Taxonomy of Recursive Self-Improvement Methods for AI Agents

This analysis introduces a nine-dimension taxonomy for recursive self-improvement (RSI) across update targets, organization, schedule, and feedback. It maps how agentic parameter and skill retention draws from traditional deep learning techniques like regularization and continual learning.