Qwen-AgentWorld: Language World Models for General Agents

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 LLMs 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

• The paper presents a unified native language world model designed to simulate agentic environments across seven diverse domains using long chain-of-thought reasoning.
• It employs a three-stage training pipeline—Continual Pre-Training, Supervised Fine-Tuning, and Reinforcement Learning—to enhance simulation fidelity and enable robust cross-domain transfer.
• Empirical results on AgentWorldBench demonstrate superior performance in both text-based and GUI domains, underpinning its potential for scalable agent simulation.

Qwen-AgentWorld: Language World Models for General Agents

Overview of Qwen-AgentWorld

Qwen-AgentWorld introduces the first family of native language world models (LWMs) capable of simulating agentic environments across seven domains---MCP, Search, Terminal, SWE, Android, Web, and OS---within a unified framework. The rationale is that general agents require robust world modeling capability: learning to predict environment state transitions is both necessary and foundational for effective action selection. Prior LLM agent research disproportionately emphasizes state-to-action policy learning, neglecting comprehensive modeling of environment dynamics. Qwen-AgentWorld directly addresses this gap by developing a native world model trained to capture diverse, realistic environment responses via long chain-of-thought reasoning.

Figure 1: Overview of Qwen-AgentWorld as a unified world model across seven interactive domains, and illustration of decoupled and unified world modeling strategies.

A major contribution is the design of a unified native language world model architecture that generalizes across domains with fundamentally disjoint action and observation spaces, ranging from textual interactions (terminal, APIs, search) to complex structured GUI state (accessibility trees, UI hierarchies, HTML). By collecting and curating more than 10M environment interaction trajectories, Qwen-AgentWorld establishes data scale sufficient for high-fidelity simulation. The model's training pipeline is distinctly multi-stage: continual pre-training (CPT) injects world knowledge, supervised fine-tuning (SFT) activates explicit next-state-prediction reasoning, and RL sharpens simulation fidelity using a hybrid rubric-and-rule reward design, all tailored for environment dynamics.

Figure 2: Qwen-AgentWorld unifies seven agentic domains within a single language world model.

Unified Trajectory Schema and Model Formulation

Qwen-AgentWorld adopts a single environment trajectory schema across all domains, representing interaction as a sequence of (action, observation) pairs with dynamic system prompts encoding the simulation context, action space, initial environment state, demonstrations, and controllable simulation instructions. This formalization enables LWM training over domain-heterogeneous data via consistent, structured input. Actions correspond to agent tool-calls, shell commands, GUI manipulations, etc., and observations reflect ground-truth environment responses.

A language world model is then trained as a conditional text generator $f_\theta$ performing next-state prediction:

$$\hat{o}_{t+1} = f_\theta(c,\, o_{\leq t},\, a_{\leq t})$$

where $c$ is the system prompt encoding environment context, $o_{\leq t}$ is the full history of observations, and $a_{\leq t}$ is the sequence of actions up to turn $t$.

Qwen-AgentWorld's coverage spans:

• Text-based domains: Terminal (bash/keystrokes $\rightarrow$ terminal output), SWE (edit/run/bash $\rightarrow$ execution output), MCP (structured tool-calls $\rightarrow$ tool/API response), Search (web search/extract $\rightarrow$ snippets/results)
• GUI domains: Android, Web, OS (touch/click/navigate/enter $$\hat{o}_{t+1} = f_\theta(c,\, o_{\leq t},\, a_{\leq t})$$0 accessibility/UI view trees, HTML, window state)

Figure 3: Example of SWE domain simulation, showing full traceback prediction including OOM error.

Training Pipeline: CPT, SFT, RL

The training scheme is tripartite:

1. Continual Pre-Training (CPT): Large-scale injection of environment dynamics, world knowledge, tool behaviors, and professional corpora. World-modeling target is explicit from CPT onward.
2. Supervised Fine-Tuning (SFT): Activation of next-state prediction as an explicit reasoning chain. Multi-template system prompt diversification and rejection sampling curate agentic reasoning traces and increase prompt-format robustness. Long context windows (256k tokens) are used to accommodate high trajectory depth.
3. Reinforcement Learning (RL): Application of Group Sequence Policy Optimization (GSPO) and rubric-based RL to sharpen simulation fidelity. Reward signals combine five-dimensional rubric ratings (Format, Factuality, Consistency, Realism, Quality) and strict rule-based verifiers, balancing open-ended evaluation with deterministic correctness. Prompt-output asymmetry (context $$\hat{o}_{t+1} = f_\theta(c,\, o_{\leq t},\, a_{\leq t})$$1 output) and reward shaping are handled to avoid collapse and reward hacking.

   Figure 4: Three-stage Qwen-AgentWorld training pipeline: CPT injects, SFT activates, RL sharpens.

Benchmarking: AgentWorldBench

The authors establish AgentWorldBench, a grounded benchmark for LWM evaluation, built from agentic trajectories on nine established datasets executed against real environments by frontier LLM agents (Claude Opus, GPT, Gemini, Qwen). Every trajectory is paired with ground-truth output, ensuring out-of-distribution evaluation.

AgentWorldBench evaluates five dimensions—Format, Factuality, Consistency, Realism, Quality—via reference-grounded rubric judging, domain-specific prompts, and differentiated content matching. The judge (GPT-5.2) is calibrated by Turing tests for discriminative evaluation, and scoring is robust to absolute model rating variance.

Figure 5: AgentWorldBench structure—domain distribution across seven domains, benchmarks mapped, multi-dimensional evaluation.

Empirical Results

Qwen-AgentWorld-397B-A17B achieves the highest overall mean score (58.71) on AgentWorldBench, with pronounced gains in text-based domains (Terminal, SWE) and competitive performance in GUI (Android/Web/OS), compared to other frontier LLMs. World-model training via the three-stage pipeline brings marked improvements relative to base checkpoints and open-weight models, demonstrating substantial advantage for explicit environment simulation training.

Cross-domain generalization experiments indicate that RL driven by data from one domain (Terminal) yields robust transfer to held-out domains (SWE, MCP, Search), confirming reinforcement of domain-agnostic environment dynamics rather than overfitting to format.

Figure 6: Main AgentWorldBench results: Qwen-AgentWorld-397B-A17B leads overall, with text-domain dominance and strong GUI competitiveness.

Figure 7: RL on Terminal data induces generalization to SWE, MCP, Search—non-trivial transfer with early convergence.

Applications: Simulation and Agent Foundation

Decoupled Simulation

Qwen-AgentWorld enables scalable, controllable simulation of thousands of agentic environments, including OpenClaw and fictional information-seeking worlds. As an independent environment simulator, it delivers:

• Zero-shot scaling: Simulates $$\hat{o}_{t+1} = f_\theta(c,\, o_{\leq t},\, a_{\leq t})$$2 OpenClaw environments, yielding +4.3 Claw-Eval and +7.1 QwenClawBench gains.
• Controllable perturbations: Systematic exposure of agent weaknesses, with +12.3 on MCPMark, +16.3 on WideSearch. Controllable simulation outperforms real-environment RL.
• Fictional world construction: Agents trained in self-consistent worlds generalize to real search.
• State tracking bottleneck: Effective simulation requires detailed initial state for fidelity.

Figure 8: WideSearch results—controllable simulation enables higher F1 scores via realistic fictional environment construction.

Unified Agent Foundation Model

LWM training acts as a meta-reasoning warm-up: next-state prediction enables internal environment simulation before action, improving agent performance across diverse downstream benchmarks (Terminal-Bench, SWE-Bench, WideSearch, Claw-Eval, QwenClawBench, BFCL-v4), with strong cross-domain transfer (+11.3 Claw-Eval, +9.7 QwenClawBench, +9.0 BFCL). RL-trained agents exhibit explicit chain-of-thought simulation and prediction-driven action refinement.

Analysis: Reasoning Patterns, RL Fidelity

Qwen-AgentWorld's reasoning traces show:

• Multi-step causal reasoning (package management, process lifecycle, tool semantics)
• Information leakage prevention (theory of mind separation in Search)
• Epistemic boundary awareness (avoiding fabrication, fallback to format-only)

RL training progressively enhances micro-level fidelity: exact arithmetic, referential integrity across API calls, realistic URL generation and snippets—even for low-salience tokens.

Figure 9: Chain-of-thought reasoning: causal chains, information separation, epistemic boundaries.

Figure 10: RL improves micro-level fidelity—URL realism, byte-arithmetic, API schema consistency.

Figure 11: GUI simulation: environment transition after file--print operation captured accurately.

Figure 12: Terminal simulation: full compiler output and error handling predicted for make command.

Implications and Future Directions

Practically, Qwen-AgentWorld provides a scalable, versatile axis for environment simulation in agent RL, opening controllable training regimes and enabling model-based transfer to domains with limited real-execution data. Theoretically, explicit world modeling capacity is validated as necessary for general agentic performance, supporting the claim that generalization requires robust internal simulation.

Anticipated future developments include agent--LWM co-evolution via self-play, multimodal fusion integrating screenshots and accessibility trees, adaptive sim-to-real routing for cost/fidelity optimization, and dynamic tool synthesis leveraging world models. There is compelling evidence that world modeling acts as a transferable foundation for robust policy learning, potentially compounding further when combined with auxiliary world-modeling losses in agent RL pipelines.

Conclusion

Qwen-AgentWorld substantiates the utility and necessity of high-fidelity language world models for general agentic intelligence, enabling both scalable simulation and foundation model transfer. Its tripartite training approach achieves superior simulation quality and downstream agent gains, and its unified architecture successfully bridges diverse agentic interaction domains. Explicit world modeling emerges as a crucial axis in scaling the capabilities of general agents, altering the traditional landscape of agent RL training beyond the constraints of real-environment interaction.