DeskForge: Dense Supervision from Desktop Environments for Computer-Use Agents

Benchmark Model Rank Results
computer-use-agents-on-ui-v-ui-visionDeskForge-Qwen3.5-4B#8Accuracy: 28.77
computer-use-agents-on-mmb-gui-mmbench-guiDeskForge-Qwen3.5-4B#11Accuracy: 76.82
computer-use-agents-on-osw-g-osworld-gDeskForge-Qwen3.5-4B#11Accuracy: 61.35
computer-use-agents-on-screenspot-v2DeskForge-Qwen3.5-4B#34Accuracy: 90.33
computer-use-agents-on-screenspot-proDeskForge-Qwen3.5-4B#39Accuracy: 41.68