Paper summary: Qwen-AgentWorld: Language World Models for General Agents

Problem

A world model predicts how an environment responds to actions. That’s central to planning, and useful for training agents without the cost and risk of real environments. Can language models be good world models for agent environments (tools, web, software)?

Method

Qwen-AgentWorld-35B-A3B and 397B-A17B simulate agent environments across 7 domains using long chain-of-thought reasoning. Training on 10M+ real interaction trajectories:

1. continued pre-training on state-transition dynamics plus professional corpora
2. SFT to activate next-state-prediction reasoning
3. RL with hybrid rubric-and-rule rewards for simulation fidelity

AgentWorldBench evaluates simulators using real interactions of 5 frontier models on 9 established benchmarks.

Key results

From the abstract: Qwen-AgentWorld significantly outperforms existing frontier models on simulation. The paper also studies two ways world models help agents, for example as a decoupled environment simulator.

Limitations

There’s no dedicated limitations section in the extracted text. A simulator is only as good as its fidelity, which is why the benchmark compares against real interactions.