Qwen-AgentWorld: Language World Models for General Agents
4 hours ago
- #General Agents
- #World Modeling
- #Language Models
- Introduces Qwen-AgentWorld language world models for simulating agentic environments across 7 domains.
- Develops models via a three-stage training pipeline: CPT, SFT, and RL with hybrid rewards.
- Presents AgentWorldBench, a benchmark built from real-world interactions to evaluate language world models.
- Demonstrates superior performance of Qwen-AgentWorld over existing frontier models on benchmarks.
- Shows Qwen-AgentWorld enhances general agents as both a decoupled environment simulator and a unified foundation model.