lobehub/awesome-rsi

What it is

A curated, community-maintained list (an "awesome list") that collects and organizes research papers and resources on Recursive Self-Improvement (RSI). RSI is the idea that an AI system can not only improve its own capabilities, but also improve the very mechanisms it uses to generate those improvements — creating a self-accelerating loop.

Why this matters now

The list argues RSI has moved from pure theory to an active empirical research field, driven by recent progress in:* self-training* agent memory* harness (outer scaffolding) optimization* embodied self-improvement* automated AI research* self-modifying coding agents* evolutionary search

What the list contains

It groups work into sections including:

• Fundamentals — introductory papers and surveys
• Model-level RSI — improving the AI's weights via self-generated data, feedback, or reasoning
• Harness-level RSI — improving prompts, memory, verification, and agent policies around the model
• Multi-agent self-improvement — several agents improving together or debating
• Coding self-improvement — agents that modify their own code
• Automated AI R&D — systems that do AI research themselves
• Embodied / physical self-improvement — improvement in physical agents
• Evolutionary & open-ended RSI
• Safety, alignment & theory
• Introspection & self-modeling
• Benchmarks & evaluations
• Frameworks & tools, blog posts, talks, and related lists

Key concepts it defines

The README introduces useful precise vocabulary:

• Self-refinement — improving the current output without a lasting change to the system
• Persistent self-improvement — changes to weights, memory, skills, prompts, or code that carry into the next round
• Recursive self-improvement — when the mechanism that produces improvements is itself the thing being improved
• RSI substrate — a system that exposes its own structure as modifiable, without necessarily forming an automatic loop

It also distinguishes three levels of evidence for claimed self-improvement:

• Task gain — better outputs, no persistent change
• Retained gain — persistent updates improve future performance
• Improver gain — the improvement-producing mechanism itself becomes more effective

The list stresses that a self-referential architecture alone does not prove real improvement — you need controlled comparisons.

Important caveats

The list is honest that not every entry demonstrates true RSI; some are bounded self-improvement or enabling techniques. The taxonomy is organizational, not strict — many systems span multiple categories. New contributors are explicitly welcome.