Recursive self-improvement.

Recursive self-improvement is a feedback process in which an AI system changes its own design or operation, and those changes help it produce subsequent improvements. The key idea is that the ability to improve can itself improve.

A system might propose changes to its code, tools, learning methods, or model architecture, test whether they help, and carry useful changes into the next round. The process is recursive when an improved version participates in making further improvements. This can happen at the agent-software level without changing the underlying model’s weights.

The Darwin Gödel Machine is a research example: it modifies a coding agent’s software, evaluates candidate versions on programming benchmarks, and explores further changes from an archive of agents. This demonstrates a bounded form of self-improvement; it does not establish that improvements will continue indefinitely or generalize to every task.

Success depends on reliable evaluation, available compute, and which parts of the system can change. Better task performance does not necessarily mean a better ability to invent the next improvement. A system can also overfit its tests or introduce regressions, so each round needs evidence that the changes actually help.

Asking a chatbot to revise an answer is not, by itself, recursive self-improvement: the underlying system may be unchanged. The term also does not imply consciousness or guarantee an intelligence explosion. Rapid, open-ended acceleration is a stronger hypothesis than repeated improvements in a limited setting.