Paper 2606.24597
Qwen-AgentWorld: Language World Models for General Agents
01 In brief
Summary
Qwen-AgentWorld introduces the first language world models (LWMs) for simulating agentic environments across seven domains (MCP, Search, Terminal, SWE, Android, Web, OS).
Trained on over 10 million real-world interaction trajectories via a three-stage pipeline (CPT, SFT, RL), the models (35B-A3B and 397B-A17B) predict next environment states given actions.
The AgentWorldBench benchmark, built from real interactions of five frontier models on nine established benchmarks, evaluates simulation quality across five dimensions.
Qwen-AgentWorld-397B-A17B achieves the highest overall score (58.71), surpassing GPT-5.4 (58.25).
As a decoupled simulator, it enables scalable and controllable simulation, yielding gains on Claw-Eval (+4.3), QwenClawBench (+7.1), MCPMark (+12.3), and WideSearch (+16.3), even exceeding real-environment training.
As a unified agent foundation model, LWM training serves as a warm-up, improving performance across seven agentic benchmarks, including out-of-domain gains on Claw-Eval (+11.3), QwenClawBench (+9.7), and BFCL v4 (+9.0).
The work demonstrates that world modeling is a crucial missing piece for general agents, offering both scalability and controllability beyond real environments and establishing next-state prediction as a transferable meta-reasoning pattern.
Future work includes agent-LWM co-evolution, multimodal extension, and adaptive sim-to-real routing.
02 From the paper
Abstract
A world model predicts environment dynamics based on current observations and actions, serving as a core cognitive mechanism for reasoning and planning. In this work, we investigate how world modeling based on language models can further push the boundaries of general agents. (i) We first focus on building foundation models for agentic environment simulation. We introduce Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B, the first language world models capable of simulating agentic environments covering 7 domains via long chain-of-thought reasoning. Leveraging more than 10M environment interaction trajectories of 7 domains in real-world environments, we develop Qwen-AgentWorld through a three-stage training pipeline: CPT injects general-purpose world modeling capabilities from the state transition dynamics and augmented professional corpora, SFT activates next-state-prediction reasoning, and RL sharpens simulation fidelity through a tailored framework with hybrid rubric-and-rule rewards. To evaluate language world models, we present AgentWorldBench, a comprehensive benchmark constructed from real-world interactions of 5 frontier models on 9 established benchmarks. Empirical results demonstrate that Qwen-AgentWorld significantly outperforms existing frontier models. (ii) Beyond foundation models, we further investigate two complementary paradigms through which world modeling enhances general agents. First, 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 real-environment training alone. Second, as a unified agent foundation model, world-model training acts as a highly effective warm-up that improves downstream performance across 7 agentic benchmarks. Code: https://github.com/QwenLM/Qwen-AgentWorld