Recursive self‑improvement (RSI) offers a pathway for AI systems to take part in their own ongoing development. Recent approaches enable language agents to retain changes in parameters, memory, skills, and code for later execution and refinement. We discuss how these approaches can inspire new research ideas from two complementary angles. First, we introduce a unified taxonomy comprising nine design dimensions arranged into four groups: update targets, update organization, update schedule and scope, and feedback and selection. This taxonomy maps existing methods into a shared design space, and examining individual dimensions and their interactions can generate novel research directions. Second, we examine how current RSI methods draw on traditional deep learning techniques such as optimization, regularization, data augmentation, continual learning, and adversarial learning, and we extend ideas by looking at finer‑grained learning mechanisms and evolutionary computation concepts. We also review how existing benchmarks assess self‑improvement under online, offline, and offline‑to‑online settings. Together, these perspectives help researchers develop new RSI ideas based on existing designs and learning mechanisms.
Review