Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds

One-line takeaway

Tsinghua’s Spatial Intelligence Lab studies continued pre-training of embodied agents across game worlds with 120 configurations (0.8B–27B models, 50–1,600 hours, five data mixtures): more data helps in-corpus play, while diversity helps transfer to related worlds.

清华空间智能实验室用 120 组配置（0.8B–27B 模型、50–1600 小时数据、5 种数据混合）研究具身智能体在多款游戏世界中的继续预训练，发现数据量提升同世界表现，数据多样性提升向相近世界的迁移。