Microsoft Research

Microsoft Research team introduces Vega zero-knowledge proof system that can verify government-issued identity information without exposing credentials themselves, supporting mobile generation within 100 milliseconds, no trusted setup required, soon to be open source.

Microsoft Research releases MagenticLite, MagenticBrain, and Fara1.5, an agentic experience optimized for small models through co-design achieving unified workflows across browser and local file system, where Fara1.5 nearly doubles web navigation performance.

SocialReasoning-Bench reveals that current frontier AI models often accept suboptimal outcomes when negotiating on behalf of users, despite completing tasks successfully.

微软研究院提出通过概率推理控制扩散模型的新方法，结合SDE和score-based技术实现对生成分布的精确控制，适用于图像生成与强化学习场景。

微软研究院提出通过编译器将数据库标量函数转换为GPU内核，实现查询性能显著提升。

MindTopo基准测试揭示多模态模型在拓扑推理任务中存在显著缺陷，尤其在动态场景规划能力不足。

Skala 1.1通过增加2.5倍训练数据和集成到多个软件，显著提升DFT计算的准确性和可访问性。

微软研究院提出二项式流模型，通过去噪和流匹配解决离散序数数据生成问题，填补连续与分类扩散模型间的空白。

EvoLib框架让AI模型通过自身经验学习，无需外部反馈，知识库持续优化提升性能。

微软发布Orchard开源框架，通过Orchard Env环境服务支持多领域智能体训练，30亿参数模型在SWE-bench测试中达到69.7%性能。

微软研究院提出扩展流生成模型，通过流映射技术实现超越固定画布的快速灵活生成。

GCT框架通过AI生成可验证的理论，揭示大脑皮层对语言的响应机制，推动神经科学与AI的结合。

Talos 是一个开源系统，可将基因组医学中的人工审核时间减少 90%，并为每位患者仅提供 1.3 个候选变异供专家审核。

Talos 是一个开源工具，通过自动化迭代基因组重分析，显著提升罕见病诊断率，同时保持低误报率。

ResNet 获得 CVPR 2026 长期影响奖，其残差连接技术已成为现代 AI 系统构建的基础。

AI模型从多样化的细胞状态中学习比单纯扩大数据集更有效，可能改变个性化治疗的匹配方式。

Microsoft Research proposes the Intervene framework that uses LLM-based projection to decompose AI agent outputs into verifiable properties and generates formal specifications in real-time for compliance assurance.

Microsoft Research releases Magenta Light, a full-stack agentic experience powered by small models.

GitHub Agentic Workflows introduces AI agents to automate the software development lifecycle, combining GitHub Actions, Copilot CLI, and a secure sandbox for end-to-end automation.

Intervene is a real-time verification framework developed by Microsoft Research that extracts verifiable properties from natural language to improve the reliability of agent systems.

Microsoft Research team built a realistic electric transmission grid model covering 48 U.S. states using open data, supporting AC-OPF analysis.

The article analyzes the economic and social impact of AI technology at different stages, proposes a three-stage model from augmentation to automation and reconstruction, and emphasizes that institutional and organizational coordination is key to achieving deep transformation.

A wavelet-based interpretable parameterization of the score function reveals that local interactions most strongly impact denoising performance, and wavelets yield greater gains at higher resolutions; this framework serves as a theoretical probe—not a replacement—for trained models.

微软研究院展示了多个AI研究进展，涵盖小语言模型谈判、非洲语言语音评估及知识转化工具。

Adding constraints to generative AI in materials inverse design significantly boosts the efficiency of generating target-performance materials, with models like DiffSCP and MatGen enabling rapid crystal structure screening.

Microsoft Research introduces the Intervene mechanism, converting AI agent behaviors into verifiable properties and generating Python validators, significantly improving small models' performance on complex tasks.

微软研究团队提出Intervene方法，在Tau Too Bench等基准测试中，小型模型的准确性可媲美前沿模型，通过提取可验证属性并自动生成Python代码进行运行时验证。

mimalloc, Microsoft's open-source memory allocator, significantly improves performance and efficiency in managing memory for modern applications and services, becoming the key tool to handle memory demands that have grown hundreds of times over the past decade.

Microsoft Research proposes a new language model fine-tuning method that improves performance by matching meaning rather than tokens.

微软研究院探讨即使个体AI代理安全，互连的AI生态系统仍可能面临风险。研究分析了AI代理交互时的问题及为何需网络层面的新策略。