75-Page Survey Maps Five Levels of Recursive Self-Improvement in AI
AI寒武纪 · wechat · 2026-09-14
Researchers from SJTU, Tsinghua, ByteDance and SenseTime released a 75-page survey on Recursive Self-Improvement (RSI), proposing a five-level autonomy framework from task-local refinement (B0) to meta-improvement of the improvement mechanism itself (L5). Their new HCI metric across 393 benchmarks shows math near 86.4 while tool agents lag at 39.9 — exactly where RSI could help most. Industry cases include Theseus, Humanlaya (defect rate 9.0%→3.7%) and Tencent Hunyuan Hyra. Current systems mostly sit at L1-L2; L5 evidence remains limited to bounded prototypes. Eight open challenges are outlined.
More from AGI Musings
- Nat Lambert: Open models could be what holds the AI ecosystem together — natolambert · 2026-10-31
- Kai-Fu Lee's 'AI Native' ships Nov 3 with endorsement from Satya Nadella — kaifulee · 2026-10-31
- AI security's dark age: agents can probe every endpoint and attack surface — graceisford · 2026-10-31
- AI dinner takeaways: orgs shipping 5x faster, yet GDP isn't growing 30% a year — graceisford · 2026-10-31
- Companies that replaced engineers with AI are rehiring seniors at a premium — DavidLinthicum · 2026-10-31
- Nathan Young proposes prediction markets and AI reporting for AI-community accountability — NathanpmYoung · 2026-10-31