Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper introduces WebWorld, a large-scale open-web simulator trained on over 1 million web interaction trajectories to address the data scarcity and practical limitations of training web agents in real-world environments. WebWorld demonstrates strong simulation performance, comparable to Gemini-3-Pro on the introduced WebWorld-Bench, and improves the performance of Qwen3-14B by 9.2% on WebArena, achieving performance comparable to GPT-4o. The simulator also exhibits cross-domain generalization capabilities and enables effective inference-time search, surpassing GPT-5 as a world model.
Training web agents in a simulator can now match real-world performance: Qwen3-14B, fine-tuned with WebWorld-synthesized trajectories, rivals GPT-4o on WebArena.
Web agents require massive trajectories to generalize, yet real-world training is constrained by network latency, rate limits, and safety risks. We introduce 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.
DAMO
Tsinghua AI