RSI researchers typically divide RSI into several levels: the shallowest is prompt optimization, the middle involves tool-calling strategies and workflow adjustments, and the deepest involves modifying training objectives, evaluation criteria, and even the architecture itself.