---
title: 'Recursive self-improvement (RSI)'
description: "A contested claim that an AI system can improve its own capability in repeated rounds, where each round's gains make the next round more effective."
canonical_url: 'https://darkfactory.dev/glossary/recursive-self-improvement'
markdown_url: 'https://darkfactory.dev/glossary/recursive-self-improvement.md'
collection: glossary
date_published: '2026-09-17T00:00:00-04:00'
date_modified: '2026-09-17T00:00:00-04:00'
---

# Recursive self-improvement (RSI)


## Definition

Recursive self-improvement is a contested claim that an AI system can improve its own capability in repeated rounds, where each round's gains make the next round more effective. The recursion carries the claim. Steady gains per cycle are ordinary optimization; the assertion here is that improving raises the rate of further improvement.

No shared threshold governs the label. Before accepting or disputing an RSI claim, establish two things: what part of the system changed, and who controlled the rules that let the change through.

## Two different claims share the name

The model-level claim is that the system improves its own weights or its general intelligence. This is the reading behind intelligence-explosion arguments, and no published work demonstrates it.

The harness-level claim is that the system rewrites the scaffolding around a fixed model: the exploration policy, the controller, the prompts, the routing, the memory. The model is untouched and the orchestration improves. This is the reading a software factory can act on, because the harness is the layer a factory operator owns.

Dream-RSI, published in September 2026 by co-authors at Google, Google DeepMind and two universities, sits in the second category. It treats a coding agent's history, the record of what it tried and what happened, as a replay simulator, then scores candidate exploration policies by re-walking that recorded tree instead of re-running the agent. The paper is explicit that the underlying coding agent is unchanged.

## What has been measured

Dream-RSI reports benchmark gains from that harness rewrite: in algorithm engineering on Gemini 3.1 Pro, 1.22x faster downstream runtime and 1.74x less discovery compute than Recursive Fixed Exploration; in GPU kernel engineering, 4 of 4 kernels improved at 2.09x higher performance for equal budget; in mathematical optimization, 2 of 3 tasks matched or beat the baseline. The figures are first-party, and full reproduction code was not released with the paper.

A 35-author roadmap published days earlier proposes five autonomy levels, from a system that executes human-specified updates to one that revises the mechanisms governing its own future improvement, and places no existing system at the top level. Its two worked cases show what the distinction rules out. The Darwin Gödel Machine raised its score on a SWE-bench subset from 20 percent to 50 percent while leaving archive-maintenance and parent-selection rules outside its own control, which buys better outputs and not a self-revised improvement process. Gödel Agent, which does rewrite its own improvement logic, ended 14 percent of its 100 recorded optimization trials below where it started. The authors read that as evidence that running the loop longer does not by itself produce progress.

## Why it is contested

The disagreement is about scope and threshold, and the two compound.

On scope, there is no agreement about which part of the system has to improve. Optimizing a controller that will steer the same agent on the next round meets a literal reading of the words without matching what most readers mean by recursion. When Dream-RSI reached Hacker News, one commenter called the work "not recursively intelligent in any way" and another argued that "if this is RSI then all RL is RSI." A co-author answered by restating the mechanism rather than defending the label: the paper "optimizes a controller/policy which will be used to agent itself in the next round."

On threshold, there is no agreed level at which iteration becomes recursion. The five-level framework exists precisely because the field lacked one, and it is a proposal rather than a standard. Two papers can therefore use RSI for a rewritten exploration heuristic and for an AI that rewrites the rules governing its own future, and a reader cannot tell which from the acronym.

This glossary treats an RSI claim as incomplete unless the speaker names the component that changes, the evaluator that judges the change, and the authority that promotes it.

## Distinguish it from nearby terms

- **Controlled self-improvement** is the bounded practice: the system may propose changes to its own prompts, skills, memory or harness, but evaluation data, promotion authority and rollback stay outside the mutation boundary. RSI makes no such commitment, and the frameworks that define it treat control over one's own improvement rules as the goal rather than the hazard.
- **Continuous improvement** uses the same loop with humans proposing every change.
- **Fine-tuning** changes model weights through a training run that the system does not govern.
- **Artificial general intelligence** is a claim about breadth of competence. RSI is a claim about the mechanism of change. Neither implies the other.
- **Reinforcement learning** already improves a policy from its own experience, which is why "all RL is RSI" is the sharpest objection to a loose definition.

## Operational significance

The harness reading brings a governance problem with it. A system that rewrites its exploration policy is changing how work gets chosen, which moves cost and failure modes without necessarily moving a headline quality score. The five-level framework's discriminator is the useful test to apply locally: does this change produce better outputs, or does it change the mechanism that decides what counts as better? The first belongs in a normal promotion gate. The second changes the gate, and needs independent evidence and a rollback path of its own.

## Check your understanding

An agent rewrites its own retrieval strategy, scores the result on a held-out suite it cannot modify, and a human promotes the winning version. Is that recursive self-improvement? Under the harness reading, yes, at the first level. The more useful observation is that the evaluator and the promotion authority stayed outside the loop. Move either one inside it and the system is no longer measured against a standard it cannot change.

## Also called

recursive self improvement, self-improving AI

## Related terms

- [Controlled self-improvement](https://darkfactory.dev/glossary/controlled-self-improvement)
- [Improvement graph](https://darkfactory.dev/glossary/improvement-graph)
- [Artificial general intelligence (AGI)](https://darkfactory.dev/glossary/artificial-general-intelligence)
- [Autonomy](https://darkfactory.dev/glossary/autonomy)
- [Agent harness](https://darkfactory.dev/glossary/agent-harness)
- [Loop engineering](https://darkfactory.dev/glossary/loop-engineering)
- [Reward hacking](https://darkfactory.dev/glossary/reward-hacking)

## Related factory areas

- [Feedback, learning & controlled self-improvement](https://darkfactory.dev/factory/feedback-self-improvement)
- [Verification, evaluation & quality truth](https://darkfactory.dev/factory/verification)

## Evidence and further reading

- [Dream-RSI: Recursive Self-Improvement through Evolving Worlds](https://arxiv.org/abs/2609.14858)
- [The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement](https://arxiv.org/abs/2609.11873)
- [Hacker News: is Dream-RSI actually recursive self-improvement?](https://news.ycombinator.com/item?id=49726955)