Reproduce “WebXSkill: Skill Learning for Autonomous 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
Zhaoyang Wang, Qianhui Wu, Xuchao Zhang, Chaoyun Zhang, Wenlin Yao, Fazle Elahi Faisal, Baolin Peng, Si Qin and 7 more
WEBXSKILL, developed by researchers from the University of North Carolina at Chapel Hill and Microsoft, enables autonomous web agents to learn and deploy 'executable skills' that pair parameterized action programs with step-level natural language guidance. This framework consistently improved task success rates on web agent benchmarks, achieving 69.5% on WebArena with GPT-5 and demonstrating higher skill adoption compared to prior methods.
Key insights
- Decoupling reusable procedural knowledge (the skill) from execution control (deployment mode) allows for flexibility, leveraging strong LLMs for efficiency (grounded mode) and weaker LLMs for robustness (guided mode).
- Acquiring skills from synthetic agent trajectories, rather than costly exploration or test-set reverse engineering, is a viable and scalable strategy, producing a diverse and reliable skill library.
- Robustness is significantly improved by thorough skill verification and context-aware skill organization (via a URL-based graph), which together ensure skills are both functional and relevant for agent deployment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.