Researchers from Shanghai Jiao Tong University, working with teams from Tsinghua University, ByteDance, Xiaohongshu and the Shanghai AI Lab, have published a review titled The Last AI Built by Humans. The paper surveys 491 related studies and tries to draw a tighter boundary around the increasingly broad idea of AI "self-evolution."
A five-level RSI framework
The paper divides recursive self-improvement, or RSI, into five levels. L1 covers systems that only execute improvement procedures predefined by humans. L2 covers systems that can decide how to improve themselves. At L3, a system can also decide what it should learn in the next round. L4 refers to systems that use feedback from real operation to keep changing memory, skills, code or the harness, with those changes carrying over into later tasks.
At L5, the system can also modify the search methods, evaluators and research strategies that generate the next round of improvements, then pass those changes forward for continued use in future iterations.
Rewriting code alone is not enough
Under the paper's framework, simply being able to rewrite one's own code does not amount to full recursive self-improvement. The authors use a Coding Agent as an example: even if it can rewrite its own source code, it only shows self-modification if humans still determine how the next generation is selected, what scoring standard is used and which changes are allowed to remain. In that case, it has not yet achieved complete RSI.
L5 is split into two layers
The paper further breaks L5 into two layers. The first is "formally recursive": the AI can already modify the mechanisms responsible for later rounds of improvement, and those mechanisms are reused by the next generation. The second asks whether the system is actually getting stronger: the modified mechanism must produce a stronger next generation under comparable resources and independent evaluation.
How the 491 studies were classified
Of the 491 papers reviewed, 43.8% were classified as L1 and 31.6% as L2. Only 29 papers, or 5.9%, fell into L5. Based on that distribution, much of the work is still concentrated on executing improvement procedures designed by humans or automatically searching for ways to improve. Far fewer studies hand the improvement mechanism itself over to AI.
The paper also says that even when research reaches L5, long-term and stable accumulation of advantage has still not been demonstrated.
A proposed framework, not an industry standard
The paper's main purpose is to establish a classification framework for the RSI field. Its L1-to-L5 scale is proposed by the authors and is not yet a universally accepted industry standard.
The review includes academic papers, technical reports, official blogs and open-source systems. According to the paper, many frontier RSI practices have not yet entered the formal academic paper pipeline.