A survey of 1,250 arXiv papers published between 2024 and 2026 analyzes AI systems that participate in their own improvement through mechanisms like self-refine, self-reward, and self-play. The authors propose a taxonomy separating bounded self-refinement, which is convergent and already used in industry, from open-ended recursive self-improvement (RSI), which remains limited by grounding requirements and compute constraints.
- The taxonomy classifies improvements based on what is improved (behavior, policy, evaluator, or research process) and the degree of loop closure.
- Self-evaluation is identified as a distinctive feature, where every improvement loop claims a signal can substitute for human judgment.
- A verification hierarchy orders signals from formal verifiers to intrinsic self-assessment, with demonstrated improvement strength tracking this order.
- Failure modes such as model collapse and diversity collapse are linked to violations of this hierarchy.
- The research direction-setting bottleneck keeps humans in the loop at the top of the hierarchy.
The work connects technical literature to RSI limits and safety governance, identifying governance-grade measurement of self-improvement as a critical underpopulated niche.