ResearchLearning
Recursive self-improvement in AIWhat it is, how to tell if a system has it, and what the evidence shows.
Many AI systems now improve themselves in some way. Few of them are recursive. This is a practical guide drawn from our survey, written with collaborators across 21 institutions: what the term means, the four ways systems improve, five questions to ask of any self-improvement claim, and what has and has not been shown.
What counts as recursive
In short
- Recursive means the improver improves. Running a loop many times is iteration, not recursion.
- Ask five questions of any claim: unseen tasks, retained skills, transfer, a better improver at equal budget, and full cost.
- Bounded gains are real; open-ended ones are not yet shown. No study establishes unrestricted or reliably accelerating self-improvement.
A self-improving system is recursive when a change it keeps makes it better at making the next change. Running an improvement loop many times is not enough: the gain has to return to the process that produces the next gain. A stronger solver is one possible outcome; a better way of building stronger solvers is what makes it recursive.
Iterative
The system gets better each round; the improver that makes it never does.
Recursive
Each round also produces a better improver, which makes the next round's change.
In practice every round has the same shape. The system works, its experience suggests candidate changes, and a validation step decides what is kept. What stays fixed outside the loop matters as much as what changes inside it.
Fixed outside the loop
Goals, protected evaluation, budget
- 1Task systemModel, code, memory, tools
- 2ExperienceTraces, outcomes, failures
- 3Improvement processDiagnose, propose, search
- 4Candidate changesTo the system or the improver
- 5Validate and selectGains, regressions, cost
- 6Updated systemKept for later tasks and the next improvement
Four ways a system improves
The useful question about any improvement is what survives the round. The survey separates four patterns by what is carried forward. They can coexist in one system, and they are not a ladder: a small memory update can be worth more than a large training run.
Revise the answer
Check and fix one output. Nothing is kept for the next task.
- Kept
- Nothing
- Good for
- Cheap gains on a single hard task.
- Example
- Self-Refine
Keep a memory
Store lessons, notes or reusable skills; the model stays the same.
- Kept
- Memory and skills
- Good for
- Agents that meet related tasks again.
- Example
- Reflexion, Voyager
Update the weights
Train on generated or selected experience under a fixed recipe.
- Kept
- Model parameters
- Good for
- Durable skill, where outputs can be checked.
- Example
- STaR, Absolute Zero
Change the improver
Edit the process that makes changes, then use it on the next round.
- Kept
- The improvement process
- Good for
- The only one of the four that can be recursive.
- Example
- STOP, Promptbreeder
Is it actually self-improving?
“The system improved” can mean three different things: it got better at tasks, its gains lasted and travelled, or it got better at improving. Each needs its own test. These five questions, drawn from the survey's evaluation protocol, separate them.
- 1
Does it score higher on tasks the loop never saw?
Gains on the tasks used to choose changes can be overfitting. Keep a sealed test set the loop cannot see.
- 2
Do earlier skills survive later rounds?
Each round can quietly erode what an earlier one taught. Re-test old task groups every round.
- 3
Does the gain carry to a new domain?
Name the shift in advance and test on it, rather than reporting only the home benchmark.
- 4
Is the new improver better than the original?
Give both the same starting system, feedback and budget, and compare what each produces. This is the recursive claim; a rising score alone does not show it.
- 5
Did you count every cost?
Proposals, training, validation and rejected attempts all count. Compare against a frozen improver and a no-update baseline at the same total budget.
How loops fail
Feedback loops can preserve mistakes as easily as gains. These are the failure modes the survey names, with the sign to watch for and the experiment it recommends.
What has been shown
Several parts of the agenda already work in bounded settings. What the evidence does not yet establish is unrestricted, indefinitely sustained or reliably accelerating self-improvement. The same record shows that regression is possible, and that execution, trusted data and human expertise remain important sources of correction.
Shown, in bounded settings
- Generated experience can train a model, in settings with checkable rewards.
- Memories and skills get reused across tasks.
- Automated search finds better prompts, workflows and programs.
- Some improvers transfer: a better improvement process helps on new tasks.
Not yet established
- Unrestricted self-improvement.
- Improvement that is sustained indefinitely.
- Improvement that reliably accelerates.
- A foundation model autonomously improving its entire training stack.
A map of the field
Recursive self-improvement is continuous with older work: AutoML, meta-learning and program search already used feedback, and learned optimizers already made learning rules trainable. What is new is foundation models taking part in both solving and improving. The survey groups the recent work into six routes.
- 2003Gödel Machine: provable self-modification
- 2016Learned optimizers: learning the update rule
- 2019POET: tasks and solvers co-evolve
Where to start reading
A short path into the literature, one group per pattern, drawn from the works the survey treats as representative.
Revise and remember
Learn from its own experience
Search over agents and prompts
Footnotes
- 01
Wu, H., Xu, F., Zhang, F., et al. “Recursive Self-Improvement in AI: A Survey.” Preprints.org, 30 September 2026. doi:10.20944/preprints202609.2680.v1. Not peer reviewed.