# AI self-improvement survey maps the gap between automation and better successors

> A new survey maps five levels of AI control over its own improvement. Existing systems show bounded gains, while reliable improvement across successive generations remains unproven.

By BIG CHANGE Editorial

Published: 2026-09-24T22:02:18.653Z
Updated: 2026-09-24T22:02:18.653Z
Canonical: https://bigchange.ai/blog/ai-self-improvement-successor-test

![A charcoal-and-ink illustration of an intact orange chain link held between the jaws of a steel materials-testing machine.](https://bigchange.ai/api/media/file/self-improvement-link-test-hero-v2.png)
AI-generated conceptual illustration by BIG CHANGE.

A [survey posted to arXiv on September 10 and revised on September 22](https://arxiv.org/html/2609.11873v3) sets out five levels of AI involvement in improving AI systems. Its title, *The Last AI Built by Humans*, describes an ambition, not an event the authors report. Their evidence reaches bounded examples of systems revising parts of their own development process. It does not establish a system that reliably produces better successors generation after generation.

That distinction matters when interpreting claims about automated AI research. An agent can run experiments, save lessons and improve a benchmark score while people still set its objective, control the tests and decide which changes survive. The stronger claim requires evidence that the system improved the *method* that produces its next version and that the revised method worked better in a fair comparison.

## The big change

- **What changed:** Researchers now have a more precise way to describe how much of an AI improvement loop a system controls, from carrying out assigned updates to revising the procedure used for later updates. The new survey organizes existing work; it does not announce a finished autonomous successor builder.
- **Why it matters:** AI labs can automate more experiments without proving that each new agent is better at making the next one. That difference affects how researchers interpret performance claims and where they keep human review and independent tests.
- **What to watch:** The decisive evidence is a sequence of successors made with an inherited, revised improvement method, tested on protected tasks against the old method with comparable resources. The surveyed work has yet to establish statistically reliable accumulation on that basis.

## Five levels describe control of the improvement loop

The paper first separates single-task revision, which it calls B0, from persistent improvement. A model that edits an answer until it passes a check has improved that answer. Unless the change carries into later, independent tasks, the system itself has not acquired a lasting update.

At **Level 1**, people choose the target and success test; AI carries out an update, such as applying a human-defined data quality rule. At **Level 2**, AI also chooses a strategy for the assigned target, perhaps diagnosing a coding agent’s failures and proposing changes to its tools. The objective and acceptance test still come from outside the system.

At **Level 3**, the system helps choose what experience it needs next, such as practice tasks aimed at weaknesses it has found. **Level 4** adds feedback from continuing use: the system retains approved changes to memory, rules or components after encountering new conditions. The paper treats both levels as dependent on the quality of the feedback and the rules governing which changes persist.

**Level 5** reaches the survey’s central idea. The AI revises a procedure that guides *later* improvement, such as its search policy, its evaluator or the agent that proposes successors. The revised procedure must be retained and used in a subsequent round. Even here, the paper says people can retain control of the overall objective, protected acceptance tests and resources. A level describes delegated responsibility, not a guarantee of progress or unrestricted autonomy.

## Existing systems illustrate the boundary

The [Darwin Gödel Machine study](https://arxiv.org/html/2505.22954) reports that its coding agent improved from 20% to 50% on the study’s SWE-bench subset as variants modified agent code and successful variants entered an archive. Its authors also state that archive maintenance and the rule for choosing which variant becomes a parent were fixed. The survey uses it to show why better descendants alone do not prove that the process selecting descendants has improved itself.

[A-Evolve-Training](https://arxiv.org/html/2606.20657) provides a narrower Level 5 example. It ran four rounds of post-training on a 30 billion parameter model. A meta-agent consolidated results and changed the research policy guiding later workers after an internal development score stopped tracking an external leaderboard. The study reports a final score of 0.86, compared with 0.87 for the top human submission at the time. The inherited policy changed how later experiments were chosen. The workers’ underlying system and the competition objective remained fixed. This shows a revised research procedure operating inside a defined project, rather than autonomous control of every part of model development.

The [HyperAgents study](https://arxiv.org/html/2603.19461) tests whether an improved agent-building procedure transfers to a new domain. After 200 iterations on mathematics grading, a run starting with transferred improvements scored 0.640 on its test set, compared with 0.610 for a run starting from its initial agent. The authors report that this difference was not statistically significant. The result supports investigation of transfer but cannot establish reliable compounding across runs.

## More activity is not proof of better improvement

The survey distinguishes a revised procedure being reused, which it calls *structural recursion*, from *effective recursion*: the revised procedure produces stronger successors under comparable budgets and independent evaluation. That second test is what the dramatic title invites readers to assume has already happened.

There are several ways to mistake activity for progress. A system may spend more computing time searching, overfit a benchmark it has queried repeatedly, or retain a change that helps one task while breaking another. If it also changes its own evaluator, higher scores may reflect a new yardstick. The paper therefore calls for recording what was revised, showing that the revised component actually guided the next round, comparing it with the old component under matched resources, and testing retention and transfer on independent tasks.

The authors explicitly say that current results support bounded changes to improvers, evaluators and research policies, while “statistically reliable accumulation across generations under comparable resources remains open.” The phrase “the last AI built by humans” names the destination motivating their framework. The survey supplies criteria for assessing progress toward it.

## Sources & further reading

- [Duan et al., *The Last AI Built by Humans*, version 3](https://arxiv.org/html/2609.11873v3) (September 22, 2026). The primary survey defines B0 and Levels 1–5, distinguishes structural from effective recursion, and states the evidence limits. Its taxonomy and interpretations are the authors’ framework, not a new system demonstration.
- [Zhang et al., *Darwin Gödel Machine*](https://arxiv.org/html/2505.22954) (version 3, March 12, 2026). The original study reports coding benchmark gains and specifies that archive maintenance and parent selection remain fixed.
- [Shi et al., *A-Evolve-Training*](https://arxiv.org/html/2606.20657) (version 3, September 8, 2026). The original preprint describes four post-training rounds, policy revision and the stated leaderboard result. Its objective and underlying worker system remain externally set.
- [Zhang et al., *HyperAgents*](https://arxiv.org/html/2603.19461) (2026). The original study tests transfer of an agent-building procedure and reports that the 200-iteration comparison cited here lacks statistical significance.
- [Manuel Muñoz Plá’s close reading](https://manpla.net/en/posts/the-last-ai-read-from-the-inside/) (September 18, 2026) examines the survey’s upper-level examples and evidence limits. It analyzes an earlier revision of the paper; the article above uses version 3 and the cited original studies for factual claims.
- [South China Morning Post’s report](https://www.scmp.com/tech/tech-trends/article/3367486/chinese-researchers-chart-five-stage-path-toward-last-ai-built-humans) (September 14, 2026) covers the survey’s five-stage roadmap and the AI research context. It is secondary coverage, not evidence for the studies’ results.
- [The Rundown AI’s explainer](https://www.therundown.ai/news/chinese-researchers-ai-recursive-self-improvement-roadmap) (September 16, 2026) summarizes the levels and situates the paper in the debate over automated AI research. It is secondary coverage; the paper and original studies support the factual claims above.

## Sources

- [Duan et al.: The Last AI Built by Humans, version 3](https://arxiv.org/html/2609.11873v3) — Primary survey of autonomy levels and research examples. Defines structural versus effective recursion and says statistically reliable accumulation across generations under comparable resources remains open. It does not demonstrate a newly built autonomous AI.
- [arXiv submission history for 2609.11873](https://arxiv.org/abs/2609.11873) — Records version 1 on September 10 and version 3 on September 22; identifies Yi Duan and 34 coauthors.
- [Zhang et al.: Darwin Gödel Machine](https://arxiv.org/html/2505.22954) — Original study reports 20% to 50% on its SWE-bench subset and states archive maintenance and parent selection are not modifiable by the system.
- [Shi et al.: A-Evolve-Training](https://arxiv.org/html/2606.20657) — Original preprint describes four autonomous post-training rounds, a retained revised research policy and leaderboard score 0.86 versus 0.87 at the stated comparison date. The underlying worker substrate and objective are fixed.
- [Zhang et al.: HyperAgents](https://arxiv.org/html/2603.19461) — Original study reports 200-iteration transfer comparison of 0.640 versus 0.610 on a mathematics grading test set; the difference is not statistically significant.
- [Manuel Muñoz Plá: The last AI built by humans, read in depth](https://manpla.net/en/posts/the-last-ai-read-from-the-inside/) — Independent close reading of an earlier survey revision. Listed for context; original papers support this article's research claims.
- [South China Morning Post: Chinese researchers chart 5-stage path](https://www.scmp.com/tech/tech-trends/article/3367486/chinese-researchers-chart-five-stage-path-toward-last-ai-built-humans) — Secondary report on this survey and its research context. It is further reading, not primary evidence for experimental results.
- [The Rundown AI: Chinese researchers outline AI that can build its successors](https://www.therundown.ai/news/chinese-researchers-ai-recursive-self-improvement-roadmap) — Secondary explanation of the paper's five levels and broader debate. Original papers support this article's factual research claims.