A field guide to AI that improves itself

What it does

Awesome RSI is a curated bibliography of research on recursive self-improvement — AI systems that improve their own capabilities, and potentially the mechanisms doing the improving. It organizes papers, benchmarks, frameworks, and talks across model-level methods (self-training, self-play, self-distillation), harness-level work (prompt, memory, and scaffold evolution), embodied self-improvement, automated AI R&D, and safety.

The interesting bit

The taxonomy is unusually honest about scope. It distinguishes self-refinement (better output, no lasting change), persistent self-improvement (changes that carry into the next round), and true RSI (the improvement mechanism itself is the thing being improved) — and admits that many listed works are only bounded self-improvement or enabling techniques, not the real recursive thing.

Key highlights

• Covers the full stack: weights, prompts, memory, scaffolds, multi-agent co-evolution, and even physical/embodied self-improvement.
• Includes a dedicated Safety, Alignment & Theory section alongside the capability work.
• Strong benchmarks section, including direct RSI evaluations and frontier-lab AI R&D evaluation frameworks.
• Fundamentals section reaches back to Gödel Machines, Schmidhuber’s 1987 meta-learning thesis, and Good’s 1965 intelligence-explosion argument.
• Community-maintained; the maintainers explicitly invite taxonomy critiques and missing papers.

Caveats

• It’s a reading list, not a tool — no code, no runnable anything.
• The taxonomy is explicitly organizational, not mutually exclusive; many systems span multiple layers, so expect overlap.

Verdict

Essential bookmark if you’re researching self-improving agents or automated AI R&D and need to know what exists. Skip it if you wanted a framework to install — this is a map, not the territory.