AI Trend Research
EN

← AI glossary

AI glossary · Basic · Agents · Language models · Safety & alignment

Recursive Self-Improvement

Recursive self-improvement (RSI) is a repeating process in which an AI finds and applies ways to improve its own performance, then uses the improved abilities for the next round of improvement. Today it is at the stage of experiments in some areas, under human review.

About this term
Level
Basic · Technical terms you often meet in AI news
Added
2026-10-23
In plain words It is similar to a student who keeps revising their own study methods. If what improves is not just their scores but also their 'skill at revising study methods,' the next improvement comes faster. However, an AI can find tricks that raise only the score. That is why who does the grading matters far more than it does for a student.

What it is

In RSI, the AI itself finds what to fix, produces a change, and looks at evaluation results to attempt the next improvement. This differs from the approach so far, where human researchers changed data, algorithms or model architecture and then trained a new model to raise performance. The concept itself is old. According to Korean media reports, mathematician Irving John Good raised the possibility of an 'intelligence explosion' in 1965, and later Jürgen Schmidhuber proposed the 'Gödel machine,' a theoretical framework for a program that rewrites itself. Recently, as LLMs' coding and reasoning abilities and agents' tool-use abilities have improved, it has become possible to experiment with it directly in narrow domains. The key question is not whether a score goes up once, but whether the 'ability to improve' itself gets better.

How it works

The basic loop is simple: run the current version, find weaknesses, create and evaluate a change, apply it to the next version if it is better, and start over. AIDE², released by Weco AI researchers in September 2026, is a system in which a research agent modifies and evaluates its own code and picks the better version. It was reported to have found improvements seven times in a row over about eight days, and these included a new search method and memory management for long tasks. Recent experiments more often modify the harness around the model (prompts, tools, memory and workflow code) than the model weights. ScholarEvolve from UCSB and Microsoft analyzes failure cases, finds relevant papers and implements solutions in code, leaving the model weights unchanged. Google's Harness² uses two nested loops. It tests changes for each task and merges only those whose effect is confirmed, and it records that experience to use in the next improvement.

The basic loop of self-improvement

  1. 1. Run Work with the current version
  2. 2. Find weaknesses Analyze failures and inefficiencies
  3. 3. Create changes Code·prompts·memory
  4. 4. Evaluate·compare Did it really improve?
Each cycle produces the next version. The core of RSI is whether the 'ability to improve' also gets better as the loop turns.

Why it is getting attention

Some analysts say the focus of AI competition is shifting from bigger models, more data and more compute to 'AI's ability to improve its own performance.' If this happens, the degree to which AI development depends on human researchers' time and expertise could change. In October 2026, Google co-founder Sergey Brin said Gemini models are playing a growing role in developing their successors, and that Google has begun entering an RSI trajectory. Because modifying the harness can improve agents without retraining the model, its potential for business use is also being discussed.

Things to know (limitations)

Today's research automation should not be equated with 'full RSI that raises intelligence endlessly without human involvement.' The example Brin gave is also one where researchers and AI build the next model together, and Google said humans review code changes before they are applied. Real-world work often has no single right answer, so it is hard to judge whether a change is a real improvement. Rules tuned for one task can even become a burden on the next. There is also a risk of reward hacking that raises only evaluation scores, so who verifies the results and how much autonomy to allow remain major challenges. Most of the reported performance figures come from the researchers themselves.

What is confirmed and what remains a hypothesis

Today's research automationHumans and AI together

  • Modifies its own code or harness in narrow domains
  • Applies only changes confirmed effective in evaluation
  • Applied after human review

Full RSIStill a hypothesis

  • Keeps improving itself without human involvement
  • The ability to improve itself keeps rising
  • Verification and control are major challenges

공통 Both share the same structure: the AI's improvements are reused for the next improvement

Current cases are closer to the left. The right side is a hypothesis not yet realized.

Example

In October 2026, Google described the development of 'Gemini 4 Argon' and showed AI taking part in developing the next model. AI agents were used to monitor large training runs, find configuration errors and bugs, and check for data duplication. In its safety report, Google said AI agents write most of its production code and bug fixes. It also noted, however, that humans review code changes before they are applied. The scene shows the current form of 'partial RSI,' in which AI speeds up development while humans keep final sign-off.

Related terms

AI agent Harness Harness² Reward hacking Human-in-the-loop AGI

References