AI News from Microsoft Research Blog

Latest coverage from Microsoft Research Blog, summarized and scored for signal.

• MindTopo reveals VLMs’ spatial reasoning abilities — Microsoft Research introduces MindTopo, a benchmark testing whether multimodal AI models understand topological relationships — connectivity, enclosure, order, separation, and knots — finding a large gap between static recognition ability and maintaining topology through interactive planning sequences.
• Introducing CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement — Microsoft Research's CARE-X unifies free-text radiology report generation with calibrated structured diagnostics in a single chest X-ray VLM, using DAPO reinforcement learning for clinical alignment and validated on real Indian hospital data including rare ICU pathologies.
• Orchard: An open framework for scalable agentic AI — Microsoft Research released Orchard, an open-source framework for training and evaluating AI agents across task types — Orchard-SWE reaches 69.7% on SWE-bench Verified (73.0% with reranking) using only ~3 billion active parameters, approaching frontier models more than 10x its size.
• Verifying Rust cryptography in SymCrypt, from standards to code — Microsoft's SymCrypt team is formally verifying post-quantum cryptographic algorithms — starting with SHA-3 and ML-KEM (NIST's 2024 post-quantum standard) — by writing code in safe Rust and machine-checking it against formal specifications in Lean, using AI agents to write proofs at scale.
• Aurora 1.5: Extending open foundation models for weather and Earth-system applications — Microsoft's Aurora 1.5 adds 22 new atmospheric variables, hourly forecast resolution, and probabilistic ensemble predictions to its open Earth-system foundation model — released on GitHub with Hugging Face checkpoints, connecting research to enterprise weather and climate intelligence.
• Flint: A visualization language for the AI era — Microsoft Research has released Flint, an open-source visualization intermediate language that lets AI agents generate polished charts from compact specs — the compiler infers scales, spacing, labels, and color schemes from semantic data types, then outputs to Vega-Lite, Apache ECharts, or Chart.js. An MCP server ships with it for direct agent workflow integration.
• SkillOpt: Agent skills as trainable parameters — Microsoft Research's SkillOpt treats agent instruction files as trainable parameters outside a frozen model, running a text-space forward-backward-update loop that tops all 52 evaluation cells across six benchmarks, seven models, and three execution modes — without updating any model weights.
• Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity — Microsoft's Memora separates memory storage from retrieval in AI agents, achieving state-of-the-art on LoCoMo and LongMemEval benchmarks while using 98% fewer context tokens than full-context inference — directly addressing the memory bottleneck limiting long-horizon agent deployments.
• Understanding the brain with AI-driven explanations and experiments — Microsoft Research and UC Berkeley introduce Generative Causal Testing (GCT), a framework that converts opaque LLM-based brain models into short verbal hypotheses about what each cortical region responds to — then verifies them in live fMRI scans, uncovering micro-regions tuned to dialogue, clock times, and measurements.
• Talos: Scaling rare disease diagnosis with automated, iterative genomic reanalysis — Microsoft Research and partners released Talos, an open-source AI system for automated rare disease genomic reanalysis that recovered 90% of diagnoses while flagging only 1.3 candidate variants per patient — delivering 241 new diagnoses across a 5,000-patient cohort with a median time-to-diagnosis of 32 days from new evidence publication.
• Ire identifies another LOTUSLITE specimen — Microsoft Research's autonomous malware-classification agent Project Ire identified a LOTUSLITE backdoor variant — attributed to Mustang Panda — that six major EDR vendors (CrowdStrike, SentinelOne, Sophos, Trellix, Palo Alto, ESET) missed, using decompiler-based behavioral analysis with no human input.
• Data Formulator 0.7: AI-powered data analytics for enterprise data — Microsoft Research released Data Formulator 0.7, an open-source AI analytics system that connects enterprise databases, warehouses, and BI tools through persistent Data Connectors, then lets context-aware agents prepare, analyze, and visualize data without requiring SQL or coding skills.
• Extending Human Intelligence Through AI — Microsoft Research argues that LLMs work precisely because they extend cognitive structures sedimented in human language — not because they replicate human intelligence — reframing AI safety as a governance challenge rather than a rogue AI problem.
• MagenticLite, MagenticBrain, Fara1.5: An agentic experience optimized for small models — Microsoft Research released MagenticLite (a browser-and-filesystem agentic app), MagenticBrain (a small orchestration model), and Fara1.5 (a 9B-parameter computer-use model that nearly doubles Fara-7B on web navigation benchmarks) — demonstrating that capable AI agents can run efficiently on small, locally-deployed models.
• Vega: Zero-knowledge proofs for digital identity in the age of AI — Microsoft Research's Vega system generates zero-knowledge proofs from government-issued credentials — proving age, professional status, or personhood without revealing the credential — in under 100ms on a commodity device, targeting EU digital identity mandates and AI agent authentication.
• Further Notes on Our Recent Research on AI Delegation and Long-Horizon Reliability — Microsoft Research's DELEGATE-52 benchmark finds frontier AI models degrade document and artifact fidelity by 19–34% over 20 delegated iterations — but Python workflows are a dramatic exception, showing under 1% degradation, quantifying a key reliability gap in agentic AI.
• mimalloc: A new, high-performance, scalable memory allocator for the modern era — Microsoft Research published a detailed overview of mimalloc, their open-source high-performance memory allocator that now ships as the default allocator for NoGIL CPython 3.13+, Unreal Engine, and games including Death Stranding — with its Rust wrapper seeing 100K+ downloads per day.
• GridSFM: A new, small foundation model for the electric grid — Microsoft Research released GridSFM, a small foundation model that solves AC optimal power flow in milliseconds — addressing decisions governing up to $20B/year in grid congestion costs and 3.4 TWh of wasted renewable energy — and releases an open transmission-topology dataset to go with it.
• Advancing AI for materials with MatterSim: experimental synthesis, faster simulation, and multi-task models — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• SocialReasoning-Bench: Measuring whether AI agents act in users’ best interests — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Building realistic electric transmission grid dataset at scale: a pipeline from open dataset — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Microsoft at NSDI 2026: Advances in large-scale networked systems — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Red-teaming a network of agents: Understanding what breaks when AI agents interact at scale — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• AutoAdapt: Automated domain adaptation for large language models — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Can we AI our way to a more sustainable world? — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.