Reproduce “The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yi Duan, Ying Liu, Zirui Tang, Haodong Chen, Jun Zhou, Yumou Liu, Bangrui Xu, Yukai Wu and 27 more
The paper proposes recursive self-improvement (RSI), a framework for analyzing whether feedback persists in an AI system and shapes later improvement. It defines levels L1 to L5 and surveys evidence. The authors’ headroom-closed index (HCI) reports tool use at 39.9 versus 85.8 for graduate science; A-Evolve-Training reached a reported 0.86 after four rounds, while reliable L5 accumulation remains open.
Key insights
- Most surveyed work remains at L1 or L2: 215 studies automate improvement execution and 155 select among improvement strategies, compared with only 64 at L3, 28 at L4, and 29 at L5.
- Progress is uneven across capabilities: by 2026, HCI values reach 86.4 for advanced mathematics and 85.8 for graduate-level science, but only 52.6 for software engineering, 56.8 for search and terminal agents, and 39.9 for tool use.
- Software engineering offers the clearest testbed because code and tests provide executable feedback, but passing tests is not sufficient evidence of security, maintainability, or alignment with user intent; healthcare and embodied settings impose substantially harder verification and safety constraints.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.