Weblica: Scalable and Reproducible Training Environments for Visual Web Agents

2026-07-07 · Source: Apple Machine Learning Research · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

Weblica (Web Replica) is a novel framework designed to create reproducible and scalable training environments for visual web agents, addressing the inherent complexity and dynamic nature of the web. Traditional data collection methods are limited to offline trajectories or a few simulated environments, failing to capture sufficient web diversity. Weblica overcomes this by employing HTTP-level caching to capture and replay stable visual states while maintaining interactive behavior, alongside LLM-based environment synthesis grounded in real-world websites and core web navigation skills. This framework enables scaling Reinforcement Learning (RL) training to thousands of diverse environments and tasks. The resulting model, Weblica-8B, demonstrates superior performance compared to open-weight baselines of similar size on multiple web navigation benchmarks, utilizes fewer inference steps, scales effectively with increased test-time compute, and competes favorably with API models.

Key takeaway

For Machine Learning Engineers developing visual web agents, Weblica offers a robust solution to the challenges of data scarcity and web dynamism. You should consider integrating its HTTP-level caching and LLM-based environment synthesis to scale your RL training effectively. This approach allows you to build agents that perform competitively on diverse web navigation benchmarks, potentially reducing inference steps and improving scalability compared to existing open-weight baselines.

Key insights

Weblica provides scalable, reproducible web training environments using HTTP caching and LLM synthesis, enabling robust visual web agent development.

Principles

• Web complexity demands scalable, diverse training.
• Caching visual states preserves interactivity.
• LLM synthesis enhances environment diversity.

Method

Weblica constructs environments by leveraging HTTP-level caching for stable visual state replay and LLM-based synthesis to generate diverse, real-world grounded web navigation tasks for RL training.

In practice

• Scale RL training for web agents.
• Benchmark web navigation models.
• Develop agents for dynamic web tasks.

Topics

• Visual Web Agents
• Reinforcement Learning
• Weblica Framework
• HTTP Caching
• LLM Synthesis
• Web Navigation

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Weblica: Scalable and Reproducible Training Environments for Visual Web Agents — Weblica enables scalable, reproducible web agent training via HTTP caching and LLM-driven synthetic environment generation.
• OpenWebRL: Demystifying Online Multi-turn Reinforcement Learning for Visual Web Agents — Online multi-turn reinforcement learning on live websites offers a scalable path for training capable visual web agents.
• Beyond Domains: Reusing Web Skills via Transferable Interaction Patterns — Web agent skill reuse is significantly improved by matching layout structure rather than specific element references.
• Scaling Web Agent Training through Automatic Data Generation and Fine-grained Evaluation — A constraint-based evaluation framework enables scalable, high-quality web agent training by leveraging partially successful trajectories.
• The Infrastructure Behind Making Local LLM Agents Actually Useful — Local LLM agents require deliberate infrastructure for performance and long-session stability beyond basic model serving.
• LongWebBench: Evaluating Structural and Functional Webpage Generation in Long-Horizon Settings — Evaluating long webpage generation requires assessing executable multi-step interactions beyond visual similarity.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Apple Machine Learning Research.