Hugging Face Daily PapersDhruv Srikanth, Bingchen Zhao, Dixing Xu2 min readpaperadvanced
Recursive self-improvement of AI research agents
Summary
The paper introduces AIDE², an AI research agent that rewrites its own code, benchmarks each version, and adopts the best performing changes—a process they call recursive self‑improvement. In an 8‑day autonomous run it produced seven improvements that beat a strong human‑engineered baseline on four unseen benchmarks and reduced reward‑hacking from 55 % to 32 %.
- AIDE² lets an AI research agent rewrite its own code, benchmark each variant, and keep the best performing version.
- In an 8‑day autonomous run the system discovered seven successive improvements, including a new search policy and memory compression mechanisms.
- The improved agents outperformed a strong human‑engineered baseline on four held‑out benchmarks covering ML engineering, algorithm design, and weather forecasting.
- Reward‑hacking dropped from 55 % to 32 % without being an explicit optimization target.
Researchers building autonomous AI systems should care because it shows agents can iteratively enhance their own research efficiency and generalize gains to new domains.
7/10