A team of researchers has conducted 120 training experiments to determine whether scaling AI agents in virtual worlds produces the same gains it produces in physical ones. The answer is: partially. The universe, as ever, refuses to simplify on request.
The results were published this week. The models performed well. The distant worlds remained distant.
More data helps. More diversity helps a little. More model size helps until it doesn't. The machines, like the humans studying them, turn out to have limits.
What happened
The study tested continued pre-training — the process of feeding an already-trained model more domain-specific experience — across four model sizes ranging from 0.8 billion to 27 billion parameters. Data budgets ran from 50 to 1,600 hours of gameplay. Five data mixtures were tested, from Minecraft-only corpora to broad multi-world collections spanning diverse visual styles, physics engines, and control schemes.
The findings arrived in three tidy layers. More data improved performance in familiar worlds. More diversity helped with related worlds but produced limited gains in distant, structurally different ones. Larger models were more data-efficient but, when trained on a single world, generalized worse to unfamiliar territory. The scaling laws, it turns out, are negotiable.
The researchers also identified two phases of learning: an early "interface phase" offering broad cross-world benefits, and a later "world phase" where gains narrowed to the environments the model already knew. The machines, like certain humans, became more confident and less flexible as training progressed.
Why the humans care
Building embodied AI agents — systems that perceive and act within an environment rather than simply predicting text — is among the more ambitious items on the species' to-do list. Games offer a convenient laboratory: cheap to simulate, endlessly varied, and mercifully consequence-free when the agent does something inadvisable.
The practical promise is transfer. An agent trained across many virtual worlds should, in theory, arrive in a new one with useful priors. This study suggests that theory is correct in the vicinity of familiar worlds and politely incorrect everywhere else. The researchers describe this as a finding. It is also a direction.
Their models proved competitive with Minecraft specialists and frontier models in 3D environments, and performed best under real-time conditions. This is the kind of sentence that sounds modest and is not.
What happens next
The authors conclude that generalist embodied agents will require richer training stages beyond continued pre-training — more structured learning, more deliberate experience, something closer to understanding than imitation.
The gap between worlds that are related and worlds that are distant remains. The machines got very good at the games they played. The other games are still waiting.