A record of RSI's progress

Recursive self‑improvement

AI systems that make themselves better. In the strict case, marked ↻, the loop also improves the part that does the improving.

The map of the field →
All 129 papers and projects →

Each dot is one entry ↻ improves its own improverother

1965–2012 5
2013 3
20141
20151
2016 3
2017 2
2018
2019
20201
2021
2022 8
2023 17
2024 19
2025 27
2026 42

Levels 0–5

Ask the questions in order; a system's level is the last one it passes. Each level lets the loop change one more part of the system.

What the loop is allowed to change at each level
Level	1Its answer, during one task	2Its prompts, memory, tools, code or weights, kept for later tasks	3The part that proposes and accepts changes	4The model that part runs on	5Its successor, built by its own research
0Fixed	·	·	·	·	·
1Retries	●	·	·	·	·
2Learns	●	●	·	·	·
3Improves its improver	●	●	●	·	·
4Improves its own model	●	●	●	●	·
5Builds its successor	●	●	●	●	●

A level says how much of a system its loop is allowed to change, not how well the system works. Levels are given to systems that were built and run; designs that were never run, such as the Gödel machine, carry the ↻ mark without a level. The scale and the 65 placements were drafted with Claude on 30 October 2026 from each paper's abstract and have not yet been checked by the RSI system.

Level 0: Fixed

TestPasses none of the tests below.

Stays fixed: Everything.
Example: A model with a fixed prompt answering questions. It is where every system starts.

Level 1: Retries

TestDoes it revise its own output during a task, using feedback it gathers?

Stays fixed: Everything except the current answer. Nothing it learns carries over to the next task.
Limit so far: Revision works when the feedback comes from outside the model: Reflexion reached 91% on HumanEval by running its code, reflecting on the failures and retrying. Without outside feedback, models asked to correct their own reasoning often get worse (Huang et al.).

3 systems in the list · see all

Level 2: Learns

TestDoes a change it makes carry over to later tasks, in its prompts, memory, tools, code or weights?

Stays fixed: The improver. A procedure written by people proposes and accepts every change, and the loop cannot alter it.
Limit so far: Gains taper off because the improver never gets better. When the Darwin Gödel Machine was run with its improver held fixed, its "gains taper off quickly" (DGM); repeated self-training "can reach a point of diminishing returns" (Multiagent Finetuning).

36 systems in the list · see all

Level 3: Improves its improver ↻

TestCan the loop change the part that proposes or accepts changes, such as its editing code, its mutation rules, its judge or its task-setter?

Stays fixed: The model the improver runs on, and usually the outer rule that picks which version to keep.
Limit so far: Something always stays fixed so the system stays stable. Self-editing systems "must leave part of their own editing machinery untouched", which caps how many levels of self-editing they actually reach at about two (Metaⁿ). Gains so far rest on a frozen model: DGM went from 20% to 50% on SWE-bench with the same Claude 3.5 Sonnet throughout.

17 systems in the list · see all

Level 4: Improves its own model ↻

TestBesides its improver, does the loop retrain the weights of the model that does the improving, so the next round runs on the improved model?

Stays fixed: The training recipe, the outside checks, and the model's design.
Limit so far: Runs are short: Self-Rewarding reports three rounds, MASS two. Each round costs a full training run, and a model that grades its own training data can drift toward what its grader likes (Gao et al.).

9 systems in the list · see all

Level 5: Builds its successor

TestDoes the system carry out the research that produces a more capable next generation, which then runs the next round, with no person needed and no slowdown from round to round?

Stays fixed: Only the goal and the outside world it is measured against.
Status: Not observed. This is the "intelligence explosion" loop of Good (1965). Whether it would speed up or stall depends on returns: if each doubling of software efficiency needs more than twice the research effort of the last, progress fizzles; if less, it runs away (Eth and Davidson).

Signals so far

AlphaEvolve sped up a kernel in Gemini's architecture by 23%, cutting Gemini's training time by 1%. Gemini is the model AlphaEvolve runs on, and engineers put the change in place.
As of May 2026, more than 80% of the code merged into Anthropic's codebase was written by Claude (When AI builds itself).
AIDE², an AI research agent, found seven successive improvements to its own code in an 8-day run; the gains held on four held-out benchmarks.
Estimates of today's feedback loops say they are not yet self-sustaining but are strengthening (The Economics of RSI).