AI Self-Improvement Taxonomy Distinguishes Bounded Refinement from RSI

AI systems are increasingly involved in their own development, from refining outputs to conducting AI research. A recent survey of 1,250 arXiv papers (2024-2026) proposes a new taxonomy to clarify the diverse approaches to AI self-improvement. This taxonomy categorizes systems based on what they improve (behavior, policy, evaluator, or research process) and the level of human involvement (human-in-the-loop to fully closed).The research differentiates "bounded self-refinement," which is characterized by convergence, evaluability, and current industrial application, from "open-ended recursive self-improvement (RSI)." RSI, despite its ambitious goals, remains constrained by grounding requirements, collapse dynamics, and computational limits across all measured aspects. A key feature of this taxonomy is its focus on self-evaluation, as every improvement loop implies that some signal can substitute for human judgment.The study maps the evaluator design space, ordering signals into a verification hierarchy from strong formal verifiers to weak intrinsic self-assessment. It observes that the strength of demonstrated self-improvement correlates with this hierarchy, and common failure modes like self-confirming loops or model collapse stem from violating its principles. The "research direction-setting" bottleneck, which keeps humans in the loop, sits at the top of this hierarchy. The paper connects technical literature to RSI limits and discusses safety and governance implications, identifying governance-grade measurement of self-improvement as a critical, under-addressed area.