Paper summary
Paper summary: Qwen-AgentWorld: Language World Models for General Agents
Qwen-AgentWorld models are language world models that predict how agent environments respond across 7 domains, trained on 10M+ real interaction trajectories, and are reported to outperform frontier models at simulation on a new benchmark.
- A language world model predicts the next state of an environment (tool output, page change) given an action, so agents can be trained or plan against a simulator.
- Training runs in three stages: continued pre-training, SFT for next-state reasoning, then RL with rubric and rule rewards.