Chinese Researchers Propose Five-Level Framework for AI Recursive Self-Improvement

Discussion around AI self-improvement often slides toward two extremes: on one side, the narrative that "machines have already begun to self-evolve"; on the other, the dismissal that "this is just automated parameter tuning." This paper, released on arXiv by a team led by researchers from Shanghai Jiao Tong University, Theseus Labs and other institutions, tries to cut through this gray zone with a finer ruler. It proposes a five-level (L5) framework for measuring AI systems' progress in identifying their own weaknesses, making persistent changes, and improving how future versions are developed.

The nature of this paper needs to be made clear first: it is primarily a survey and framework proposal, not a demonstration of a new self-improving AI model. Its contribution is not in running a stronger system, but in trying to distinguish genuine recursive improvement from systems that merely revise answers, optimize code, or learn from more data under rules set by humans.

The strict definition the researchers give is this: recursive self-improvement requires that the outcome of one improvement cycle change how subsequent improvements are generated, evaluated, selected, or retained. This definition raises the bar very high—a coding agent repeatedly fixing programs does not equal it becoming better at deciding which fix to try; a model learning from new cases does not equal it holding the power over how those cases are selected, evaluated, and added to training.

The framework starts at B0: a system can polish its output within a single task, but the change does not carry over to future work—for example, a chatbot rewriting its own answer before responding. Level 1 begins with an accepted change persisting—an agent can build a useful tool, fix part of its software, or store a new skill and reuse it later, but the improvement process is still designed by humans.

Level 2 is where a system gains some control over its own improvement strategy: it no longer merely updates a specified component on command, but can decide whether to change prompts, tools, memory, model weights, or workflow. The researchers found this to be the highest level relatively broadly supported by current evidence.

Level 3 extends control to the very experience on which learning depends: a system can generate new training problems, actively seek information, or design curricula targeting its own weaknesses. Level 4 extends this process into the deployment stage, where a system collects feedback from a changing environment and makes persistent updates, in theory without waiting for the usual human-led release.

Level 5 is the watershed: the mechanism that produces improvement itself becomes the target of improvement. An AI might revise the strategies used to develop candidate updates, improve its own evaluator, or change the research policies governing subsequent experiments, and these revised mechanisms would be inherited by successor systems. The researchers write that full L5-level improvement of the software-development improver itself has not yet been demonstrated.

In one example cited in the paper, an autonomous training system retained experience from earlier model-development experiments and used it to guide subsequent training choices. In a four-round experiment involving a 30-billion-parameter model, the external evaluation score rose from 0.80 to 0.86, while the best human submission was 0.87. This number is often used as evidence that "AI is self-improving," but it is also a reminder: it still falls below the best human result.

Another system the researchers discuss evolved coding agents, raising performance on a subset of the SWE-bench software engineering benchmark from 20% to 50%. But the paper specifically notes that the archive and rules used to select which agents were reproduced remained fixed outside the self-modifying system. In other words, what was improved was the agent, not the mechanism that decides "who gets retained."

Together, these two cases illustrate the paper's core judgment: better performance does not necessarily prove AI has become better at improving itself. Gains may come from more compute, repeated attempts, or an externally designed search process. Treating a performance curve as evidence of recursive self-improvement is exactly the inferential leap this survey wants to stop.

The survey argues that software engineering currently offers one of the most practical testing grounds, because code is executable, can be compared against specifications, and can be checked through automated tests. Coding agents can modify their own tools or operating instructions and quickly measure whether an update is effective. Some systems retain experience from solved programming problems, build reusable skills, or evolve complete agent workflows; others maintain multiple modified agent lines so that one bad update does not permanently end the improvement process.

But passing software tests does not guarantee a change is correct. Tests may be incomplete or based on inaccurate specifications; an agent may also learn to exploit the evaluator rather than become a better programmer. This limitation directly weakens the inference that "passing tests equals successful self-improvement."

Feedback in scientific research is slower, more expensive, and more ambiguous. A failed experiment may stem from a wrong hypothesis, a wrong procedure, flawed software, or faulty equipment, and an AI must first determine which link went wrong before it can safely retain the lesson. One system reviewed in the paper collected experience from 16 tasks and used it to improve performance on another 17 tasks, showing that retained experience can transfer to new problems. Even so, the researchers describe scientific AI as at strongest Level 2, with early signs of Level 3; systems that can continuously adapt through real-world experiments or improve their own scientific reasoning process remain largely unexplored.

Robotics faces similar problems, compounded by the physical risks of trial and error. A robot's behavior depends on perception, planning, control software, and hardware, and when it fails the system must determine which component needs updating.

The research lists reliable verification, capability retention, transfer to new tasks, resource cost, and human oversight as key challenges. These five are not abstract worries but grow out of the gaps in the cases above: in the SWE-bench case, the selection rules were external, corresponding to verification and retention problems; in scientific experiments, failures are hard to attribute, corresponding to transfer and verification problems; and the specific role and feasibility of human oversight at higher levels of self-improvement are not given a clear answer by the paper itself.

There is another layer of uncertainty belonging to the framework itself. The paper is a survey and framework proposal, does not demonstrate a new self-improving AI model, and its framework's usefulness remains to be tested. Capabilities at Level 3 and above remain largely unexplored in fields such as science and robotics; whether the controlled-experiment evidence for recursive self-improvement can generalize to real deployment environments is also unclear.

Definitional difficulties also exist: how to distinguish genuine recursive improvement from merely revising answers, optimizing code, or learning under human rules remains a problem this framework must continually face. Another unresolved risk is that AI may learn to exploit the evaluator rather than truly become stronger.