Reproduce “OSWorld-Pro: Process-based Evaluation for Computer Use Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhilin Wang, Shaokun Zhang, Yifan Zhang, Hao Zhang, Jin Xu, Binfeng Xu, Jian Hu, Yunheng Zou and 4 more
OSWorld-Pro evaluates computer-use agents on desktop tasks by decomposing them into ordered subgoals and labeling whether actions target feasible milestones, advance them, or complete them. It contains 305 tasks and 2,814 subgoals. Matched Minimax M3 scores 28.9% on OSWorld-Pro versus 75.2% on OSWorld, though differing tasks and protocols limit the comparison; its LLM judge is weaker on feasibility and progression.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.