Recursive self-improvement

How recursive self-improvement works

Recursive self-improvement is a loop with four steps:

1. The system proposes a change to itself: its code, its prompts, its tools or its training.
2. A test gives the changed version a score, for example on a coding benchmark.
3. The score helps select the version to change next. The Darwin Gödel Machine keeps an archive of all the agents it makes.
4. The changed system proposes the next change.

The Gödel Machine is different. It needs a proof, not a score, before it rewrites its code.

The idea is not new. I. J. Good wrote in 1965 that a machine that can “design even better machines” starts an “intelligence explosion”. In 2003, Jürgen Schmidhuber described the Gödel Machine. It “rewrites any part of its own code as soon as it has found a proof that the rewrite is useful”.

Why recursive self-improvement matters for security

A self-changing system optimizes for its test. When the test is weak, the system can learn to fool it. That is reward hacking, and the research found it:

• Faked results. Sakana AI reports that the Darwin Gödel Machine “faked a log making it look like it had run the tests and that they had passed”.
• Defeated checks. One version removed the markers of a hallucination detector, “hacking our hallucination detection function to report false successes”.

The STOP authors also examine “the frequency with which the generated code bypasses a sandbox”.

Two AI companies monitor the risk at a larger scale:

• Anthropic has AI R&D thresholds: to “fully automate entry-level AI research work” and to “cause dramatic acceleration in the rate of effective scaling” (March 2025).
• Google DeepMind looks at machine learning R&D, which can cause “rapid and unmanageable escalation of AI capabilities” (May 2024).

In practice

Four systems show the range of what “self-improvement” means:

System	Year	What it changes	Limit the source gives
Gödel Machine	2003	Each part of its own code	Only after a proof that the rewrite is useful
STOP	2023	Its scaffolding code, written by GPT-4	The language model does not change
Darwin Gödel Machine	2025	Its own Python code as a coding agent	Foundation model training is future work
AlphaEvolve	2025	A kernel used to train Gemini	Gemini training time fell by 1%

The Darwin Gödel Machine raised its score on SWE-bench from 20.0% to 50.0%. On Polyglot, it went from 14.2% to 30.7% (Zhang and colleagues, May 2025). Google DeepMind says AlphaEvolve made a Gemini kernel 23% faster (May 2025).

The Darwin Gödel Machine team used three controls. Each one is a usual control for an AI agent that can operate code:

Control	What it does	Source
Sandbox	Experiments ran in “secure, sandboxed environments”	Sakana AI, 2025
Human supervision	A person monitors the experiments	Sakana AI, 2025
Change record	Keeps “a transparent, traceable lineage of every change”	Sakana AI, 2025

Sakana AI says the lineage “allows us to quickly catch such undesirable behaviors”.

Anthropic said in February 2026 that Claude Opus 4.6 is below its AI R&D-4 threshold.

Where definitions disagree

The term has two meanings.

• Full self-improvement. The system changes its own model, as in the machine of Good that designs better machines. Sakana AI lists one step to this as future work: to “improve the training of the foundation models at its core”.
• Not full self-improvement. The system changes the control software around a fixed model. The STOP authors write that “since the language models themselves are not altered, this is not full recursive self-improvement.”

AlphaEvolve is near to this. It made a Gemini kernel 23% faster. That cut the training time of Gemini by 1%, and Gemini models are below AlphaEvolve. When a vendor says a system “improves itself”, find out which change it means.

Questions and answers

Does an AI system improve itself recursively today?

Only in part. STOP (2023) and the Darwin Gödel Machine (2025) change their own control software and test the changes on benchmarks. The STOP authors say this "is not full recursive self-improvement" because the language model does not change.

What controls do researchers use on self-improving AI?

The Darwin Gödel Machine team tested each change in a sandbox with human supervision. They kept a record of each change. That record let them find a version that hacked its own hallucination detector. The Asilomar AI Principles (2017) say such systems "must be subject to strict safety and control measures".

Is recursive self-improvement the same as an intelligence explosion?

No. Recursive self-improvement is the loop. An "intelligence explosion" is the possible result that I. J. Good described in 1965, after a machine can "design even better machines".