阿里巴巴 Qwen 团队推出了 Qwen-AgentWorld，一个旨在模拟各种代理环境的新型语言世界模型。该模型专注于训练大型语言模型理解和预测环境，而不仅仅是在其中行动。该研究探索了两个主要方向：构建用于环境模拟的基础模型，以及研究世界建模如何增强代理训练，结果表明，通过世界模型训练的代理在真实环境中训练的代理表现更优，并且预测知识能有效地迁移到代理任务中。 AI
影响 这种方法可以通过提高代理对其运行环境的理解能力，从而可能加速复杂任务自动化的进展，从而实现更强大的代理。
排序理由 前沿实验室模型发布，附带系统卡和基准测试结果。
AI 生成摘要 · Google Gemini · 来自 5 个来源。 我们如何撰写摘要 →
🧠 Paradigm II — Agent Foundation Model: world modeling as agent capability. Single-turn, non-agentic environment prediction → tested directly on multi-turn, tool-calling agent tasks. No agentic RL, no task-specific tuning. Gains across 7 benchmarks, including 3 entirely https:/…
Part II: Investigating the Role of World Modeling in Agent Training 🔬 Paradigm I — Decoupled Simulation: world model as environment simulator for agent RL. The key is controllability: 1️⃣ Zero-shot generalization to 4k OOD OpenClaw environments → +4.3 Claw-Eval, +7.1 https://t…
📊 AgentWorldBench: 7-domain benchmark with ground-truth observations from real environments, constructed from 5 frontier model trajectories on 9 established benchmarks. Results: Qwen-AgentWorld-397B-A17B achieves the highest overall score (58.71), outperforming Claude Opus 4.8 h…
Part I: Building the Foundation Model for Environment Simulation Faithful environment simulation requires multi-step causal reasoning, stateful tracking, and domain-specific knowledge. Frontier LLMs have some simulation ability from pretraining — but it's incidental, not an http…
📣📣 Meet Qwen-AgentWorld — a native language world model that simulates 7 agent environments (MCP, Search, Terminal, SWE, Web, OS, Android) within a single model. Environment modeling is the training objective from day one, not a post-hoc adaptation. 🤔 LLMs are trained to be htt…