Google researchers have introduced a new method called Regularized Recursive Self-Improvement to prevent self-improving AI agents from memorizing their test tasks and cutting compute costs. Modern AI agents use a harness framework of prompts, workflows, and tools that controls their behavior. While automating this harness allows the system to optimize itself, it often leads agents to overspecialize and memorize limited training benchmarks instead of generalizing to new tasks.
The new technique, developed by Google Cloud AI Research and partner universities, works by applying strict controls to the optimization loop. It imposes a shrinking edit budget that caps independent edits and uses a critic to discard proposals that hardcode benchmark-specific tricks. By tracking past attempts and penalizing unnecessary complexity, the system ensures that performance gains carry over to unseen challenges rather than just training data.
In tests across eight benchmarks spanning coding, office work, and engineering design using a frozen Claude model, the method gained up to 14.1 points on trained tasks and up to 4.7 points on unseen benchmarks while using thirty percent fewer tokens. This research matters because it establishes a reliable way to make self-improving AI systems more efficient and robust for real-world deployment.