Modifying one's own code does not count as true self-evolution: Chinese research team divides RSI into 5 levels.

Beating AI News Flash: Shanghai Jiao Tong University, together with Tsinghua University, ByteDance, Xiaohongshu, Shanghai AI Lab, and other teams, released a survey titled "The Last AI Built by Humans," reviewing 491 related studies in an attempt to draw a stricter line around the increasingly broad concept of "AI self-evolution."

The paper divides recursive self-improvement (RSI) into 5 levels. L1 only executes an improvement process specified by humans; L2 can decide on its own how to improve; L3 can even decide on its own what to learn in the next round; L4 continuously modifies memory, skills, code, or harness based on feedback from actual operation, allowing these changes to affect subsequent tasks; by L5, even the search methods, evaluators, and research strategies responsible for generating the next round of improvement can themselves be modified and carried forward into the next round.

So simply "being able to modify its own code" is not enough. For example, if a Coding Agent can rewrite its own source code, but how the next generation is selected, by what standards it is scored, and which modifications can persist are still hard-coded by humans, then it can only prove that it can self-modify, and has not yet mastered complete recursive self-improvement.

The paper further splits the highest level, L5, into two layers. The first layer is "formal recursion": AI can already modify the mechanisms responsible for subsequent improvement and have the next round continue to use them. The second layer is "actually getting stronger through iteration": the modified mechanisms must also, under comparable resources and independent evaluation, genuinely produce a stronger next generation.

Among the 491 papers, 43.8% were classified as L1, 31.6% as L2, and only 29 papers were at L5, accounting for 5.9%. A large amount of research is still concentrated on executing improvements designed by humans and automatically finding ways to improve; very little work truly hands the mechanism of "how to improve" itself over to AI. Even when L5 is reached, long-term stable accumulation of advantages has not yet been proven.

This paper is mainly establishing a classification framework for the field of RSI. L1 to L5 are the standards proposed by the authors and are not yet an industry-recognized unified grading system. The paper also includes academic papers, technical reports, official blogs, and open-source systems, because many frontier RSI practices have not yet entered formal papers.