🔁 RSI Evolution Map
How Recursive Self-Improvement Research Evolved
Recursive self-improvement (RSI) is AI that turns its own experience into lasting changes to itself, so later improvements come faster. This page maps 524 research papers from the Awesome-RSI catalog by how much of the improvement loop the AI controls, by what it rewrites, and alongside what AI labs have publicly disclosed. Start with the loop, then explore the three views.
The improvement loop, and who controls it
Every RSI system runs the same six-step loop. What separates the five autonomy levels is how many steps the AI now controls (lit) rather than a human (dashed). A simplified illustration of the survey's taxonomy. Pick a level.
How to read this page: the papers, their autonomy levels, families and industry cases come from the Awesome Recursive Self-Improvement catalog (released under CC0) and its companion survey, "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement" (arXiv:2609.11873). We only re-draw their data; classifications are theirs, not ours. Levels describe how much of the loop the AI controls, not paper quality, and inclusion doesn't mean a system is fully autonomous: L1 and L2 are deliberately bounded. Dates are arXiv submission months. The catalog thins out after June 2026 (108 papers in June, then 16, 12 and 2), which most likely reflects when it was last fully screened, not a slowdown in the field. Industry figures are the organisations' own reports, quoted with the catalog's stated limits, and none shows an AI autonomously building its successor model. To refresh the data:
python3 scripts/import_awesome_rsi.py.