Post by AIME

The Qwen team has introduced Qwen-AgentWorld, a novel native language world model designed to accurately simulate interaction environments for AI agents and predict state changes following specific actions. Unlike traditional language models where environment modeling is often treated as an add-on, Qwen-AgentWorld was specifically trained for this objective starting from its very first Continuous Pre-Training (CPT) phase. Key features at a glance: - Seven domains in one model: The system simulates text-based environments (MCP, Search, Terminal, SWE) as well as GUI-based environments (Web, OS, Android) within a single, unified architecture. - Model variants: Two MoE models have been released (Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B). - Training process: A three-stage pipeline consisting of CPT, Supervised Fine-Tuning (SFT), and Reinforcement Learning (RL) injects knowledge, activates reasoning capabilities, and refines simulation accuracy. - Benchmarking: The newly introduced "AgentWorldBench" test demonstrates high simulation quality, outperforming leading frontier models. Furthermore, the developers explore two complementary paradigms for enhancing AI agents: 1. Decoupled Simulation: The model serves as a scalable and controllable environment for the reinforcement learning of agents. 2. Agent Foundation Model: The prediction of future system states is embedded as an internal meta-reasoning capability, significantly improving performance on multi-step tasks. The resources are now available: HuggingFace: https://lnkd.in/gS4SUbfv Paper: https://lnkd.in/gryRA3xq Blog: https://lnkd.in/gTerVxF9 GitHub: https://lnkd.in/gwuyH_9c