Official Microsoft Research & Azure Blog RSS + GitHub + Hugging Face.
39 updates · page 1 of 3
Microsoft Research introduces Magnetar, proof-based computing where a short zero-knowledge proof shows a program ran correctly and can be checked in milliseconds without rerunning it.
Why it matters: Anyone who needs to trust a computed result can check a short proof in milliseconds instead of rerunning the program.
Microsoft Agent Framework python v1.22.0 adds Python 3.15 support for the core framework, with some packages and extras still limited by upstream dependencies.
Why it matters: Python developers on 3.15 can use the core agent-framework package after checking each extra they rely on.
Microsoft Agent Framework dotnet v1.25.0 adds approval reuse considerations, ends chat completions SSE streams with data [DONE], includes citations and error details in purview checks, and fixes reasoning remnants and handoff bounce-back.
Why it matters: .NET developers on the Agent Framework can pick up the v1.25.0 stream, purview, and handoff fixes.
We’re proud to share that Microsoft has been named a Leader in the 2026 Gartner® Magic Quadrant™ for Server Virtualization Platforms .
Why it matters: A new model option to evaluate — compare quality, price, and context length against what you use today.
StudentSim is a Microsoft research system that trains personalized AI student simulators so tutoring tools can be evaluated before tests with real learners. It is not a classroom product and is not meant to replace teachers.
Why it matters: Tutor builders can try changes on simulated students before running slow, costly studies with real learners.
Azure Maia 300, the third generation, is a full system of software, silicon, racks, networking, and datacenters co-designed so each watt and dollar delivers more LLM inference.
Why it matters: People running models on Azure can read how Maia 300 is built around LLM inference.
Microsoft Agent Framework Python v1.21.0 adds tool-result standing guidance, rewritten variable-expansion arguments, full-buffer Purview checks for streamed responses, a maximum reasoning-effort option, and an alpha Oracle vector-store connector.
Why it matters: Python agents can cap reasoning effort and evaluate Purview policy on a full streamed buffer.
Microsoft Research Asia introduces Agent Lightning v1.0, a roughly 3,500-line agentic RL framework that trains agents with the same harness used in deployment, including native Kubernetes support.
Why it matters: Teams can train against the real harness instead of reimplementing the agent inside the trainer.
Microsoft describes AI across the infrastructure lifecycle, from supply chain to fleet operations, where silicon choices affect how hardware is sourced, deployed, and run.
Why it matters: Infrastructure teams can see how lessons from a running fleet feed the next build.
Microsoft Agent Framework .NET v1.24.0 keeps explicit null session state across roundtrips, honors disabled recent-search memory, preserves lazy request messages, and stabilizes the Anthropic agent package.
Why it matters: .NET agents can store an explicit null session and skip recent-search memory when it is turned off.
A Microsoft Research conversation with Jennifer Neville looks at what AI gets wrong and what those failures teach about real-world use.
Why it matters: Readers can hear how a research manager uses failure cases to judge AI outside the lab.
Microsoft Agent Framework Python v1.20.0 hosts native workflows over Foundry Responses and Invocations, and adds DuckDB, SQL Server, and TypeSafe AI vector-store connectors plus response-stream gates.
Why it matters: Python agents can store vectors in DuckDB or SQL Server and call TypeSafe AI from the same framework.
Microsoft Agent Framework .NET v1.23.0 pins origins on the Foundry toolbox MCP client, fixes workflow HTTP variables and topology checks, stores external input messages, and changes how tool updates apply between runs.
Why it matters: .NET workflow authors get origin pinning and clearer behavior when tools change between runs.
A Microsoft machine-learning pipeline uses forecast-time solar-wind data to estimate risk for 66,935 substations in the continental United States, combining AE and Dst forecasts with local latitude, geology, and ground conductivity. The pipeline detected nearly 80%.
Why it matters: Grid operators can get location-specific space-weather risk for those substations before a storm hits.
Microsoft introduced Quine, a research effort for a multimodal world model of biology and a harness linking models, scientific tools, literature, and researchers. With the Broad Institute, the system prioritized compounds predicted to drive therapeutic tumor-state shifts.
Why it matters: Biology researchers can use Quine to connect models and literature and rank compounds for tumor-state shifts.