Continued pre-training for game agents helps unevenly across virtual worlds, study finds
A new arXiv paper (2610.28698) asks whether the scaling gains seen in embodied learning in physical environments carry over to virtual game worlds. In physical settings, diverse experiences share the same physical laws. Virtual worlds differ in visual style, dynamics, controls and objectives, and the paper says it remains unclear whether the same gains hold there.
To test this, the authors run a controlled study of continued pre-training (CPT) along three axes: data volume, data diversity and model size. The study covers 120 training configurations. They span four model sizes from 0.8B to 27B parameters, six data budgets from 50 to 1,600 hours, and five data mixtures ranging from Minecraft-only data to diverse multi-world corpora. Performance and generalization are assessed with three-level validation losses and with gameplay benchmarks, running from worlds inside the training corpus to unseen worlds.
The results come in three parts. First, the three axes behave differently. More data improves performance in worlds inside the corpus. Diversity improves transfer to related worlds, but gains in distant worlds stay limited. Larger models are more data efficient, yet under single-world training they generalize worse to distant worlds.
Second, the authors see two phases during training: an early interface phase with broad benefits across worlds, and a later world phase in which gains narrow to in-corpus and related worlds. They also fit an empirical scaling law that relates validation loss to data volume, diversity and model size.
Third, the authors report that their models are competitive with Minecraft specialists and with frontier models in 3D worlds, and that they perform best in real-time settings.
The conclusion is a caution. The limits of imitation and the persistent gaps in distant worlds suggest, in the authors' words, that building generalist embodied agents requires richer training stages beyond continued pre-training.
Key facts
- The study covers 120 training configurations: four model sizes (0.8B to 27B), six data budgets (50 to 1,600 hours) and five data mixtures, from Minecraft-only to diverse multi-world corpora.
- More data improves performance in the training corpus; diversity helps transfer to related worlds but gives limited gains in distant ones; larger models are more data efficient but generalize worse to distant worlds under single-world training.
- Training shows an early interface phase with broad cross-world benefits, then a later world phase where gains narrow to in-corpus and related worlds.
- The authors fit an empirical scaling law linking validation loss to data volume, diversity and model size.
- They report their models are competitive with Minecraft specialists and frontier models in 3D worlds, best in real-time settings, but argue generalist agents need richer training stages beyond continued pre-training.
Why it matters
Scaling data, diversity and model size has helped embodied learning in physical environments. This paper tests whether the same recipe works across virtual worlds that differ in visuals, dynamics, controls and objectives. The answer is mixed: gains are real near the training data and in related worlds, but they thin out in distant ones. That pushes against the idea that simply scaling continued pre-training yields a generalist game agent.
Who it affects
Researchers building embodied and game-playing agents are the direct audience, especially those choosing how to mix data from multiple worlds or how large a model to train. The findings on diversity and model size bear on those choices.
How to use it
The practical reading comes from the reported trade-offs. If the target is performance in worlds already in the corpus, more data helps. If the target is transfer to related worlds, diversity of the training mix helps. Larger models save data but, under single-world training, generalize worse to distant worlds. The authors also fit a scaling law relating validation loss to data volume, diversity and model size. No code, data or model release is mentioned.
How solid is it
The design is a controlled study: 120 configurations varying three axes, evaluated with three-level validation losses and gameplay benchmarks from in-corpus to unseen worlds. The claims here come from the authors' own summary of their results. The abstract does not name which frontier models or Minecraft specialists were compared against, nor give benchmark scores or margins, so the 'competitive' claim cannot be checked from it.
Risks and caveats
The authors' conclusion is worded as a suggestion, not a proof: the limits of imitation and the persistent gaps in distant worlds suggest that richer training stages are needed. The abstract does not say which worlds count as related versus distant, does not list the worlds besides Minecraft, and does not give the base model, architecture or the form of the scaling law. The findings apply to continued pre-training as studied, not to other training approaches.
“the limits of imitation and persistent gaps in distant worlds suggest that building generalist embodied agents requires richer training stages beyond continued pre-training”
— From the paper's abstract (arXiv 2610.28698)