Reproduce “AsyncWebRL: Efficient Asynchronous Reinforcement Learning for Multi-Step Visual Web 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
Hao Bai, Rui Yang, Chenlu Ye, Spencer Whitehead, Aviral Kumar, Tong Zhang
AsyncWebRL combines an asynchronous training system with a trajectory-weighting change for visual web agents. It overlaps rollouts, gradient updates, and policy refresh, while the loss replaces step-count normalization with a constant. On WebGym’s Instruct out-of-distribution test set, it reaches 45.4% versus 42.9%, with trajectory throughput and shorter trajectories at tied reward, but results are single-seed and combine changes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.