Qwen-AgentWorld: Language World Models for General Agents

Researchers have introduced Qwen-AgentWorld, a language world model designed to simulate agentic environments across seven domains using long chain-of-thought reasoning, according to a paper posted on arXiv [1]. The model, developed in two sizes — 35B-A3B and 397B-A17B — is described as the first language world model capable of such broad simulation [1][2]. A world model predicts environment dynamics based on current observations and actions, functioning as a core cognitive mechanism for reasoning and planning [2]. The Qwen-AgentWorld system was built using more than 10 million environment interaction trajectories gathered from real-world settings [1][2]. The training process follows a three-stage pipeline: continued pre-training (CPT) injects general-purpose world modeling capabilities from state transition dynamics and augmented professional corpora; supervised fine-tuning (SFT) activates next-state-prediction reasoning; and reinforcement learning (RL) sharpens simulation fidelity through a tailored framework with hybrid rubric-and-rule rewards [1][2]. To evaluate the model, the team constructed AgentWorldBench, a benchmark built from real-world interactions of five frontier models on nine established benchmarks [1][2]. Empirical results show Qwen-AgentWorld significantly outperforms existing frontier models [1][2]. The work explores two complementary paradigms for enhancing general agents. As a decoupled environment simulator, Qwen-AgentWorld supports scalable and controllable simulation of thousands of real-world environments for agentic RL, yielding gains that surpass training on real environments alone [1][2]. As a unified agent foundation model, world-model training acts as a warm-up that improves downstream performance across seven agentic benchmarks [1][2]. Intelligent agents are entities that perceive their environment and take autonomous actions to achieve goals, with agentic AI expanding this concept by proactively pursuing objectives over extended periods [5]. Reasoning models, a related class of large language models, are specifically trained to solve complex tasks requiring multiple logical steps and can revise earlier reasoning during inference [3]. The Qwen-AgentWorld paper builds on these concepts by using language models as the backbone for world simulation, a departure from traditional approaches that rely on symbolic or physics-based simulators [2][5]. Code for the project has been made available on GitHub [2].