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

TL;DR

WebWorld is a large-scale open-web simulator trained on over a million interactions, enabling advanced reasoning, multi-format data handling, and long-horizon simulations, significantly improving web agent training and generalization.

Contribution

Introduces WebWorld, the first large-scale open-web simulator supporting extensive interactions and long-horizon simulations, with new benchmarks and cross-domain generalization capabilities.

Findings

WebWorld achieves simulation performance comparable to Gemini-3-Pro.

Training Qwen3-14B on WebWorld data improves WebArena performance by +9.2%.

WebWorld outperforms GPT-5 in inference-time search and generalizes across domains.

Abstract

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.