Research at Microsoft 2025: Translating ideas into action

ExACT teaches AI agents to explore more effectively, gather and evaluate information, and then find the best decision-making strategies.

Since its opening in 2005, Microsoft Research India has achieved an extraordinary record of innovation—in areas ranging from health and education to agriculture and accessibility.

Microsoft is re-engineering the quantum computer with the world’s first topoconductor and first quantum processing unit (QPU) architecture with a topological core.

Microsoft Research makes BioEmu-1 open source. The new deep learning model can generate thousands of protein structures per hour on a single GPU.

This new model takes a significant step toward AI agents that can serve as versatile, general-purpose assistants.

Podcast guests discuss state-of-the-art foundation models as a source of creative inspiration and how the Accelerating Foundation Models Research initiative is partnering with Minority Serving Institutions.

A Microsoft Research study published in Nature Scientific Reports identifies pain points in the biomedical data life cycle and offers recommendations on secure data sharing, interoperability, and collaboration among researchers. 

In this special Microsoft Research Podcast series, guests from across the healthcare ecosystem join Peter Lee, president of Microsoft Research, to discuss the current state and future of AI in medicine. This first episode examines how clinicians are using the technology.

A new approach makes visual data more compact and manageable, giving developers in visual world modeling a flexible and efficient new tool.

Host Peter Lee, Microsoft’s Dr. Matthew Lungren, and Epic’s Seth Hain discuss implementing AI in healthcare, from automating administrative tasks to integrating AI into electronic health records at global scale.

Developers spend a lot of time debugging code instead of writing it. Debug-gym aims to maximize the potential time savings from AI tools by teaching them to debug code.

The competition is organized by leading mathematicians and computer scientists to encourage the development of AI models capable of mathematical reasoning and problem-solving.

A new position paper explores Societal AI, an interdisciplinary framework for guiding AI development that examines how AI intersects with social systems and public life.

The Fusion Summit brought together scientists and researchers to explore how AI can accelerate fusion energy research and help bring this energy source to the grid sooner.

This research prototype of a human-centered agent helps researchers study open questions on human-in-the-loop approaches and oversight mechanisms for AI agents.

This new key encapsulation mechanism is based on the Learning with Errors (LWE) problem, a cornerstone of lattice-based cryptography, and built on generic, unstructured lattices.

ASHABot provides rural India’s frontline health workers with reliable, localized medical guidance, enabling them to better serve underserved communities more confidently and effectively.

Bill Gates and Sébastien Bubeck discuss how advancing AI is already influencing clinical work, global health, and scientific discovery and what breakthroughs, risks, and shifts they expect to define the future of AI-enabled medicine.

Lessons from established sectors show that strengthening rigor, standardization, and interpretability are essential for building reliable AI evaluation frameworks.

A team at Microsoft Research is rewriting parts of the SymCrypt cryptographic library in Rust to improve memory safety, defend against side-channel attacks, and enable formal verification.

The paper “PoWER  Never Corrupts: Tool-Agnostic Verification of Crash Consistency and Corruption Detection” introduces two new techniques for verifying storage systems using only widely supported constructs, addressing a longstanding challenge in system design.

Daphne Koller, Noubar Afeyan, and Dr. Eric Topol—leaders in AI-driven medicine—examine how AI is accelerating drug discovery, transforming biotech research and development, and advancing disease prevention.

Recipient of an ICML 2025 Outstanding Paper Award, CollabLLM improves how LLMs collaborate with users. Using reinforcement learning, it trains models to ask questions and adapt its tone to different situations.

The lab, Microsoft’s first in Southeast Asia, will drive innovation on intertwined goals: deploying industry-transforming AI, pursuing frontier breakthroughs in AI foundations, and advancing responsible, socially beneficial applications.

A research collaboration is developing a low-carbon cement alternative made from intact seaweed biomass. The work uses a machine learning–driven design process to rapidly test and refine blends and reduce development time from months to 28 days.

Project Ire is an autonomous AI agent that can analyze and classify software without assistance and even reverse engineer a software file with no clues about its origin or purpose—a big step forward in cybersecurity and malware detection.

Microsoft researchers demonstrate that an optimization-based, self-adaptive AI system developed without post-training reinforcement learning can rival post-trained models in domains where adaptability, explainability, and control matter most.

MindJourney helps AI agents explore 3D spaces they can’t directly detect by addressing a key limitation in vision-language models, which allow AI agents to interpret and describe visual scenes.

Crescent is a cryptographic library that helps people protect their personal privacy by preventing widely used identity formats from automatically linking their separate digital interactions and building a detailed user behavior profile.

Model Context Protocol (MCP) and a growing society of tools promise to extend AI-agent capabilities through cross-provider horizontal integration. But the copresence of these cross-provider tools or agents could lead to failures. Microsoft Research offers steps to avoid that.

MOSAIC is a novel optical link technology that uses microLEDs and a wide-and-slow design to improve datacenter performance and GPU communication by simultaneously providing low power and cost, high reliability, and long reach.

Working with Drexel University and the Broad Institute, Microsoft Research developed strategies for using generative AI to help synthesize biomedical data, enabling AI-expert collaboration to increase the diagnoses of rare diseases.

Developers who are blind or have low vision are often limited to back-end programming. But with vibe coding, an AI-assisted software development technique that uses natural language, developers who use screen readers can now perform tasks like UI development.

Concerned that open-source AI tools could be used to re-engineer toxins capable of evading existing biosecurity software, Microsoft Chief Scientific Officer Eric Horvitz initiated a project to show how AI could be used to strengthen biosecurity tools and prevent potentially deadly vulnerabilities.

SentinelStep, a new mechanism that enables AI agents to complete monitoring tasks over hours or days without getting sidetracked, is a first step toward practical, proactive, longer‑running agents.

In a wide-ranging interview with AI educator Harper Carroll, Bonnie Kruft, partner and deputy director of AI for Science at Microsoft Research, discusses how her team is using AI to address real-world problems from climate change to drug discovery.

RedCodeAgent is the first fully automated and adaptive agent for red-teaming LLM-based code agents to identify safety vulnerabilities. BlueCodeAgent is an end-to-end framework that strengthens code security by using automated red-teaming processes, data, and safety rules to guide LLM defenses.

Magentic Marketplace helps explore how AI agents impact the digital economy and how to guide them toward outcomes that serve the common good.

A collaboration between Microsoft Research and Signify—a leader in connected LED lighting products, systems, and services—improves the company’s enterprise knowledge systems, delivering faster, more reliable answers to customer questions.

The newest version of Data Formulator takes data exploration to the next level with Agentic Mode. People can update charts, get recommendations, and create reports grounded in their explorations.

Microsoft Principal Researcher Esha Choukse is recognized for foundational contributions to hardware memory compression and sustainable and efficient datacenter systems.

This native agentic small language model from Microsoft Research makes it possible to run computer use models directly on devices at a fraction of the cost of larger competitors.

As the Women in Machine Learning Workshop (WiML) marks its 20th annual gathering, cofounders, friends, and collaborators Jenn Wortman Vaughan and Hanna Wallach reflect on WiML’s evolution, navigating the field of ML, and their work in responsible AI.

Microsoft Research releases its Holoportation™ 3D telecommunications technology under an open-source license to facilitate future development by external researchers and organizations. 

From designing new materials to mapping flood risks through clouds, Microsoft researchers are using AI to solve problems more quickly and effectively than ever before.

This cutting-edge generative model creates high-fidelity 3D assets from a single image with full PBR materials in seconds. It is powered by a highly-capable O-Voxel representation for complex geometry and rich textures.