Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds
Paper · https://arxiv.org/abs/2610.28698
Scaling data, diversity and model capacity has improved embodied learning in physical environments, where diverse experiences remain grounded in shared physical laws. Whether these gains extend to virtual worlds with different visual styles, dynamics, controls, and objectives remains unclear. We investigate this question through a controlled study of continued pre-training (CPT) across three axes: data volume, diversity, and model size. Our 120 training configurations span four model sizes from 0.8B to 27B, six data budgets from 50 to 1,600 hours, and five mixtures ranging from Minecraft-only data to diverse multi-world corpora. We assess performance and generalization using three-level validation losses and gameplay benchmarks from in-corpus worlds to unseen worlds. Our results show: (1) more data improves in-corpus performance, diversity improves transfer to related worlds with limited gains in distant worlds, and model size improves data efficiency but generalizes worse to distant worlds under single-world training; (2) we observe an early interface phase with broad cross-world benefits and a later world phase with gains narrowed to in-corpus and related worlds, and fit an empirical scaling law relating validation loss to data volume, diversity and model size; and (3) our models are competitive with Minecraft specialists and frontier models in 3D worlds, and perform best in real-time settings. However, the limits of imitation and persistent gaps in distant worlds suggest that building generalist embodied agents requires richer training stages beyond continued pre-training.
First seen by Oriel on .
AI summary
A controlled study of continued pre-training for embodied agents finds that scaling data, diversity, and model size improves performance unevenly across virtual worlds, with persistent limits on distant-world generalization.
- The study evaluates 120 configurations across four model sizes (0.8B–27B), six data budgets (50–1,600 hours), and five training mixtures.
- More data improves in-corpus performance; diverse data helps transfer to related worlds but yields limited gains in distant worlds.
- Larger models use data more efficiently, yet single-world training leads to worse generalization to distant worlds.
- Training shows an early interface phase with broad cross-world benefits, followed by a world phase whose gains concentrate in in-corpus and related worlds.
- The models are competitive with Minecraft specialists and frontier models in 3D worlds, perform best in real-time settings, and remain limited by imitation and distant-world gaps.
#embodied agents #continued pre-training #scaling laws #virtual worlds
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Points
Score reported by the source.
- Hugging Face Papers: 86, latest observation ; +37 over observed period since
Comments
Comment count reported by the source.
- Hugging Face Papers: 2, latest observation ; +1 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| arXiv AI Papers | oai:arXiv.org:2610.28698v1 | https://arxiv.org/abs/2610.28698 Read discussion |
| Hugging Face Papers | 2610.28698 | https://arxiv.org/abs/2610.28698 Read discussion |