Qwen-AgentWorld: Language World Models for General Agents

Qwen-AgentWorld introduces a new class of language world models designed to predict environment dynamics based on observations and actions, enabling general agents to reason and plan more effectively. By simulating agentic environments across seven distinct domains, Qwen-AgentWorld serves as both a scalable environment simulator for reinforcement learning and a foundation model that improves downstream agent performance.

Foundation Models for Environment Simulation

Qwen-AgentWorld provides two model scales, Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B, which are the first language world models capable of simulating agentic environments using long chain-of-thought reasoning. These models were trained on over 10 million environment interaction trajectories across seven domains.

The development of Qwen-AgentWorld followed a three-stage training pipeline:

• Continual Pre-training (CPT): This stage injects general-purpose world modeling capabilities by utilizing state transition dynamics and augmented professional corpora.
• Supervised Fine-Tuning (SFT): This stage activates the model's ability to perform next-state-prediction reasoning.
• Reinforcement Learning (RL): This stage improves simulation fidelity using a tailored framework that employs hybrid rubric-and-rule rewards.

AgentWorldBench Evaluation

To measure the performance of language world models, the researchers introduced AgentWorldBench. This comprehensive benchmark is constructed from real-world interactions of five frontier models across nine established benchmarks. Empirical results indicate that Qwen-AgentWorld significantly outperforms existing frontier models in its ability to simulate environment dynamics.

Enhancing General Agents via World Modeling

Qwen-AgentWorld enhances general agents through two primary paradigms:

Decoupled Environment Simulation

Qwen-AgentWorld can act as a decoupled environment simulator, allowing for the scalable and controllable simulation of thousands of real-world environments. When used for agentic reinforcement learning, this approach yields performance gains that surpass those achieved through training in real environments alone.

Unified Agent Foundation Model

Training as a world model serves as a highly effective warm-up for general agents. When Qwen-AgentWorld is used as a unified agent foundation model, it demonstrates improved downstream performance across seven different agentic benchmarks.