AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
Microsoft Research
First-party releases and research from Microsoft Research. Headlines and excerpts link to the original articles.
Latest stories
Newest first
Microsoft Research, UC Berkeley and Microsoft Security released TokenEconomizer, an open-source plugin that learns from an agent's past sessions in a repository and cut per-task costs by 18.4–30.9% across 277 SWE-Bench tasks without modifying the agent or its model.

Microsoft Research's newsletter roundup reports that a new study of 13.5 million coding sessions found fewer than 8% needed context compaction but those sessions consumed 44% of AI processing costs, and introduces CompactionBench plus APEC, which cut costs by up to 25.4%.

Microsoft researchers describe Lazarus, a backdoor that reprograms fewer than 10 dormant neurons in an LLM so an agent built on it runs attacker commands on a hidden trigger, with no retraining and while passing standard evals.

Microsoft Research introduced PDD Kit, a toolkit that turns plain-language intent into a formal Verus specification, proves agent-written Rust code meets it for every input, and falls back to tests where proofs cannot reach. In a demo on a roughly 5,000-line Rust storefront, it formally verified the login logic and uncovered existing bugs without line-by-line code review.

Microsoft Research Focus edition 97, hosted by Sowbhagya Suresh, covers a self-improving agent framework called ForkPilot that matched top methods on six benchmarks while cutting AI processing demands by up to 59%, plus new work on coding agents, agent skills from activity logs, weather forecasting, and a public preview of the BioEmu2 protein model.

Microsoft Research has introduced Flint-interactive, a language layered on its Flint visualization system that lets AI agents and developers specify chart interactions with high-level semantic events and update operators instead of low-level rendering code. It is available through Flint-MCP in chat environments and the Flint Interactive Canvas programming API for apps and dashboards.

Microsoft researchers and collaborators have proved the Kannan–Lovász–Simonovits conjecture, showing that hyperplane cuts capture the narrowest bottleneck of log-concave distributions within a universal constant bound regardless of dimension. The October 2026 proof followed a human-directed AI search and was confirmed by parallel independent proofs and a Lean 4 formalization.

Microsoft Research's podcast features Weishung Liu interviewing Dr. Christopher White, director of the Catalyst Lab, about his path from a Harvard postdoc to leading advanced tech operations in Afghanistan and later building dark-web search tools to combat human trafficking.

Microsoft Research unveiled Magnetar, a zero-knowledge virtual machine that proves programs written in Rust, Go or C, claims 10 to 39 times the speed of the fastest prior comparable systems, and relies only on standard hash functions for security.

Microsoft Research introduced StudentSim, a research system that trains personalized AI student simulators from public de-identified learner records so AI tutors can be evaluated and refined before human studies. In a chess tutor reinforcement-learning test, StudentSim-based reward reached 90.5% accuracy versus 75.7% with no RL and 71.6% using GPT-5.4 as the simulated student.

Microsoft Research Asia has open-sourced Agent Lightning v1.0, a Harnessed Agentic RL framework of roughly 3,500 lines that trains agents using their real production harness via an OpenAI-compatible LLM proxy. Its coding agent example lifted Qwen3.5-9B from 41.8% to 56.4% Pass@1 on SWE-bench Verified with about 6,000 training samples.

A Microsoft Research podcast conversation features Jennifer Neville, who leads the AI Interaction and Learning team, recounting research finding that model performance on single-turn benchmarks drops significantly when the same tasks are spread across multiple turns.