Qwen-AgentWorld releases language world models for agent training

Category: Paper
Status: developing
Sources: arXiv 2606.24597, discussion
Summary: Researchers from the Alibaba Qwen team posted Qwen-AgentWorld (arXiv 2606.24597, dated 2026-06-23), two large language world models (35B and 397B parameters) that simulate agentic environments through long chain-of-thought reasoning. They are trained in a three-stage pipeline (continued pre-training on state-transition dynamics, supervised fine-tuning for next-state prediction, reinforcement learning with hybrid rubric-and-rule rewards) over more than 10 million environment-interaction trajectories across seven domains. The models serve both as decoupled environment simulators for scalable reinforcement-learning training and as unified agent foundation models, and the authors report gains over frontier models on their AgentWorldBench and across seven agentic benchmarks.
Why it matters: Using a world model as a synthetic environment simulator targets the data and infrastructure bottleneck in training agentic models without live environment access.
Follow-up: Independent reproduction of the AgentWorldBench results and the released weights and license terms.