EvoCUA: Evolving Computer Use Agents via Learning from Scalable Synthetic Experience

Benchmark Model Rank Results
computer-use-agents-on-osworld-verified-369-tasks-50-stepsEvoCUA-32B#1Success Rate: 56.7
computer-use-agents-on-osw-g-osworld-gEvoCUA-OpenCUA-72B#6Accuracy: 67.65
computer-use-agents-on-screenspot-proEvoCUA-OpenCUA-72B#10Accuracy: 63.24
computer-use-agents-on-screenspot-v2EvoCUA-OpenCUA-72B#17Accuracy: 93.47
multimodal-reasoning-on-mathvistaEvoCUA-32B#28Accuracy: 80.40
vision-language-understanding-on-mmstarEvoCUA-32B#28Accuracy: 73.20
multimodal-reasoning-on-mmmu-proEvoCUA-32B#31Accuracy: 59.16
optical-character-recognition-ocr-on-ocrbenchEvoCUA-32B#36Score: 853.5
multimodal-reasoning-on-mmmuEvoCUA-32B#39Accuracy: 68.11