Weblica: Scalable and Reproducible Training Environments for Visual Web Agents

Short summary

Weblica introduces a framework for constructing reproducible and scalable web training environments by combining HTTP-level caching with LLM-based environment synthesis. The resulting Weblica-8B model outperforms open-weight baselines of similar size on web navigation benchmarks while using fewer inference steps. It demonstrates favorable scaling with additional test-time compute and remains competitive with larger API-based models.

• •Framework scales web agent training across thousands of diverse environments using HTTP caching and LLM synthesis
• •Weblica-8B outperforms open-weight baselines on navigation benchmarks with fewer inference steps
• •Competitive with larger API models while improving efficiency and scaling with additional compute

Generated with AI, which can make mistakes.