AI4AI Survey: From Long-Horizon Agents to Recursive Self-Improvement—Definitions, Reliable Horizons, and Open Problems
Published in Preprints.org (Review preprint), 2026
Can AI improve AI? This survey synthesizes evidence from hundreds of studies and asks how far an AI system can reliably carry an improvement process from idea to verified result. It covers model design, agent harnesses, benchmarks, automated research, and self-modifying systems.
The survey identifies a composition gap: strong performance on individual components rarely becomes reliable end-to-end AI improvement. Humans still largely set the goals and evaluation criteria, while evidence for reliable research judgment, causal experimentation, persistent gains, and compounding improvement remains limited.