- 资讯公开站Digitimes
US unveils South Korea's US$20B nuclear, Alaska LNG investment
The US has unilaterally announced South Korea's investment package in the US, including Alaska LNG development and the construction of eight large nuclear power plants. Seoul had previously listed those two projects as i…
- 资讯公开站Digitimes
Qualcomm, Samsung redesign chip packaging as heat limits phone performance
Qualcomm and Samsung Electronics are redesigning how memory is packaged with flagship smartphone processors, as rising chip performance puts greater pressure on the limited cooling capacity of mobile devices.
- 资讯公开站Digitimes
Jabil targets US$17.5 billion in data center revenue as it builds out power, cooling, and networking
Jabil's plan for fiscal 2027 puts a price tag on its move from building servers to building the power, cooling, and networking systems around them. The company expects about US$17.5 billion of cloud and data-center infra…
- 资讯公开站Digitimes
Synopsys raises long-term growth and margin targets, guides fiscal 2027 revenue to US$11.15 billion
Synopsys is betting that custom silicon and AI-driven chip design will lift its growth above the double-digit pace it previously targeted. At its investor day in New York on September 30, 2026, the design software and IP…
- 论文公开站arXiv
Phase-Field Modeling of Liquid Phases with Ordering
A strong negative enthalpy of mixing can lead to ordering in liquid solutions. To describe ordering in liquids, different thermodynamic descriptions of liquid phases have been developed in the CALPHAD approach, including…
- 论文公开站arXiv
Compression Footprints as Security Signals for Model-Poisoning Defense in Federated Learning
Lossy compression is widely used in Federated Learning (FL) but is generally treated as an error source, while conventional poisoning defenses inspect update geometry. In this work, we instead treat the compressor's resp…
- 论文公开站arXiv
Scaling Laws for Looped Mixture of Experts
Looped transformers and Mixture-of-Experts (MoE) offer complementary routes to efficient scaling: recurrence increases computational depth at fixed parameters, while MoE sparsity expands total capacity at fixed active co…
- 论文公开站arXiv
MatLoom: Layered Text-to-Material Generation in a Compact Program Space
Material generation should produce not only an appearance, but also the rules that construct it. We introduce MatLoom, a compact, layer-oriented language for text-to-material generation with pretrained language models. E…
- 论文公开站arXiv
Cogentic：面向自动证明发现的多智能体编排
Cogentic: Multi-Agent Orchestration for Automated Proof Discovery
提出 Cogentic，一个用于开放研究问题自动证明发现的多智能体框架。前沿语言模型虽能单次生成优质数学想法，但面对需多路径探索的开放问题，单次生成往往不足。
- 论文公开站arXiv
WorldAuditBench：多模态智能体的交互式3D世界审计
WorldAuditBench: Interactive 3D World Auditing with Multimodal Agents
随着交互式3D世界被广泛用于研究智能行为，识别其中异常（如悬浮物体、可穿墙、与环境不一致的物体）的高效流程变得重要，本文提出相应基准。
- 论文公开站arXiv
Turbo Harness：实例自适应框架优化
Turbo Harness: Instance-Adaptive Harness Optimization
自动化搜索有效框架是实现智能体递归自我改进的重要一步。现有优化通常产出统一应用于所有任务的单一全局框架，但适用于某类任务的框架未必通用。
- 论文公开站arXiv
DP-SGD 下权重绑定对纯解码器 LLM 是否仍有益？
Is Weight Tying Still Beneficial for Decoder-Only LLMs in Private Settings Under DP-SGD?
DP-SGD 是大模型隐私保护微调的主流方法。许多纯解码器 LLM 采用输入输出嵌入权重绑定，本文在差分隐私设定下重新审视这一设计是否仍然有利。
- 论文公开站arXiv
Ego4WAM：扩展第一人称人类数据用于机器人学习的关键因素
Ego4WAM: What Matters When Scaling Egocentric Human Data for Robot Learning?
第一人称人类数据为机器人学习提供可扩展经验，但在人机对齐、行为覆盖和监督信号上差异显著。已有工作显示数据越多效果越好，但哪些数据属性真正关键仍不明确。
- 论文公开站arXiv
图像分类器是高效的自监督视频表征学习器
Image Classifiers are Efficient Self-Supervised Video Representation Learners
提出 VideoMSN，一种用于视频高效自监督时空表征学习的掩码孪生网络框架。它不依赖重型3D架构或重建式自编码器，而是复用标准图像分类器处理无标注数据。
- 论文公开站arXiv
AssemblyWorld：用通用智能体重思3D装配
AssemblyWorld: Rethinking 3D Assembly with General-Purpose Agents
3D装配需将零件及其关系的理解转化为精确空间排布。本文探究预训练通用智能体能否仅通过视觉交互完成物体装配，而无需针对装配任务额外微调。
- 论文公开站arXiv
ViTeX-Bench：高保真视频场景文本编辑基准
ViTeX-Bench: Benchmarking High-Fidelity Video Scene Text Editing
视频生成日益逼真可控，但视频编辑仍欠成熟，尤其是需保持原始场景动态的精确局部编辑。视频场景文本编辑需替换店面招牌等场景表面上的文字。
- 论文公开站arXiv
去除时间捷径可改进非侵入式脑机文本解码
Removing Timing Shortcuts Improves Non-Invasive Brain-to-Text
研究发现，多项报道的非侵入式脑信号词解码性能提升，在完全不用脑数据的情况下也能复现。在 d'Ascoli 等人（2025）的工作中，受试者感知连续语音的脑活动时间序列被分段处理。
- 论文公开站arXiv
半事实信用增强策略优化
Semifactual Credit-Augmented Policy Optimization
可验证奖励强化学习（RLVR）提升了 LLM 推理能力，但其预测仍对任务无关的提示特征敏感。本文通过半事实提示干预研究这种敏感性。
- 论文公开站arXiv
面向多模态临床诊断的排序感知提示优化
Ranking-Aware Prompt Optimization for Multimodal Clinical Diagnosis
多模态大模型正快速推进临床诊断，但其适配流程仍以准确率为目标。临床数据类别严重不平衡，恒定多数类预测器准确率可超90%却无临床价值。
- 资讯公开站Ars Technica
恐龙灭绝撞击坑或曾生机勃勃
Dinosaur-killing impact crater might have been teeming with life
- 资讯公开站Ars Technica
本地植入物网络以人体作为线路
A local network of implants uses your body as the wiring
- 资讯公开站Ars Technica
攻击者持续利用 Zimbra 严重漏洞窃取邮件
Attackers have been exploiting critical Zimbra flaw to steal emails
- 资讯公开站Electrek
芝加哥 Metra 称其新型电池电动机车为首创
Chicago Metra calls its new battery-electric locomotive a first of its kind
Metra 的柴油通勤列车在芝加哥周边很常见，柴油废气气味亦然，但 Stadler 新型电池电动 FLIRT 机车即将改变这一切，本周到货。
- 资讯公开站CNX Software
面向 CM5 的 30 美元树莓派智能显示模块，瞄准数字标牌与自助终端
$30 Raspberry Pi Smart Display Module for CM5 targets digital signage displays and kiosks
树莓派智能显示模块是兼容英特尔智能显示模块（SDM）规范的 CM5 载板，具体为 SDM-L 规格。英特尔 SDM 标准已存在多年。
- 资讯公开站Energy Storage News
昆士兰 CleanCo 归功储能热潮，该州批发电价降40%
Queensland’s CleanCo credits battery storage boom as Australian state records 40% wholesale electricity price drop
昆士兰2025-26财年平均批发电价下降40%，CleanCo Queensland 将这一降幅归因于电池储能系统（BESS）。
