WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents
Summary: arXiv:2603.05044v2
Abstract: Current paradigms for training GUI agents are fundamentally limited by a reliance on either unsafe, non-reproducible live web interactions or costly, scarce human-crafted data and environments. We argue this focus on data volume overlooks a more critical factor: the efficiency of compressing a large language model’s (LLM) latent knowledge into actionable agent behavior.
We introduce WebFactory, a novel, fully automated closed-loop reinforcement learning pipeline for GUI agents, systematically compressing LLM-encoded internet intelligence into efficient, grounded actions. Our pipeline features a process of scalable environment synthesis, knowledge-aware task generation, LLM-powered trajectory collection, decomposed reward RL training, and systematic agent evaluation.
Remarkably, our agent demonstrates exceptional data efficiency and generalization. Trained on synthetic data from only 10 websites within WebFactory, it achieves performance comparable to GUI agents trained on the same amount of human-annotated data from a much larger set of environments. This superior performance is consistent across our internal offline and online transfer benchmarks, where our agent also significantly outperforms the base foundation model.
We further provide critical insights into the “embodiment potential” of different LLM foundations, offering a new axis for model evaluation. This work presents a scalable and cost-effective paradigm for transforming passive internet knowledge into active, grounded intelligence, marking a critical step towards general-purpose interactive agents.
Key Features of WebFactory
- Automated Pipeline: WebFactory operates through a fully automated reinforcement learning pipeline that eliminates the need for manual data crafting.
- Scalable Environment Synthesis: The system generates diverse and complex environments automatically, ensuring high variability without the overhead of human input.
- Knowledge-Aware Task Generation: Tasks are designed with an understanding of the knowledge embedded in the LLM, promoting efficient learning.
- LLM-Powered Trajectory Collection: The agent collects trajectories using the latent knowledge from LLM, enhancing the quality of interaction data.
- Decomposed Reward RL Training: The training process breaks down rewards into manageable components, facilitating better learning outcomes.
- Systematic Agent Evaluation: Continuous evaluation allows for real-time adjustments and improvements to the agent’s performance.
Significance and Future Implications
The introduction of WebFactory stands to revolutionize the training of GUI agents, highlighting the potential for using existing internet knowledge in a more efficient manner. This new approach not only reduces reliance on expensive human-labeled data but also enhances the scalability of developing intelligent agents. As the field continues to evolve, the insights provided by this study may pave the way for the next generation of interactive AI systems capable of operating autonomously across various domains.