Microsoft Research

Microsoft Research is Microsoft’s research organization, conducting fundamental science and technology research and collaborating with academic researchers and Microsoft product teams. Its AI work includes TinyStories, the Phi family of language models and AutoGen, a framework for coordinating agents powered by language models, humans and tools.[4][2][3][7]

Ronen Eldan and Yuanzhi Li’s TinyStories, listed as a May 2023 publication, uses GPT-3.5 and GPT-4 to generate short stories with vocabulary typically understood by three- to four-year-olds. The authors reported fluent, consistent multi-paragraph stories from models below 10 million parameters or architectures with just one transformer block. These were results within the simplified story setting. They also proposed using GPT-4 to grade generated stories on dimensions including grammar, creativity and consistency.[2]

The June 2023 Textbooks Are All You Need introduced phi-1, a 1.3-billion-parameter Transformer for code. Its larger author team included Suriya Gunasekar, Sébastien Bubeck, Eldan and Li. The team reported training for four days on eight A100 GPUs, using 6 billion tokens of selected “textbook quality” web data and 1 billion tokens of GPT-3.5-generated textbooks and exercises. Reported pass@1 accuracy was 50.6% on HumanEval and 55.5% on MBPP; the 350-million-parameter phi-1-small achieved 45% on HumanEval with the same pipeline.[3]

The December 2024 Phi-4 Technical Report extended the emphasis on data quality to a 14-billion-parameter language model, incorporating synthetic data throughout training while making minimal changes to the phi-3 architecture. Its authors attributed performance gains to improved data, training curriculum and post-training techniques. They reported that phi-4 surpassed its GPT-4 teacher on STEM-focused question answering—a task-specific comparison.[8]

Microsoft Research announced AutoGen on September 25, 2023, as freely available preview Python software. Developers define agents’ specialized capabilities and roles, then specify how they respond to messages from other agents. Automated conversations connect language models with tools and human participation, allowing workflows to include code generation, execution, feedback and troubleshooting.[7]

A supply-chain question-answering example makes the division of work concrete: a Commander coordinates the workflow and executes code, a Writer produces code and interpretations, and a Safeguard checks safety. Steps can repeat when problems arise. AutoGen emerged as an open-source, community-driven spinoff from FLAML, with contributors including Pennsylvania State University, the University of Washington, and Microsoft’s Fabric and ML.NET product teams.[7]

Microsoft Research founded AI Frontiers in October 2023, after AutoGen’s September announcement. The mission-driven lab works on model optimization, synthetic data, small language models and agent frameworks. Its research agenda includes agents that learn and adapt, reliability under real-world conditions, and collaboration between autonomous agents and humans.[6][7]

On March 4, 2026, Microsoft Research announced Phi-4-reasoning-vision-15B, a distinct 15-billion-parameter open-weight multimodal model. It combines the SigLIP-2 Naflex vision encoder with the Phi-4-Reasoning backbone. Its mid-fusion architecture converts images into visual tokens and projects them into the language model’s embedding space, reusing pretrained components. The team selected the dynamic-resolution encoder after experiments with a smaller proxy model, addressing perception problems such as locating small interactive elements in dense screenshots.[9]

Training combined filtered and repaired open-source data with internal and acquired datasets. Incorrect answers and poor captions were regenerated with GPT-4o and o4-mini; persistently flawed datasets were excluded. Supervised fine-tuning mixed approximately 20% reasoning examples with non-reasoning examples, aiming for direct responses on perception tasks and longer reasoning on mathematics and science. The authors cautioned that the boundary between modes is learned implicitly and that the 20/80 split may not suit every domain.[9]

The announcement’s authors were Jyoti Aneja, Michael Harrison, Neel Joshi, Tyler LaBonte, John Langford and Eduardo Salinas. They credited Rachel Ward for data collection and curation; GenDatasets, PhiGround, SimCity and Fara-7B for training data; and Harkirat Behl, Mojan Javaheripi and Suriya Gunasekar for checkpoints and training guidance. The release made weights, fine-tuning code and benchmark logs available under a permissive license. Its efficiency comparison averaged four benchmarks, and the authors presented their measurements as comparison and analysis rather than leaderboard claims.[9]

These AI projects sit within a longer institutional history. Microsoft Research’s combination of academic collaboration and research transfer into products was already established by its twentieth anniversary in 2011.[5]

Bill Gates, Rick Rashid and Nathan Myhrvold created Microsoft Research in 1991, with a mission to advance computing through basic and applied research. The organization paired an open academic model with a process for transferring results into Microsoft products.[5]

Research contributions included Office speech recognition and spelling and grammar tools, Bing search, spam filtering, Xbox online gaming, and Kinect body tracking and voice technology. Kinect Fusion, associated with the UK lab, used Kinect sensor data to create 3D models. Tiger, from MSR Asia, supplied new search-indexing capabilities to Bing.[5]

In September 2011, Microsoft reported 12 research facilities, more than 850 researchers across over 60 fields, and more than 30 joint research institutes. These are historical figures. Microsoft’s annual corporate R&D investment was described as $9 billion, but the vast majority funded product development; that amount was not Microsoft Research’s standalone budget.[5]

Peter Lee recalled that Microsoft Research gained access to GPT-4 through its OpenAI collaboration toward the end of 2022 and had to investigate it in secret, an unusual constraint for an organization accustomed to external research collaboration. Public discussion and publication of its initial findings became possible after OpenAI released GPT-4 in March 2023, leading to Sparks of Artificial General Intelligence. Lee also described Microsoft Research’s work with the Bing team on guardrails and risks surrounding GPT-4’s integration into products.[3]

Microsoft Research operates across regional labs and mission-oriented organizations. In its organization description, it reports that its research community and collaborators published tens of thousands of papers and released hundreds of open-source projects and datasets over thirty years. It also maintains an index of datasets, models, SDKs, APIs and other research code shared with the academic community.[4]