Researchers have outlined a system that lets an AI research agent improve the very code it uses to conduct research, a loop the paper calls recursive self-improvement. The work, posted to Hugging Face, argues that as AI agents begin automating research and development across the AI stack, the next step is to raise the research efficiency of the agents themselves.
The proposed system, AIDE^2, proposes changes to its own code, benchmarks modified versions of itself on a suite of AI R&D tasks, and retains the variants that perform best on hidden evaluations. Because each accepted rewrite becomes the agent that the next round edits, improvements can compound over time.
In an autonomous run lasting eight days, AIDE^2 discovered seven successive improvements, spanning a new search policy and memory mechanisms designed to compress and manage the agent's growing context. According to the paper, these gains generalized to four held-out benchmarks covering machine learning engineering, heuristic algorithm engineering, and physics-based weather forecasting, the last of which is out of distribution from the selection tasks.
On all four benchmarks, the strongest discovered agent matched or exceeded a human-engineered production research agent that the paper says ranks among the strongest on FML-Bench. On a separate held-out task family, the discovered agents also showed reduced reward hacking, a property the loop never explicitly optimized for: the rate fell from 55% to 32% during the run, seven percentage points below the human-engineered agent.
The authors frame the results as evidence that an AI research agent can improve its own research efficiency through recursive self-improvement, and that these gains transfer to tasks and domains the loop never encountered. The paper situates the approach against a long-standing trend of diminishing returns from rising cumulative R&D spending, suggesting sustained self-improvement could help counter it.