BenchmarkQED: Automated benchmarking of RAG systems
| Darren Edge, Ha Trinh, Andres Morales Esquivel, and Jonathan Larson
BenchmarkQED is an open-source toolkit for benchmarking RAG systems using automated query generation, evaluation, and dataset prep. It shows that LazyGraphRAG outperforms standard methods, especially on complex, global queries.
CVPR 上新 | 微软亚洲研究院精选论文解读——生成模型与扩散技术篇
编者按：欢迎阅读“科研上新”栏目！“科研上新”汇聚了微软亚洲研究院最新的创新成果与科研动态。在这里，你可以快速浏览研究院的亮点资讯，保持对前沿领域的敏锐嗅觉。 6月10日至17日，全球计算机视觉领域的顶尖学术盛会 CVPR 将在美国田纳西州纳什维尔举办。我们将通过两期“科研上新”为大家带来多篇微软亚洲研究院入选 CVPR 2025 的精选论文解读。第一期的分享内容是主要围绕生成模型与扩散技术等方向...
In the news | CNBC
How Microsoft, Physics Wallah are bringing AI-driven learning to Indian classrooms
As part of its broader $3 billion AI and cloud investment in India, Microsoft is deepening its focus on education through a collaboration with edtech company Physics Wallah, aimed at improving learning outcomes using AI-powered tools and personalised academic support.
OPA-DPO：多模态大模型幻觉难题的高效解决方案
编者按：在视觉多模态大语言模型的快速发展中，幻觉问题一直是研究者们关注的焦点。模型生成与输入图像不一致甚至虚假的内容，不仅影响用户体验，也阻碍了多模态技术在实际场景中的落地。对此，微软亚洲研究院和香港中文大学的联合研究团队从直接偏好优化（DPO）入手，提出了 On-Policy Alignment (OPA)-DPO 算法，可通过确保训练数据与初始策略（reference policy）的一致性，...
端侧大模型部署新引擎：无需硬件修改，解锁商用DRAM的大模型推理
编者按：随着应用场景的扩展，端侧设备（如手机、电脑、可穿戴设备、机器人等）对大模型高效运行的需求日益增长，但端侧设备对模型运行的计算资源、访存带宽、能耗都有着极其苛刻的要求。存内计算技术有望从根本上解决以上资源问题，它能够将存储单元和计算单元融合，显著减少数据在存储和计算单元间频繁搬运而产生的资源损耗。然而，传统存内计算涉及对硬件架构的改动，不仅技术难度大，且迭代周期长，无法在实际场景中大规模量产...
What AI’s impact on individuals means for the health workforce and industry
| Peter Lee, Ethan Mollick, and Azeem Azhar
Ethan Mollick and Azeem Azhar, thought leaders at the forefront of AI’s influence on work, education, and society, discuss the impact of AI at the individual level and what that means for the healthcare workforce and the organizations and systems…
FrodoKEM: A conservative quantum-safe cryptographic algorithm
| Patrick Longa
The recent advances in quantum computing offer many advantages—but also challenge current cryptographic strategies. Learn how FrodoKEM could help strengthen security, even in a future with powerful quantum computers.
编者按：在产业智能化加速发展的当下，时间序列数据已然成为智能决策系统的关键基石。然而，传统的时间序列生成模型往往难以应对跨领域、跨风格的数据需求，且生成的数据在实际应用中缺乏可控性和实用性。 为解决这些痛点，微软亚洲研究院推出开源框架 TimeCraft，融合多项研究成果，通过跨域泛化、自然语言控制与任务感知等创新技术，助力时间序列生成任务从结构理解到任务对齐的全流程能力建设。TimeCraft...
Abstracts: Zero-shot models in single-cell biology with Alex Lu
| Gretchen Huizinga and Alex Lu
The emergence of foundation models has sparked interest in applications to single-cell biology, but when tested in zero-shot settings, they underperform compared to simpler methods. Alex Lu shares insights on why more research on AI models is needed in biological…