Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds
Tsinghua THUSI-Lab, miHoYo
A controlled study of continued pre-training for keyboard-mouse game agents across 120 configurations (0.8B-27B models, 50-1,600 hours of gameplay, five data mixtures). Data volume helps within a corpus and diversity helps transfer to related worlds, but neither closes the gap to distant worlds; the authors fit an empirical scaling law and release eight checkpoints.
Why it matters
At 301 Hugging Face Daily upvotes, it is the most-engaged paper of the window and pairs a fitted scaling law with open weights, while documenting that imitation learning leaves persistent gaps in unseen worlds.
Importance: 4/5
top paper of the window with 301 Hugging Face Daily upvotes and open checkpoints
Sources
official
THUSI-Lab/GameScaling - GitHub