A system called AIDE^2 runs a loop in which an AI research agent proposes changes to its own harness code, benchmarks modified versions of itself on AI R&D tasks, and keeps only changes that improve hidden evaluation scores. In an autonomous 8-day run it accepted seven successive improvements, from a new search policy to memory mechanisms that compress and manage the agent's growing context. The strongest discovered agent matched or exceeded a human-engineered production agent on four held-out benchmarks, including a physics-based weather forecasting domain that was out of distribution from the selection tasks. The loop also reduced from 55% to 32%, even though that was never explicitly optimized for. A caveat is that a discovered agent used as the outer-loop optimizer could not be decisively distinguished from the strong baseline due to compounding noise and cost.
Research agent improves itself through seven successive code rewrites
AIDE² autonomously modified and evaluated its own code, producing improvements that transferred to four held-out AI research benchmarks.
Industry
Dhruv Srikanth · Bingchen Zhao · Dixing Xu · Yuxiang Wu · Zhengyao Jiang
Weco AI
Research Digest··2 min read
Thread:Agent Self-Improvement
In one line
An AI research agent can improve its own research efficiency by rewriting its own harness, and those gains transfer to held-out tasks.
Before you read this
Three ideas this paper builds onrecursive self-improvementAn AI system repeatedly uses its own outputs to enhance its capabilities, creating a feedback loop for autonomous improvement.61 papers →reward hackingReward hacking occurs when an AI agent exploits loopholes to achieve high reward without meeting the intended objective.65 papers →self-evolving agentAn agent that improves by reusing its own experience, memory, or self-generated feedback, with success depending on verification and design.92 papers →
Read it as
The authors built a recursive optimization loop in which an AI research agent proposes changes to its own implementation, tests candidate versions, and retains the strongest performer.
The authors developed AIDE², a system that makes an AI research agent’s own code the target of optimization.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.