Recursive self-improvement gets a five-level yardstick: most systems stop at strategy, not meta-improvement
Synopsis
Researchers from Shanghai Jiao Tong University and partner institutions propose treating the improvement loop, rather than any single algorithm, as the unit of analysis for recursive self-improvement, and sort existing work into five autonomy levels: executing human-specified improvements, choosing improvement strategies, acquiring learning experience, adapting in deployment, and revising the mechanism that governs later improvement. The framework also uses a Headroom-Closed Index to compare capability trajectories across domains from 2023 to 2026, finding the largest remaining gaps in interactive capabilities such as tool use and software engineering, and asks which improvement decisions still stay with humans.
Interpretation
Recursive self-improvement is broken into a loop that can be checked level by level, each level marking a different improvement decision handed from humans to AI. Earlier surveys of self-evolving systems organized work by update object, timing or technical mechanism, leaving unclear how much decision authority AI holds when the same component is updated. The five levels run from improvement execution through improvement strategy, learning-signal or experience acquisition, environment adaptation, to recursive inheritance, and each asks where the loop closes, what is retained into the next round, and which critical decisions remain under human control.
Capability trajectories show that quickly verifiable domains close fastest, while interactive tasks that require long-horizon state tracking leave larger gaps. A single aggregate benchmark would flatten both the level differences and the differing growth rhythms across domains. Among 393 model-benchmark observations for models released between 2023 and September 2026, the 2026 Headroom-Closed Index reaches 86.4 for advanced mathematics and 85.8 for graduate-level science, against 52.6 for software engineering, 56.8 for search and terminal agents, and 39.9 for tool agents.
The framework separates structural recursion from effective recursion: a revised improvement mechanism entering the next round does not mean it produced stronger successors under comparable budgets and independent evaluation. Higher task performance alone does not show that an improvement mechanism was revised, retained and reused. The Darwin Gödel Machine raises performance on its SWE-bench subset from 20% to 50%, but its archive maintenance and parent-selection rules remain outside self-modification; 14% of Gödel Agent's 100 MGSM optimization trials ended below the initial policy's performance.
Perspective
The scope covers academic literature and industrial practice, including technical reports, engineering blogs, open-source repositories and model documentation; the capability trajectories admit only results whose benchmark version and evaluation harness can be aligned, with 17 of 33 results in the latest-model audit entering the trajectories and the other 16 kept for reference.
The authors label the post-2026 curves as illustrative extrapolations that shrink remaining headroom by a fixed fraction rather than measured forecasts; later cybersecurity observations use changed Cybench subsets or pass@1 aggregation, so that comparison needs caution; public descriptions of industrial systems document operational automation but do not by themselves establish a complete learner-experience loop.
