‘Huge Breakthrough’ in the Math of Imbalance
（翻译）不平衡数学中的“巨大突破”
For the first time in 30 years, computer scientists have found a better way to allocate objects evenly between two groups.
烧掉500块实测H3、Seedance 2.0 fast，意外挖出了性价比之王（附全部提示词）
H3与Seedance 2.0 fast在同一提示词、分辨率和平台下完成多场景视频横评，比较真人表演、短剧、动效、广告和科幻等场景的指令遵循、镜头连贯与画面质感。实测Seedance 2.0 fast整体更稳，H3在视觉动效和广告场景各有优势。
李飞飞押注的空间智能：生成的世界，如何「经得起折腾」？
作者丨 刘紫东 编辑丨 幸丽娟 &nbs
李泽湘投过的商业园林机器人完成数千万融资，瞄准海外绿地智能运维｜硬氪首发
作者 | 黄楠 编辑 | 袁斯来 硬氪获悉，商业园林机器人公司食铁兽科技（PANDAG）近日完成数千万元A轮融资，本轮由云时资本独家投资，星⾠资本担任财务顾问。资金将重点投入欧美澳市场交付体系搭建及全球渠道布局。此前，公司已获得李泽湘教授、奇绩创坛、险峰⻓⻘等机构投资。
IJCAI 2026 四项论文奖、三项个人成就奖，回答 AI 的「来路」与「去路」
作者丨 幸丽娟 编辑丨岑 峰  
Introducing Gemini 3.7 Flash
（翻译）介绍 Gemini 3.7 Flash
Gemini 3.7 Flash is our most intelligent workhorse model yet for coding and agents.
从汽车到机器人、太空，五一视界复制十年仿真经验
近几年，自动驾驶技术路线加速变化，产业链上的公司也在重新寻找自己的位置。 其中，也包括长期为智驾行业提供仿真测试服务的五一视界。这家成立于2015年的数字孪生公司，通过仿真技术切入自动驾驶领域，随后延伸到更多领域。2025年12月，五一视界成功登陆港股。
Celebrating 25 years of visual search innovation
（翻译）庆祝视觉搜索创新25周年
Google Images is turning 25. Here’s a look back at some major milestones — and new ways to explore and create visual content.
How Much Memory Does Your Agent Actually Need?
（翻译）你的智能体到底需要多少记忆？
A Blog post by IBM Research on Hugging Face
Developing Nemotron 3.5 Lightning NVFP4 with QAD Using NVIDIA Model Optimizer
（翻译）使用 NVIDIA Model Optimizer 通过 QAD 开发 Nemotron 3.5 Lightning NVFP4
Teams customize their models to hit their targets for latency, speed, memory, and compute. With the open NVIDIA Nemotron family of models…
RENDER: Controlling Reader-Facing Evidence in LLM Memory Evaluation
（翻译）RENDER：在LLM记忆评估中控制面向读者的证据
Abstract page for arXiv paper 2608.23568: RENDER: Controlling Reader-Facing Evidence in LLM Memory Evaluation
A survey detection channel overrides the pixels in an astronomical foundation model, and biases tomographic mean redshifts
（翻译）巡天检测通道覆盖天文学基础模型中的像素，并使层析平均红移产生偏置
arXiv:2608.23626v1 Announce Type: new Abstract: Foundation models for astronomy are trained on survey pixels together with the catalogue products derived from those pixels. Those catalogues are incomplete at a measurable rate, and a model trained on both inherits that incompleteness as a systematic.
TRACE: Transition-Aware Residual Control for Multi-Objective Materials Discovery
（翻译）TRACE：面向多目标材料发现的转换感知残差控制
Abstract page for arXiv paper 2608.23631: TRACE: Transition-Aware Residual Control for Multi-Objective Materials Discovery
How much of a measured AI preference is the model, and how much is the instrument?
（翻译）测得的AI偏好有多少来自模型本身，又有多少来自测量工具？
Abstract page for arXiv paper 2608.23641: How much of a measured AI preference is the model, and how much is the instrument?