Loading...
Two white line icons on a gradient background transitioning from blue to pink. From left to right: icon representing a set of gears; an icon representing three connected nodes each containing a user icon
Microsoft Research Blog

Breaking the networking wall in AI infrastructure  

September 9, 2025 | Paolo Costa

Datacenter memory and network limits are restraining AI system performance. MOSAIC uses microLEDs and a wide-and-slow optical architecture to deliver faster, longer, more reliable, and energy efficient connections that could transform AI cluster designs.

shape
Articles

BRIDGE:突破SFT-RL两阶段训练瓶颈,开辟推理模型协同优化新路径 

September 9, 2025

编者按:在数学、编程等高度挑战的推理任务中,大语言模型正在展现前所未有的突破。然而,支撑其背后的传统训练范式仍存在效率低下与“灾难性遗忘”的困境。为此,微软亚洲研究院与香港中文大学联合提出了 BRIDGE 框架,通过双层优化机制让监督微调(SFT)与强化学习(RL)“携手并进”,突破了长期以来的两阶段训练瓶颈。BRIDGE 不仅在多个权威测试中实现了显著的性能提升,还为未来大模型后训练开辟了新方向...

Awards | MIT Technology Review

Tian Xie named to MIT Technology Review’s 2025 Innovators Under 35 

September 8, 2025

Tian led the development of MatterGen, our generative AI model for materials discovery.

Baining Guo Microsoft Technical Fellow
Articles

祝贺微软亚洲研究院常务副院长郭百宁博士荣膺微软技术院士! 

September 5, 2025

近日,微软亚洲研究院常务副院长郭百宁博士荣获了微软技术院士(Microsoft Technical Fellow)称号。作为微软公司技术领域的最高荣誉,微软技术院士是对技术人员在技术创新与行业影响力方面的至高认可。更值得庆贺的是,郭百宁博士是微软亚洲研究院以及微软亚太研发集团史上首位获此殊荣的研究员! 郭百宁博士长期深耕于计算机视觉、计算机图形学、多媒体以及多模态人工智能等领域。他带领团队取得了多...

In the news | Open to Debate

E. Glen Weyl debate Curtis Yarvin 

September 5, 2025

Glen Weyl debate prominent anti-democratic technologist and philosopher Curtis Yarvin on whether the US should be ruled by a CEO dictator.

In the news | NVIDIA Blog

Scene it to believe it: Populate 3D worlds quickly with NVIDIA AI blueprints 

September 3, 2025

NVIDIA released the AI Blueprint for 3D object generation — a workflow that lets 3D artists create up to 20 3D objects to prototype a scene from a simple text prompt. Microsoft Research Asia’s TRELLIS NVIDIA NIM microservice works within…

In the news | Microsoft Source

Microsoft’s analog optical computer cracks two practical problems and shows AI promise 

September 3, 2025

A small Microsoft Research team had lofty goals when it set out four years ago to create an analog optical computer that would use light as a medium for solving complex problems. From the beginning, they wanted to build it using…

Screen Reader Users in the Vibe Coding Era
Articles

Vibe Coding能否赋予屏幕阅读器开发者“编程超能力”? 

September 2, 2025

编者按:生成式人工智能的持续发展正在改变人与数字系统交互、执行任务及协同工作的模式。在众多领域中,软件开发无疑是变革最为显著的一个。AI 编程工具大大提升了编码效率,尤其对依赖屏幕阅读器的视力受损开发者而言,新兴的“Vibe Coding(氛围编程)”模式更是带来了前所未有的可能。 然而,屏幕阅读器用户如何与 AI 编程工具协作?这些工具又将如何赋能他们?他们在使用 AI 编程工具时是否会遇到新的...

graphical user interface
Articles

VibeVoice实现90分钟、多角色播客生成,拓展语音合成新边界 

August 27, 2025

编者按:AI 自动生成播客早已不是新鲜事,但常见的 AI 播客只局限于几分钟的双人对话,这是因为传统语音生成模型大多基于离散化方法,更擅长生成短句、单一音色、结构规整的语音内容。近日,微软亚洲研究院提出了一种全新的语音生成模型 VibeVoice。该模型采用 next-token diffusion 机制,实现了长达90分钟、支持最多4人自然对话的高质量语音生成。更重要的是,VibeVoice 生...

  • Previous
  • 1
  • …
  • 14
  • 15
  • 16
  • 17
  • 18
  • …
  • 573
  • Next