SAI-V: A Multidimensional Framework for Self-Improving Artificial Intelligence

Source abstract

Self-improving artificial intelligence (AI) refers to systems that participate in identifying limitations, generating or selecting candidate modifications, evaluating their effects, and retaining validated changes that improve subsequent behavior. Existing work spans output revision, memory adaptation, synthetic-data generation, workflow optimization, parameter updates, code modification, and recursive improvement, but these directions remain fragmented in their definitions, units of analysis, and evaluation criteria. This paper introduces the SAI-V framework for characterizing and validating self-improving AI across six dimensions: target, trigger, mechanism, evaluator, persistence, and autonomy. The framework is complemented by a target-based taxonomy, a classification of improvement mechanisms, a seven-level autonomy scale, and a multidimensional evaluation model covering capability, efficiency, generalization, stability, autonomy, and safety. A controlled lifecycle based on Variation, Assessment, Verification, and Integration separates candidate modification from validated improvement and incorporates monitoring and rollback for persistent change. The framework emphasizes that modification alone does not establish improvement: retained changes must demonstrate net benefit relative to an appropriate baseline while preserving relevant operational constraints. The paper further analyzes representative application domains, technical and conceptual limitations, safety and control challenges, and open research problems involving evaluator independence, self-generated data, reproducibility, repeated improvement cycles, and recursive self-improvement.