Weblica framework scales visual web agent training with HTTP caching and LLMs

Web agents—AI systems that can visually navigate and interact with web pages—are notoriously hard to train because the web is dynamic, diverse, and constantly changing. Existing approaches rely on limited offline trajectories for supervised fine-tuning or a handful of simulated environments for reinforcement learning. A new paper from researchers at ETH Zurich and Apple proposes Weblica (Web Replica), a framework that constructs reproducible and scalable web environments by combining two key techniques. First, HTTP-level caching captures and replays stable visual states while preserving interactive behavior—essentially freezing a realistic web page for repeated agent training. Second, LLM-based environment synthesis generates new tasks and environments grounded in real-world websites and core navigation skills, enabling RL training across thousands of diverse environments.

The team used Weblica to train Weblica-8B, an 8-billion-parameter model that sets a new bar for open-weight web agents. On benchmarks like Mind2Web, WebVoyager, and WebArena, Weblica-8B outperforms comparably sized open models (e.g., CogAgent-18B, SeeClick) while using fewer inference steps. It also scales favorably with additional test-time compute and competes with proprietary API models like GPT-4V. The paper, submitted to arXiv on May 7, 2026, includes 28 pages and 19 figures detailing the architecture and experiments. This work opens the door to more reliable, generalizable web automation for tasks ranging from data extraction to complex multi-step workflows.