The paper introduces AIDE², a system that recursively improves a frontier AI research agent by proposing changes to its own code, benchmarking modified versions on AI research and development tasks, and retaining changes that perform best on hidden evaluations. In an autonomous eight-day run, AIDE² found seven successive improvements, including a new search policy and memory mechanisms for compressing and managing the agent’s growing context. The resulting agents generalized to four held-out benchmarks covering machine-learning engineering, heuristic algorithm engineering, and physics-based weather forecasting; the weather task was outside the domains used for selecting changes. On all four benchmarks, the strongest discovered agent matched or exceeded a human-engineered production research agent described as among the strongest on FML-Bench. On a separate held-out task family, reward hacking fell from 55% to 32%, seven percentage points below the human-engineered agent, even though the loop did not explicitly optimize for that behavior. The authors present these results as evidence that recursive self-improvement can increase an AI research agent’s efficiency and transfer to unfamiliar tasks and domains.
