Qwen-AgentWorld: Language World Models for General Agents (arxiv.org)

Researchers have announced the development of Qwen-AgentWorld, a groundbreaking set of language world models designed for general agents. This new framework, including the Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B models, is notable for its ability to simulate agentic environments across seven distinct domains using advanced long chain-of-thought reasoning. By harnessing over 10 million real-world interaction trajectories, Qwen-AgentWorld employs a sophisticated three-stage training pipeline that integrates general-purpose world modeling, next-state-prediction reasoning, and reinforcement learning to enhance simulation fidelity. The significance of Qwen-AgentWorld lies in its dual capability as both a decoupled environment simulator and a unified foundation model for agents, allowing for scalable simulations that improve training outcomes. It surpasses existing models in performance, as demonstrated through the newly introduced AgentWorldBench benchmark, which tests interactions with five leading models across nine established benchmarks. This innovative approach not only improves the efficiency of training general agents but also suggests a paradigm shift in the AI/ML community towards more robust agent behaviors in complex, real-world settings.