Dream-RSI: Recursive Self-Improvement through Evolving Worlds
Summary
Dream-RSI presents a scalable framework for recursive self-improvement in AI exploration using a replay simulator built from historical discovery trees, plus a lightweight orchestration layer that keeps the core agent unchanged. It enables immediate, low-cost off-policy feedback to evaluate and refine exploration policies, reducing expensive online evaluations while expanding the simulator pool over time. The approach achieves competitive or improved discovery quality at substantially lower costs across multiple settings.