Recursive Self-Improvement
Studying the mechanisms of improvement themselves.
An agent may improve its policies, skills, or system implementation. We ask a further question: can the methods that propose, test, and select these modifications also improve under independent evaluation?
From system changes to improvement mechanisms
A system can adjust its behavior through experience or propose changes to itself. Recursive self-improvement brings the process of proposing and selecting those changes into the research question.
This requires distinguishing two levels: the system being modified, and the methods that generate candidates, design evaluations, and select modifications. Better task performance alone does not establish improvement at the second level.
Questions we study
- Which changes produce repeatable improvements?
- How can we distinguish accumulated experience, policy updates, and changes to the improvement mechanism?
- Under comparable resource budgets, can the improvement process itself become more effective?
Making feedback testable
Candidate modifications require evaluation before there is a basis for retaining or reverting them. Evaluation criteria need appropriate independence from the modification process and should cover new tasks and settings.
We examine where modifications come from, what information and resources evaluation uses, and whether improvements hold across repeated experiments.
Evaluation and limits
A single increase in a score does not demonstrate recursive self-improvement. Candidate changes, evaluation criteria, and resource budgets must be explicit, and improvements must withstand testing on new tasks and in new settings.
Three-body civilization experiment
Observe how agents understand their environment, learn and exchange discoveries in the three-body civilization experiment.
Explore the civilization experiment