Scaling Computer-Use Grounding via User Interface Decomposition and Synthesis

Benchmark Model Rank Results
computer-use-agents-on-windowsagentarena-15-stepsJEDI-7B w/ GPT-4o#2Success Rate: 30.17
computer-use-agents-on-ui-v-ui-visionJedi-7B#10Accuracy: 24.8
computer-use-agents-on-windowsagentarena-50-stepsJEDI-7B w/ GPT-4o#10Success Rate: 32.8
computer-use-agents-on-osw-g-osworld-gJedi-7B#15Accuracy: 54.1
computer-use-agents-on-screenspot-v2Jedi-7B#27Accuracy: 91.7
computer-use-agents-on-screenspot-proJedi-7B#41Accuracy: 39.5