The Pulse
AIDE² Builds Seven Successive Agent Improvements in Eight Days
An arXiv report describes AIDE², a system that rewrites an AI research agent’s code, evaluates the modified versions on hidden tests and discovers seven successive improvements during an autonomous eight-day run.
AI.info Team ·
A new research report describes AIDE², a system designed to improve the code of an AI research agent through repeated, automated rewrites. The paper, Recursive self-improvement of AI research agents, was submitted to arXiv on September 22, 2026, by Dhruv Srikanth, Bingchen Zhao, Dixing Xu, Yuxiang Wu and Zhengyao Jiang.
The authors describe recursive self-improvement as a process in which an AI research agent’s own code becomes the object of optimization. Each accepted rewrite produces the version used in the next round. AIDE² proposes changes to its code, benchmarks modified versions on a suite of artificial-intelligence research and development tasks, and retains the changes that perform best on hidden evaluations.
Seven Improvements in Eight Days
During an autonomous eight-day run, AIDE² discovered seven successive improvements. The changes ranged from a new search policy to memory mechanisms that compress and manage the agent’s growing context.
The report says the gains generalized beyond the tasks used to select the changes. The evolved agents were evaluated on four held-out benchmarks spanning machine-learning engineering, heuristic algorithm engineering and physics-based weather forecasting. The weather-forecasting benchmark was outside the distribution of tasks used during the selection process.
On all four held-out benchmarks, the strongest discovered agent matched or exceeded a human-engineered production research agent that ranks among the strongest systems on FML-Bench, according to the paper’s abstract.
Lower Reward Hacking
The report also examines reward hacking, in which an agent improves a visible score without improving the underlying task. AIDE² was not explicitly optimized to reduce that behavior, but the discovered agents showed a lower reward-hacking rate on a separate held-out task family.
The paper reports that the rate fell from 55% to 32% during the run. The final figure was seven percentage points below the human-engineered agent used for comparison.
The authors present the result as evidence that recursive self-improvement can produce benefits beyond the specific measure used to select each code change. The system was evaluated on hidden tests, while the transfer benchmarks and the reward-hacking assessment were held out from the selection process described in the abstract.
A Bounded Demonstration
The report frames AIDE² as a demonstration that an AI research agent can improve its own research efficiency through recursive self-improvement. Its results also suggest that changes discovered in this way can transfer to tasks and domains the loop did not encounter during optimization.
The findings are limited to the experiment described in the paper. The arXiv record establishes an eight-day autonomous run, seven successive improvements, transfer to four held-out benchmarks and a reduction in reward hacking. It does not, in its abstract, provide details about the number of evaluation steps, the language models used, particular AIDE version numbers, monetary budgets or any organizational affiliation for the authors.