WebWorld: A Large-Scale World Model for Web Agent Training

Web agents require massive trajectories to generalize, yet real-world training is constrained by network latency, rate limits, and safety risks. We introduce \textbf{WebWorld} series, the first open-web simulator trained at scale. While existing simulators are restricted to closed environments with thousands of trajectories, WebWorld leverages a scalable data pipeline to train on 1M+ open-web interactions, supporting reasoning, multi-format data, and long-horizon simulations of 30+ steps. For intrinsic evaluation, we introduce WebWorld-Bench with dual metrics spanning nine dimensions, where WebWorld achieves simulation performance comparable to Gemini-3-Pro. For extrinsic evaluation, Qwen3-14B trained on WebWorld-synthesized trajectories improves by +9.2\% on WebArena, reaching performance comparable to GPT-4o. WebWorld enables effective inference-time search, outperforming GPT-5 as a world model. Beyond web simulation, WebWorld exhibits cross-domain generalization to code, GUI, and game environments, providing a replicable recipe for world-model construction.

Imagine teaching an AI assistant to handle everyday web tasks like booking flights or filling forms. The best way is to let it practice on real websites, but this is slow and risky. One wrong click could submit a real payment. As a result, today's web assistants get too little practice and often fail at routine tasks. We built WebWorld, a flight simulator for the web. It learns from over one million real browsing sessions, then imagines how any webpage would respond to any action. An AI assistant can safely rehearse millions of tasks inside this simulated web at almost no cost. Assistants trained inside WebWorld become substantially better at real web tasks. We are releasing WebWorld openly to make safer and more capable AI assistants accessible to everyone.