Where Transfer Ends: Data Scale, World Diversity, and Model Size for Game Agents
Across 120 configurations, the study tested whether the benefits of extended pretraining persist when transferring between virtual environments with different graphics, physics, controls, and tasks. Expanding the corpus helps in familiar worlds, greater variety broadens transfer to similar environments, and scaling up the model increases the returns on data—but training in a single world hurts adaptation to distant ones. The authors conclude that pretraining alone is not enough to create universal agents.

What Was Tested and How Transfer Was Measured
The study focuses on continued pre-training (CPT) for embodied agents in virtual worlds. The authors tested whether improvements from scaling data volume, data diversity, and model size transfer to other worlds. The study covers 120 training configurations.

| Factor | What Was Varied |
|---|---|
| Model size | four values from 0.8B to 27B |
| Data volume | six values from 50 to 1,600 hours |
| Data mixture | five options: from data from a single world to diverse multi-world corpora |
| Evaluation | three-level validation losses and game benchmarks |
Evaluation takes place at three levels: worlds seen during training, related worlds, and unseen worlds. This design separates gains in familiar environments from transfer. Practical criterion: include all three groups of worlds in validation. Otherwise, gains in trained worlds will hide the gap on distant worlds.
Data Volume: Effects in Familiar Worlds
Increasing data volume improves performance in worlds included in the training data. This is a confirmed scaling effect in familiar territory. For distant worlds, the authors do not reduce transfer to data volume alone.
Data volume is not a universal lever for every goal. It delivers a clear gain where a world is already represented in training. For a new world, separately assess how much of the gap data volume closes.
Selection criteria:
- target world is included in training — increase data volume;
- target world is distant — do not rely on data volume alone;
- world composition is unclear — first group worlds by similarity.
Diversity: Transfer to Related, but Not Distant, Worlds
Data diversity improves transfer to related worlds. Gains on distant worlds are limited. This effect distinguishes diversity from simply increasing data volume.
| Degree of world similarity | Observed effect of diversity | Practical takeaway |
|---|---|---|
| Related worlds | improves transfer | a diverse mixture is worthwhile |
| Distant worlds | limited gains | diversity is insufficient as a standalone approach |
Diversity expands the zone of transfer, but does not make the agent universal. A data mixture acts as a bridge to similar worlds. For distant worlds, this bridge is shorter than needed.
Practical criterion: assess how similar target worlds are to the training worlds. For related worlds, a diverse mixture is a useful tool. For distant worlds, diversity is not enough.
Model Size: Data Efficiency and the Risk of Narrow Generalization
Increasing model size improves data efficiency. When training on a single world, model size worsens generalization to distant worlds. This makes the effect of model size dependent on the data composition.
A larger model makes better use of available data. But with narrow training, it limits transfer. The same increase in parameters has different effects depending on the mixture of worlds.
Selection criteria:
- training on a single world — do not expect model size to increase transfer;
- diverse data — model size boosts data use;
- choosing model size — do so only alongside choosing the data mixture.
Two Training Phases: Interface and World
The authors observe an early interface phase. In this phase, the model gains broad benefits across different worlds. Later, a later world phase begins. In this phase, improvements narrow to worlds seen during training and related worlds.
The phases have different transfer ranges. The early phase builds a general interface for different worlds. The later phase ties improvements to familiar and similar environments. This is not a training error, but a shift in the zone of applicability.
Selection criteria:
- need a broad interface — early phase;
- need a target world — later phase;
- need distant worlds — the early phase does not sustain transfer.
Limits of Continued Pre-Training
The authors note the limitations of imitation learning and the persistent gap on distant worlds. They conclude that universal embodied agents require more substantive training stages beyond CPT. CPT delivers competitive results, but does not close the gap at every level of transfer.
The models in the study are competitive with specialized models for 3D game worlds and frontier models. They show better results in real-time settings. These results do not eliminate the gap on distant worlds. They define the limits of CPT's applicability.
Selection criteria:
- universal agent — plan stages beyond CPT;
- distant worlds — do not treat CPT as the final solution;
- worlds similar to those in training — CPT remains a useful stage.
Frequently asked questions
Related materials
All materials
SpecEvo: Aligning Lightweight and Advanced LLMs to Reduce the Cost of Evolutionary Search
The SpecEvo framework redistributes evolutionary search: routine candidates refine inexpensive models, while stronger ones are brought in only at strategic checkpoints, guided by a history of successes and dead ends. In experiments across 176 tasks, this approach matched or outperformed costly alternatives while reducing API costs by 80%.
21 November 2026
Algorithms and Robots in the Field: Targeted Treatments and Crop Monitoring
The article shows how sensor data, aerial imagery, and automated platforms enable early threat detection, target treatment only to problem areas, and refine harvest forecasts. It also examines barriers to adoption: equipment costs, data reliability, staff training, and access to connectivity.
21 November 2026
Fig: Controlled game interventions for testing the social behavior of AI agents
Fig takes a natural-language description of the behavior being targeted, divides it into variants, and strategically changes the rules of the game to elicit the desired actions from models. In valid scenarios across six social patterns, the interventions increased the average share of attempts at the target behavior by 10.9% compared with controls, and the authors propose treating multi-agent environments as a configurable testbed for reproducible evaluation.
21 November 2026