AI self-improvement research survey

A 75-page paper with 30+ researchers classified AI self-improvement into 5 levels, from human-directed changes (Level 1) to AI rewriting the next generation of AI (Level 5). Reviewing 491 papers, 75% remain at Levels 1-2, while less than 6% reach Level 5.
Key insights
Five-level taxonomy reveals current AI self-improvement status
Researchers defined AI self-improvement in five levels: Level 1 where humans decide what to change, through Level 5 where AI rewrites the next generation of AI creation. The taxonomy is based on a survey of 491 papers, with the paper asserting everything AI has achieved is a drop in the ocean.
Most research remains at early self-improvement stages
75% of surveyed papers stay at Levels 1-2 (human-directed improvement), while fewer than 6% reach Level 5 (AI rewriting next-generation AI). This distribution indicates the field has not yet achieved autonomous recursive self-improvement at scale.