Researchers proposed Dream-RSI, a framework for recursive self-improvement in AI agents. It uses a lightweight orchestration layer to make exploration explicit. Historical discovery trees act as a replay simulator for off-policy feedback. This feedback refines exploration policies without costly online evaluations. The improved policy is redeployed to drive further discovery. The loop continuously expands the simulator pool. It achieves competitive discovery quality in algorithm engineering, math optimization and GPU kernel engineering. It also reduces discovery cost in these settings.