UI-Mate: Advancing Open-Weight Foundation GUI Agents with In-Context Demonstrations
A 27B open-weight GUI agent roughly doubled strict success when given one relevant demonstration.
UI-Mate pairs an environment-grounded training stack with in-context demonstrations that can be reused as subtask-level workflows. The paper reports 77.0% on OSWorld-Verified and 66.2% on WindowsAgentArena. On its OSWorkerBench benchmark, UI-Mate-27B reached 41.0% strict success and 76.9% progress. In the 33-task self-demo subset, one demo lifted strict success from 17.2% to 35.4%. HF Daily Papers' note
UI-Mate pairs an environment-grounded training stack with in-context demonstrations that can be reused as subtask-level workflows. The paper reports 77.0% on OSWorld-Verified and 66.2% on WindowsAgentArena. On its OSWorkerBench benchmark, UI-Mate-27B reached 41.0% strict success and 76.9% progress. In the 33-task self-demo subset, one demo lifted strict success from 17.2% to 35.4%. HF Daily Papers' note
score 5