STEVE: A Step Verification Pipeline for Computer-use Agent Training

Benchmark Model Rank Results
computer-use-agents-on-mind2web-cross-domainSTEVE UI-Grounding + GPT-4o (extra data)#3Element Accuracy: 46.0Step Success Rate: 41.2
computer-use-agents-on-mind2web-cross-taskSTEVE UI-Grounding + GPT-4o (extra data)#3Element Accuracy: 46.2Step Success Rate: 40.0
computer-use-agents-on-mind2web-cross-websiteSTEVE UI-Grounding + GPT-4o (extra data)#3Element Accuracy: 44.4Step Success Rate: 37.7