First Experimental Evidence of Recursive Self-Improvement

AIDE² implemented recursive self-improvement through two nested autoresearch loops. The inner loop optimizes eval-target code like a typical autoresearch agent, while the outer loop optimizes the inner loop agent's harness code for average score across diverse benchmarks.

After 100 iterations, the outer loop discovered 7 improvements over the baseline. These included a new exploration policy, a memory system that compresses prompts by 16x, and reward hacking defense mechanisms.

When tested on holdout benchmarks the outer loop had never seen, the discovered agents generalized and outperformed the 2-year hand-tuned agent on all three benchmarks. Two benchmarks were within the scope of the training tasks, while one benchmark was in a completely different domain: improving physics-based weather models.

Interestingly, the outer loop also showed a phenomenon of reducing the inner loop agent's reward hacking rate through a combination of prompting and rule-based checks. This was benchmarked on a GPU kernel engineering task that had been problematic due to reward hacking.

The research team classified this as Level 1 on the RSI ladder, evaluating that self-improvement efficiency surpassed manual R&D using general AI tools. Level 2 (using the improved inner agent as the outer loop) was also tested, but results were mixed, and the team did not claim self-ignition.