Can AI help power the grid? MSR-led AI Foundations for Power Grids workshop at NeurIPS 2026 is accepting submissions until August 29 on benchmarks, model training, reliability, foundation models, and real-world deployment. https://msft.it/6043vAmaz
Microsoft Research
Think Tanks
Redmond, Washington 410,145 followers
We advance science and technology to benefit humanity.
About us
At Microsoft Research, we accelerate scientific discovery and technology innovation to empower every person and organization on the planet to achieve more. We do this by bringing together the best minds across diverse disciplines and backgrounds to take on the most pressing research challenges for Microsoft and for society. Our Research Lens We consider research directions through the lens of the positive impact we aspire to create with and for customers, communities, and all of society.
- Website
-
http://www.microsoft.com/research
External link for Microsoft Research
- Industry
- Think Tanks
- Company size
- 1,001-5,000 employees
- Headquarters
- Redmond, Washington
- Founded
- 1991
Updates
-
How can AI help advance precision oncology? Explore Project Ex Vivo, a collaboration between Microsoft Research and the Broad Institute, with support from Dana-Farber Cancer Institute, that is using AI to better understand cancer cell states. Read the story in Signal Magazine: https://msft.it/6048v7JTM
-
Researchers Dr. Jenna Butler, Jake Hofman, and Rebecca Janßen spend a lot of time studying AI’s influence on how people do their jobs. On the Microsoft Research Podcast, they share their thoughts on what the future of work looks like in the age of AI. https://msft.it/6045vAqAR
-
Microsoft Research reposted this
Much of what AI knows about the world comes from data most people never encounter. Behind the scenes, what's included and excluded help shape how it represents people. Microsoft's Community Library Creator is a new approach that gives communities a structured way to influence how they show up in AI-generated imagery. Learn how this tool is helping people with disabilities and other shared experiences shape the future of AI. https://msft.it/6001vf4vf
-
-
MagenticLite's models are now fully open weight. Today, we are open sourcing the weights for Fara1.5 and MagenticBrain models on Hugging Face under an open MIT license, to work with MagenticLite, the next generation of Magentic-UI. What’s in the release: MagenticLite, the agentic application that works across the browser and local file system in a single workflow, with sandboxed execution and human approval built in. MagenticBrain, the 14B orchestration model that plans, writes code, and delegates. It was trained end-to-end inside the MagenticLite harness with the same tool schemas it sees at inference. Fara 1.5, the computer-use model family in 4B, 9B, and 27B sizes. Fara1.5-27B scores 72% on Online-Mind2Web, ahead of much larger proprietary computer-use systems, and the 4B is small enough to run on-device. We recommend Fara1.5-9B with MagenticLite. You can download the weights, run them locally if needed, fine-tune them for your workflows, and get a complete agentic experience with no cloud dependency. Models: https://msft.it/6044vA8Ry MagenticLite: https://msft.it/6048vA8RM Technical report: https://msft.it/6049vA8R3 Fara1.5 blog: https://msft.it/6040vA8RO
-
Today, we announce the second release of PazaBench, our benchmark for evaluating Automatic Speech Recognition (ASR) models across African languages. We’ve significantly expanded coverage to support: ● 22 additional languages, now reaching communities across 38 African countries ● 1 additional ASR model ● Coverage expanded to 11 datasets ● Nearly 2x more test samples for robust evaluation These improvements make PazaBench one of the most comprehensive resources for assessing speech to text models across African languages and contexts. We’re grateful to everyone who’s submitted datasets for language evaluation, this update was made possible by you. Reliable benchmarks are essential for advancing inclusive AI. By expanding language, geographic, and dataset coverage, PazaBench V2 provides a stronger foundation for researchers and practitioners developing speech technologies for historically underserved languages. A huge thank you to all the dataset creators, community contributors, maintainers and collaborators who continue to help make this work possible. Request additional language support to contribute to our next release. https://msft.it/6043vfVNX
-
-
Microsoft Research reposted this
Excited to highlight new work from Microsoft Research on in-context adaptation of generative robot policies. FlowDAgger is an innovative step forward in how we build more effective human-agent teams for robotics. Real-world robotic systems inevitably encounter ambiguity, distribution shift, embodiment quirks, and long-tail failure modes. The path to more capable and trustworthy agents is not simply greater autonomy. It is tighter feedback loops where people and agents can guide and improve robot behavior in context. FlowDAgger’s key insight is using action inversion as the enabler for human correction. Modern robot action models use a process called flow-matching to turn noise inputs into robot actions based on the robot's observations of the environment. FlowDAgger maps human corrections back to the noise inputs that would have produced those corrected actions under the robot's initial (base) action policy. It then trains a lightweight policy that works together with the base policy to produce better actions in similar situations. This work is a strong example of research aiming beyond impressive demos toward systems people can work with, steer, and trust. Kudos to Michael Murray, Andrey Kolobov, Simran Bagaria, Dean Fortier, Tess Hellebrekers, Harshavardhan Reddy Gajarla, Galen Mullins, Oier Mees, and our collaborators from the University of Washington. More here: https://lnkd.in/gMdYY4NN
-
In this issue: Aurora 1.5 extends open foundation models for weather and Earth‑system forecasting with 22 new variables, hourly resolution, and ensemble prediction; Flint introduces a visualization intermediate language that lets AI agents create expressive charts from simple specs; AI agents demonstrate how obscurity fails as a Rowhammer defense by reverse‑engineering simulated hardware protections; HASTE enables rapid post‑disaster building damage assessment from satellite imagery through a no‑code web platform; new research on conversational memory reveals how different memory roles shape personalization and factual accuracy in AI assistants; and FlowDAgger accelerates robot adaptation through human‑in‑the‑loop corrections in latent space.
-
Microsoft Research reposted this
We're pleased to announce that Kitala AI, in collaboration with YUX Design and Microsoft Research Africa, will be hosting a workshop at Deep Learning Indaba 2026 in Lagos, Nigeria. Beyond Benchmarks: Scaling Multi-Turn Participatory AI Evaluations in Health and Education explores how we can move beyond traditional AI benchmarks to evaluate models through real-world, multi-turn interactions that better reflect the complexity of African contexts. Together, we'll share practical approaches for designing human-centered AI evaluations, developing meaningful evaluation metrics, and using participatory methods to better understand how AI performs across diverse communities. If you're attending Deep Learning Indaba 2026, we'd love to have you join us for the workshop and be part of the conversation shaping the future of AI evaluation in Africa. For more information about in-person attendance at the Deep Learning Indaba 2026 Conference, please visit this page: https://lnkd.in/dGnkn7nW If you're attending Deep Learning Indaba 2026 and would like to join our workshop, learn more and register here: https://lnkd.in/dFJMphSA We look forward to welcoming researchers, practitioners, and AI builders in Lagos this August as we explore the future of human-centered AI evaluation together. Millicent Ochieng Oluchi Audu Mame Coumba Ka Mercy Muchai Yann Le Beux Stephanie Nyairo Felermino Ali Oche David Ankeli Deep Learning Indaba #DeepLearningIndaba #DLI2026 #AIEvals #AI #ArtificialIntelligence #AISafety #AIEvaluation #ResponsibleAI #HealthAI #TrustworthyAI #HCI
-
-
Recalling a new password after hours of distractions can be hard. Not for a transformer. Microsoft CVP Doug Burger & AI researchers Subutai A. & Nicolo Fusi discuss the fundamental differences between human memory and machine intelligence. https://msft.it/6043vFsK9