A new open-weight AI agent gets dramatically better at office tasks when you show it, rather than just tell it, what to do.
UI-Mate pairs a large-scale training pipeline that auto-generates tasks, runs them across parallel virtual environments, and mixes supervised and reinforcement learning, with a mechanism for learning from demonstrations. Feed it a recorded example of a task and it breaks that recording into reusable subtask steps, follows the relevant ones, and improvises when the live screen does not match the recording. Its creators also built OSWorkerBench, a benchmark of 100 long, multi-step office tasks spread across 41 different applications, designed to test agents both with plain instructions and with demonstrations attached. On general computer-use benchmarks, the 27-billion-parameter UI-Mate model scores 77.0% on OSWorld-Verified and 66.2% on WindowsAgentArena, the best results reported yet for an open-weight model. On OSWorkerBench itself, it hits 41.0% strict success and 76.9% partial progress, beating its own Qwen3.6-27B base model by 17.7 and 24.5 points respectively.
The demonstration piece is the real story. On OSWorkerBench's 33-task subset, where the demo shows the exact same task the agent is then asked to do, a single example nearly doubles strict success, from 17.2% to 35.4%, and pushes progress from 67.9% to 81.1%. That tracks with why GUI agents actually fail in the wild: not because they cannot reason, but because unstated, person-specific workflow habits, like which button, which order, which shortcut, are impossible to guess from a text instruction alone.
Worth noting: OSWorkerBench also includes a 45-task setting built from human recordings of related but different tasks, a harder and more realistic test of whether an agent can generalize from a demo rather than just replay one. The paper does not report scores for that setting, so how well this actually holds up outside near-identical repeats is still an open question.