Microsoft Research Lab – Asia
Nouvelles et reportages
SkillOpt: Agent skills as trainable parameters
| Yifan Yang, Xuemei Gao, Qi Dai, Bei Liu, Kai Qiu, Dongdong Chen, et Chong Luo
AI agents often fail because their instructions, or skills, are manually modified with no guarantee of improvement. Learn how SkillOpt turns skill editing into a training process, making agent behavior more reliable without changing model weights.
By Kinam Kim, Namiko Saito, Heecheol Kim, Katsushi Ikeuchi, Jaegul Choo and Yasuyuki Matsushita Vision-Language-Action (VLA) models enable broad manipulation capabilities by leveraging large-scale pretraining and robot demonstrations. However, imitation learning can cause small execution errors to accumulate over time, pushing…
When you see robots participating in running races or performing folk dances on stage, you might envision a future where a simple natural language command is all it takes for a robot to tidy up a desk, clean a room,…
Microsoft at NSDI 2026: Advances in large-scale networked systems
| Sujata Banerjee
Microsoft researchers share advances in building and operating large-scale distributed systems, spanning datacenters, networking, and the growing intersection with AI during NSDI ’26.
Imagine an AI assistant that can navigate a computer the same way humans do—clicking buttons, filling out forms, and moving between applications—all by simply interpreting what’s on the screen. This vision is becoming a reality through computer use agents—AI systems…
Extracting useful information from long videos, whether meeting recordings, experimental data, or lecture content, requires painstaking manual review. AI tools offer some help: language-vision models can summarize short clips or answer questions when videos are divided into clear scenes or…
Agent Lightning: Adding reinforcement learning to AI agents without code rewrites
| Xufang Luo, Yuge Zhang, Zhiyuan He, Zilong Wang, Dongsheng Li, Luna K. Qiu, et Yuqing Yang
By decoupling how agents work from how they’re trained, Agent Lightning turns each step an agent takes into data for reinforcement learning. This makes it easy for developers to improve agent performance with almost zero code changes.
“Curiosity drives scientific breakthroughs, and the tools we create often reflect the human motivations behind that curiosity.” For Yansen Wang, a senior researcher at Microsoft Research Asia, this philosophy has guided his work at the intersection of AI and neuroscience.…
AI assistants, designed to perform actions on behalf of users, may not be as capable as current benchmarks suggest. New research reveals that existing tests for UI grounding—the ability of assistants to locate elements in the graphical user interface (GUI)—have…