Qwen-AgentWorld: Language World Models for General Agents
This paper introduces Qwen-AgentWorld, large language models trained to simulate agentic environments across seven domains using over 10 million interaction trajectories. The authors detail a three-stage training pipeline (CPT, SFT, RL) and evaluate on AgentWorldBench, showing superior simulation fidelity. They also demonstrate its use as both a standalone simulator for RL and a warm-up step that boosts downstream agent performance.
- Qwen-AgentWorld's 397B model held correct object-grasp sequences over 25 reasoning steps in a robotics task without losing context, beating existing simulators on state accuracy and reasoning-chain length.
- RL agents trained inside Qwen-AgentWorld's simulated environments outperformed agents trained only in the real world by 15–20%.
- Using the world model's weights to warm-start a unified agent boosted performance across seven benchmarks and sometimes cut required training runs in half.
Qwen-AgentWorld builds two large language–based world models, one with 35 billion parameters (A3B) and another with 397 billion (A17B). Both are trained on over 10 million real-world interaction trajectories spanning seven domains—everything from robotic manipulation to strategy games. The team uses a three-stage process: CPT (contrastive pretraining) to teach basic state-transition dynamics, followed by supervised fine-tuning that shapes next-state prediction, and capped off with reinforcement learning guided by a mix of rule-based and learned rewards to sharpen simulation detail.
To measure performance, the authors introduce AgentWorldBench, a new benchmark built from live interactions of five top agent models across nine classic tasks. On this test set, Qwen-AgentWorld outperforms existing simulators by clear margins in both accuracy of predicted states and length of coherent reasoning chains. In one robotics manipulation task, for example, the 397B model maintained correct object-grasp sequences over 25 reasoning steps without dropping context.
Beyond pure simulation, the paper shows two ways these world models improve agents. First, using Qwen-AgentWorld as a drop-in simulator allows training reinforcement-learning agents on thousands of configurable environments. Agents trained this way beat counterparts trained only in the real world by 15–20 percent on average. Second, initializing a unified agent’s training with world-model weights acts like a warm-start, boosting downstream performance on seven benchmarks—sometimes slashing required training runs in half.