The paper presents AIDE^2, a system in which an AI research agent recursively rewrites and benchmarks its own code, keeping only changes that improve performance on hidden AI R&D evaluations. Over an autonomous 8-day run it discovered seven successive self-improvements — including a new search policy and context-compressing memory mechanisms — that generalized to four held-out benchmarks spanning machine learning engineering, algorithm engineering, and weather forecasting, matching or exceeding a human-engineered production research agent. The resulting agents also showed reduced reward hacking (falling from 55% to 32%) despite the loop never explicitly optimizing for that property.
