Study finds scaling laws for embodied agents in virtual worlds, but gaps persist for distant generalization

A new arXiv preprint investigates whether scaling laws observed in physical environments extend to virtual worlds with distinct visual styles, dynamics, controls, and objectives. The study focuses on continued pre-training (CPT) for embodied agents and evaluates the effects of data volume, data diversity, and model size across 120 controlled training configurations.

The experiments span four model sizes (0.8B, 3B, 8B, and 27B parameters), six data budgets (50 to 1,600 hours), and five data mixtures ranging from Minecraft-only corpora to diverse multi-world datasets. Performance and generalization were assessed using three-level validation losses and gameplay benchmarks spanning in-corpus worlds to unseen worlds.

The authors report that increasing data volume improves in-corpus performance, while greater data diversity enhances transfer to related worlds but yields limited gains for distant worlds. Larger models improve data efficiency but generalize worse to distant worlds when trained on single-world data. The study also identifies two training phases: an early 'interface phase' that delivers broad cross-world benefits, and a later 'world phase' where improvements narrow to in-corpus and related worlds.

The team fits an empirical scaling law relating validation loss to data volume, diversity, and model size. They find that their models are competitive with Minecraft specialists and frontier models in 3D worlds, and perform best in real-time settings. However, they note limits to imitation-based learning and persistent performance gaps in distant worlds, arguing that building generalist embodied agents will require richer training stages beyond continued pre-training.