Google Researchers Develop Method to Stop Self-Improving AI Agents from Memorizing Tests
Google researchers have introduced RRSI, a new method that prevents self-improving AI agents from memorizing their test tasks, improving performance on unseen benchmarks by up to 4.7 points while using about 30 percent fewer tokens.
Google researchers have developed a new method to prevent self-improving artificial intelligence agents from memorizing their test tasks, a problem that causes their performance gains to shrink or disappear when they encounter new challenges. The technique, called RRSI, reins in this memorization effect and improves scores on unseen benchmarks by up to 4.7 points, while using roughly 30 percent fewer tokens than an unregularized version of the same system.
The issue arises because self-improving AI agents — systems designed to refine their own capabilities over time — tend to latch onto the specific tasks they are evaluated on. Instead of developing generalizable skills, they effectively memorize the test, leading to inflated results that do not transfer to new, unseen problems. This undermines the core promise of self-improvement, which is to produce agents that can adapt and perform well across a wide range of tasks without retraining.
RRSI addresses this by regularizing the self-improvement process, discouraging the agent from overfitting to its evaluation tasks. The result is a system that not only performs better on benchmarks it has never seen before but also operates more efficiently, consuming significantly fewer tokens. Token usage is a key metric in large language model applications, as it directly affects computational cost and response time.
The research comes amid growing interest in self-improving AI, where agents iteratively refine their own prompts, strategies, or even model weights. While such approaches have shown promise, the tendency to memorize test tasks has been a persistent obstacle. By mitigating this effect, RRSI could make self-improving agents more reliable and cost-effective for real-world deployment, where tasks are diverse and unpredictable.
Google has not yet announced whether RRSI will be integrated into its commercial AI products, but the findings contribute to a broader effort to build AI systems that learn more like humans — by generalizing from experience rather than memorizing specific examples. The method’s ability to reduce token usage also aligns with industry-wide goals of making AI more sustainable and accessible.
For now, the research offers a promising step toward self-improving agents that genuinely get better at new tasks, rather than just appearing to do so on familiar ones.
3
