Alibaba's Qwen-AgentWorld uses language models to simulate agent environments

Alibaba's Qwen team has released Qwen-AgentWorld, a pair of language world models designed to simulate complex agentic environments purely through text-based reasoning. The models—Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B—are the first language models capable of simulating environments across 7 distinct domains (including robotics, navigation, and tool use) using long chain-of-thought reasoning. They were trained on over 10 million environment interaction trajectories from real-world environments using a three-stage pipeline: continual pretraining (CPT) injects general world modeling from state transitions and domain-specific corpora, supervised fine-tuning (SFT) activates next-state prediction logic, and reinforcement learning (RL) sharpens simulation fidelity via a hybrid rubric-and-rule reward framework.

To evaluate performance, the team created AgentWorldBench, a comprehensive benchmark built from real interactions by 5 frontier models across 9 established agent benchmarks. Qwen-AgentWorld significantly outperformed existing frontier models like GPT-4o and Claude 3.5 in predicting environment dynamics. Beyond simulation, the team demonstrated two complementary use cases: using Qwen-AgentWorld as a decoupled environment simulator to generate thousands of synthetic training scenarios for agentic RL—yielding gains beyond real-environment training alone—and using world-model training as a warm-up phase for unified agent foundation models, improving downstream performance across 7 agentic benchmarks.