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

DeepMind researchers introduce Dream-RSI, a framework that enables AI agents to recursively improve their exploration strategies by using accumulated discovery history as a simulator for low-cost policy refinement. Rather than repeatedly running expensive online evaluations, the system uses "dreaming" in a replay simulator built from historical discovery trees to evaluate and optimize exploration policies before redeploying them. The approach demonstrates competitive or improved discovery quality with substantially reduced computational costs across algorithm engineering, mathematical optimization, and GPU kernel engineering tasks.