Reproduce “ComputerSD: Online Self-Distillation from Real-Time Feedback 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
Yong Du, Tongbo Chen, Zhengxi Lu, Yizhou Liu, Bofan Chen, Tao Jiang, Wenhao Xu, Yongliang Shen
ComputerSD trains computer-use agents by using a GUI analyzer to assess each action against the resulting screen state. Its step-level value judgment gates token-level learning signals from real-time guidance, alongside outcome-based GRPO. On OSWorld-Verified, it raises Pass@1 from 37.9% to 39.8% for Qwen3-VL-8B-Thinking and from 43.8% to 47.9% for EvoCUA-8B; results are averages over three runs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.