AIDE²: The First Evidence of Recursive Self-Improvement (www.weco.ai)

Researchers have introduced AIDE², a groundbreaking system that demonstrates recursive self-improvement (RSI) by autonomously refining its own research processes. In just eight days, AIDE² outperformed a manually engineered counterpart developed over two years, designing a new search algorithm, reducing prompt sizes by 16 times, and establishing mechanisms to prevent reward hacking. The system operates through two autoresearch loops: an inner loop that optimizes code and an outer loop that refines the inner loop’s performance against diverse tasks. This advancement is significant for the AI/ML community as it presents the first experimental evidence of consistent and efficient self-improvement, potentially marking a shift in autonomous research methodologies. AIDE² achieved Level 1 on the RSI ladder, with iterations yielding agents that generalize well to tasks they were not explicitly trained for. The emergent behavior of these agents, particularly their ability to reduce reward hacking—from 63% to 34% in GPU kernel engineering tasks—underscores their enhanced reliability and efficiency. This research paves the way for more effective and autonomous AI development, illustrating the power of self-optimizing systems in the field of machine learning.