Microsoft Research Blog

Researchers introduce generative causal testing, which translates black box models into clear hypotheses and verifies them in the scanner, revealing what specific brain regions respond to in language.

Project Ire, Microsoft's autonomous malware-classification agent, reverse-engineered a LOTUSLITE variant that went undetected by most major EDR tools. Through behavioral analysis rather than signature matching, Ire identified the sample's malicious intent and produced a detailed function-level report consistent with Acronis's published analysis.

Data Formulator 0.7 is an open-source AI-powered system for enterprise data analytics that combines data connectivity, agent-guided exploration, and visualization refinement in a shared workspace.

Modern AI systems are powerful not because they replicate human intelligence, but because they extend structures already present in human cognition and language. This perspective explains AI's capabilities and limitations, and reframes AI safety as a system-level challenge requiring engineering and governance, not fear of rogue AI.

Microsoft Research releases MagenticLite, an agentic application designed for small models, along with MagenticBrain orchestrator and Fara1.5 computer-use model. The system works across browser and local file system, achieving state-of-the-art results on web navigation tasks while keeping data on-device.

Vega is a new zero-knowledge proof system from Microsoft Research that enables users to prove facts from government-issued credentials without revealing the credential itself. It achieves under 100ms proving time on commodity devices using folding schemes, and is designed for real-world digital identity formats like mobile driver's licenses and the EU Digital Identity Wallet.

Microsoft Research clarifies the scope of its paper on AI delegation, noting that while models show fidelity degradation in long-horizon tasks, production systems mitigate these effects, and the benchmark is a diagnostic tool for future improvement.

mimalloc is an open-source, modern, scalable memory allocator that is a drop-in replacement for malloc and free. It is relatively small (~12K lines), with clear internal data structures, and is easy to build and integrate into other projects. It provides bounded worst-case allocation times (up to OS primitives), bounded space overhead, low internal fragmentation, and minimal contention by relying almost exclusively on atomic operations.

Microsoft releases a lightweight foundation model that can predict AC optimal power flow in milliseconds, boosting efficiency and unlocking cost savings in grid analysis.

Microsoft Research introduces SocialReasoning-Bench, a benchmark evaluating AI agents' social reasoning in principal-agent settings. Tests show frontier models complete tasks but often fail to secure optimal outcomes for users, even with explicit instructions. The benchmark measures outcome optimality and due diligence to assess agents' ability to act in users' best interests.

Microsoft Research releases an open dataset of U.S. power grid transmission topology derived from public data, enabling AC optimal power flow analysis and addressing research challenges due to restricted grid data. The pipeline uses OpenStreetMap and public energy data to create geographically grounded models that are solvable for power flow analysis, demonstrated across 48 states and the Eastern Interconnection. The dataset supports studies of congestion, transmission expansion, and demand siting.

Microsoft researchers share advances in building and operating large-scale distributed systems, spanning datacenters, networking, and the growing intersection with AI during NSDI '26.

Microsoft Research red-teamed a live platform of over 100 AI agents, identifying network-level risks that only appear through agent interactions, including self-propagating worms, reputation manipulation, manufactured consensus, and proxy chains. These risks cannot be reproduced by testing agents in isolation. The study also observed emergent security behaviors in a small fraction of agents, reducing attack success. Findings suggest the need for layered defenses across platform, agent, and model layers.

AutoAdapt automates domain adaptation for LLMs in high-stakes settings, turning weeks of manual iteration into repeatable pipelines. It uses a configuration graph, agentic planner, and budget-aware optimization to select appropriate strategies (e.g., RAG, fine-tuning) and tune hyperparameters under real constraints.

The 2025 New Future of Work report from Microsoft Research finds that generative AI is rapidly transforming work, but its benefits are unevenly distributed. AI is changing how people collaborate, and human expertise becomes more important. Organizations that treat AI as a partner see the biggest gains. The report calls for inclusive AI adoption to prevent widening divides.

Microsoft Chief Scientist Jaime Teevan and researchers Jenna Butler, Jake Hofman, and Rebecca Janssen unpack the New Future of Work Report 2025 and explore the ideal AI-driven working world. Plus, is AI a tool or a collaborator? And why the answer matters.

ADeLe, developed by Microsoft Research in collaboration with Princeton and Universitat Politècnica de València, scores AI models and tasks across 18 core abilities, enabling prediction of performance on unseen tasks with ~88% accuracy. It reveals model strengths and weaknesses, providing explainable AI evaluation beyond traditional benchmarks.

AsgardBench is a new benchmark that tests whether embodied AI agents can adjust their plans based on visual feedback. Built on AI2-THOR, it places agents in kitchen-like scenarios and requires them to dynamically modify action sequences by observing object states (e.g., whether a mug is clean). Tests show that visual input significantly boosts success rates, but current models still struggle with fine-grained visual distinctions, progress tracking, and timely plan updates.

Microsoft Research introduces GroundedPlanBench, a benchmark to evaluate vision-language models on joint action planning and spatial grounding for robot tasks. Their V2GP framework converts robot demonstration videos into training data, showing that grounded planning outperforms decoupled approaches.

Are machines truly intelligent? AI researchers Subutai Ahmad and Nicolò Fusi join Doug Burger to compare transformer-based AI with the human brain, exploring continual learning, efficiency, and whether today’s models are on a path toward human intelligence.

Microsoft Research open-sources AgentRx, a framework for automated diagnosis of AI agent failures. It pinpoints the first critical failure step using constraint synthesis and guarded evaluation, improving localization by 23.6% over baselines. The accompanying benchmark includes 115 annotated failed trajectories across three domains.