# Training Agents Inside of Scalable World Models

*Source* · `src-hafner-2025-dreamer-4` · status: **n/a**

| field | value |
|---|---|
| kind | Preprint |
| authors | Danijar Hafner, Wilson Yan, Timothy Lillicrap |
| venue | arXiv |
| date | 2025-09-29 |
| arxiv | 2509.24527 |
| accessed | 2026-10-24 |

## Notes

Dreamer 4 learns behavior by RL inside a world model trained mostly on unlabeled video; model runs interactively on one GPU; first agent to obtain Minecraft diamonds from offline data only.

## Backlinks (computed)

**cited by:** [Self-play planning with learned models (AlphaZero, MuZero, Dreamer)](../technique/model-based-planning-muzero.md)

## Citation

Danijar Hafner, Wilson Yan, Timothy Lillicrap (2025-09-29). *Training Agents Inside of Scalable World Models*. arXiv https://arxiv.org/abs/2509.24527

---
Provenance: created in run `2026-10-24-techniques-rl-design`.