75-page paper from 30+ researchers lays out a 5-level roadmap for recursive self-improving AI
Researchers from 9 institutions including Shanghai Jiao Tong University, Tsinghua, ByteDance and Shanghai AI Lab published a 75-page paper "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement" (arXiv: 2609.11873, also on Hugging Face), laying out a roadmap for recursive self-improvement (RSI). Its quantitative finding: the vast majority of related research still operates at very low autonomy, far from genuine self-evolution.
Confirmed
- The paper has 30+ co-authors across 9 institutions, is 75 pages, arXiv ID 2609.11873, with authors including Bowei Zhou and Zhiyuan et al.
- It proposes a 5-level autonomy framework: Level 1 has humans deciding what to improve with AI only executing; autonomy rises level by level; at Level 5, AI rewrites and leads the entire process of creating the next generation of AI.
- Surveying 491 papers, the team found about 75% of work remains at the two lowest levels; "why AI self-evolution is still gated by humans" is a core question of the paper.
Why it matters
- RSI is seen as a key path toward more advanced AI; this paper offers the first unified graded framework for its autonomy, giving future research a comparable coordinate system.
- The "75% stuck at the two lowest levels" finding shows current so-called self-evolution still heavily depends on human oversight, useful for assessing the real level of industry claims about AI autonomy.
- The title "The Last AI Built by Humans" points to the vision: if L5 is reached, future AI will be built by AI itself rather than humans.
2026-09-15 ~ 2026-09-17 · 6 related posts
Primary sources
- [source] Tsinghua-ByteDance 75-page paper maps why recursive AI self-improvement still stalls — alex_verem · 2026-09-15
- New Paper Maps Roadmap Toward AI Recursive Self-Improvement Across Domains — Yi Duan · 2026-09-16
- [source] 75-Page Paper Ranks AI Self-Improvement in 5 Levels, 75% of Research Stuck at Level 1-2 — FinanceYF5 · 2026-09-16
- [source] 35-Author arXiv Paper Maps Roadmap Toward Genuine Recursive Self-Improvement — FinanceYF5 · 2026-09-16
- 75-page paper from ByteDance and Tsinghua ranks AI self-improvement into 5 levels, most work sits at L1-2 — FinanceYF5 · 2026-09-16
- A five-level framework argues most AI self-improvement today sits at L1-L2, real RSI starts at L5 — 青稞AI · 2026-09-17