Reproduce “WebChain: A Large-Scale Human-Annotated Dataset of Real-World Web Interaction Traces” 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
Sicheng Fan, Rui Wan, Yifei Leng, Gaoning Liang, Li Ling, Yanyi Shang, Dehan Kong
Researchers introduced WebChain, a large-scale, open-source dataset of over 318,000 human-annotated web interaction traces from real-world websites, designed to advance web agent development. The dataset's multi-modal "Triple Alignment" and associated training paradigms demonstrate improved spatial grounding and long-horizon planning capabilities, achieving competitive performance on various GUI control benchmarks.
Key insights
- The scale of training data positively correlates with the development of robust long-horizon planning capabilities in web agent models.
- Combining Visual Grounding Densification (VGD) and Reasoner Prompting (RP) significantly enhances spatial grounding performance by diversifying training patterns and reducing spatial hallucinations.
- A "Dual Mid-Training" paradigm, which disentangles spatial perception (SGRL/VGD) from temporal planning (CoT-SFT) before final Reinforcement Learning (LCRL), effectively optimizes web agents for complex, long-horizon tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.