Weblica: Scalable and Reproducible Training Environments for Visual Web Agents

Researchers introduce Weblica, a framework for creating reproducible and scalable web environments to train visual web agents at scale. The system uses HTTP-level caching and LLM-based synthesis to generate thousands of diverse training environments, with the resulting Weblica-8B model achieving competitive performance against larger API-based models on web navigation benchmarks.

Weblica addresses a fundamental challenge in AI development: scaling training data for agents that must interact with the open-ended, constantly evolving web. Traditional approaches rely on offline trajectory data or limited simulated environments, neither of which adequately captures the diversity required for robust web agents. This framework's innovation lies in combining HTTP-level caching—which preserves interactive behavior while maintaining reproducibility—with LLM-driven environment synthesis grounded in real websites.

The research builds on the growing recognition that web agents represent a valuable AI frontier. As applications increasingly automate web-based tasks, the ability to train agents at scale becomes commercially significant. Previous limitations stemmed from the technical difficulty of capturing web interactions in reproducible ways without massive manual annotation. Weblica's synthetic approach sidesteps these bottlenecks by leveraging existing LLMs to generate diverse, realistic training scenarios.

The performance results carry practical implications for developers building autonomous AI systems. Weblica-8B, a relatively modest 8-billion parameter model, matches or exceeds larger open-weight competitors while requiring fewer inference steps, suggesting more efficient resource utilization. Its competitiveness with proprietary API models demonstrates that thoughtful training data construction can offset raw model size—a key finding for organizations optimizing cost and latency.

The framework's scalability to thousands of environments points toward future applications in robotics, automation, and human-AI collaboration tools. However, questions remain about real-world generalization beyond synthetic environments and how well these agents handle novel website designs or edge cases not represented in training data.

• →Weblica enables reproducible, scalable web training environments using HTTP caching and LLM-based synthesis, addressing data scarcity in web agent research.
• →Weblica-8B achieves competitive performance with larger models and API-based systems while using fewer inference steps, improving efficiency.
• →The framework scales to thousands of diverse environments and tasks, significantly expanding training data diversity compared to existing approaches.
• →HTTP-level caching preserves interactive behavior while maintaining reproducibility, a technical advantage over previous web simulation methods.
• →The approach suggests that thoughtful training data construction can reduce model size requirements for web navigation tasks.