Leakage-Free Evaluation of Frontier LLMs on a Live Tournament· 世界杯竞技场：实时无泄漏评估前沿大模型
Benchmarks that measure the forecasting ability of large language models are almost always…
领域：cs.CL作者：Zhenran Wang、Zhonghan Bian、Jinsong Li
论文精选 82
TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated Reasoning· TurnSight: 回合级事后自我蒸馏的工具集成推理
Tool-Integrated Reasoning (TIR) enables LLMs to solve complex tasks through iterative tool…
领域：cs.CL作者：Changle Qu、Sunhao Dai、Hengyi Cai
论文精选 85
PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents· PAST-Bench: 个人代理递归自我改进基准测试
Recursive self-improvement requires agents to turn accumulated experience into better futu…
领域：cs.CL作者：Shuhan Xue、Zixin Ding、Yichen Shen
论文精选 83
When Attention Goes Blind: Numerical Failure in ALiBi Positional Encodings· 当注意力失明：ALiBi位置编码的数值失败
We identify a previously overlooked failure mode of ALiBi positional encoding: its linear …
领域：cs.CL作者：Christopher Schröder、Lukas Gienapp、Ferdinand Schlatt
论文精选 65已解读
ReflectRL: Learning from Golden Negative Trajectories via Reflective-to-Direct Reasoning· ReflectRL: 从黄金负轨迹学习
On-policy training has emerged as a powerful post-training paradigm for improving the reas…
领域：cs.AI作者：Jinhe Bi、Chennan Zhou、Zengjie Jin
论文精选 65已解读
Should We Type or Talk to LLM Agents? A Comprehensive Study of Voice and Keyboard Input Perturbations· 语音与键盘输入对大模型的影响
Human input reaches language models by typing or speaking, and each channel leaves a disti…
领域：cs.AI作者：Zizhao Hu、Nathan Elijah Segura、Mohammad Rostami
论文精选 65已解读
Interpretable Adaptive Sampling for LLM Test-Time Scaling· LLM测试时自适应缩放
Test-time scaling improves LLM reasoning by generating and aggregating multiple candidate …
领域：cs.AI作者：Mobina Kashaniyan、Ali Jannesari
论文精选 65已解读
A game theory for foundation models shows new paths to rational cooperation through similarity inference· 大模型代理的游戏理论
As autonomous agents powered by foundation models are increasingly integrated into social …
领域：cs.AI作者：Alexander Meulemans、Maciej Wołczyk、Marissa A. Weis
论文精选 65已解读
TACT: Taxonomy-Aligned Post-Training for Pedagogically Adaptive English Tutoring· TACT：适应性英语辅导框架
Large language models (LLMs) are increasingly used to provide conversational practice for …
领域：cs.AI作者：Dongjie Yang、Siyan Lin、Leixian Shen
论文精选 65已解读
The Transformer Revolution, Part 1: Dynamic Processing through Output- Weight Interconnections· Transformer革命：输出权重互联的动态处理
This paper offers a new interpretation of the Transformer during inference. Against the "s…
领域：cs.AI作者：Marco Giunti、Fabrizia Giulia Garavaglia
论文精选 65已解读
Implementing Causal Perception: Competing SCMs and Situated Fairness· 实现因果感知：竞争的SCM和定位的公平性
Causal perception occurs when agents with competing Structural Causal Models (SCMs) of the…
领域：cs.AI作者：Jose M. Álvarez
论文精选 65已解读
Socially Grounded Agentic AI: Coordinating Plural Perspectives through Social Theory· 基于社会的代理AI：通过社会理论协调多元视角
As AI systems are deployed across increasingly diverse social contexts, alignment can no l…
领域：cs.AI作者：Matt Ratto、Abhishek Moturu、Daniel Silver
论文精选 65已解读
When Efficiency Becomes Fragility: Exploiting Dynamic Routing Vulnerabilities in Adaptive UAV Tracking· 当效率成为脆弱性：自适应 UAV 跟踪中的动态路由漏洞
Resource constraints on UAV platforms have driven a paradigm shift in aerial tracking, fro…
领域：cs.AI作者：Shaofeng Liang、Runwei Guan、Wenshuo Chen
论文精选 65已解读
Intertemporal Preference Steering in Qwen3 via Contrastive Activation Addition· Qwen3中的时间偏好调整
We study linear representations of temporal horizon in the large language model Qwen3-32B …
领域：cs.AI作者：Michal Mráz、Justin Shenk
论文精选 65已解读
ContinualSkillBench: Can LLM Agents Truly Evolve Their Capabilities?· LLM代理能否真正进化其能力
Modern agent frameworks equip large language models with external skill libraries to solve…
领域：cs.AI作者：Tianyi Guan、Yiding Wang、Haotong Yang
论文精选 65已解读
ADMITBench: A Safety-Governed Reference Framework for Evaluating the Admissibility of Industrial LLM Advisories· ADMITBench：工业 LLM 咨询的安全评估框架
This white paper presents ADMITBench, a reference framework for evaluating industrial LLM …
领域：cs.AI作者：Yash Misra、Javal Vyas、Siddharth Gutta
论文精选 82
MAFIA: Query-Only Memory Attacks via Probing and Factual Injection against Audited LLM Agents· MAFIA：通过探测和事实注入对抗审核的查询式内存攻击
Memory-augmented LLM agents rely on rich context for long-horizon reasoning and acting, ye…
领域：cs.AI作者：Jiaming Chen、Yisen Gao、Yanping Li
论文精选 83
Oilbird: Training-Free Speculative Decoding with Keys the Verifier Already Computes· Oilbird：利用验证者已计算键的免训练推测解码
Training-free speculative decoding drafts by matching an exact suffix of the context again…
领域：cs.AI作者：Tao Jin、Phuong Minh Nguyen、Zhenzhu Yan
论文精选 82
LatentGuard: Efficient and Inspectable Latent Reasoning for LLM Safeguards· LatentGuard：高效可检潜变量推理框架
Reasoning-based guard models improve LLM safeguards, but decoding explicit rationales for …
领域：cs.AI作者：Zhinan Liu、Jie Li、Mingyu Kang
论文精选 82
Does Forgetting Transfer Across Modalities? A Real-World Benchmark for Cross-Modal Knowledge Unlearning Evaluation· 跨模态知识遗忘的现实世界基准
Vision-Language Models (VLMs), like Large Language Models (LLMs), may memorize sensitive, …
领域：cs.AI作者：Chunlin Liu、Junnian Chen、Haitong Jiang
论文精选 82
KnowHal: A Knowledge-Driven Benchmark for Comprehensive Multimodal Hallucination Evaluation· 多模态幻觉评估基准 KnowHal
Hallucination remains a critical challenge for developing trustworthy Multimodal Large Lan…
领域：cs.AI作者：Ruihan Li、Jiyang Tan、Kailin Jiang
论文精选 83
Computing Actual Causes for Neural Network Predictions under Structured Causal Inputs· 结构因果输入下的神经网络预测因果计算
Explaining the predictions of neural networks is a central challenge in trustworthy AI. Ex…
领域：cs.AI作者：Jannick Strobel、Muqsit Azeem、Stefan Leue
论文精选 85
GDPevo: Evaluating Agent Self-Evolution on Real Business Tasks· GDPevo: 评估代理自我演化
Agent self-evolution updates an agent's persistent state from prior experience and reuses …
领域：cs.AI作者：Leijun Zhou、Zhihao Liu、Xiang Qu
论文精选 82
Risky Business: Measuring The Faithfulness-Safety Tension· 风险业务：测量忠实性-安全性张力
Chain-of-Thought (CoT) reasoning offers a promising window into model monitoring. However,…
领域：cs.AI作者：Dominik Meier、Luca Joshua Francis、Marco Bernhard Kaiser
GitHub 热门开源项目
开源项目精选 82
waiterve/wai-play· WAI Play
WAI Play - AI web game testing and quality evaluation platform
⭐ 70 stars语言 Python
开源项目精选 80
HELPMEEADICE/TE-Speed-MiniMaxH3-OSS
MiniMax-H3超级缓存加速插件
⭐ 113 stars语言 Python
开源项目精选 85
foodpm/timecut
跑在 NAS Docker 上的智能监控录像系统：FFmpeg 循环录像 + go2rtc 实时画面，大模型自动筛选一天回顾精华并生成日记，Web 面板一键管理。
⭐ 79 stars语言 Python
精选 AI 行业资讯
行业资讯精选 82
Into the Omniverse: How Open World Models Push the Frontier of Physical AI· 进入 Omniverse：开放世界模型如何推动物理 AI 前沿
In July, NVIDIA joined more than 200 companies and organizations in signing “Open Weights …
行业资讯精选 82
Ex-Spotify employees raise $10M to bring the AI behind its recommendations to e-commerce· 前 Spotify 员工为电商带来 AI 推荐引擎
The startup's platform predicts what product a shopper wants next, learn their general tas…
行业资讯精选 85
Exclusive: Mirendil inks $100M+ Google Cloud deal to scale self-improving AI· Mirendil 与 Google Cloud 达成合作推进自改进 AI
Mirendil has signed a $100 million-plus Google Cloud partnership to expand its compute inf…
行业资讯精选 85
Artificial Analysis榜单：阿里Qwen3.8Agentic能力得分全球第一· 阿里Qwen3.8领先全球
阿里Qwen3.8在国际榜单上领先
行业资讯精选 82
金刚GC3芯片重新定义视频AI算力规则，国产RISC-V高端算力商业化提速
近日，第三届中国计算机学会芯片大会（CCF Chip 2026）在太湖之滨无锡成功举办。 会场中，两院院士与顶尖工程师们的讨论话题直指底层架构的“根技术”创新。一个核心追问被反复推…
行业资讯精选 82
蚂蚁集团 IJCAI 2026 论文盘点：让 AI 学会「随机应变」
<img class="rich_pages wxw-img" data-aistatus="1" data-croporisrc="https://mmbiz.qpic.cn/m…
行业资讯精选 80
对话 IDEA 张磊：「不以动作为输入条件，就不叫世界模型」
<img class="rich_pages wxw-img" data-aistatus="1" data-croporisrc="https://mmbiz.qpic.cn/s…
行业资讯精选 82
华为 IJCAI 2026 论文盘点：从「规模密度」转向「设计密度」
<img class="rich_pages wxw-img" data-aistatus="1" data-croporisrc="https://mmbiz.qpic.cn/m…
行业资讯精选 65
这人谁啊？哈萨比斯都让位了· AI领域人物地位变化
没想到还有LeCun的事儿
行业资讯精选 82
独家丨联手千问AI大牛 ，觅光联创郦轲杀入女性AI健康硬件江湖
<img data-src="https://static.leiphone.com/6a6c4981e5bcb.jpg" alt="图片" class="rich_pages w…
行业资讯精选 82
把512 GiB闪存搬到xPU旁边，HBF能打破推理内存墙？
<section data-mpa-powered-by="yiban.io" data-pm-slice="0 0 []" style="-webkit-tap-highligh…
行业资讯精选 82
MiniMax H3登顶开源社区第一，定义视频模型领域“斩杀线”
MiniMax H3在视频模型领域取得显著进展。
行业资讯精选 85
超级算力枢纽远景乌兰察布星河基地投产，全球最大AI算力超级单体落地
全球最大AI算力超级单体在乌兰察布投产。
行业资讯精选 65
“千问办公”上架鸿蒙电脑应用市场，全面支持三大主流操作系统
8月6日，阿里巴巴旗下企业级Agent产品“千问办公”推出鸿蒙电脑版，用户可在华为应用市场下载体验。目前，“千问办公”已全面支持Windows、macOS和HarmonyOS三大主…
行业资讯精选 82
国内首个，千问办公通过中国信通院办公智能体能力评估
8月6日消息，中国信通院公布“互联网智能体治理·办公智能体能力评估”结果，“千问办公”成为国内首个通过评估的办公智能体（Agent）产品，意味着其在办公任务执行、企业数据安全、风险…
行业资讯精选 85
突发，Jeff Dean离职创业！谷歌股价应声蒸发1.34万亿
与此同时，哈萨比斯卸任谷歌DeepMind CEO
行业资讯精选 82
Meta launches Muse Code, an AI agent for large code bases· Meta 推出 Muse Code，AI 代码助手
Meta expanded its AI coding offerings with a new agent that, it promises, can handle compl…
行业资讯精选 65
Klaviyo acquires Elias Torres’ Agency in full-circle reunion for tech founders· Klaviyo 全面收购 Elias Torres 的代理机构
The serial entrepreneur joins the e-commerce company as CPO to lead its AI agents.
行业资讯精选 85
Jeff Dean and other top AI researchers are leaving Google to launch their own startup· 谷歌顶级AI研究人员离职创业
The legendary Google executive is joined by other outgoing Google execs in a joint mission…
行业资讯精选 82
Shopify says AI search is driving more traffic and sales, not replacing Google· Shopify 称 AI 搜索带动流量和销售增长
Shopify says AI isn’t cannibalizing search traffic the way it has for publishers. Instead,…
行业资讯精选 65
Hark previews its browser use agent for completing tasks· Hark 浏览器代理完成任务的预览
Hark claims that its browser use agent is faster and cheaper than competition.
行业资讯精选 80
TechCrunch Disrupt 2026’s Real World AI Stage features robots, automated factories, and extinct animals· TechCrunch 2026 实际应用人工智能舞台
On our new Real World AI stage, we’ll be focusing on the intersection between the digital …
行业资讯精选 82
Anthropic is hiring an AI chip design team· Anthropic招聘AI芯片设计团队
Anthropic is building a team for designing its own custom AI chips. The Claude maker said …
行业资讯精选 80
别再吹AI生图了，不能图层编辑的AI都是“画饼”
兔展智能发布RabbitVis，让AI真正走完设计全流程
行业资讯精选 62
等等，MiniMax H3不是刚发布吗？怎么就卷到几分钱的价格了……
MiniMax H3也能开箱即用了
YouTube 热门 AI 视频
AI 视频精选 65
How do you explain world models to your sibling? 🤖· 世界模型通俗解释
YouTube · NVIDIA
📺 NVIDIA▶ 562 次观看
AI 视频精选 80
What is RTX Spark?· RTX Spark 是什么？
YouTube · NVIDIA
📺 NVIDIA▶ 33 次观看
AI 视频精选 80
Voice Agent observability with LangSmith 🌟· 用LangSmith提升语音代理可观测性
YouTube · Google for Developers
📺 Google for Developers▶ 351 次观看
AI 视频精选 85
3rd Place Winner: Voice AI Prevents Data Loss. Coding Agent Calls Developer Before Deleting Records· 语音AI预防数据丢失
YouTube · DeepLearningAI
📺 DeepLearningAI▶ 91 次观看
AI 视频精选 75
2nd Place Winner: Coding Agent Calls Developer to Pitch Launch Strategy· 编程代理赢得比赛
YouTube · DeepLearningAI
📺 DeepLearningAI▶ 60 次观看
AI 视频精选 75
1st Place Winner: Coding Agent Calls Developer to Resolve Code Block· AI 编码代理解决代码问题
YouTube · DeepLearningAI
📺 DeepLearningAI▶ 150 次观看