Reproduce “Computer-Use Agents Gain Little Benefit from a World Model When Conditioned on Tutorials” 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
Anh Nguyen, Tianyi Xu, Phat Tran, Harry Yu, Huazheng Wang, Stefan Lee
The authors introduce RAG-PolicyJudge (RAG-PJ), a computer-use agent that gives task-specific tutorials to both an action-proposing policy and an action-selecting judge, then tests whether world-model previews improve task success. Across OSWorld-Tutorial and ScienceBoard-Tutorial, tested rollouts do not significantly improve success when tutorials are present, though rollouts can help without them; this finding is limited to the evaluated agents and settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.