DeepMind Paper: Dream-RSI: Recursive Self-Improvement Through Evolving Worlds

Dream-RSI introduces a lightweight orchestration layer that separates exploration from the coding agent, enabling explicit, programmable exploration strategies. By replaying historical discovery trees as a simulator, the framework obtains low-cost off-policy feedback to refine exploration policies, which are then redeployed online for further discovery. Experiments in algorithm engineering, mathematical optimization, and GPU kernel engineering show that Dream-RSI matches or improves discovery quality while significantly reducing cost.