游戏中的缩放：面向多样虚拟世界具身智能体的持续预训练
原标题：Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds
AI 摘要
该研究对虚拟世界中的具身智能体持续预训练（CPT）进行了受控网格实验，覆盖 0.8B 至 27B 四种模型规模、50 至 1600 小时六档数据预算，以及从仅 Minecraft 到多游戏混合的五种数据配比，共 120 组配置。结果显示：数据量提升语料内表现，多样性改善向相关世界的迁移但对远距离世界增益有限，大模型提升数据效率但在单一世界训练下对远距离世界泛化更差。作者还观察到早期“接口阶段”与后期“世界阶段”的两阶段学习动态，并拟合了验证损失与数据量、多样性、模型规模的经验缩放律；模型在 3D 世界可与 Minecraft 专用模型及前沿模型竞争，实时场景表现更优，但远距离世界仍存在持续差距。
正文节选
Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds Abstract Scaling data, diversity and model capacity has improved embodied learning in physical environments, where diverse experiences remain grounded in shared physical laws. Whether these gains extend to virtual worlds with different visual styles, dynamics, controls, and objectives remains unclear. We investigate this question through a controlled study of continued pre-training (CPT) across three axes: dat