HeurAgenix：大语言模型驱动的组合优化“全能教练”
编者按：在物流规划、生产排程、能源调度等行业应用场景中，组合优化问题往往规模庞大、约束复杂，长期依赖专家手工调参的启发式算法已难以应对动态多变的现实需求。微软亚洲研究院最新提出的 HeurAgenix 框架，将大语言模型视作“全能教练”，为启发式算法带来了自进化与自适应的能力，并通过蒸馏压缩实现快速响应。HeurAgenix 不仅全面超越了现有 LLMs 超启发式方法，在多数任务中甚至优于专业求解...
CausCell：虚拟细胞进入“白盒时代”
编者按：在生命科学研究中，逼真且可解释的细胞模型一直备受期待。但现有的单细胞 AI 建模多依赖于深度神经网络等“黑盒”方法，难以解开高维组学数据背后的生物学因果机制。近日，微软亚洲研究院（上海）与同济大学合作在 《自然-通讯》（Nature Communications）杂志上发表了最新成果 CausCell。这一框架首次将结构因果模型与扩散模型深度融合，实现了细胞尺度的因果解耦表征与可控的反事实...
Dion: the distributed orthonormal update revolution is here
| Kwangjun Ahn and John Langford
Dion is a new AI model optimization method that boosts scalability and performance over existing leading methods by orthonormalizing only a top rank subset of singular vectors, enabling more efficient training of large models such as LLaMA-3 with reduced overhead.
Reimagining healthcare delivery and public health with AI
| Peter Lee, Dr. Umair Shah, and Dr. Gianrico Farrugia
Former Washington State Secretary of Health Dr. Umair Shah and Mayo Clinic CEO Dr. Gianrico Farrugia explore how healthcare leaders are approaching AI when it comes to public health, care delivery, the healthcare-research connection, and the patient experience.
Self-adaptive reasoning for science
| Newman Cheng, Gordon Broadbent, Steven Truitt, and William Chappell
Microsoft is pioneering a vision for a self-adapting AI system that can adapt to the dynamic nature of scientific discovery, promoting deeper, more refined reasoning in complex scientific domains.
ACL上新 | 公平与责任：语言模型的另一种进化
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。 作为自然语言处理领域全球顶级的学术盛会之一，ACL 2025 于近日在维也纳召开。来自微软亚洲研究院的多篇论文入选，我们通过两期“科研上新”为大家分享研究院入选 ACL 2025 的精选论文解读。第一期聚焦了使大语言模型和语音模型在预训练、...
Project Ire autonomously identifies malware at scale
| Brian Caswell, Dustin Fraze, Sarah Smith, Rodrigo Racanicci, Tim Middleton-Sally, Shelby Hayes, Stanley He, Katy Smith, Bhakta Pradhan, and Mike Walker
Designed to classify software without context, Project Ire replicates the gold standard in malware analysis through reverse engineering. It streamlines a complex, expert-driven process, making large-scale malware detection faster & more consistent.
VeriTrail: Detecting hallucination and tracing provenance in multi-step AI workflows
| Dasha Metropolitansky
VeriTrail, new from Microsoft Research, can detect AI-generated content that is not supported by the source text, trace the provenance of content from final output back to the source, and locate where errors were likely introduced.
Time-series data—measurements collected at regular intervals, like stock prices or traffic flows—has become a key driver of intelligent decision-making systems across industries. From medical monitoring to financial risk control, identifying patterns in this data is essential to many important operations.…