Publisher
Stories with reporting from Microsoft Research, ranked by what is gaining ground now. Each story links to the original articles.
Microsoft ResearchAgent Lightning v1.0 connects existing AI agents to reinforcement-learning training without requiring developers to rebuild them. The lightweight framework is described as having 3,500 lines and is designed to work with agents’ existing tools and decision processes.
Microsoft ResearchMicrosoft Research introduces CARE-X, an approach to radiology vision-language models for chest X-ray interpretation. It combines auxiliary supervision, reward-aligned learning, calibrated predictions, and tool-based measurement.
Microsoft ResearchMicrosoft Research describes a machine-learning system that predicts where extreme space weather may damage power grids 30–60 minutes before a storm arrives. Such events can also reduce GPS accuracy and disrupt satellite operations.
Microsoft ResearchQuine is an early-stage research effort to build a multimodal world model that connects biological scales and data types. The system aims to help scientists search a broad hypothesis space computationally and prioritize ideas before laboratory experiments.
Microsoft ResearchMicrosoft Research releases Skala 1.1, an updated deep-learning exchange-correlation functional for predictive density functional theory. The update aims to improve accuracy, broaden access across computational chemistry tools, and provide a living benchmark for performance.
Microsoft ResearchMicrosoft Research describes RetroChimera, a predictive model for improving small-molecule synthesis prediction, in connection with a Nature paper. The model aims to help researchers explore molecules and address the slow, costly process of producing them.
Microsoft ResearchMicrosoft Research reports that moving AI inference off robots can improve task success and efficiency while supporting more advanced physical AI workloads. The work addresses how robotic hardware can keep pace with increasingly capable AI.
Microsoft ResearchMicrosoft Research presents GigaPath-Flash and GigaTIME-Flash, pathology foundation models designed to use fewer computational resources while maintaining strong performance. The reduced demands are intended to make larger studies and broader exploration more practical.
Microsoft ResearchMicrosoft Research introduces MindTopo, a benchmark for evaluating whether vision-language models understand topological relationships. The work also points to opportunities to improve models’ spatial reasoning and planning.
Microsoft ResearchMicrosoft Research Asia–Singapore marks its first year by building partnerships across government, academia and industry while exploring real-world applications for frontier AI research. The lab says its work focuses on research collaboration and talent development.
Microsoft ResearchMicrosoft Research profiles Jennifer Neville's work on identifying surprising failures that make it difficult for AI to handle complexity. The post also recounts how she came to work in computer science.
