Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds
Training across more game worlds improved transfer to related worlds, but offered limited gains in distant ones.
The study tested 120 training configurations, spanning models from 0.8 billion to 27 billion parameters and 50 to 1,600 hours of data. More data improved performance in training worlds; larger models used data more efficiently but generalized worse to distant worlds when trained on a single world. The researchers observed an early phase of broad cross-world improvement, followed by gains concentrated in training worlds and related environments. They conclude that persistent gaps in distant worlds and the limits of imitation point to a need for training beyond continued pre-training.
HF Daily Papers' note
The study tested 120 training configurations, spanning models from 0.8 billion to 27 billion parameters and 50 to 1,600 hours of data. More data improved performance in training worlds; larger models used data more efficiently but generalized worse to distant worlds when trained on a single world. The researchers observed an early phase of broad cross-world improvement, followed by gains concentrated in training worlds and related environments. They conclude that persistent gaps in distant worlds and the limits of imitation point to a need for training beyond continued pre-training.
HF Daily Papers' note
score 5