Reproduce “Back to Basics: A Skill-Centric Approach to Generalist 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
Minh Pham-Dinh, Md Farhan Ishmam, Hernan Yupanqui, Reaz Hassan Joarder, Kenneth Marino
MiniWeb evaluates browser agents at reusable-interaction level, checks each required interaction, and trains compact agents on verified skill practice for a planner to compose into tasks. For Qwen3.5-4B, a trained executor with an untrained planner raises held-out task success from 45.2% to 58.1% on 62 human-authored tasks. Skill training alone barely improves full-task success, so evidence supports combining practice with planning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.