Genuine recursive self-improvement requires AI systems to progress through four levels of autonomy—from executing given improvements to independently identifying what needs improving—before achieving true self-directed capability enhancement.
This paper proposes recursive self-improvement (RSI) as a framework for AI systems to autonomously enhance their own capabilities through experience and feedback. It introduces a roadmap progressing from executing improvements to autonomously discovering what to improve, and examines how RSI applies differently across domains like scientific discovery and robotics.