QwenLM/Qwen-AgentWorld

What it solves

Qwen-AgentWorld is designed to simulate agentic environments, allowing AI agents to be trained or tested without needing access to real-world systems. It addresses the problem of scaling agent training by providing a high-fidelity, controllable simulation of various digital environments, which can be used to generate synthetic data for reinforcement learning (RL) and improve agent performance across both in-domain and out-of-domain tasks.

How it works

It is a native language world model trained on over 10 million real-world interaction trajectories. The model uses a three-stage training pipeline:

1. CPT (Continual Pre-training): Injects environment knowledge.
2. SFT (Supervised Fine-Tuning): Activates next-state-prediction reasoning via long chain-of-thought.
3. RL (Reinforcement Learning): Sharpens the simulation fidelity.

Unlike traditional simulators, it is a single model that can simulate seven unified domains: MCP, Search, Terminal, SWE (Software Engineering), Android, Web, and OS.

Who it’s for

This project is for AI researchers and developers building autonomous agents who need a scalable, controllable environment for training (Sim RL) and evaluation (AgentWorldBench) without the risks or costs of interacting with live systems.

Highlights

• Seven Unified Domains: Simulates MCP, Search, Terminal, SWE, Android, Web, and OS in one model.
• Native World Model: Environment modeling is the training objective from the start, not a post-hoc addition.
• Controllable Simulation: Supports injecting perturbations or creating fictional worlds to stress-test agents.
• Agent Foundation Model: Training as a world model improves the model's own ability to perform multi-turn tool-calling tasks.
• AgentWorldBench: Includes a comprehensive evaluation benchmark across seven domains scoring on format, factuality, consistency, realism, and quality.