Qwen-AgentWorld: Language World Models for General Agents
Summary
Qwen-AgentWorld introduces language world models for general agents, presenting large language-model-based environment simulation and a three-stage training pipeline (CPT, SFT, RL). It also introduces AgentWorldBench for evaluating agentic performance across multiple benchmarks and domains, with claims of improvements over frontier models. The work provides a GitHub link for code and emphasizes scalable, language-driven world modeling.