Aurora 1.5 scales open forecasting; Flint redefines visualization; FlowDAgger accelerates adaptation. Plus: AI agents crack Rowhammer defenses and memory reshapes conversational AI. https://t.co/aEuukEm5dB https://t.co/s1eEBczlkJ

22 new weather variables, rapid post‑disaster building damage assessment, and robots that learn from human corrections.

Source: https://www.linkedin.com/pulse/22-new-weather-variables-rapid-postdisaster-building-lkhue/ Microsoft Research

Microsoft Research

We advance science and technology to benefit humanity.

Published Jul 20, 2026

In this issue:

• Aurora 1.5 extends open foundation models for weather and Earth‑system forecasting with 22 new variables, hourly resolution, and ensemble prediction
• Flint introduces a visualization intermediate language that lets AI agents create expressive charts from simple specs
• AI agents demonstrate how obscurity fails as a Rowhammer defense by reverse‑engineering simulated hardware protections
• HASTE enables rapid post‑disaster building damage assessment from satellite imagery through a no‑code web platform
• New research on conversational memory reveals how different memory roles shape personalization and factual accuracy in AI assistants
• FlowDAgger accelerates robot adaptation through human‑in‑the‑loop corrections in latent space

NEW RESEARCH

https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Emicrosoft%2Ecom%2Fen-us%2Fresearch%2Fblog%2Faurora-1-5-extending-open-foundation-models-for-weather-and-earth-system-applications%2F&urlhash=J80U&trk=article-ssr-frontend-pulse_publishing-image-blockAurora 1.5 is an update to Microsoft’s Aurora Earth System model. It introduces new capabilities that make open Earth-system forecasting models more practical and scalable for organizations that rely on atmospheric and environmental intelligence. As climate and weather-related risks increasingly affect communities, infrastructure, and economies, more accurate forecasting can help improve preparedness and decision-making.

The update adds 22 new weather variables relevant to a multitude of applications including energy, agriculture, transportation, and climate-risk analysis, along with hourly temporal resolution and probabilistic ensemble forecasting—a capability frequently requested by users. Ensemble forecasting runs multiple simulations to estimate the range and likelihood of possible weather outcomes.

NEW RESEARCH

https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Emicrosoft%2Ecom%2Fen-us%2Fresearch%2Fblog%2Fflint-a-visualization-language-for-the-ai-era%2F&urlhash=gCXD&trk=article-ssr-frontend-pulse_publishing-image-blockFlint is a visualization intermediate language for AI-driven chart creation. It helps AI agents create expressive, visually polished charts from simple, human-editable chart specs. Instead of requiring verbose low-level parameters for scales, axes, spacing, and layout, the Flint compiler derives optimized chart settings from the data, semantic types, chart type, and encodings.

Flint automatically manages sizing, spacing, labels, and layout so charts remain readable as cardinality and density change, without explicit user configurations. A single Flint specification can compile to multiple back-end systems without rewriting the chart from scratch. This open-source project includes a library and MCP server, so agents can create, validate, and render charts directly in chat or coding environments.

NEW RESEARCH

https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Emicrosoft%2Ecom%2Fen-us%2Fresearch%2Fpublication%2Fai-agents-can-defeat-security-by-obscurity-for-rowhammer-defenses%2F&urlhash=2cJH&trk=article-ssr-frontend-pulse_publishing-image-blockSecurity researchers have long understood that security by obscurity is not good security. Nevertheless, the hardware industry continues to keep the designs and configurations of many Rowhammer defenses secret. This research shows that AI agents can substantially reduce the value of this obscurity by reverse engineering a broad range of host-side Rowhammer defenses implemented in a simulator, even in a black box setting where the agent has no prior knowledge of the defense.

Rowhammer is a hardware security vulnerability in which accessing one DRAM row repeatedly can disturb nearby rows and flip bits without directly accessing them. These bit flips can corrupt data and potentially undermine system security.

This research evaluates simulated Rowhammer defenses under different threat models. The results show that agents can find effective Rowhammer attack patterns even when all implementation and configuration details are kept secret and unavailable to the agent. Although this work does not evaluate real hardware, it is a first step toward the automatic reverse engineering of hardware defenses and a wake-up call for an industry that continues to rely on design secrecy for their Rowhammer defenses.

NEW RESEARCH

https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Emicrosoft%2Ecom%2Fen-us%2Fresearch%2Fpublication%2Fhaste-a-platform-for-rapid-post-disaster-building-damage-assessment%2F&urlhash=VQWr&trk=article-ssr-frontend-pulse_publishing-image-blockThis research presents HASTE (High-speed Assessment and Satellite Tracking for Emergencies). It’s a no-code web platform that lets analysts produce building-specific damage maps from post-disaster satellite imagery, without expertise in machine learning. The ultimate goal is to help responders react more quickly to disasters, including earthquakes, hurricanes, cyclones, floods, wildfires, and tornadoes.

HASTE and its predecessors have supported more than thirty real-world disaster responses since 2023, delivering results to humanitarian partners within days or sometimes hours of imagery becoming available. The authors highlight promising avenues for improvement of this open-source technology, including vision-language assessment, active learning, and damage models for roads and other infrastructure.

NEW RESEARCH

https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Emicrosoft%2Ecom%2Fen-us%2Fresearch%2Fpublication%2Fmemory-makes-the-difference-evaluating-how-different-memory-roles-shape-conversational-agents%2F&urlhash=G2vB&trk=article-ssr-frontend-pulse_publishing-image-blockThis research provides new insight into how memory can be used to build more effective and personalized conversational AI systems. The authors examine how different types of conversational memory—information carried forward from previous interactions—shape the responses generated by AI assistants. While most prior work has focused on how memories are stored and retrieved, this research investigates whether different kinds of memories have distinct effects on response quality. Their findings show that performance depends not simply on retrieving relevant information, but on retrieving theright kindof memory. For example, clarifying memories improve factual accuracy and adherence to user constraints, while irrelevant memories can reduce relevance and personalization.

NEW RESEARCH

https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fmicrosoft%2Egithub%2Eio%2FFlowDAgger%2F&urlhash=bobw&trk=article-ssr-frontend-pulse_publishing-image-blockFlowDAgger helps robot foundation models adapt to new real-world situations from just a handful of human corrections, reducing the need for costly data collection and retraining. Today, when robots encounter conditions outside their training experience, adaptation often requires large amounts of new data or reinforcement learning on physical hardware. This process is slow, expensive, and can introduce safety risks. By enabling efficient adaptation while preserving existing capabilities, FlowDAgger allows robots to acquire new skills more quickly, operate more reliably, and be deployed in changing environments with less effort and risk.

Research Focus

Research Focus

145,148 followers

More articles by Microsoft Research

• Jul 7, 2026 ### Highlights from ICML 2026 Microsoft at ICML 2026 Microsoft is sponsoring The International Conference on Machine Learning (ICML), a premier…
• Jun 15, 2026 ### Hardware-Accelerated Analytics, AI-Powered Cancer Research, and the Future of Knowledge Work In this issue: CoddSpeed: A hardware acceleration framework delivering up to 30x faster analytics across GPUs, FPGAs…
• Jun 1, 2026 ### Microsoft Build, plus Agentic Evaluation, Repositories as the Medium, and AI Value Alignment Challenge In this issue: AgentPex: An open-source system that automatically detects procedural failures in agentic traces that…
• May 18, 2026 ### Cloud Efficiency, Developer Productivity, and LINGUA Africa call for proposals In this issue: HarvestContainers: A system that reclaims up to 75% of idle cloud CPU capacity without violating latency…
• May 4, 2026 ### Human-AI Collaboration, Enterprise Privacy, and Efficient Attention In this issue: Scaffolding Human-AI Collaboration: A field study with 388 employees finds that how AI use is structured…
• Apr 20, 2026 ### Microsoft at ICLR 2026: Deep Learning, LLM Reasoning, Generative Models Microsoft is proud to have more than 100 accepted papers and be a sponsor of the 14th International Conference on…
• Apr 6, 2026 ### Sentiment in context, robot dexterity, smarter AI agents, and CHI 2026 In this issue: Reasoning Beyond Labels: A new diagnostic framework tests how LLMs interpret sentiment in low-resource…
• Mar 23, 2026 ### MicroLED in datacenters, multimodal AI at scale, agent memory, and interactive planning In this issue: MicroLED datacenter networking: Lower-power, lower-cost optical communication for next-generation…
• Mar 9, 2026 ### Phi-4 multimodal reasoning, scalable LLM systems, cinematic video generation, and AI for health In this issue: Phi-4-reasoning-vision-15B: A compact multimodal reasoning model for vision-language tasks, now…
• Feb 23, 2026 ### Glass storage breakthroughs, robust reasoning, scalable safety testing, multilingual AI, and evaluating how models think In this issue: Project Silica: Advances in glass storage technology for long-term data preservation RLTR: Training LLMs…

Explore content categories