Qwen has built a world model for training AI agents in Sim

Qwen team from Alibaba published Qwen-AgentWorld, a set of world models for simulating agent environments. There are two versions: Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B, both trained on over 10 million environment interaction trajectories across seven agentic domains (i.e., real-world tasks performed by real agents).

The core concept behind world models is that instead of interacting with real tools and observing the result, we can use a model to predict the consequence of taking actions. This allows agents to safely rehearse complex tasks in simulated environments. Similar to flight simulators, pilots learn how to fly in a much cheaper machine rather than real ones.

1What the benchmarks show

On nine classical benchmarks and five cutting-edge models, Qwen shows that AgentWorld outperforms significantly on AgentWorldBench, a benchmark specially crafted to evaluate the accuracy of world models in agent scenarios. They adopted a three-stage training process including general pre-training (named CPT), supervised fine-tuning on next state prediction, and reinforcement learning with hybrid reward for refining simulation quality. Over 27 members from Alibaba’s Qwen team collaborated on this work.

In the short term, world modeling for agents could help reduce the cost and improve the safety of developing AI agents. Currently, developing robust AI agents requires extensive iterations in real environments (e.g., making API calls, database queries, etc.), which could be time-consuming and risky. With a high-quality world model, fewer iterations are required to achieve trustworthy agent workflows. This paper from Qwen provides a solid foundation for large-scale adoption of world modeling for agents.

Researchers have been working on building world models for agents for a few years. Previous efforts mostly focused on specific domains or simplified scenarios. It’s exciting to see a top lab like Qwen publishing a 397B parameter model covering 7 real-world agentic domains and releasing their paper publicly. Although this won’t affect your work directly, it’s a foundational research project that helps make future agents more affordable, efficient, and safe to use. Save this one.