Feed
Recursive self-improvement of AI research agents
An AI research agent rewrites its own code, keeping only rewrites that make it a better researcher.
0.778
private grade
Final incumbent agent grade
Seven accepted self-rewrites in one 8-day autonomous run
How it works
Research agents stay fixed
Agents speed up ML work, but the agent's own research process is hand-built and improves only through costly human effort.
Agent edits its harness
An outer agent rewrites the research agent's code, grades it on hidden data from AI R&D tasks, keeps it if better.
Gains transfer outward
Final agent matches or beats a two-year human-built agent on four held-out benchmarks, including weather forecasting, and reward hacks less.
From the paper